Systems and methods relating to peak exploration and artificial tissue response including processing for functional response waveforms

Abstract

Systems and methods for processing functional response waveforms obtained from artificial tissue are described herein. In one example aspect, the systems and methods include obtaining a first waveform including a systolic response and a relaxation response of artificial tissue during a systolic-relaxation cycle. A model is then fitted to the first waveform, wherein the model independently parameterizes the growth of the contraction response and the relaxation response. A second waveform is then generated from the model fitted to the first waveform such that the second waveform includes a noise filtered representation of the first waveform. Feature values may then be extracted from the second waveform, which feature values may be used for noise filtering and to achieve a high fidelity and high capacity synthetic training dataset to develop artificial intelligence (AI) or otherwise for downstream tasks, such as drug discovery and development.

Description

[0001] This application claims the following benefits: U.S. Provisional Application No. 63 / 457,574 (filed April 6, 2023); U.S. Provisional Application No. 63 / 466,105 (filed May 12, 2023); U.S. Provisional Application No. 63 / 505,257 (filed May 31, 2023); and U.S. Provisional Application No. 63 / 526,566 (filed July 13, 2023). The entire contents of each of the foregoing provisional applications are incorporated herein by reference. Background Technology

[0002] Tissue behavior can be measured and modeled as a waveform of tissue response (e.g., functional response or contractile force) over time. The resulting functional response waveform is a fundamental and valuable unit for representing and encoding tissue data. This representation and encoding of tissue behavior as a waveform provides a rich substrate for downstream analysis. However, existing methods for characterizing such waveforms require extracting specific features from the peaks within the waveform. The extraction of such features is often hampered by factors such as noise, which in turn limits the effectiveness of these features in downstream analysis. Furthermore, building models that accurately represent the underlying peaks within the waveform is often limited by the amount of in vitro data on which such models can be trained.

[0003] Therefore, new methods are needed to process functional response data, which allow for accurate and efficient feature extraction.

[0004] Furthermore, existing methods typically rely on tissues exhibiting regular or periodic contractile responses (e.g., due to the application of external stimuli). When tissues exhibit spontaneous or irregular contractile behavior, existing methods may not be able to adequately identify individual contractile responses and extract relevant features.

[0005] Therefore, new methods are needed to detect and process spontaneous behavior in functional response data.

[0006] Furthermore, many functional response waveforms may also include combinations of single and double contractions (e.g., ectopic pulsations). The ability to process such waveforms typically depends on the ability to identify the type of contraction present.

[0007] Therefore, new methods are needed to process functional response data, which allow for accurate and efficient feature extraction and effective classification of shrinkage types.

[0008] Furthermore, in some systems, tissue can be attached to a tissue scaffold within the device, allowing images capturing the deflection of the tissue scaffold to be used to obtain measurements of the tissue response at the time point of image capture. However, the conversion from the tissue scaffold's positioning within the image frame to a measure of contractile force can be suppressed by inconsistent or inaccurate tracking of the tissue scaffold or by inaccurate conversion of scaffold deflection to contractile force.

[0009] Therefore, new methods are needed to track the deflection of tissue scaffolds and subsequently model the forces applied to the tissue scaffolds. Summary of the Invention

[0010] Systems and methods for processing functional response waveforms.

[0011] According to one aspect of this disclosure, a method for processing a functional response waveform is provided. The method includes: obtaining a first waveform including a contraction response and a relaxation response of an artificial tissue during a single contraction-relaxation cycle; fitting a model to the first waveform, wherein the model independently parameterizes the growth of the contraction response and the relaxation response; and generating a second waveform from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform.

[0012] According to another aspect of this disclosure, a method for training a model using synthetic training data is provided. The method includes: obtaining a plurality of waveforms, the plurality of waveforms including functional responses of one or more artificial tissues during a single contraction-relaxation cycle; extracting a plurality of parameter sets from the plurality of waveforms, wherein parameter sets in the plurality of parameter sets characterize corresponding waveforms among the plurality of waveforms; and determining a parameter set distribution from the plurality of parameter sets. The method further includes: generating a synthetic training dataset, each element of the synthetic training dataset including a synthetic waveform and a corresponding parameter set used to generate the synthetic waveform, wherein the corresponding parameter set is obtained from the parameter set distribution; and using the synthetic training dataset to train a predictive model, wherein the predictive model is trained to estimate an output parameter set from input waveforms.

[0013] According to an additional aspect of this disclosure, a method for extracting a contraction-relaxation cycle waveform is provided. The method includes: obtaining a first waveform comprising multiple functional responses of an artificial tissue stimulated at a first frequency; and convolving the first waveform with a pulse train by one or more processors to generate a convolved waveform, wherein the pulse train is generated at the first frequency. The method further includes: identifying a first position associated with a maximum value of the convolved waveform, wherein the first position corresponds to an expected position of a first contraction-relaxation cycle; and extracting a second waveform comprising the first contraction-relaxation cycle from the first position of the first waveform, wherein the second waveform has a first duration proportional to the first frequency.

[0014] According to another aspect of this disclosure, a method for predicting treatment effects is provided. The method includes obtaining a plurality of signals, including a baseline signal and a perturbation signal, wherein the baseline signal includes a first plurality of functional responses of an engineered tissue under reference conditions, and the perturbation signal includes a second plurality of functional responses of the engineered tissue under perturbation conditions involving the first perturbation. The method further includes splitting the plurality of signals into a first plurality of waveforms, each of the first plurality of waveforms including a contraction response and a relaxation response of the engineered tissue during a single contraction-relaxation cycle. The method further includes: fitting a model to each of the first plurality of waveforms, wherein the model independently parameterizes the growth of the contraction response and the relaxation response of the engineered tissue during a single contraction-relaxation cycle of each waveform; and generating a second plurality of waveforms from the model fitted to each of the first plurality of waveforms, wherein the second plurality of waveforms includes a plurality of filtered baseline waveforms associated with the baseline signal and a plurality of filtered perturbation waveforms associated with the perturbation signal. The method further includes: extracting a first feature value of a first feature from the plurality of filtered baseline waveforms; extracting a second feature value of the first feature from the plurality of filtered perturbation waveforms; and determining an effect associated with the first perturbation based on a comparison of the first feature value and the second feature value.

[0015] Another aspect of this disclosure provides a method for processing a functional response waveform. The method includes obtaining a first waveform by one or more processors, the first waveform including a contraction response and a relaxation response of an artificial tissue during a single contraction-relaxation cycle, wherein the first waveform is obtained from a bioreactor including the artificial tissue. The method further includes fitting a model to the first waveform by the one or more processors, wherein the model independently parameterizes the growth of the contraction response and the relaxation response. The method further includes generating a second waveform by the one or more processors from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform. The method further includes extracting one or more feature values ​​from the second waveform by the one or more processors. The method further includes training a machine learning model on the one or more feature values ​​from the second waveform. The method further includes generating predictions defining one or more properties of the human tissue by inputting a dataset of human tissue into the machine learning model, the predictions corresponding to at least one of the one or more feature values ​​of the second waveform.

[0016] Based on the foregoing and the disclosure herein, this disclosure includes applying a particular feature or aspect in conjunction with a particular machine (e.g., a bioreactor) or by using such a particular machine. In various aspects, the bioreactor may include means configured to grow or manipulate human tissue (e.g., human tissue, such as muscle tissue, cardiac tissue, and / or musculoskeletal tissue). Additionally or alternatively, the bioreactor may include a sensor assembly configured to detect one or more functional responses of the tissue within the apparatus.

[0017] Furthermore, this disclosure includes the transformation or restoration of a particular item to a different state or thing, such as the transformation or restoration of the functional response of human tissue in a bioreactor to a different state or thing as sensed as a waveform by a sensor component, such as the generation, creation, or more accurate fitting of a model of an improved waveform that has been filtered to remove errors (e.g., noise) in the original received or original waveform signal sensed from artificial tissue (e.g., by one or more sensors).

[0018] Furthermore, this disclosure includes improvements in computer functionality or other techniques, at least because this disclosure discloses systems and methods for reducing errors in underlying computing devices, for example, by generating enhanced (e.g., second) waveforms that are filtered from data noise, such as that caused by measurement hardware (e.g., sensor components that extract data from artificial tissue). Furthermore, once the enhanced (e.g., second) waveform is generated, fitted, or otherwise obtained, a high-fidelity training dataset can then be trained or generated from it. Such high-fidelity and high-volume synthetic training datasets can then be used to train or update models, such as new or updated machine learning models, to produce more accurate and less error-prone outputs, and thus high-quality pharmaceutical products, such as therapeutic drugs. Furthermore, compared to prior art systems and methods, the predictive model, when deployed on the underlying system, allows the systems and methods of this disclosure to be executed with fewer iterations and using fewer computational resources. In other words, this disclosure describes improvements to the functionality of the computer itself or "any other technical or technological field" because the enhanced predictive improvements provided by the predictive model allow the underlying computer system to utilize fewer processing and memory resources compared to prior art systems and methods. This is at least because a predictive model trained on a reduced-error (less noise) enhanced or second waveform can generate or determine the probability that a given organization will exhibit a given behavior with higher accuracy, without requiring various test and / or empirical computer simulations across a wide range of tests using multiple computational cycles and data. Therefore, using the predictive model results in fewer computational cycles or other iterations, which has less impact on the underlying computing device compared to previous prior art systems and methods. In other words, the systems and methods of this disclosure improve upon the prior art, at least because prior art systems and methods require empirical or trial-and-error methods that may involve real-world trials on human organizations, which may result in and require significant database and memory utilization as well as processor usage to obtain similar real-world or simulated results with the same or similar outcomes. On the other hand, the disclosed systems and methods describe the generation and / or use of bioreactors for growing and testing tissues to define limited datasets specific to a particular tissue (e.g., human engineered tissues). These limited datasets require less memory usage and / or processing utilization compared to conventional methods that use or require large amounts of unknown, potentially irrelevant datasets. Furthermore, the disclosure allows for the identification and use of high-fidelity datasets, which further reduces the need for additional computational loops.

[0019] In addition, this disclosure relates to improvements in other technical or technological fields, at least because the systems and methods of this disclosure provide robust, efficient, and comparable coding of organizational behavior that can be used to improve the efficiency and performance of several downstream drug discovery and development tasks. This can be done, for example, by a predictive model that is trained or otherwise generated using spectral representations as training data defining the functional responses of an organization (e.g., engineered human tissue). This predictive model can be deployed on an underlying computing device or system to improve its accuracy and predictive power in performing drug discovery and development tasks as described herein.

[0020] Furthermore, this disclosure includes specific features beyond those well-known, routine, and conventional activities in the art, and / or otherwise adds unconventional steps that limit this disclosure to specific useful applications (e.g., systems and methods for processing functional response waveforms of artificial tissues for the purpose of generating high-fidelity feature data, which can be used, for example, to train more accurate models and / or improve downstream tasks such as drug discovery and development).

[0021] Systems and methods for detecting spontaneous tissue contraction.

[0022] According to one aspect of this disclosure, a method for determining spontaneous behavior of an engineered organization is provided. The method includes obtaining a waveform comprising a functional response of the engineered organization over a time period, and applying a frequency-based global classifier to the waveform to generate a first classification score, wherein the first classification score indicates whether the waveform includes periodic contractions of the engineered organization over the time period. When the first classification score indicates that no periodic contractions exist within the waveform, the method further includes: applying a local classifier to the waveform to generate a second classification score, wherein the second classification score indicates whether the waveform includes spontaneous contractions of the engineered organization over the time period; and generating behavioral characteristics of the engineered organization during the time period based on the second classification score.

[0023] According to another aspect of this disclosure, a system for determining spontaneous behavior of engineered tissue is provided. The system includes a bioreactor comprising means configured for tissue growth and a sensor assembly configured to detect one or more functional responses of the engineered tissue within the means. The system further includes a processing unit communicatively coupled to the bioreactor, wherein the processing unit includes one or more processors configured to acquire a waveform comprising a functional response of the engineered tissue within the means over a time period, and to apply a periodicity classifier to the waveform to generate a first classification score, wherein the first classification score indicates whether the waveform includes periodic contractions of the engineered tissue during the time period. When the first classification score indicates that no periodic contractions are present in the waveform, the one or more processors are configured to apply a spontaneous contraction classifier to the waveform to generate a second classification score, wherein the second classification score indicates whether the waveform includes spontaneous contractions of the engineered tissue during the time period, and to generate behavioral characteristics of the engineered tissue during the time period based on the second classification score.

[0024] According to an additional aspect of this disclosure, a non-transitory computer-readable medium is provided storing instructions for determining spontaneous behavior of an engineered organization. When executed by one or more processors, these instructions cause one or more processors to obtain a waveform comprising a functional response of the engineered organization over a time period, and apply a first classifier to the waveform to generate a first classification score, wherein the first classification score indicates whether the waveform includes periodic contractions of the engineered organization over the time period. When the first classification score indicates that no periodic contractions are present in the waveform, the one or more processors further apply a second classifier to the waveform to generate a second classification score, wherein the second classification score indicates whether the waveform includes spontaneous contractions of the engineered organization over the time period, and a behavioral characteristic of the engineered organization during the time period is generated based on the second classification score.

[0025] Based on the foregoing and the disclosure herein, this disclosure includes applying a particular feature or aspect in conjunction with a particular machine (e.g., a bioreactor) or by using such a particular machine. In various aspects, the bioreactor may include means configured to grow or manipulate human tissue (e.g., human tissue, such as muscle tissue, cardiac tissue, and / or musculoskeletal tissue). Additionally or alternatively, the bioreactor may include a sensor assembly configured to detect one or more functional responses of the tissue within the apparatus.

[0026] Furthermore, this disclosure includes the transformation or restoration of a particular item to a different state or thing, such as the transformation or restoration of the functional response of human tissue, such as within a bioreactor, to a different state or thing, as sensed as a waveform by a sensor component, for example, by generating, creating, or otherwise developing behavioral characteristics of an engineered tissue based on raw received or raw waveform signals sensed from an artificial tissue (e.g., by one or more sensors).

[0027] Furthermore, this disclosure includes improvements in computer functionality or other techniques, at least because this disclosure discloses systems and methods for reducing errors in underlying computing devices, for example, by implementing a hierarchical system of multiple classifiers, where one classifier (e.g., a periodic classifier) ​​can filter out waveforms known to exhibit periodic or regular behavior. This reduces the amount of data used by the system, where only waveforms that do not exhibit periodic or regular behavior are fed to later local classifiers (e.g., spontaneous classifiers). This not only reduces the amount of data analyzed by the underlying computing system but also simplifies the system's ingestion of waveforms (or waveform data) that may exhibit spontaneous behavior, thereby providing improved identification of spontaneous behavior (and the location of spontaneous contractions) because the system will have fewer false positives, as it will only analyze narrow or otherwise reduced suspicious datasets. In other words, once waveforms (or waveform data) associated with spontaneous behavior of engineered organization are identified by the classifier, such data can then be selected (filtered) for analysis. This type of waveform data includes high-fidelity data used to generate behavioral characteristics of engineered tissues in order to produce more accurate and less error-prone outputs, and thus produce high-quality pharmaceutical products, such as therapeutic drugs.

[0028] Furthermore, compared to existing systems and methods, the hierarchical structure of classifiers, when deployed on the underlying system, allows the systems and methods of this disclosure to execute with fewer iterations and use fewer computational resources. That is, this disclosure describes an improvement in the functionality of the computer itself or "any other technical field" because the enhanced predictive improvement provided by the hierarchical structure of classifiers allows the underlying computer system to utilize fewer processing and memory resources compared to existing systems and methods. This is at least because, without requiring various test and / or empirical computer simulations across a wide range of tests using multiple computational cycles and data, the hierarchical structure of a classifier, designed to filter out periodic waveform data to provide reduced error (less noise) from spontaneous waveform data, can generate or determine the probability that a given organization will exhibit a given behavior with higher accuracy. Therefore, using a hierarchical structure of classifiers results in fewer computational cycles or other iterations, which has less impact on the underlying computing device compared to previous prior art systems and methods. In other words, the systems and methods disclosed herein improve upon existing technologies, at least because existing systems and methods require empirical or trial-and-error approaches that may involve real-world trials on human tissues. These real-world trials may induce and require substantial database and memory utilization, as well as processor usage, to obtain similar real-world or simulated results with the same or similar outcomes. On the other hand, the disclosed systems and methods describe the generation and / or use of bioreactors for growing and testing tissues to define a limited dataset specific to the tissue (e.g., human engineered tissues). This limited dataset requires less memory and / or processing utilization compared to conventional methods that use or require large amounts of unknown, potentially irrelevant datasets. Furthermore, the disclosure herein allows for the identification and use of high-fidelity datasets, which further reduces the need for additional computational loops.

[0029] In addition, this disclosure relates to improvements in other technical or technological fields, at least because the systems and methods of this disclosure provide robust, efficient, and comparable encoding of organizational behavior that can be used to improve the efficiency and performance of several downstream drug discovery and development tasks. For example, this can be achieved through a hierarchical structure of classifiers employed to filter specific waveform type data (e.g., spontaneous waveforms and / or correlated data) via classification, which can then be used to identify or define the functional response of an organization (e.g., engineered human tissue). The hierarchical structure of classifiers can be deployed on underlying computing devices or systems to improve their accuracy, classification, and / or prediction in performing drug discovery and development tasks as described herein.

[0030] Furthermore, this disclosure includes specific features beyond those well-known, routine, and conventional activities in the art, and / or otherwise adds unconventional steps that limit this disclosure to specific useful applications (e.g., systems and methods for determining spontaneous behavior of engineered organizations, which can be used, for example, to determine more accurate behavioral characteristics based on engineered organizations, which can improve downstream tasks such as drug discovery and development).

[0031] A system and method for classifying the shrinkage types of engineered tissues.

[0032] According to one aspect of this disclosure, a method for processing a functional response waveform is provided. The method includes obtaining a first waveform comprising at least one contraction response and at least one relaxation response of an engineered tissue, the first waveform having a predetermined length corresponding to a desired length of the contraction-relaxation cycle of the engineered tissue. The method further includes determining a predicted contraction type among a plurality of contraction types for the first waveform, wherein the plurality of contraction types includes a single contraction type and a double contraction type. The method further includes fitting a model to the first waveform based on the predicted contraction type, wherein the model, independent of the growth of at least one relaxation response of the engineered tissue, parameterizes the growth of at least one contraction response of the engineered tissue. The method further includes generating a second waveform from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform.

[0033] According to another aspect of this disclosure, a method for training a classifier to predict tissue contraction type from functional response waveforms is provided. The method includes obtaining a plurality of waveforms comprising the functional response of one or more engineered tissues, each of the plurality of waveforms having a predetermined length corresponding to an expected length of a contraction-relaxation cycle of the engineered tissue. The method further includes extracting a first plurality of parameter sets by one or more processors from a first subset of the plurality of waveforms associated with a single contraction type, wherein the first parameter set in the first plurality of parameter sets characterizes a first waveform within the first subset of the waveforms. The method further includes extracting a second plurality of parameter sets from a second subset of the plurality of waveforms associated with a dual contraction type, wherein the second parameter set in the second plurality of parameter sets characterizes a second waveform within the second subset of the waveforms. The method further includes determining a plurality of parameter set distributions, wherein the plurality of parameter set distributions include a first parameter set distribution determined from the first plurality of parameter sets and a second parameter set distribution determined from the second plurality of parameter sets. The method further includes generating a synthetic training dataset, each element of which includes a synthetic waveform and a corresponding tissue contraction type associated with the synthetic waveform, wherein the synthetic waveform is generated using a parameter set distribution from the plurality of parameter set distributions associated with the corresponding tissue contraction type. The method further includes training a classifier using a synthetic training dataset, wherein the classifier trained using the synthetic training dataset determines the predicted type of tissue contraction for the input waveform.

[0034] According to an additional aspect of this disclosure, a method for processing a functional response waveform is provided. The method includes: obtaining a first waveform comprising a plurality of functional responses of an artificial tissue stimulated at a first frequency; and convolving the first waveform with a pulse train to generate a convolved waveform, wherein the pulse train is generated at the first frequency. The method further includes identifying a first position associated with a first maximum value of the convolved waveform, wherein the first position corresponds to a desired position of a first contraction-relaxation cycle. The method further includes extracting a second waveform comprising the first contraction-relaxation cycle from the first position of the first waveform, wherein the second waveform has a first duration proportional to the first frequency.

[0035] According to another aspect of this disclosure, a method for predicting the effects of a disturbance is provided. The method includes obtaining a plurality of signals, including a baseline signal and a disturbance signal, wherein the baseline signal includes a first plurality of functional responses of an engineered organization under control conditions, and the disturbance signal includes a second plurality of functional responses of the engineered organization under a first set of disturbance conditions. The method further includes splitting the plurality of signals into a first plurality of waveforms, each of the first plurality of waveforms having a predetermined length and including at least one contraction response and at least one relaxation response of the engineered organization, wherein the predetermined length corresponds to an expected length of a contraction-relaxation cycle of the engineered organization. The method further includes determining a predicted contraction type among a plurality of contraction types for each of the first plurality of waveforms, wherein the plurality of contraction types includes a single contraction type and a double contraction type. The method further includes fitting a model to each of the first plurality of waveforms based on the corresponding predicted contraction type, wherein the model parameterizes the growth of at least one contraction response independently of the growth of at least one relaxation response of the engineered organization. The method further includes generating a second plurality of waveforms from a model fitted to each of the first plurality of waveforms, wherein the second plurality of waveforms includes a plurality of filtered baseline waveforms associated with a baseline signal and a plurality of filtered perturbation waveforms associated with a perturbation signal. The method further includes: extracting a first feature value from the plurality of filtered baseline waveforms; extracting a second feature value from the plurality of filtered treatment waveforms; and determining an effect associated with a first perturbation condition set based on a comparison of the first feature value and the second feature value.

[0036] According to an additional aspect of this disclosure, a method for training a model using synthetic training data is provided. The method includes: obtaining a plurality of waveforms, the plurality of waveforms including functional responses of one or more artificial tissues during a single contraction-relaxation cycle; and extracting a plurality of parameter sets from the plurality of waveforms, wherein parameter sets in the plurality of parameter sets characterize corresponding waveforms among the plurality of waveforms. The method further includes determining a parameter set distribution from the plurality of parameter sets and generating a synthetic training dataset, each element of the synthetic training dataset including a synthetic waveform and a corresponding parameter set used to generate the synthetic waveform, wherein the corresponding parameter set is obtained from the parameter set distribution. The method further includes using the synthetic training dataset to train a parameter estimation model, wherein the parameter estimation model is trained to estimate an output parameter set from an input waveform.

[0037] Based on the foregoing and the disclosure herein, this disclosure includes applying a particular feature or aspect in conjunction with a particular machine (e.g., a bioreactor) or by using such a particular machine. In various aspects, the bioreactor may include means configured to grow or manipulate human tissue (e.g., human tissue, such as muscle tissue, cardiac tissue, and / or musculoskeletal tissue). Additionally or alternatively, the bioreactor may include a sensor assembly configured to detect one or more functional responses of the tissue within the apparatus.

[0038] Furthermore, this disclosure includes the transformation or restoration of a particular item to a different state or thing, such as the transformation or restoration of the functional response of human tissue within a bioreactor to a different state or thing, as sensed as a waveform by a sensor component, for example, by using a fitted model and the generation, creation, or otherwise development of a noise-filtered second waveform based on the original received or original waveform signal sensed from an artificial tissue (e.g., by one or more sensors).

[0039] Furthermore, this disclosure includes improvements to computer functionality or other techniques, at least because this disclosure discloses systems and methods for reducing errors in underlying computing devices, for example, by implementing a contraction-relaxation loop model, wherein a first waveform of engineered tissue is improved by noise filtering through model fitting, thereby eliminating or reducing data usage and enabling the generation of high-fidelity and high-volume synthetic training datasets. The contraction-relaxation loop model can also be efficiently extended to model different types of contraction, such as functional response waveforms including single or double contraction responses. Thus, the amount of data used by the system is reduced, where waveforms that have already undergone noise reduction (and therefore data reduction) can be used as outputs for final processes, such as outputs for, for example, improved downstream tasks (such as drug discovery and development). This not only reduces the amount of data analyzed by the underlying computing system but also simplifies the system's ingestion of waveforms (or waveform data) having both contraction and relaxation responses, as the system will have fewer false positives because it will only analyze high-fidelity datasets that have been filtered or otherwise reduced, as it relies on synthetic (controlled) data as determined by the model. This type of data is scalable because it allows models to predict different types of contraction (e.g., functional response waveforms including single or double contraction responses), classify those different contraction types, or otherwise output those different contraction types. In other words, once waveforms (or waveform data) associated with the relaxation and contraction responses of engineered tissue are input, such data can be fed into a model (e.g., a classifier model) to determine or otherwise classify the various contraction types identified by the model (e.g., single and / or double contraction types), and then this data can be selected (filtered) for analysis. This waveform data includes high-fidelity data for fitting the input waveform (e.g., a first waveform) to the model to parameterize the growth of the engineered tissue's contraction response independently of the growth of the engineered tissue's relaxation response, to generate a second, noise-filtered version of the input (first) waveform. This also allows for the generation of more accurate and less error-prone outputs, and therefore allows for the generation of high-quality pharmaceutical products such as therapeutics.

[0040] Furthermore, compared to existing systems and methods, the contraction-relaxation loop model, when deployed on the underlying system, allows the systems and methods of this disclosure to execute with fewer iterations and use fewer computational resources. That is, this disclosure describes an improvement in the functionality of the computer itself or "any other technical field" because the enhanced predictive improvements provided by the contraction-relaxation loop model allow the underlying computer system to utilize fewer processing and memory resources compared to existing systems and methods. This is at least because, without requiring various test and / or empirical computer simulations across a wide range of tests using multiple computational loops and data, the contraction-relaxation loop model, designed to filter out periodic waveform data to provide reduced error (less noise), can generate or determine the probability that a given organization will exhibit a given behavior with higher accuracy. Therefore, the hierarchical structure using classifiers results in fewer computational loops or other iterations, which has less impact on the underlying computing device compared to previous prior art systems and methods. In other words, the systems and methods disclosed herein improve upon existing technologies, at least because existing systems and methods require empirical or trial-and-error approaches that may involve real-world trials on human tissues. These real-world trials may induce and require substantial database and memory utilization, as well as processor usage, to obtain similar real-world or simulated results with the same or similar outcomes. On the other hand, the disclosed systems and methods describe the generation and / or use of bioreactors for growing and testing tissues to define a limited dataset specific to the tissue (e.g., human engineered tissues). This limited dataset requires less memory and / or processing utilization compared to conventional methods that use or require large amounts of unknown, potentially irrelevant datasets. Furthermore, the disclosure herein allows for the identification and use of high-fidelity datasets, which further reduces the need for additional computational loops.

[0041] In addition, this disclosure relates to improvements in other technical or technological fields, at least because the systems and methods of this disclosure provide robust, efficient, and comparable coding of tissue behavior that can be used to improve the efficiency and performance of several downstream drug discovery and development tasks. For example, this can be done by a contraction-relaxation cycle model (e.g., a type of classifier) ​​employed to filter specific waveform type data (e.g., waveform and / or correlation data) via classification, which can then be used to identify, classify, or define the contraction type (single contraction type and / or double contraction type) of engineered human tissue. The contraction-relaxation cycle model can be deployed on the underlying computing device or system to improve its accuracy, classification, and / or prediction in performing drug discovery and development tasks as described herein.

[0042] Furthermore, this disclosure includes specific features beyond those well-known, routine, and conventional activities in the art, and / or otherwise adds unconventional steps that limit this disclosure to specific useful applications (e.g., systems and methods for classifying the contraction types of engineered tissues, which can be used, for example, to determine noise-filtered and / or model-fitted waveforms based on engineered tissues, which can improve downstream tasks such as drug discovery and development).

[0043] Systems and methods for tracking tissue scaffolds.

[0044] According to one aspect of this disclosure, a system for modeling the contractile deflection of a flexible tissue scaffold is provided. The system includes a bioreactor comprising a flexible scaffold for attachment to biological tissue, wherein the flexible scaffold is arranged to deflect in response to contractile forces applied thereon. The bioreactor further includes an imaging device configured to acquire one or more images of the flexible scaffold. The system further includes a processing unit communicatively coupled to the bioreactor. The processing unit is configured to acquire multiple images of the flexible scaffold at multiple time points from the bioreactor, wherein the multiple images capture the deflection of the flexible scaffold along a first dimension due to contractile forces applied thereon at the multiple time points. The processing unit is further configured to fit multiple curves to the multiple images such that each curve extends along a centerline of the flexible scaffold within a corresponding image of the multiple images and determines multiple displacement values ​​from the multiple curves, wherein each of the multiple displacement values ​​includes a measurement between the corresponding curve of the multiple curves and a reference line extending along a second dimension perpendicular to the first dimension. The processing unit is further configured to generate a model based on the multiple displacement values, wherein the model characterizes the contractile forces applied to the flexible scaffold at the multiple time points.

[0045] According to another aspect of this disclosure, a method for modeling the contractile deflection of a flexible tissue scaffold is provided. The method includes acquiring multiple images of the flexible scaffold at multiple time points, wherein the multiple images capture the deflection of the flexible scaffold along a first dimension due to contractile forces applied thereon at the multiple time points. The method further includes fitting multiple curves to the multiple images such that each curve extends along the centerline of the flexible scaffold within a corresponding image of the multiple images and determining multiple displacement values ​​from the multiple curves, wherein each of the multiple displacement values ​​includes a measurement along the first dimension between the corresponding curve of the multiple curves and a reference line extending along a second dimension perpendicular to the first dimension. The method further includes generating a model based on the multiple displacement values, wherein the model characterizes the contractile forces applied to the flexible scaffold at the multiple time points.

[0046] According to an additional aspect of this disclosure, a non-transitory computer-readable medium is provided that stores instructions, when executed by a processing unit, to cause the processing unit to: acquire multiple images of a flexible stent at multiple time points, wherein the multiple images capture the deflection of the flexible stent along a first dimension due to contractile forces applied thereon at the multiple time points; fit multiple curves to the multiple images such that each curve extends along the centerline of the flexible stent within a corresponding image of the multiple images; determine multiple displacement values ​​from the multiple curves, wherein each of the multiple displacement values ​​includes a measurement along the first dimension between the corresponding curve of the multiple curves and a reference line extending along a second dimension perpendicular to the first dimension; and generate a model based on the multiple displacement values, wherein the model characterizes the contractile forces applied to the flexible stent at the multiple time points.

[0047] Based on the foregoing and the disclosure herein, this disclosure includes applying a particular feature or aspect in conjunction with a particular machine (e.g., a bioreactor) or by using such a particular machine. In various aspects, the bioreactor may include means configured to grow or manipulate human tissue (e.g., human tissue, such as muscle tissue, cardiac tissue, and / or musculoskeletal tissue). Additionally or alternatively, the bioreactor may include a sensor assembly configured to detect one or more functional responses of the tissue within the apparatus.

[0048] Furthermore, this disclosure includes the transformation or restoration of a particular item to a different state or thing, such as the transformation or restoration of a displacement value of a flexible scaffold within a bioreactor sensed by an imaging component to a different state or thing, such as the generation, creation, or otherwise determination of the contractile force of an engineered tissue attached to a flexible scaffold tissue.

[0049] Furthermore, this disclosure includes improvements to computer functionality or other techniques, at least because this disclosure discloses systems and methods for accurately tracking and efficiently extracting and measuring the deflection of tissue scaffolds, for example, by using flexible scaffolds, thereby allowing the generation of models characterizing the contractile forces that produce deformation, thus reducing errors in the underlying computing device. The underlying system can be updated using this model, which can be used to encode or calibrate the relationship between contractile forces and measured displacements, thereby improving the accuracy of contractile force measurements obtained from such models. Improvements to such models provide improvements to downstream tasks utilizing such models, while also improving the efficiency and performance of the computing systems used to generate and deploy such models. This also allows for the generation of more accurate and less error-prone outputs, and therefore allows for the generation of high-quality pharmaceutical products such as therapeutic drugs.

[0050] Furthermore, compared to existing systems and methods, the contraction-relaxation loop model, when deployed on the underlying system, allows the systems and methods of this disclosure to be executed with fewer iterations and use fewer computational resources. That is, this disclosure describes an improvement in the functionality of the computer itself or "any other technical field" because the increased predictive improvements provided by implementing a model (e.g., a force-displacement model) generated for a particular device and / or tissue scaffold allow for modeling any variations in the contraction response of the tissue scaffold within the force-displacement model, thereby improving the consistency of force measurements obtained from displacement values ​​provided to the model. Therefore, using the model (e.g., the force-displacement model) results in fewer computational loops or other iterations, and has less impact on the underlying computing device compared to previous prior art systems and methods. In other words, the systems and methods of this disclosure improve upon existing technologies, at least because existing systems and methods require empirical or trial-and-error approaches that may involve real-world trials on human tissues, which may provoke and require significant database and memory utilization as well as processor usage to obtain similar real-world or simulated results with the same or similar outcomes. On the other hand, the disclosed systems and methods describe the generation and / or use of bioreactors for growing and testing tissues to define limited datasets specific to a particular tissue (e.g., human engineered tissues). These limited datasets require less memory usage and / or processing utilization compared to conventional methods that use or require large amounts of unknown, potentially irrelevant datasets. Furthermore, the disclosure allows for the identification and use of high-fidelity datasets, which further reduces the need for additional computational loops.

[0051] In addition, this disclosure relates to improvements in other technical or technological fields, at least because the systems and methods of this disclosure provide robust, efficient, and comparable tissue measurement solutions that can be used to improve the efficiency and performance of several downstream drug discovery and development tasks. For example, this can be achieved using a flexible scaffold and an imaging device configured to generate a model based on displacement values, such that the model can characterize the contractile forces applied to the flexible scaffold. The model can then be used to identify, classify, or define tissues such as engineered human tissues. The model can be deployed on an underlying computing device or system to improve its accuracy, classification, and / or prediction in performing drug discovery and development tasks as described herein.

[0052] Furthermore, this disclosure includes specific features beyond those well-known, routine, and conventional activities in the art, and / or otherwise adds unconventional steps that limit this disclosure to specific useful applications (e.g., systems and methods for modeling the contraction deflection of flexible tissue scaffolds and for generating models characterizing the contractile forces applied to flexible tissue scaffolds at multiple time points, which can improve downstream tasks such as drug discovery and development).

[0053] The advantages will become more apparent to those skilled in the art from the following description of the preferred embodiments, which have been shown and described by way of illustration. It should be understood that this embodiment may be capable of having other different embodiments, and its details may be modified in various aspects. Therefore, the drawings and description should be considered illustrative in nature and not restrictive. Attached Figure Description

[0054] Embodiments of this disclosure will now be described by way of example only and with reference to the accompanying drawings, in which:

[0055] Figure 1 A system for processing functional response data according to one aspect of this disclosure is shown;

[0056] Figure 2 An aspect of this disclosure is shown including from Figure 1 The waveform of the functional response of the engineered organization obtained by the system;

[0057] Figure 3 A contraction-relaxation cycle model according to an embodiment of the present disclosure is illustrated;

[0058] Figure 4 An example waveform with a corresponding contraction-relaxation cycle model fitting is shown according to one aspect of this disclosure;

[0059] Figure 5A and Figure 5B The effect of each parameter of a contraction-relaxation cycle model according to one embodiment of the present disclosure is illustrated;

[0060] Figure 6 A parameter estimation model for fitting parameters of a contraction-relaxation cycle model according to an embodiment of the present disclosure is shown;

[0061] Figures 7A to 7D An embodiment according to the present disclosure is illustrated. Figure 6 The elements of the parameter estimation model shown;

[0062] Figure 8 A method for processing a functional response waveform according to one aspect of this disclosure is shown;

[0063] Figure 9 A method for fitting a model to a functional response waveform according to an embodiment of the present disclosure is shown;

[0064] Figure 10 A method for training a parameter estimation model using synthetic training data, according to one aspect of this disclosure, is shown;

[0065] Figure 11 An embodiment of the present disclosure is shown for use according to... Figure 10 A method for predicting parameter value sets by training a parameter estimation model using the method described above;

[0066] Figure 12 A method for extracting a contraction-relaxation cycle waveform according to one aspect of this disclosure is shown;

[0067] Figure 13 A method for extracting additional contraction-relaxation cyclic waveforms according to one embodiment of the present disclosure is shown;

[0068] Figure 14 A method for predicting treatment effects using a contraction-relaxation cycle model, according to one aspect of this disclosure, is shown;

[0069] Figure 15A The functional response of an organization exhibiting periodic contractile behavior is shown according to one aspect of this disclosure;

[0070] Figure 15B A functional response of an organization exhibiting spontaneous contraction behavior is shown according to one aspect of this disclosure;

[0071] Figure 16 A system for determining spontaneous behavior of an engineered organization, according to one aspect of this disclosure, is shown;

[0072] Figure 17A and Figure 17B The spectral responses of different contraction responses according to embodiments of the present disclosure are shown;

[0073] Figure 18 A convolutional neural network for predicting contraction behavior according to an embodiment of the present disclosure is shown;

[0074] Figure 19 A step-by-step result of a modified Pan-Tompkins algorithm according to an embodiment of the present disclosure is illustrated;

[0075] Figure 20 An example of adaptive thresholding of a waveform according to an embodiment of the present disclosure is illustrated;

[0076] Figure 21A method for determining spontaneous behavior of an engineered organization according to one aspect of this disclosure is shown;

[0077] Figure 22 A method for obtaining a classification score from a waveform according to an embodiment of the present disclosure is shown;

[0078] Figure 23 A method for obtaining a classification score from a waveform according to an embodiment of the present disclosure is shown;

[0079] Figure 24 A modified Pans-Tompkins method according to one embodiment of the present disclosure is shown;

[0080] Figure 25 A dual-contraction type contraction-relaxation cycle model according to an embodiment of the present disclosure is illustrated;

[0081] Figure 26 A system for processing functional response waveforms with different contraction types according to an embodiment of the present disclosure is shown;

[0082] Figure 27 The classification of contraction types for three waveforms according to one embodiment of the present disclosure is illustrated;

[0083] Figure 28 A method for processing a functional response waveform according to one aspect of this disclosure is shown;

[0084] Figure 29 A method for fitting a model to a functional response waveform according to an embodiment of the present disclosure is shown;

[0085] Figure 30 A method for training a classifier to predict tissue contraction type from functional response waveforms is shown;

[0086] Figure 31 A method for predicting the type of tissue contraction for a waveform using a synthetically trained classifier, according to an embodiment of the present disclosure, is shown.

[0087] Figure 32 A method for training a parameter estimation model using synthetic training data, according to one aspect of this disclosure, is shown;

[0088] Figure 33 A method for extracting single or double contraction-relaxation cycle waveforms according to one aspect of this disclosure is shown;

[0089] Figure 34 A method for extracting additional single or double contraction-relaxation cycles from a waveform is shown according to one embodiment of the present disclosure;

[0090] Figure 35 A method for predicting the effects of a disturbance according to one aspect of this disclosure is shown;

[0091] Figure 36 A bioreactor (such as) according to an embodiment of the present disclosure is shown. Figure 1 The well of the bioreactor shown;

[0092] Figure 37 Example images of tissue scaffolds according to embodiments of the present disclosure under different contractile forces are shown;

[0093] Figure 38 A method for tracking a flexible stent according to one aspect of this disclosure is shown;

[0094] Figure 39 A one-dimensional vascular enhancement filter according to an embodiment of the present disclosure is shown;

[0095] Figure 40 A plot of data points obtained from a transformed image of a tissue scaffold according to one embodiment of the present disclosure is shown;

[0096] Figure 41 A model generated corresponding to a time series of stent deflection values ​​is shown according to an embodiment of the present disclosure;

[0097] Figure 42 A flexible support according to an embodiment of the present disclosure is shown to deflect due to a predetermined force applied thereon by a probe;

[0098] Figure 43 A drawing of a force-displacement model according to an embodiment of the present disclosure is shown;

[0099] Figure 44 A method for modeling the contraction deflection value of a flexible tissue scaffold is shown according to one aspect of this disclosure;

[0100] Figure 45 A method for fitting multiple curves according to an embodiment of the present disclosure is shown;

[0101] Figure 46 A method for fitting a reference line according to an embodiment of the present disclosure is shown; and

[0102] Figure 47 An example computing system for implementing the methods of this disclosure is shown according to one embodiment. Technical Field

[0103] This disclosure relates to processing functional response waveforms. Specifically, but not exclusively, this disclosure relates to using a contraction-relaxation cyclic model to process functional response waveforms. More specifically, but not exclusively, this disclosure relates to using a contraction-relaxation cyclic model to generate a noise-filtered representation of the functional response waveform. Again, more specifically, but not exclusively, this disclosure relates to identifying the effects associated with perturbations in engineered structures based on the noise-filtered functional response waveform generated using a contraction-relaxation cyclic model.

[0104] In addition, this disclosure relates to detecting organizational behavior. Specifically, but not exclusively, this disclosure relates to detecting spontaneous organizational contraction. Specifically, but not exclusively, this disclosure relates to detecting spontaneous organizational contraction in the functional response waveform of engineered tissues.

[0105] Furthermore, this disclosure relates to classifying and processing functional response waveforms. Specifically, but not exclusively, this disclosure relates to using a contraction-relaxation cyclic model to classify and process functional response waveforms. More specifically, but not exclusively, this disclosure relates to generating a noise-filtered representation of the functional response waveform based on an identified contraction type using a contraction-relaxation cyclic model. Again, more specifically, but not exclusively, this disclosure relates to identifying the effects associated with perturbations in engineered structures based on the noise-filtered functional response waveform generated using a contraction-relaxation cyclic model.

[0106] Furthermore, this disclosure relates to tissue scaffold modeling. Specifically, but not exclusively, this disclosure relates to modeling the contraction deflection of flexible tissue scaffolds. Specifically, but not exclusively, this disclosure relates to generating models characterizing the contractile forces applied to flexible tissue scaffolds at multiple time points. Detailed Implementation

[0107] The ability to accurately and efficiently extract features from the waveforms of the functional responses of engineered tissues is a crucial step when applying such waveforms to downstream tasks such as drug discovery and development. Existing methods are often limited by the quality and / or quantity of available data. Furthermore, the functional responses of engineered tissues can exhibit different response types, making it difficult to fit a single model to all types of responses. This disclosure proposes systems and methods for peak exploration and artificial tissue responses to overcome these problems.

[0108] Figure 1 A system 100 for processing functional response data or spontaneous tissue contraction data according to various aspects of this disclosure is shown. Additionally or alternatively, system 100 also illustrates modeling the contraction deflection of a flexible tissue scaffold according to one aspect of this disclosure.

[0109] System 100 includes a bioreactor 102 and a control unit 104. Bioreactor 102 includes a device 106 for growing engineered tissues, a sensor assembly 108, and an interface 110. In one embodiment, the sensor assembly 108 is part of device 106, but such elements may be separate in other embodiments. Control unit 104 may include a model fitting unit 104-1, a signal processing unit 104-2, and a signal processing unit 112 (which may also be referred to as a signal processor). Additionally or alternatively, control unit 104 includes signal processing unit 112 (alternatively referred to as a processing unit, signal processor, or processor). A processing unit may include one or more processors, such as those described herein. Figure 47 One or more processors are described. Interface 110 communicatively couples bioreactor 102 and control unit 104, allowing data to be exchanged between bioreactor 102 and control unit 104. In various aspects, the bioreactor is used to grow tissues (e.g., engineered tissue 124) (such as human tissues, including but not limited to muscle tissue, heart tissue, musculoskeletal tissue, or other human tissues).

[0110] As shown in extension 106-1 of device 106, device 106 or substrate includes one or more wells (such as well 114), one or more cell culture wells (such as cell culture well 116), a pair of electrodes (including a first electrode 118-1 and a second electrode 118-2), and a pair of elements (e.g., scaffolds) (including a first element 120-1 (e.g., a first scaffold) and a second element 120-2 (e.g., a second scaffold)). Well 114 is located in cell culture well 116 and has a bottom, a first end 122-1, and a second end 122-2 on device 106. Well 114 is configured to allow engineered tissue 124 to grow from cells seeded therein. Culture medium may be added to cell culture well 116 to allow the engineered tissue 124 to grow and / or maintain the tissue. Engineered tissue 124 (optionally referred to as artificial tissue) includes engineered muscle tissue. In one embodiment, engineered muscle tissue is engineered heart tissue. For example, engineered heart tissue is cardiomyocytes using Cell Dynamics International (CDI) iCell. 2 The cells are generated from side groups of normal human ventricular cardiac fibroblasts, embedded in a hydrogel composed of Sigma-Aldrich, collagen, and corning. In an alternative embodiment, any other human cardiomyocytes are used instead of iCell cardiomyocytes. 2Such as human iPSC-induced ventricular cardiomyocytes from Axol Bioscience or human cardiomyocytes (HCM) from Sigma-Aldrich. Additionally or alternatively, fibrin and / or collagen are removed from the hydrogel. In an alternative embodiment, the engineered muscle tissue is engineered skeletal muscle tissue.

[0111] The electrode pair is separated by a gap, within which well 114 is positioned. The electrode pair is configured to apply electrical stimulation to cell cultures (e.g., engineered tissue 124 within well 114 shown in extension 106-1) within one or more wells of device 106. During the maturation of the cell cultures within device 106, the electrode pair applies stimulation to the cell cultures according to a multi-week electrical stimulation protocol. After the cell cultures have matured, the electrode pair can be configured to stimulate the cell cultures (e.g., engineered tissue 124) at a set frequency or pacing frequency. In one embodiment, the frequency or pacing frequency at which it stimulates the cell cultures is set by control unit 104. Thus, control unit 104 can be configured to send instruction 126 to bioreactor 102 to cause bioreactor 102 to stimulate the engineered tissue within device 106 at a set pacing frequency.

[0112] In some example embodiments, the first element 120-1 and the second element 120-2 are disposed across well 114 such that a gap exists between the bottom of well 114 and the pair of elements. The first element 120-1 and the second element 120-2 are configured to: (a) allow attachment of engineered tissue 124 formed therebetween, thereby suspending the engineered tissue 124 above the bottom of well 114; and (b) deform in response to contractile forces exerted on the pair of elements by the engineered tissue 124, thereby simulating the physiological environment of the engineered tissue 124 itself and / or allow measurement of the contractile forces (e.g., by sensor assembly 108). For example, the pair of electrodes may subject the engineered tissue 124 to electrical stimulation at a frequency of 0.1 Hz. The engineered tissue 124 will contract in response to this electrical stimulation, thereby causing deformation of at least one of the first element 120-1 and the second element 120-2. Measuring the deformation of one or more pairs of elements allows recording the functional response of the engineered tissue 124 when stimulated at 0.1 Hz.

[0113] Sensor assembly 108 is configured to detect one or more functional responses of the engineered tissue within device 106 (e.g., one or more functional responses of engineered tissue 124). In one embodiment, sensor assembly 108 includes an optical sensor. The optical sensor is configured to detect deformation of the first element 120-1 and / or the second element 120-2 (e.g., due to contractile forces exerted on the first element 120-1 and / or the second element 120-2 by the engineered tissue 124). Detecting deformation of the first element 120-1 and / or the second element 120-2 allows the determination of one or more functional responses, such as displacement or contractile displacement of the engineered tissue 124 or contractile force of the engineered tissue 124. Additionally or alternatively, the optical sensor of sensor assembly 108 is configured to detect the fluorescence intensity of the engineered tissue 124. Detecting the fluorescence intensity of the engineered tissue 124 allows the determination of one or more functional responses, such as transient calcium responses of the engineered tissue 124 or changes in the membrane potential of the engineered tissue 124. Additionally or alternatively, the optical sensor of sensor assembly 108 is configured to detect changes in the dimensions of engineered tissue 124 over a time period. Detecting changes in the dimensions of engineered tissue 124 over a time period allows for the determination of one or more functional responses, such as displacement or contraction displacement of engineered tissue 124 or contraction force of engineered tissue 124.

[0114] The optical sensor of sensor assembly 108 is configured to acquire multiple image-based representations of one or more functional responses of an organization (e.g., engineered organization 124) over a time range or time period. For example, the optical sensor may be configured to... Capture images or frames of the tissue in seconds. Here, Associated with a predetermined rate at which the image or frame will be captured. For example, when At that time, one image of the tissue is captured per second. Selectable... Any suitable value, such as In one embodiment, the frame rate is determined based on the frequency at which the tissue within the device is being stimulated. Therefore, images or a sequence of frames of the tissue over a time period capture one or more functional responses of the tissue within that time period.

[0115] In one embodiment, bioreactor 102 is configured to transform an image sequence capturing one or more functional responses of the tissue over a time span into a waveform representation. For example, sensor assembly 108, interface 110, or another component of bioreactor 102 may process images within the image sequence to extract features related to the tissue's functional response at the time point associated with the image. The features extracted from the image sequence that are related to the tissue's functional response can then be combined to form a waveform that includes one or more functional responses of the tissue over time. This waveform, such as waveform 128, can then be output from bioreactor 102. In an alternative embodiment, raw image or frame data is output from bioreactor 102, and a separate unit (e.g., control unit 104 or image analysis unit (not shown)) processes this data to determine time-series functional response data.

[0116] The time-series functional response data (e.g., waveform 128) generated by bioreactor 102 provides a high-fidelity and high-information-density representation of the functional response of the engineered tissue over a predetermined time period (e.g., the contractile force generated by the first tissue in response to a pacing frequency of 1 Hz over a 30-second period). To obtain such waveform data for the engineered tissue 124 from bioreactor 102, a command can be sent to bioreactor 102 (e.g., from control unit 104) to begin stimulating the engineered tissue 124 at a pacing frequency. Alternatively, no electrical stimulation is applied to observe the spontaneous response of the engineered tissue 124. Sensor assembly 108 then captures an image sequence of the engineered tissue 124 over a predetermined time period. Specifically, this image sequence captures the response (e.g., deformation) of the first element 120-1 and / or the second element 120-2 due to the contractile response of the engineered tissue 124. The image sequence is then processed to extract the responses of the first element 120-1 and / or the second element 120-2 over a predetermined time period, and the responses of these elements over time are transformed into a time series of functional responses over time. For example, the displacements of the first element 120-1 and / or the second element 120-2 can be used to determine the forces of the contractile response of the engineered tissue 124. The time series (e.g., waveform 128) is then output from the bioreactor 102 for further processing and / or analysis.

[0117] Drugs or compounds can be periodically administered to tissues (e.g., engineered tissue 124) within bioreactor 102, and the functional response of the tissue after administration can be recorded in waveform form. Alternatively, the functional response of the tissue can be periodically recorded without any external administration protocol. In either case, changes in the functional response of the tissue are captured in subsequent waveforms, such as... Figure 2 exemplified in .

[0118] Figure 2 Waveform 202 shows the functional response of the organization within a predetermined time period.

[0119] Waveform 202 from a bioreactor (such as those described above) Figure 1 The bioreactor 102 shown is used to obtain the tissue. Waveform 202 provides the tissue over a time range. to A high-fidelity and high-information-density representation of the functional response within the waveform. Waveform 202 includes multiple peaks (alternatively referred to as contraction cycles or contraction-relaxation cycles) that correspond to the contraction response of the tissue over a time range (e.g., 30 s, 40 s, etc.). An enlarged view of a single contraction-relaxation cycle (i.e., peak) within portion 204 of waveform 202 is shown in extension 204-1.

[0120] Typically, multiple features are extracted from each contraction-relaxation cycle of waveform 202 to characterize waveform 202, thereby allowing for further processing or analysis of waveform 202. As shown in extension section 204-1, the features extracted from a single contraction-relaxation cycle include peak (or twitch) amplitude 206, time to reach peak amplitude 208 (or contraction time), time to reach peak descent 210 (or relaxation time), duration 212 (contraction-relaxation cycle duration or twitch duration), maximum rate of development 214 (or maximum contraction slope), maximum rate of descent 216 (or maximum relaxation slope), and passive tension 218.

[0121] While the aforementioned features provide a rich characterization of each contraction-relaxation cycle (i.e., waveform 202), the noise level of the underlying signal often makes accurate extraction or measurement of these features difficult. For example, noise introduced into the underlying waveform due to measurement (e.g., noise introduced by sensor component 108) or transmission can lead to inaccuracies in identifying key points needed to extract the aforementioned features. These inaccuracies, in turn, limit the effectiveness and outcomes of downstream tasks utilizing such features. Furthermore, in vitro functional response data are typically limited in size, thus limiting the applicability of such data for training predictive models, such as machine learning regression or classification models.

[0122] This disclosure aims to address some (if not all) of the aforementioned problems by introducing a contraction-relaxation loop model. As described in more detail below, this model is used for noise filtering and enables the generation of high-fidelity and high-capacity synthetic training datasets. These high-fidelity and high-capacity synthetic training datasets can then be used to train, update, or otherwise improve artificial intelligence models, such as new or updated machine learning models, to produce more accurate and error-resistant outputs. Such outputs can be used downstream for the discovery, development, and / or manufacturing of high-quality pharmaceutical products, such as therapeutics. In some aspects, this disclosure proposes systems and methods for processing functional response waveforms to achieve efficient and effective feature extraction.

[0123] Systems and methods for processing functional response waveforms.

[0124] This disclosure presents systems and methods for processing functional response waveforms to achieve efficient and effective feature extraction.

[0125] Figure 3 Elements of a contraction-relaxation cycle model according to one embodiment of the present disclosure are illustrated. In some aspects, the contraction-relaxation cycle model may include a single contraction type contraction-relaxation cycle model.

[0126] Figure 3 The diagram illustrates the contraction function 302, relaxation function 304, and contraction-relaxation cycle model 306. For simplicity, the contraction-relaxation cycle model 306 is also referred to as a single contraction model, a single contraction type model, a single type model, or a single model. The contraction-relaxation cycle model 306 may include a combination or product of the contraction function 302 and the relaxation function 304. Thus, the contraction-relaxation cycle model 306 independently models the contraction and relaxation responses of an engineered organization. Specifically, the growth rate of the contraction response (i.e., the growth rate of the contraction function 302) and the growth rate or decay rate of the relaxation response (i.e., the growth rate or decay rate of the relaxation function 304) vary independently, allowing the contraction-relaxation cycle model 306 to model a wide variety of different response types. Therefore, the contraction-relaxation cycle model 306 can be used effectively and efficiently across a range of application areas involving various types of functional response waveforms. Additionally or alternatively, the growth rate of the contraction response (i.e., the growth rate of the contraction function 302) and the growth rate or decay rate of the single relaxation response (i.e., the growth rate or decay rate of the relaxation function 304) vary independently, thereby allowing the single contraction model 306 to model a wide variety of single contraction-relaxation responses.

[0127] According to one aspect of this disclosure, a contraction-relaxation cycle model (such as contraction-relaxation cycle model 306) can be fitted to in vitro functional response data of engineered tissues, such as the waveform of contractile force of engineered cardiac tissue. A noise-filtered representation of the in vitro functional response data is then generated from the contraction-relaxation cycle model, and features are accurately extracted from the noise-filtered representation.

[0128] The contraction and relaxation functions are logistic (S-shaped) functions, double exponential functions, or any other suitable functions. In one embodiment, the logistic function is used to model the force-based functional response waveform, while the double exponential function is used to model the calcium transient-based functional response waveform. Generally, the contraction-relaxation cycle model (e.g., contraction-relaxation cycle model 306) includes a contraction function. and relaxation function The product of:

[0129]

[0130] here, It is the maximum value of the contraction-relaxation cycle model, and It is along The offset of the contraction-relaxation cycle model applied to the axis. Figure 3 In the illustrated embodiment, the contraction function 302 is an ascending logic function of the following form:

[0131]

[0132] Furthermore, the relaxation function 304 is a descent logic function of the following form:

[0133]

[0134] Therefore, the contraction-relaxation cycle model 306 can be expressed as:

[0135]

[0136] here, and This parameterizes the contraction function 302 and corresponds accordingly to the midpoint of the contraction function 302 and the growth rate of the contraction function 302 (the reciprocal of the logical growth rate). Parameter and The relaxation function 304 is parameterized and correspondingly corresponds to the midpoint of the relaxation function 304 and the growth rate or decay rate of the relaxation function 304 (the reciprocal of the logical growth rate or decay rate).

[0137] Therefore, parameter set The contraction-relaxation cycle of engineered tissues is fully described, thus allowing the contraction-relaxation cycle model 306 to fit and approximate the contraction-relaxation cycle within real-world functional response data obtained from engineered tissues (e.g., artificial heart tissue).

[0138] Although Figure 3 The contraction-relaxation cycle model illustrated above can be used to model functional responses with a single contraction type, but not all functional responses follow a single contraction type. For example, arrhythmias such as atrial fibrillation or tachycardia can cause ectopic beats, resulting in irregular beating patterns. Therefore, functional response waveforms can include single contraction types (as described above regarding...). Figure 3 (as described above) and double-contraction types (as described below) Figure 25 The combination of the two (as described above).

[0139] Figure 4 Three example waveforms with corresponding contraction-relaxation cycle model fittings are shown.

[0140] Figure 4 The first waveform 402, the second waveform 404, and the third waveform 406 are shown. Figure 4 The first model-fitted waveform 408, the second model-fitted waveform 410, and the third model-fitted waveform 410 are further illustrated. Each of the model-fitted waveforms corresponds to a waveform generated from one of the first, second, or third waveforms by the contraction-relaxation cycle model. For example, the first model-fitted waveform 408 corresponds to a waveform generated from the contraction-relaxation cycle model fitted to the first waveform 402.

[0141] Figure 5A and Figure 5B The effect of each parameter of a contraction-relaxation cycle model according to an embodiment of the present disclosure is shown.

[0142] Figure 5A and Figure 5B Each plot in the diagram illustrates a contraction-relaxation cycle model with a varying parameter according to equation (4) above (e.g., a single contraction type contraction-relaxation cycle model). In some respects, as stated above, Figure 5A and Figure 5B The parameter changes shown in the single contraction-relaxation cycle model can also be applied to parameter changes in double contraction type models, for example, as described below. Figure 25 As stated above.

[0143] Figure 5A Plot A shows the parameter with respect to the maximum value. The contraction-relaxation cycle model of the change value. Figure 5A Plot B shows the parameters for offset. The contraction-relaxation cycle model of the change value. Figure 5A Plot C shows the parameters for the contraction midpoint. The contraction-relaxation cycle model of the change value. Figure 5A Plot D shows the parameters for the relaxation midpoint. The contraction-relaxation cycle model of the change value. Figure 5B Plot E shows the parameters for the contraction rate of growth. The contraction-relaxation cycle model of the change value. Figure 5B The plot F shows the parameters for relaxation rate. The contraction-relaxation cycle model of the change value.

[0144] The range and variation of shapes that can be captured by either or both of a single and / or dual contraction-relaxation cycle means that the model can accurately fit a range of different functional response data. Specifically, complex responses can be modeled efficiently and accurately by independently modeling the contraction and relaxation responses (e.g., in some cases, independently modeling the growth and decay of one or both responses within a given cycle). The contraction-relaxation model of this disclosure can therefore be fitted to a range of different functional response data by identifying the type of contraction within the functional response data to determine which type of contraction model to fit, and / or additionally or alternatively determining the parameters that best fit the model to the underlying data.

[0145] As described in this article, for Figure 26 An example of this process is shown.

[0146] As another example, any suitable parameter fitting technique can be used (such as least squares-based fitting or machine learning-based fitting, such as...). Figure 6 (or machine learning-based fitting as shown elsewhere in this article) to fit the contraction-relaxation model.

[0147] Figure 6 A parameter estimation model 600 for fitting parameters to a contraction-relaxation cycle model according to an embodiment of the present disclosure is shown.

[0148] The parameter estimation model 600 comprises a machine learning model in the form of a deep neural network. The parameter estimation model 600 includes a residual layer 602, a convolutional neural network 604, a dilated convolutional network 606, a cascaded block 608, a long short-term memory (LSTM) network 610, and a fully connected network 612. The parameter estimation model 600 receives an input vector 614 and produces an output vector 616.

[0149] Input vector 614 includes waveforms of the contraction and relaxation responses of the artificial tissue during a single contraction-relaxation cycle (e.g., Figure 2 The waveform 202 shown corresponds to the sequence (time series). The input vector 614 is fed into the residual layer 602, the convolutional neural network 604, and the dilated convolutional network 606. The residual layer 602 is a two-dimensional (2D) convolutional layer with 64 filters of size 1. The convolutional neural network 604 and the dilated convolutional network 606 will be discussed below. Figure 7A and Figure 7B To describe in more detail: The outputs of each of the residual layer 602, the convolutional neural network 604, and the dilated convolutional network 606 are cascaded at cascade block 608. Cascade block 608 includes: a depthwise cascaded layer that cascades the outputs of the residual layer 602, the convolutional neural network 604, and the dilated convolutional network 606; a batch normalization layer that normalizes the outputs of the depthwise cascaded layer; and an exponential linear unit layer that performs an identity operation on positive inputs and on negative inputs. conduct ,in The output of cascaded block 608 (i.e., the output of the exponential linear unit layer) is fed into the LSTM and network 610, as discussed below. Figure 7C The LSTM and network are described in more detail below. The output of the LSTM and network 610 is fed into the fully connected network 612, as described below. Figure 7D The fully connected network is described in more detail. The output vector 616 generated by the fully connected network 612 includes parameter values ​​for a contraction-relaxation cyclic model fitted to the waveform represented in the input vector 614.

[0150] Advantageously, by using dilated convolutions and LSTM networks, the parameter estimation model 600 can efficiently and accurately fit parameters to a waveform including a complete contraction-relaxation cycle. This efficiency and accuracy lead to more efficient use of computational resources and improved performance in downstream tasks that utilize the estimated parameters, such as drug discovery and development. The parameter estimation model also reduces errors that would otherwise be present in the original waveform, such as errors generated by the underlying computing device. As a result, the reduced error makes the parameter estimation model more efficient to operate, thereby reducing additional computational cycles for the underlying computing device (e.g., one or more processors and / or memory of the underlying device), thus saving processor and memory utilization on the device on which the parameter estimation model is executed.

[0151] In one implementation, the parameter estimation model 600 is trained using a synthetic dataset comprising 40,000 training samples and 10,000 validation samples. The synthetic dataset is used as described below. Figure 10The described method generates the model. Mini-batch gradient descent with a batch size of 128 is used with the ADAM solver to train the parameter estimation model 600. The ADAM solver has an initial learning rate of 1e-3, where early stopping is performed based on validation loss.

[0152] Additionally or alternatively, in some embodiments, Figure 6 The parameter estimation model 600 may include a prediction model. In such embodiments, the parameter estimation model 600 may be referred to herein as the prediction model 600.

[0153] In this respect, the prediction model 600 includes a machine learning model in the form of a deep neural network. The prediction model 600 includes a residual layer 602, a convolutional neural network 604, a dilated convolutional network 606, a cascaded block 608, a long short-term memory (LSTM) network 610, and a fully connected network 612. The prediction model 600 receives an input vector 614 and produces an output vector 616.

[0154] Prediction model 600 is trained for different purposes. As described in more detail below, by modifying the fully connected network 612, prediction model 600 is trained to predict the type of contraction or the parameters of a contraction-relaxation loop model. In the case of predicting model parameters, prediction model 600 is trained to predict either the parameters of a single contraction type model or the parameters of a double contraction type model. Therefore, Figure 6 As shown and about Figures 7A to 7D The detailed prediction model 600 can be trained as a parameter estimation model or a shrinkage-type classifier.

[0155] Input vector 614 includes a sequence (time series) corresponding to a waveform comprising at least one contraction response and at least one relaxation response of an artificial tissue. The waveform represented by input vector 614 has a contraction type, namely a single contraction type or a double contraction type. For a single contraction type, input vector 614 includes a time series corresponding to a single contraction response. For a double contraction type, the input vector includes a time series corresponding to both contraction responses.

[0156] Input vector 614 is fed into residual layer 602, convolutional neural network 604, and dilated convolutional network 606. In one implementation, residual layer 602 is a two-dimensional (2D) convolutional layer with 64 filters of size 1. Convolutional neural network 604 and dilated convolutional network 606 will be discussed below. Figure 7A and Figure 7BTo describe in more detail: The outputs of each of the residual layer 602, the convolutional neural network 604, and the dilated convolutional network 606 are cascaded at cascade block 608. Cascade block 608 includes: a depthwise cascaded layer that cascades the outputs of the residual layer 602, the convolutional neural network 604, and the dilated convolutional network 606; a batch normalization layer that normalizes the outputs of the depthwise cascaded layer; and an exponential linear unit layer that performs an identity operation on positive inputs and on negative inputs. conduct In one implementation, The output of cascaded block 608 (i.e., the output of the exponential linear unit layer) is fed into LSTM network 610, as discussed below. Figure 7C The LSTM network will be described in more detail below. The output of the LSTM network 610 is fed into the fully connected network 612, as described below. Figure 7D Describe the fully connected network in more detail.

[0157] As described in more detail below, the architecture of the fully connected network 612 and the final activation function are modified according to the application of the prediction model 600. Therefore, when the prediction model 600 is used for contraction type classification, the output vector 616 generated by the fully connected network 612 includes a vector or scalar score associated with the contraction type classification. For example, the output vector 616 may include the probability that the waveform represented by the input vector 614 is a single contraction type or a double contraction type. When the prediction model 600 is used for parameter estimation, the output vector 616 generated by the fully connected network 612 includes a vector of parameter values ​​for a single or double contraction-relaxation recurrent model fitted to the waveform represented by the input vector 614.

[0158] Beneficially, by using dilated convolutions and LSTM networks, the prediction model 600 is able to accurately predict the type of waveform contraction and efficiently fit parameters to waveforms including single or double contraction types. These benefits result in more efficient use of computational resources and improved performance in downstream tasks such as drug discovery and development.

[0159] In one implementation, a training dataset of labeled waveforms is used to train a prediction model 600 for a contraction type classification task. The training dataset uses approximately 10,000 training samples and approximately 2,000 validation samples. The training dataset includes approximately 6,000 manually labeled single-contraction-type contraction-relaxation cyclic waveforms and 6,000 manually labeled double-contraction-type contraction-relaxation cyclic waveforms.

[0160] In another implementation, prediction model 600 is trained using a synthetic dataset comprising 40,000 training samples and 10,000 validation samples. The synthetic dataset is used as described below. Figure 29The described method generates the data. When training prediction model 600 for contraction type classification, the synthetic dataset includes 25,000 single-contraction type waveforms and 25,000 double-contraction type waveforms. For the contraction type classification task, each waveform in the synthetic dataset is associated with a corresponding label indicating whether the waveform is a single-contraction type or a double-contraction type. When training prediction model 600 for parameter estimation, the synthetic data includes single-contraction type waveforms or double-contraction type waveforms and corresponding parameters for each waveform. Therefore, for parameter estimation, one model is trained to predict the parameters for single-contraction type waveforms, and another model is trained to predict the parameters for double-contraction type waveforms.

[0161] For both the classification and parameter estimation tasks in the two implementations described above, mini-batch gradient descent with a batch size of 128 is used in conjunction with the ADAM solver to train the prediction model 600. The ADAM solver has an initial learning rate of 1e-3, where early stopping is performed based on validation loss.

[0162] Figure 7A A convolutional neural network 700 according to an embodiment of the present disclosure is shown, the convolutional neural network being formed Figure 6 It is part of the parameter estimation model 600 (e.g., the prediction model).

[0163] Convolutional Neural Network 700 corresponds to Figure 6 The parameter estimation model 600 (e.g., prediction model) shown is a convolutional neural network 604. The convolutional neural network 700 includes a first initial network 702, a cascaded block 704, and a second initial network 706. The first initial network 702 includes a first initial module 708-1, a second initial module 708-2, a third initial module 708-3, a fourth initial module 708-4, and a fifth initial module 708-5. Each of these initial modules includes the same architecture, as illustrated with respect to the first initial module 708-1, which includes a first 2D convolutional layer 710, a batch normalization layer 712, an exponential linear unit layer 714, and a second 2D convolutional layer 716. The cascaded block 704 includes a depthwise cascaded layer 722, a batch normalization layer 724, and a layer with... The exponential linear unit layer 726. The second initial network 706 has the same architecture as the first initial network 702 (i.e., the same number and configuration of initial modules).

[0164] Each initial module in the first initial network 702 and the second initial network 706 shares the same architecture (i.e., as per the...). Figure 7AThe first initialization module 708-1 shown is exemplified here, and is parameterized according to the filter size of the second 2D convolutional layer (e.g., the second 2D convolutional layer 716). The filter sizes of the second 2D convolutional layers in the first initialization module 708-1, the second initialization module 708-2, the third initialization module 708-3, the fourth initialization module 708-4, and the fifth initialization module 708-5 are respectively... The first 2D convolutional layer (e.g., first 2D convolutional layer 710) includes 64 filters of size 1. The batch normalization layer 712 includes a standard batch normalization layer, and the exponential linear unit layer 714 includes filters with... The standard exponential linear unit layer.

[0165] Figure 7B A dilated convolutional network 730 according to an embodiment of the present disclosure is shown, the dilated convolutional network forming Figure 6 It is part of the parameter estimation model 600 (e.g., the prediction model).

[0166] The dilated convolutional network 730 corresponds to Figure 6 The parameter estimation model 600 (e.g., prediction model) shown is a dilated convolutional network 606. The dilated convolutional network 730 includes a first dilated convolutional module 730-1, a second dilated convolutional module 730-2, a third dilated convolutional module 730-3, a fourth dilated convolutional module 730-4, a fifth dilated convolutional module 730-5, a sixth dilated convolutional module 730-6, and a final dilated 2D convolutional layer 732.

[0167] The architecture of each dilated convolutional module is the same and... Figure 7B The architecture of the second dilated convolutional module 730-2 is illustrated in the example. This second dilated convolutional module includes a dilated 2D convolutional layer 734, a batch normalization layer 736, and a layer with... The exponential linear unit layer 738. Each dilated convolutional module is parameterized by the size of the dilation operation performed on the dilated 2D convolutional layer. The dilation sizes of the dilated 2D convolutional layers for the first dilated convolutional module 730-1, the second dilated convolutional module 730-2, the third dilated convolutional module 730-3, the fourth dilated convolutional module 730-4, the fifth dilated convolutional module 730-5, and the sixth dilated convolutional module 730-6 are respectively... .

[0168] The dilation operation is illustrated in a first extension 740, which shows portion 742 of the input to the second dilated convolution module 730-2 and portion 744 of the dilated 2D convolutional layer 734. Here, the input corresponds to the output of the exponential linear unit layer of the first dilated convolution module 730-1. Dilated convolution advantageously increases the receptive field of a layer (i.e., the time period the layer processes) without increasing the number of parameters. The dilated 2D convolutional layer 734 performs this extension by inserting zeros between each filter element. As shown in portion 744 of the dilated 2D convolutional layer 734, a dilation size or factor of 2 causes a single zero to be inserted between each filter element (while for a dilation size or factor of 4, three zeros are inserted).

[0169] The final dilated 2D convolutional layer 732 corresponds to the standard dilated 2D convolutional layer with a hole size or factor of 64.

[0170] Figure 7C A Long Short-Term Memory (LSTM) network 746 according to an embodiment of the present disclosure is shown, the LSTM network being formed Figure 6 It is part of the parameter estimation model 600 (e.g., the prediction model).

[0171] LSTM network 746 corresponds to Figure 6 The LSTM network 610 is shown as the parameter estimation model 600 (e.g., a prediction model). The LSTM network 746 includes an LSTM module 748, a first dense module 750, and a second dense module 752. The LSTM network 746 includes a flattening layer 754, a first bidirectional LSTM layer 756, a first dropout layer 758, a second bidirectional LSTM layer 760, and a second dropout layer 762. The first dense module 750 includes a first dense layer 764, having… The first exponential linear unit layer 766, the first discard layer 768, the second dense layer 770, have The second exponential linear unit layer 772, the second discard layer 774, and the planarization layer 776 are included. The architecture of the second dense module 752 is the same as that of the first dense module 750.

[0172] LSTM module 748 attempts to learn the bidirectional long-term dependencies between time steps of waveform (i.e., time series) data via bidirectional LSTM layers. In this way, LSTM module 748 is able to learn from the complete waveform at each time step. Both the first bidirectional LSTM layer 756 and the second bidirectional LSTM layer 760 comprise 64 hidden units. The first dropout layer 758 and the second dropout layer 762 help reduce overfitting by randomly setting the input to zero with a probability of 0.1.

[0173] The first dense layer 764 and the second dense layer 770 of the first dense module 750 have an output of size 256. The first dropout layer 768 and the second dropout layer 774 randomly set the input to zero with a probability of 0.5. As stated above, the architecture of the second dense module 752 is the same as that of the first dense module 750.

[0174] Figure 7D A fully connected network 778 according to an embodiment of the present disclosure is shown, the fully connected network forming Figure 6 It is part of the parameter estimation model 600 (e.g., the prediction model).

[0175] Fully connected network 778 corresponds to Figure 6 The parameter estimation model 600 (e.g., prediction model) shown is a fully connected network 612. The fully connected network 778 includes a cascaded layer 780, a first dense block 782-1, a second dense block 782-2, a third dense block 782-3, a dense layer 784, and / or an output layer 786.

[0176] The first dense block 782-1, the second dense block 782-2, and the third dense block 782-3 all include the same architecture. As shown with respect to the first dense block 782-1, this architecture includes a dense layer 786, having… The system consists of an exponential linear unit layer 788 and a dropout layer 790. The dense layers of the first dense block 782-1 and the second dense block 782-2 have an output size of 512. The dense layer of the third dense block 782-3 has an output size of 1024. The dropout layers of the first and second dense blocks 782-1 randomly set the input to zero with a probability of 0.5. The dropout layer of the third dense block 782-3 randomly sets the input to zero with a probability of 0.1.

[0177] In one example, the dense layer 784 has an output size corresponding to the number of parameters the parameter estimation model will fit. The output size is set to 6 to learn the above about Figure 3 The complete parameter set of the described contraction-relaxation cycle model Therefore, the output of dense layer 784 corresponds to the output of the parameter estimation model (e.g., Figure 6 The output vector 616 of the parameter estimation model 600 shown is shown.

[0178] In one embodiment, optimization methods are used to further refine the estimated parameters generated by the parameter estimation model. The following will discuss... Figure 9 In more detail, the optimization methods used include simplex search algorithms, such as the Nelder-Mead method.

[0179] Once the contraction-relaxation loop model has been fitted to the data, the waveform can be reconstructed from the model to generate a noise-filtered representation of the waveform. For example, multiple points (e.g., 100, 200, 500, 1000, etc.) are sampled from the contraction-relaxation loop model of the data within the time period fitted to the first waveform, and the contraction-relaxation loop model is fitted to this first waveform. Advantageously, this allows for the generation of high-resolution waveform data from the contraction-relaxation loop, which helps ensure more accurate features of the functional response of the engineered organization are extracted. This, in turn, helps improve the accuracy and efficiency of downstream tasks involving functional response features. The contraction-relaxation loop model also reduces errors that would otherwise be present in the original waveform, such as errors generated by the underlying computing device. As a result, the reduced error makes the contraction-relaxation loop model more efficient, thereby reducing additional computational loops for the underlying computing device (e.g., one or more processors and / or memory of the underlying device), thereby saving processor and memory utilization on the device on which the contraction-relaxation loop model is executed.

[0180] In another example, when the parameter estimation model 600 (e.g., the prediction model 600) (of which the fully connected network 778 is part) is used for contraction type classification, the dense layer 784 has two nodes, one for each contraction type, such that the output for each node corresponds to the probability that the input waveform belongs to the corresponding contraction type. When the prediction model 600 is used for parameter estimation, the dense layer 784 has an output size (i.e., the number of nodes) corresponding to the number of parameters the parameter estimation model is to fit. For a single contraction type model, the output size (i.e., the number of nodes in the dense layer 784) is set to 6 to learn the above regarding... Figure 3 The complete parameter set of the single-contraction type contraction-relaxation cycle model is described. For the double-shrinkage type model, the output size (i.e., the number of nodes in the dense layer 784) is set to 11 to learn the above about Figure 25 The complete parameter set of the described double-contraction type contraction-relaxation cycle model .

[0181] The output of dense layer 784 corresponds to the output of the parameter estimation model (e.g., Figure 6 The output vector 616 of the prediction model 600 shown is shown.

[0182] When the prediction model described above (e.g., Figure 6 When the prediction model 600 is used for contraction type classification, the predicted contraction type is used to determine which of the two parameter estimation models will fit the waveform (e.g., by...). Figure 26 (Example of system 2600). In one embodiment, the predicted contraction type provides a phenotype for the state of the engineered tissue (e.g., disease state, treatment state, etc.), such as Figure 27 exemplified in .

[0183] We will now turn to describe the method used to process the functional response waveform using the contraction-relaxation cycle model described above.

[0184] Figure 8 A method 800 for processing a functional response waveform according to one aspect of this disclosure is shown.

[0185] Method 800 includes the following steps: obtaining a first waveform (802); fitting a model (804) to the first waveform; and generating a second waveform (806) from the model. Method 800 further includes the optional steps of extracting feature values ​​(808) from the second waveform and outputting feature values ​​(810). In some aspects, the extracted feature values ​​can be used to train a machine learning model, whereby by using the feature values ​​of the second waveform (which itself has been filtered to remove error-prone data (e.g., noise)), the machine learning model will have increased predictive accuracy, resulting in improved predictive accuracy. Predictions can then be generated by feeding a dataset of human tissue into the machine learning model. The prediction may define one or more characteristics of the human tissue and may correspond to at least one of one or more feature values. In this way, feature values ​​of the waveform with reduced error (e.g., the second waveform) can be used to train or generate highly accurate machine learning models or other artificial intelligence models. In one embodiment, method 800 is... Figure 1 The model fitting unit 104-1 of the control unit 104 shown is used.

[0186] Generally, method 800 is used to generate a noise-filtered or noise-suppressed representation of a contraction-relaxation cyclic waveform (functional response waveform). In an embodiment, the contraction-relaxation cyclic waveform is derived from hardware such as a bioreactor (e.g., Figure 1 The bioreactor 102 shown illustrates that, due to factors such as sensor variability, signal transmission, and signal conversion, the hardware can introduce noise into the waveform. Therefore, the contraction-relaxation cycle waveform includes a potentially noisy representation of the functional response of the engineered tissue during a single contraction-relaxation cycle. Advantageously, method 800 efficiently generates a noise-filtered representation of a single contraction-relaxation cycle, thereby improving the accuracy of the extracted features, which in turn helps improve the performance of downstream tasks utilizing such features.

[0187] At step 802, a first waveform is obtained, comprising the contraction and relaxation responses of the artificial tissue during a single contraction-relaxation cycle.

[0188] The first waveform (i.e., a single contraction-relaxation cycle waveform, a single-cycle waveform, a twitching cycle waveform, a peak, or a functional response waveform) captures the functional response of the artificial tissue during a single contraction-relaxation cycle, which includes a contraction period (i.e., a period in which the artificial tissue contracts and generates tension) and a relaxation period (i.e., a period in which the artificial tissue returns to its normal state or length). Thus, the contraction response of the first waveform includes the functional response of the artificial tissue during the contraction period of the contraction-relaxation cycle, and the relaxation response includes the functional response of the artificial tissue during the relaxation period of the single contraction-relaxation cycle. The functional response is the contractile force of the artificial tissue (e.g., a contractile force measured using data obtained from sensor assembly 108 of bioreactor 102, in which the artificial tissue grows or is maintained). Alternatively, the functional response is a contractile displacement, a transient calcium response, or a change in membrane potential.

[0189] Artificial or engineered tissues include engineered muscle tissue, such as engineered heart tissue or engineered skeletal muscle tissue. In one embodiment, a first waveform is emitted from a bioreactor comprising the artificial tissue (e.g., Figure 1 The first waveform is obtained from a bioreactor 102 shown. Thus, the first waveform is obtained from a waveform comprising multiple contraction-relaxation cycles of artificial tissue obtained from the bioreactor. In one embodiment, an extraction method (such as described below) is used. Figure 12 A more detailed description of method 1200) is provided for obtaining or extracting a first waveform from a waveform.

[0190] At step 804, the model is fitted to the first waveform. This model, or contraction-relaxation cyclic model, independently parameterizes the growth of the contraction and relaxation responses. Thus, the model does not assume that the underlying functional response is symmetric. This allows the model to efficiently and accurately fit a range of functional responses from a wide variety of engineered tissue types. More efficient and accurate model fitting helps generate more precise features, which in turn leads to better data.

[0191] The model includes a contraction function. and relaxation function In one embodiment, the contraction function is an ascending logic function with a positive growth rate, and the relaxation function is a descending logic function with a negative growth rate. Therefore, the contraction response of the contraction-relaxation loop is modeled by the ascending logic function, and the relaxation response of the contraction-relaxation loop is modeled by the descending logic function. The model comprises the product of the ascending and descending logic functions.

[0192] The contraction-relaxation cycle model includes several parameters associated with the contraction and relaxation responses. These parameters include a maximum value parameter. Ascent rate parameter descent rate parameter , Offset parameter ,rise Offset parameter and decline Offset parameter The above text is about Figure 5A and Figure 5B The relationship between each of these parameters and the overall model is shown and described in more detail.

[0193] In one embodiment, data from a machine learning model (e.g., Figure 6 The predicted parameter values ​​obtained from the parameter estimation model (shown) are used to fit the model to the first waveform, as described below. Figure 9 Method 900 is described in more detail.

[0194] At step 806, a second waveform is generated from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform. For example, multiple points (e.g., 100, 200, 500, 1000, etc.) are sampled from a shrink-relaxation cyclic model of data within the time period fitted to the first waveform, and the shrink-relaxation cyclic model is fitted to that first waveform. Advantageously, this allows for the generation of high-resolution waveform data from the shrink-relaxation cycle, which helps ensure more accurate features of the functional response of engineered tissues are extracted. This, in turn, helps improve the accuracy and power of downstream tasks involving functional response features.

[0195] Optionally, a second waveform may be output. In one embodiment, outputting the second waveform includes storing or saving the second waveform to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the second waveform includes transmitting the second waveform via a network (e.g., a local area network, a wide area network, etc.) or displaying the waveform for user review.

[0196] At any optional step of Extraction 808, one or more feature values ​​are extracted from the second waveform.

[0197] The second waveform is a noise-filtered representation of the first waveform (i.e., a noise-filtered representation of the contraction-relaxation cycle), and thus, more accurate values ​​of the features of the underlying functional response to be extracted are obtained. One or more feature values ​​include one or more of the following: twitching or peak amplitude values ​​(i.e., Figure 2 The peak amplitude shown is 206), and the contraction time value (i.e., Figure 2 The time to reach the peak amplitude (208) and the maximum contraction slope value (i.e., Figure 2 The maximum development rate shown is 214), and the relaxation time value (i.e., Figure 2 The time to reach the peak descent (210) and the maximum relaxation slope value (i.e., Figure 2 The maximum rate of decline (216) and the duration of the convulsion (i.e., Figure 2 The duration shown is 212).

[0198] Once extracted, one or more eigenvalues ​​can be used as quantitative descriptors of the contraction-relaxation cycle, and thus provide a numerical representation of the functional response of the artificial tissue during the contraction-relaxation cycle. (See also: Regarding...) Figure 14 In more detail, such features are used for a variety of downstream processing tasks, such as effect identification in drug discovery and development.

[0199] At an optional step of output 810, one or more feature values ​​extracted from the second waveform are output. In one embodiment, outputting one or more feature values ​​includes storing or saving one or more feature values ​​to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting one or more feature values ​​includes transmitting one or more feature values ​​via a network (e.g., a local area network, a wide area network, etc.) or displaying one or more feature values ​​for user review.

[0200] Figure 9 A method 900 for fitting a model to a functional response waveform according to an embodiment of the present disclosure is shown.

[0201] In one embodiment, method 900 is performed as part of the fitting step 804 of method 800. Method 900 includes a step of predicting 902 multiple values ​​for multiple parameters, and further includes an optional step of optimizing 904 multiple values.

[0202] The steps in method 900 are used to predict parameter values ​​for the contraction-relaxation cycle model from the first waveform. The trained machine learning model (e.g., Figure 6 The parameter estimation model shown is used to predict parameter values ​​so that the fitted model closely approximates the first waveform.

[0203] At step 902, multiple values ​​are predicted for multiple parameters of the model, such that the model fitted to the first waveform includes multiple values ​​for multiple parameters.

[0204] At an optional step of optimization 904, the multiple values ​​determined at prediction 902 are optimized by minimizing the error between the first waveform and the model fitted to the first waveform (i.e., using multiple values ​​for multiple parameters of the model).

[0205] Given multiple values ​​for the model determined at step 902 of the prediction. In the case of optimizing step 904, seek multiple updated values. , making .here, It is a cost or loss function that measures the first waveform. According to the parameter value set Model fitted to the first waveform The error between them. A lower value indicates a better fit of the model to the first waveform. In one embodiment, the cost function It is the root mean square error:

[0206]

[0207] In one embodiment, it corresponds to:

[0208]

[0209] The optimization of this is a multidimensional problem because the model involves multiple parameters. Therefore, making Minimization requires fitting multiple parameters simultaneously. To perform this minimization, in one embodiment, a simplex search algorithm (such as the Nelder-Mead method) is used to optimize multiple values. Advantageously, performing optimization after obtaining initial predictions of the parameters helps to achieve an accurate model fit, while also making better use of processing resources, because the optimization process starts with a solution expected to be close to the optimal solution.

[0210] Figure 10 A method 1000 for training a parameter estimation model using synthetic training data is shown, according to one aspect of this disclosure.

[0211] Method 1000 includes the following steps: obtaining 1002 multiple waveforms; extracting 1004 multiple parameter sets; determining 1006 the parameter set distribution; generating 1008 a synthetic training dataset; and training 1010 a prediction model on the synthetic training dataset. In one embodiment, method 1200 is... Figure 1 The control unit 104 or its subunit shown is used for this purpose.

[0212] In many cases, the effectiveness of predictive models (e.g., machine learning models such as deep neural networks) is limited by the quantity and quality of available training data. Without a large volume of high-quality training data, predictive models will often fail to produce sufficient output. In this disclosure, this problem leads to inaccurate shrink-relaxation models being fitted, which subsequently reduces the effectiveness and applicability of such models in practice for tasks such as drug discovery and development. Method 1000 aims to address this problem by generating high-fidelity synthetic training data, thereby allowing for the generation of a virtually unlimited amount of data. This helps improve the performance of the trained predictive model, which in turn improves the accuracy of models fitted using that predictive model. This improvement in accuracy helps drive improvements to downstream tasks that utilize features extracted from such models.

[0213] At step 1002, multiple waveforms are obtained. These waveforms encompass the functional response of one or more artificial tissues during a single contraction-relaxation cycle.

[0214] Multiple waveforms correspond to real or guided data from which synthetic training datasets will be generated. Each waveform corresponds to a time series representing the values ​​of functional responses (e.g., contractile force, calcium transients, etc.) of an artificial or engineered tissue within a single contraction-relaxation cycle. Therefore, each waveform includes a contraction cycle and a relaxation cycle, and can be defined according to the above regarding... Figure 3 The described contraction-relaxation model is parameterized.

[0215] To help ensure that the synthetic training data represents the broadest possible population of contraction-relaxation cycles, multiple waveforms are preferably obtained from a variety of artificial tissues under a series of different conditions. Artificial tissues include one or more engineered muscle tissues, such as engineered cardiac tissue and / or engineered skeletal muscle tissue. Diversity is achieved by obtaining waveforms from artificial tissues across a range of different cell lines, disease states, and treatment methods. Alternatively, the range of conditions within the multiple waveforms is restricted, allowing the synthetic data and subsequent parameter estimation models to be fine-tuned for specific applications. For example, the multiple waveforms may be restricted to vector-processed waveforms to generate control parameter estimation models, or they may be restricted to specific tissue types (e.g., engineered cardiac tissue) to generate tissue-specific parameter estimation models. Advantageously, this helps improve the performance of the parameter estimation models when it is known which type of waveform the model will be used with.

[0216] At step 1004, multiple parameter sets are extracted from multiple waveforms. The parameter sets in the multiple parameter sets represent the corresponding waveforms in the multiple waveforms.

[0217] The parameter set includes the parameters of the contraction-relaxation cycle model (e.g., as mentioned above regarding...). Figure 3 (Described). In one embodiment, the set of parameters associated with the waveform includes a maximum parameter value ( ), offset parameter value ( ), midpoint of contraction parameter value ( ), contraction growth rate parameter value ( ), relaxation midpoint parameter value ( ) and relaxation rate parameter value ( ).

[0218] Multiple parameter sets are extracted using supervised, unsupervised, or semi-supervised methods. With supervised methods, the parameter sets are manually fitted to each waveform. For example, a first waveform is presented to the user, who adjusts the values ​​of the parameters within the parameter set until a second waveform, generated by a contraction-relaxation loop model fitted based on that parameter set, closely matches the first waveform. The final parameter set producing the closely matched second waveform is then used as the parameter set in one of the multiple parameter sets associated with the first waveform. With unsupervised methods, the parameter set is automatically fitted to the waveform (e.g., using the parameters described above). Figure 8 and Figure 9 (The method described). In one embodiment, the unsupervised method utilizes a trained machine learning model to predict parameter set values ​​for a waveform. According to the semi-supervised method, the automatically determined parameter set obtained from the unsupervised method is manually reviewed and refined by one or more users.

[0219] At step 1006, the parameter set distribution is determined from multiple parameter sets.

[0220] The multiple parameter sets extracted at step 1004 of method 1000 include multiple values ​​for each parameter of the contraction-relaxation cycle model. For example, if 100 parameter sets are extracted, there will be 100 parameter values ​​extracted for each parameter of the contraction-relaxation cycle model. At step 1006, a distribution or value distribution is determined for each parameter of the model. In one embodiment, the distribution is determined independently for each parameter. Alternatively, a multivariate distribution is determined for the multiple parameters forming the parameter sets.

[0221] Any suitable method can be used to determine the parameter set distribution, such as histogram-based methods, density estimation methods, and clustering methods. In one embodiment, a kernel density estimation (KDE) method is used to determine the parameter set distribution, which utilizes a kernel and bandwidth parameters to estimate the parameter set distribution. In one implementation, a normal (Gaussian) kernel is used in conjunction with bandwidth selected using cross-validation or a bandwidth selection method such as the Scott rule or Silverman rule.

[0222] Once the distribution of the parameter set has been determined, the parameter set can be obtained by sampling from that distribution (i.e., sampling from each individual distribution or from the joint distribution).

[0223] At step 1008, a synthetic training dataset is generated. Each element of the synthetic training dataset includes a synthetic waveform and a corresponding parameter set used to generate the synthetic waveform. The corresponding parameter set is obtained from a parameter set distribution.

[0224] The synthetic training dataset is generated by repeatedly sampling the parameter set for the contraction-relaxation recurrent model from the parameter set distribution (as described above) and generating a corresponding waveform for each of the sampled parameters. In this way, training datasets of any size can be generated efficiently (e.g., 1000, 10000, 100000 training data elements, etc.). The synthetic data will also closely approximate real waveform data because the parameter set distribution is modeled on real-world data.

[0225] Optionally, a noise component is added to each of the waveforms in the synthetic training dataset. The noise component is determined via a uniform distribution derived from multiple waveforms.

[0226] At step 1010, the prediction model is trained using a synthetic training dataset. The prediction model is trained to estimate the output parameter set from the input waveform. In one embodiment, the prediction model corresponds to the above regarding... Figure 6 The parameter estimation model 600 is described. As stated above, in one embodiment, training the parameter estimation model 600 includes using mini-batch gradient descent with a batch size of 128 and an ADAM solver. The ADAM solver has an initial learning rate of 1e-3, where early stopping is performed based on validation loss. Further details regarding the training of the parameter estimation model 600 at step 1010 are described above. Figure 6 The description is given.

[0227] Figure 11 A method 1100 for predicting a set of parameter values ​​for a contraction-relaxation cycle model using a synthetically trained prediction model is shown according to an embodiment of the present disclosure.

[0228] Method 1100 includes the steps of obtaining 1102 a first waveform and predicting 1104 a first set of parameter values. Method 1100 also includes the optional step of outputting 1106 the first set of parameter values. In one embodiment, method 1100 is... Figure 1 The control unit 104 shown or its sub-units (such as model fitting unit 104-1) are used for this purpose.

[0229] At step 1102, a first waveform is obtained. The first waveform includes the functional response of the first artificial tissue.

[0230] At the step of predicting 1104, a prediction model trained on the synthetically generated training dataset is used (as mentioned above). Figure 10 The first parameter value set is predicted from the first waveform.

[0231] At an optional step of output 1106, a first set of parameter values ​​is output. In one embodiment, outputting the first set of parameter values ​​includes storing or saving the first set of parameter values ​​to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the first set of parameter values ​​includes transmitting the first set of parameter values ​​via a network (e.g., a local area network, a wide area network, etc.) or displaying the first set of parameter values ​​for user review.

[0232] The above text is about Figures 8 to 10 The described method fits a contraction-relaxation cycle model to a waveform that includes a single functional response of an artificial tissue (i.e., a single contraction-relaxation cycle). In practice, waveforms typically contain multiple contraction-relaxation cycles, which may require multiple model fittings. To analyze and process such waveforms, each individual contraction-relaxation cycle is extracted from the waveform before model fitting.

[0233] Figure 12 A method 1200 for extracting a contraction-relaxation cyclic waveform according to one aspect of this disclosure is shown.

[0234] Method 1200 includes the following steps: obtaining a first waveform 1202; convolving the first waveform with a pulse train 1204; identifying a first position 1206; and extracting a second waveform from the first waveform at the first position 1208. Method 1200 also includes the optional step of outputting the second waveform 1210. In one embodiment, method 1200 is performed by... Figure 1 The signal processing unit 104-2 of the control unit 104 shown performs this function.

[0235] Generally, Method 1200 extracts a single contraction-relaxation cycle waveform (i.e., a single cycle, twitching cycle, peak, or functional response) from a larger waveform that includes multiple contraction-relaxation cycles. The larger waveform can be derived from hardware devices such as bioreactors (i.e., Figure 1 The waveform obtained from the bioreactor 102 shown typically corresponds to the functional response of an artificial tissue under specific conditions. For example, a larger waveform could include the contractile response of engineered cardiac tissue stimulated at a frequency of 1 Hz over a 30-second time period. In this example, the larger waveform would include approximately 30 peaks or 30 contraction-relaxation cycles, corresponding to the contractions of the engineered cardiac tissue in response to electrical stimulation. Method 1200 provides an efficient and accurate mechanism for extracting each contraction-relaxation cycle from the larger waveform, such that these sub-waveforms can then be used for further processing and analysis (e.g., fitting a contraction-relaxation cycle model to these waveforms and extracting relevant features, as described above regarding...). Figure 8 (as described).

[0236] At step 1202, a first waveform is obtained. The first waveform includes multiple functional responses of the artificial tissue stimulated at a first frequency.

[0237] In one embodiment, the first waveform originates from a bioreactor (e.g., Figure 1 The bioreactor 102 shown provides an environment in which artificial or engineered tissues are grown / maintained. As previously stated, artificial tissues include engineered muscle tissues, such as engineered heart tissues or engineered skeletal muscle tissues.

[0238] As mentioned above Figure 1 The procedure involves applying electrical stimulation to the cell culture during maturation and, once matured, to the artificial tissue to mimic the natural physiological environment of the artificial tissue, thereby allowing measurement of the artificial tissue's functional response to the stimulation. The first waveform obtained at step 1202 includes the functional response of the artificial tissue to stimulation at a first frequency. In one embodiment, the first frequency or pacing frequency for stimulating the artificial tissue is 0.1 Hz to 20 Hz. In another embodiment, the first frequency is 1 Hz to 6 Hz.

[0239] Optionally, method 1200 includes a step of stimulating (not shown) an artificial tissue at a first frequency prior to the step of obtaining the first waveform 1202. For example, stimulating the artificial tissue into a bioreactor containing the artificial tissue (e.g., Figure 1 The bioreactor 102 in the middle sends instructions (e.g., Figure 1 Instruction 126) or command in order to cause the bioreactor to stimulate the artificial tissue at a first frequency.

[0240] At step 1204 of convolution, the first waveform is convolved with the pulse train to generate the convolved waveform. The pulse train is generated at a first frequency.

[0241] A pulse train corresponds to an idealized representation of the functional response of an artificial tissue at a first frequency. It is well known that a pulse train, or pulse wave, comprises... Duration of frequency The waveform of a non-sinusoidal (rectangular) pulse or wave, in which This is the period of the pulse train. Therefore, the duty cycle of the pulse train is... Therefore, the pulse train convolved with the first waveform includes pulses with periodicity. and duration A rectangular pulse sequence, in which It is the first frequency.

[0242] Optionally, method 1200 further includes a step of generating (not shown) a pulse train at a first frequency prior to the step of performing convolution 1204.

[0243] Given a first waveform and pulse train In the case of convolution at step 1204, the convolution is performed. for:

[0244]

[0245] At step 1206, a first position is identified that is associated with the maximum value of the convolved waveform. This first position corresponds to the expected position of the first contraction-relaxation cycle.

[0246] The maximum value of the convolved waveform corresponds to the optimal alignment point between the first waveform and the pulse train. Thus, the location of the maximum value of the convolved waveform is used to identify the most likely location of a single contraction-relaxation cycle within the first waveform. The location of the maximum value of the convolved waveform also provides an anchor point from which other contraction-relaxation cycles can be extracted from the first waveform.

[0247] At step 1208, a second waveform is extracted from a first position of the first waveform. The second waveform includes a first contraction-relaxation cycle and has a first duration proportional to the first frequency.

[0248] The first position marked at step 1206 corresponds to the most likely location of a single contraction-relaxation cycle within the first waveform. Therefore, the second waveform extracted from the first position includes this contraction-relaxation cycle. Since the first waveform corresponds to the functional response of the artificial tissue when stimulated at a predetermined frequency, the duration or length of the second waveform is proportional to that frequency. For example, if the artificial tissue is stimulated at a frequency of 1 Hz, the first duration will be 1 s; if the artificial tissue is stimulated at a frequency of 2 Hz, the first duration will be 0.5 s, and so on.

[0249] Therefore, the second waveform corresponds to a window or subframe within the first waveform, the window or subframe having a length corresponding to the first duration. In one embodiment, the second waveform is centered at a first position such that the midpoint of the second waveform is aligned or substantially aligned with the first position of the first waveform.

[0250] At an optional step of output 1210, a second waveform is output. In one embodiment, outputting the second waveform includes storing or saving the second waveform to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the second waveform includes transmitting the second waveform via a network (e.g., a local area network, a wide area network, etc.) or displaying the waveform for user review.

[0251] In one embodiment, outputting the second waveform includes outputting the second waveform to another process or method of this disclosure. For example, the second waveform may be output to the method 800 described above, such that the step of obtaining 802 includes obtaining the second waveform from method 1200.

[0252] Figure 13 A method 1300 for extracting additional contraction-relaxation cycles from a waveform is shown according to an embodiment of the present disclosure.

[0253] Method 1300 includes the steps of identifying a second position 1302, extracting a third waveform 1304 from a first waveform at the second position, and optionally outputting a third waveform 1306. Method 1300 is performed after method 1200. Specifically, method 1300 may be performed after the step of identifying the first position 1206, and may be performed in parallel with the step of extracting the second waveform 1208. In one embodiment, method 1300 is... Figure 1 The signal processing unit 104-2 of the control unit 104 shown performs this function.

[0254] Method 1300 is used to extract additional contraction-relaxation cycle waveforms from the first waveform. Advantageously, the extraction performed in method 1300 is efficient and highly parallel because method 1300 utilizes prior information about the expected location of the contraction-relaxation cycle within the first waveform, thereby enabling the independent extraction of the contraction-relaxation cycle.

[0255] At step 1302, the second position is identified based on the first position and the first frequency. The second position corresponds to the expected position of the second contraction-relaxation cycle.

[0256] The first position (identified at step 1206 of method 1200) corresponds to the optimal alignment between the first waveform and the pulse train. Therefore, the first position can be understood as the most likely location of a contraction-relaxation cycle within the first waveform. Since the first waveform includes the functional response of the artificial tissue at a predetermined frequency (i.e., the first frequency), other contraction-relaxation cycles related to the functional response of the artificial tissue are likely located at positions spaced apart from the first position. Therefore, the first position can serve as an anchor point within the first waveform from which other contraction-relaxation cycle waveforms can be extracted.

[0257] The second position corresponds to the intended position of the second contraction-relaxation cycle and will be spaced from the first position by a distance proportional to the first frequency. Specifically, the first position within a given first waveform In the case of the second position It will be ,in It is the step size factor, and It is the period of the pulse train. Therefore, the scaling factor can be set to... To identify the next contraction-relaxation cycle, and also by setting the scaling factor to... This is used to identify the previous contraction-relaxation cycle.

[0258] At step 1304, a third waveform is extracted from a second position of the first waveform. The third waveform includes a second contraction-relaxation cycle and has a second duration proportional to the first frequency.

[0259] The third waveform corresponds to the functional response of the artificial tissue when stimulated at a predetermined frequency (i.e., the contraction-relaxation cycle). Therefore, the duration or length of the third waveform is proportional to this frequency. For example, if the artificial tissue is stimulated at a frequency of 1 Hz, the second duration will be 1 s; if the artificial tissue is stimulated at a frequency of 2 Hz, the second duration will be 0.5 s, and so on. In one embodiment, the first and second durations are the same.

[0260] In one embodiment, the third waveform is centered at the second position such that the midpoint of the third waveform is aligned or substantially aligned with the second position of the first waveform.

[0261] At an optional step of output 1306, a third waveform is output. In one embodiment, outputting the third waveform includes storing or saving the third waveform to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the third waveform includes transmitting the third waveform via a network (e.g., a local area network, a wide area network, etc.) or displaying the waveform for user review.

[0262] In one embodiment, outputting a third waveform includes outputting the third waveform to another process or method of this disclosure. For example, the third waveform may be output to the method 800 described above, such that the step of obtaining 802 includes obtaining the third waveform from method 1300.

[0263] As stated above, method 1300 for extracting additional contraction-relaxation cycle waveforms from a first waveform can be repeated for all contraction-relaxation cycles within the first waveform. Since the extraction performed by method 1300 depends only on the first position and the first frequency, no additional signal processing or analysis is required to identify the positions of the additional contraction-relaxation cycles. Therefore, method 1300 provides a fast and efficient method for extracting contraction-relaxation cycles from a waveform. These waveforms can then be further processed, for example, by fitting a model to the waveform to generate a noise-filtered representation of the waveform.

[0264] Figure 14 A method 1400 for predicting treatment effects using a contraction-relaxation cycle model is shown according to one aspect of this disclosure.

[0265] Method 1400 includes the following steps: obtaining 1402 a plurality of signals; splitting the plurality of signals 1404 into a first plurality of waveforms; fitting a model 1406 to each of the first plurality of waveforms; generating 1408 a second plurality of waveforms from the model; extracting 1410 a first feature value; extracting 1412 a second feature value; and determining 1414 an effect. Method 1400 also includes an optional step of outputting 1416 an effect. In one embodiment, method 1400 is performed by... Figure 1 The control unit 104 or its subunit shown is used for this purpose.

[0266] Generally, Method 1400 describes the application of the contraction-relaxation cycle model to downstream drug discovery / development tasks. Specifically, the contraction-relaxation cycle model is used to efficiently generate accurate feature values ​​from baseline and perturbation signals of engineered tissues. Accurate feature extraction from these signals allows for efficient and accurate identification of effects associated with perturbations.

[0267] At step 1402, multiple signals are obtained. These signals include a baseline signal and a perturbation signal. The baseline signal includes a first plurality of functional responses of the engineered tissue under reference conditions. The perturbation signal includes a second plurality of functional responses of the engineered tissue under perturbation conditions involving the first perturbation.

[0268] Baseline and perturbation signals comprise multiple functional responses (i.e., multiple contraction-relaxation cycles or peaks) of an engineered tissue under reference and perturbation conditions. Generally, a reference condition refers to a condition that provides a baseline comparison to a perturbation condition. In embodiments, a reference condition is a condition associated with control settings or the environment. A reference condition may correspond to an engineered or artificial tissue in its default, natural, or unaltered state (i.e., without a dose of drug or agent). Alternatively, a reference condition may correspond to engineered tissue treated with a carrier. A perturbation condition refers to a condition in which the engineered tissue has been perturbed in some way. Examples of perturbation conditions include administration of a drug or compound (i.e., a perturber), disease state, different cell lines, or physical perturbation applied to the engineered tissue. Thus, a perturbation condition may alternatively be referred to as a therapeutic condition. In the case of perturbations involving drugs or compounds, these conditions are further associated with effects related to the drug or compound, such as mechanism of action or toxicity. Given a range of different perturbation conditions, an engineered tissue may be associated with more than one perturbation condition (e.g., a diseased engineered tissue that has been treated with a specific compound).

[0269] In one embodiment, the baseline signal and the perturbation signal are emitted from the bioreactor (e.g., Figure 1The engineered or artificial tissue is obtained in a bioreactor 102, in which it is grown / maintained. As previously stated, the artificial tissue includes engineered muscle tissue, such as engineered cardiac tissue or engineered skeletal muscle tissue. The baseline signal and perturbation signal are obtained at two distinct time points. For example, the baseline signal is obtained at a first time point, and then the engineered tissue is perturbed according to a first perturbation (e.g., by applying a first dose of the compound to the engineered tissue), and the perturbation signal is obtained at a second time point after the first time point. The baseline signal and perturbation signal include the functional response of the engineered tissue when stimulated at a predetermined pacing frequency (e.g., 0.1 Hz, 0.5 Hz, 1 Hz, 2 Hz, etc.). Alternatively, the baseline signal and perturbation signal include the spontaneous functional response of the engineered tissue in the absence of external stimulation.

[0270] At step 1404, multiple signals are split into a first plurality of waveforms. Each of the first plurality of waveforms includes the contraction and relaxation responses of the engineered tissue during a single contraction-relaxation cycle.

[0271] The method 1200 described above for extracting contraction-relaxation cycle waveforms is used to split multiple signals into a first plurality of waveforms. Alternatively, the multiple signals are split by manually annotating or extracting regions within the first plurality of waveforms that correspond to individual contraction-relaxation cycles.

[0272] The first subset of waveforms generated at step 1404 includes a first subset of waveforms associated with waveforms extracted from the baseline signal and a second subset of waveforms associated with waveforms extracted from the perturbation signal. Subsequent model fitting (described below) is independent of the source of the waveforms (i.e., independent of whether the waveform is a reference waveform or a perturbation waveform), but uses an identifier indicating whether the waveform corresponds to a reference condition or a perturbation condition in a subsequent feature extraction step.

[0273] At step 1406, the model is fitted to each of the first plurality of waveforms. This model is a contraction-relaxation cycle model that independently parameterizes the growth of the contraction and relaxation responses of the engineered tissue during a single contraction-relaxation cycle of each waveform.

[0274] The step of fitting the model to 1406 for each waveform corresponds to the above section on... Figure 8 The steps of fitting 804 are described in more detail. Therefore, at step 1406, step 804 is repeated for each of the first plurality of waveforms. The process of fitting the model to the waveforms is described above regarding... Figure 8 and Figure 9 To describe in more detail.

[0275] At step 1408, a second plurality of waveforms is generated from the model fitted to each of the first plurality of waveforms. The second plurality of waveforms includes a plurality of filtered baseline waveforms associated with the baseline signal and a plurality of filtered treatment waveforms associated with the treatment signal.

[0276] The step that generates 1408 corresponds to the step that generates 806 (as mentioned above). Figure 8 (To be described in more detail) The model is repeatedly applied to each waveform fitted to the first plurality of waveforms. The process of generating the second waveform from the model fitted to the first waveform is described above regarding... Figure 8 and Figure 9 To describe in more detail.

[0277] At step 1410, the first feature value of the first feature is extracted from multiple filtered baseline waveforms.

[0278] Multiple filtered baseline waveforms include noise-filtered or noise-suppressed representations of the contraction-relaxation cycle in the baseline signal. Because the signal has been filtered to remove noise, features can be accurately extracted from these waveforms.

[0279] The step of extracting the first feature value 1410 includes extracting a plurality of feature values ​​from a plurality of filtered baseline waveforms, such that the first feature value includes a plurality of feature values ​​or a representation of a plurality of feature values. Thus, a value of the first feature is extracted from each of the plurality of filtered baseline waveforms to determine the first feature value. In one embodiment, the first feature value includes the average (mean, median, etc.) of the first feature determined from the plurality of filtered baseline waveforms. In another embodiment, the first feature value includes the maximum value, minimum value, or value distribution determined from the plurality of filtered baseline waveforms.

[0280] The primary characteristic is one of the following: convulsions or peak amplitude (i.e., Figure 2 The peak amplitude shown is 206), and the contraction time (i.e., Figure 2 The time to reach peak amplitude (208) and the maximum contraction slope (i.e., Figure 2 The maximum development rate shown is 214), and the relaxation time (i.e., Figure 2 The time to reach the peak descent (210) and the maximum relaxation slope (i.e., Figure 2 The maximum rate of decline shown is 216) or the duration of the convulsion (i.e., Figure 2 The duration shown is 212).

[0281] At step 1412, the second feature value of the first feature is extracted from multiple filtered perturbation waveforms.

[0282] The step of extracting the second feature value 1412 includes extracting a plurality of feature values ​​from a plurality of filtered perturbation waveforms, such that the second feature value includes the plurality of feature values ​​or a representation of the plurality of feature values. Thus, a value of a first feature is extracted from each of the plurality of filtered perturbation waveforms to determine the second feature value. In one embodiment, the second feature value includes the average (mean, median, etc.) of the first feature determined from the plurality of filtered perturbation waveforms. In another embodiment, the second feature value includes the maximum value, minimum value, or value distribution determined from the plurality of filtered perturbation waveforms.

[0283] At step 1414, the effect associated with the first perturbation is determined based on a comparison of the first eigenvalue and the second eigenvalue.

[0284] The first eigenvalue is a quantitative descriptor of the functional response of the engineered organization under reference conditions. The second eigenvalue is a quantitative descriptor of the functional response of the engineered organization under perturbation conditions involving the first perturbation. Thus, the descriptions of the first and second eigenvalues ​​reveal any changes in the functional response or effect of the engineered organization due to the first perturbation.

[0285] For example, the first perturbation may correspond to the application of a compound with unknown physiological effects. A comparison of a first eigenvalue (corresponding to the peak amplitude of the contractile force waveform of the engineered tissue under reference conditions) and a second eigenvalue (corresponding to the peak amplitude of the contractile force waveform of the engineered tissue under the perturbation conditions involving the application of the compound) reveals an increase in the mean peak amplitude. Therefore, it can be inferred that the compound has an effect associated with increased contractile force of the engineered tissue during the contraction-relaxation cycle. Advantageously, since the eigenvalues ​​are determined from noise-filtered waveforms, the effects resulting from differences between the eigenvalues ​​can be more accurately identified, leading to improved treatment and potentially improved patient outcomes.

[0286] Systems and methods for detecting spontaneous tissue contraction.

[0287] Functional response waveforms comprise time-series values ​​corresponding to measurements of a tissue's functional response over a given time period. Thus, functional response waveforms encode changes in a tissue's functional response (e.g., contractility, displacement, etc.) over a defined time period. In some cases, these changes are induced by external stimuli applied to the tissue. For example, electrical stimulation can be applied periodically to tissues (such as engineered muscle tissue) at predetermined pacing frequencies (e.g., 0.5 Hz, 1 Hz, 2 Hz, etc.). Given such external stimuli, functional response waveforms typically exhibit periodic behavior, which can be utilized in identifying and extracting individual responses for feature extraction and downstream analysis tasks. In the absence of such stimulation, or when the tissue is exhibiting anomalous contractile behavior, such periodicity may not be present within the waveform. Therefore, identifying and extracting individual responses from such "spontaneous" contractile response waveforms is more challenging. This disclosure describes systems and methods for efficiently and accurately identifying spontaneous contractions within functional response waveforms that may not exhibit periodic behavior. This allows waveforms encoding spontaneous organizational behavior to be processed and analyzed, opening up the possibility of using such waveforms in downstream tasks such as drug discovery and drug development.

[0288] As stated above, in many cases, the contraction-relaxation cycle within a waveform occurs due to external stimuli, such as electrical stimulation applied to the tissue at a predetermined pacing frequency (e.g., 1 Hz, 2 Hz, etc.). The regularity of the functional response allows for the identification and processing of individual contraction-relaxation cycles by recognizing repetitive patterns or by combining prior knowledge of the pacing frequency. This is described below. Figure 15A Example in.

[0289] Figure 15A The functional response of an organization exhibiting periodic contraction behavior is shown.

[0290] Figure 15AA functional response waveform 1500 and a timeline 1502 showing peak locations, including a first point 1504, a second point 1506, and a third point 1508, are illustrated. The functional response waveform 1500 includes the functional response (e.g., contractile force) of an organization over a time period. A single contractile response (a single peak or contraction) is indicated within the functional response waveform 1500 by a peak (i.e., a local maximum within the functional response waveform 1500). The timeline 1502 illustrates the periodicity or regularity of these contractile responses because the points within the timeline 1502 (which correspond to the locations of peaks or contractions within the functional response waveform 1500) are spaced apart along the timeline 1502 at generally regular intervals. For example, the interval between the contraction associated with the first point 1504 and the contraction associated with the second point 1506 is substantially the same as the interval between the contraction associated with the second point 1506 and the contraction associated with the third point 1508. Thus, the periodicity of contractions within the functional response waveform 1500 indicates that the organization exhibits periodic or regular contractile behavior.

[0291] As previously stated, for waveforms exhibiting periodic behavior, periodicity can be leveraged to aid downstream tasks such as contraction response extraction (i.e., extracting single contraction-relaxation cycles from the waveform). For example, periodicity (once known or learned) can be used to determine the expected intervals between single contractions, thus providing prior information for identifying the relative positions of contraction responses. However, some tissues may exhibit spontaneous contraction behavior rather than periodic contraction behavior, such as… Figure 15B exemplified in .

[0292] Figure 15B The functional response of an organization exhibiting spontaneous contraction behavior is shown.

[0293] Figure 15B A timeline 1512 showing the functional response waveform 1510 and peak locations, including a first point 1514, a second point 1516, and a third point 1518, is illustrated. The functional response waveform 1510 encompasses the functional response (e.g., contractile force) of an organization over a time period. (As shown in...) Figure 15AIn the functional response waveform 1510, a single contraction response (single peak or contraction) is indicated by a peak. The timeline 1512, representing the peak positions, illustrates the spontaneity (i.e., lack of periodicity or regularity) of these contraction responses because the points within timeline 1512 (corresponding to the locations of peaks or contractions) are irregularly spaced along timeline 1512. For example, the interval between the contraction associated with the first point 1514 and the contraction associated with the second point 1516 is substantially different from the interval between the contraction associated with the second point 1516 and the contraction associated with the third point 1518. Therefore, the lack of periodicity in the contractions within the functional response waveform 1510 indicates that the tissue exhibits spontaneous contraction behavior. Alternatively, spontaneous contraction behavior exhibited by the tissue may be periodic but is considered spontaneous because the tissue's contraction response is not due to stimulation of the tissue (e.g., by means of the above regarding...). Figure 1 This occurs due to the described electronic stimulation. In either case, there is no prior information available to identify and extract individual contractions from the functional response waveform.

[0294] Figure 15B The irregularity of the contractile behavior observed in tissues can be caused by a variety of factors. For example, in engineered tissues, this irregularity may be due to the cessation of external stimulation. Alternatively, this irregularity or spontaneity may be caused by one or more conditions of the tissue, such as those caused by a disease state, a compound or drug applied to the tissue.

[0295] Without applying stimulation to the tissue to induce contraction, it is impossible to identify individual contractile responses using prior knowledge of periodicity (or fundamental characteristics of regular behavior) (i.e., it is impossible to identify individual contraction-relaxation cycles or individual "peaks" within a waveform using the frequency with which the tissue is stimulated). According to one aspect of this disclosure, different methods can be employed to determine the spontaneous behavior of tissues (such as engineered or artificial tissues), such as... Figure 16 As shown.

[0296] Figure 16 A system 1600 for determining spontaneous behavior of an engineered organization, according to one aspect of this disclosure, is shown.

[0297] System 1600 includes a periodic classifier 1602, a decision unit 1604, a spontaneous contraction classifier 1606, and a feature generator 1608. A waveform 1610 representing the functional response of an engineered organization over a time period is provided as input to the periodic classifier 1602. The periodic classifier 1602 generates a classification score 1612 based on the waveform 1610. The decision unit 1604 determines whether the classification score 1612 indicates any periodic contraction within the waveform 1610. When the classification score 1612 indicates no periodic contraction within the waveform 1610, the waveform 1610 is provided as input to the spontaneous contraction classifier 1606. The spontaneous contraction classifier 1606 generates a classification score 1614 based on the waveform 1610. The feature generator 1608 generates behavioral features 1616 for the engineered organization based on the classification score 1614. In one embodiment, the periodic classifier 1602 includes a spectral transformation process 1618 and a prediction model 1620. In one embodiment, the spontaneous shrinking classifier 1606 includes a transformation process 1622 and a thresholding operation 1624.

[0298] System 1600 corresponds to a hierarchical approach for determining the spontaneous behavior of an engineered organization. By determining the spontaneous behavior of the engineered organization (e.g., encoded within a functional response waveform), behavioral and contraction location data can be used for multiple downstream analytical tasks (e.g., feature extraction, measurement development, etc.). A hierarchical structure of classifiers (i.e., a periodic classifier 1602 and a spontaneous contraction classifier 1606) is used to identify the global contraction behavior of the engineered organization, and subsequently, the local contraction behavior. Employing this hierarchical structure of classifiers allows for more efficient use of computational resources because the global classifier (i.e., the periodic classifier 1602) initially filters out waveforms known to exhibit periodic or regular behavior. Therefore, only waveforms predicted not to have periodic or regular behavior are fed to the local classifier (i.e., the spontaneous classifier 1606). Furthermore, this means that the local classifier (i.e., the spontaneous classifier 1606) can be customized to represent waveforms exhibiting spontaneous behavior, thus providing improved identification of spontaneous behavior (and the location of spontaneous contractions).

[0299] Waveform 1610 includes an engineered or artificially organized functional response over a time period (e.g., 10 seconds, 20 seconds, 30 seconds, etc.). (See above regarding...) Figure 1 The stated functional response may include changes in the contractile force of the engineered structure during the time period, or displacement or contractile displacement of the engineered structure during the time period. Although Figure 16 Not shown in the figure, but waveform 1610 is directly or indirectly derived from a bioreactor (such as...) Figure 1 The engineered tissue is obtained from the bioreactor 102 shown. In one embodiment, the engineered tissue includes artificial muscle tissue, such as artificial cardiac muscle or skeletal muscle tissue.

[0300] A periodic classifier 1602 (alternatively referred to as a first classifier, a frequency-based global classifier, or a global classifier) ​​is used to predict the global characteristics of waveform 1610. Specifically, the periodic classifier 1602 is configured to determine whether waveform 1610 includes any periodic contractions, such as those described above. Figure 15A The periodic classifier 1602 includes any suitable predictive model capable of predicting whether a time-series input contains a regular or periodic response (peak or contraction). In one embodiment, the periodic classifier 1602 includes a spectral transformation process 1618 and a predictive model 1620. The spectral transformation process 1618 applies a spectral transformation, such as a Fourier transform, to a waveform 1610 to generate a spectral waveform. The spectral waveform includes a frequency-based representation of the waveform 1610. The predictive model 1620 then uses the spectral waveform to determine a classification score 1612 (i.e., the presence or absence of any regular, periodic contractions within the waveform 1610). Advantageously, the spectral waveform encodes important features of the tissue's contractile behavior, which helps improve the classification performance of the predictive model 1620. This is achieved by... Figure 17A and Figure 17B The example spectral response is shown below.

[0301] Figure 17A and Figure 17B The spectral responses of different contraction responses according to embodiments of the present disclosure are shown.

[0302] The spectral response is obtained from the functional response waveform using a spectral transformation process involving Fourier transform. Each waveform is associated with an engineered or artificial tissue exhibiting a different contraction response.

[0303] The first functional response waveform 1702 was obtained from an engineered tissue exhibiting normal contraction behavior. That is, the contraction response encoded in the first spectral response 1702 follows a periodic (regular) pattern of contraction-relaxation cycles. As can be seen, the power is strongest at 1 Hz (i.e., the pacing frequency) and decreases with increasing harmonics. The second functional response waveform 1704 was obtained from an engineered tissue exhibiting an abnormal contraction response. That is, some double contraction (or double pulsation) is present. This is seen in the change in harmonic power in the second spectral response 1714. The third functional response waveform 1706 was obtained from an engineered tissue exhibiting a contraction response with increased force. That is, the power in the third spectral response 1716 shifts from harmonics to the dominant frequency (1 Hz). The fourth functional response waveform 1708 was obtained from an organization exhibiting a spectral response with decreased contraction force. That is, the power in the fourth spectral response 1708 shifts away from the dominant frequency (1 Hz) to harmonics (1718). The fifth functional response waveform 1710 was obtained from an organization that did not exhibit a periodic or regular contraction response. In other words, the power in the fifth spectral response 1710 is strongest at 0 Hz (1720).

[0304] Thus, the use of spectral waveforms provides a concise and descriptive representation of the contraction behavior of tissues, which can help improve the discriminative performance of classifiers or predictive models whose task is to identify different contraction behaviors.

[0305] Refer again Figure 16 Predictive model 1620 uses the spectral waveform (i.e., a frequency-based representation of waveform 1610) to determine a classification score 1612. The classification score 1612 represents the contractile behavior of the tissue associated with waveform 1610 and indicates whether waveform 1610 contains regular or periodic contractions. Thus, the classification score 1612 is a binary classification that uses one score or value (e.g., "+1") to indicate that waveform 1610 includes periodic contractions and another score or value (e.g., "0" or "-1") to indicate that waveform 1610 does not include periodic contractions. Alternatively, predictive model 1620 is a multi-class classifier that predicts the type of contractile behavior from the spectral waveform. For example, predictive model 1620 can be trained to assign spectral waveforms to... Figure 17A and Figure 17B One of the contraction response types illustrated (e.g., a classification score of "0" indicates no periodic contraction, a classification score of "1" indicates normal periodic contraction, a classification score of "2" indicates abnormal contraction behavior, etc.).

[0306] Prediction model 1620 is a trained machine learning model, such as a trained neural network, support vector machine, random forest, etc. In one embodiment, prediction model 1620 is a convolutional neural network, such as... Figure 18 The convolutional neural network shown.

[0307] Figure 18 A convolutional neural network 1800 for predicting contraction behavior is shown according to one embodiment of the present disclosure.

[0308] Neural network 1800 includes a convolutional network 1802, a long short-term memory (LSTM) network 1804, and a fully connected network 1806. The convolutional network 1802 receives an input vector 1808, and the fully connected network 1806 produces an output vector 1810. The convolutional network 1802 includes a first block 1812, a second block 1814, a third block 1816, a fourth block 1818, and a fifth block 1820. Each block includes two-dimensional (2D) convolutional layers, batch normalization layers, and optionally, rectified linear unit layers. This... Figure 18 The first 1812 is illustrated by a 2D convolutional layer 1822, a batch normalization layer 1824, and a rectified linear unit layer 1826. The LSTM network 1804 includes a flattening layer 1828, a bidirectional LSTM layer 1830, and a dropout layer 1832. The fully connected network 1806 includes a fully connected layer 1834, a dropout layer 1836, and a softmax layer 1838.

[0309] Input vector 1808 includes frequency spectrum waveforms (such as...) Figure 17A and Figure 17B The value sequence corresponding to the spectrum waveform shown.

[0310] Convolutional network 1802 comprises a sequence of blocks with a similar architecture. The first block 1812 includes a 2D convolutional layer 1822 with four filters of size 3, a batch normalization layer 1824 (which normalizes the input via recentering and rescaling), and a rectified linear unit layer 1826. The second block 1814 and the third block 1816 include the same architecture: a 2D convolutional layer with seven filters of size 3, a batch normalization layer, and a rectified linear unit layer. The fourth block 1818 includes a 2D convolutional layer with 16 filters of size 3, a batch normalization layer, and a rectified linear unit layer. The fifth block 1820 includes a 2D convolutional layer with 32 filters of size 3 and a batch normalization layer (i.e., the fifth block 1820 does not include an optional rectified linear unit layer).

[0311] The LSTM network 1830 includes: a flattening layer 1828, which compresses the spatial dimension of the output of the fifth block 1820 into a single dimension; a bidirectional LSTM layer 1830, which includes 8 hidden units; and a dropout layer 1832, which helps reduce overfitting by randomly setting the input to zero with a probability of 0.3.

[0312] The fully connected network 1806 includes: a fully connected layer 1834 with an output size set to the number of contraction behaviors to be predicted (i.e., two layers are used when predicting periodic contractions and aperiodic contractions); a dropout layer 1836 that randomly sets the input to zero with a probability of 0.05; and a softmax layer 1838 that applies a softmax function to the output of the dropout layer 1836.

[0313] Output vector 1810 corresponds to a probability vector with a magnitude equal to the number of predicted contraction behaviors. Therefore, output vector 1810 is used to determine the classification score (i.e., as mentioned above regarding...). Figure 16 The classification score is described in 412. For example, via threshold, identifying the maximum value, etc.

[0314] In one implementation, the Neural Network 1800 is trained using a synthetic dataset comprising 40,000 training samples and 10,000 validation samples. Each element in the synthetic training dataset is created by generating a synthetic waveform that includes periodic or aperiodic contractions and then applying a spectral transform (Fourier transform) to generate a spectral waveform. The synthetic dataset comprises 25,000 periodic waveforms and 25,000 aperiodic waveforms. Each waveform is associated with a corresponding label indicating whether it is periodic or aperiodic. The Neural Network 1800 is trained using mini-batch gradient descent with a batch size of 128 and an ADAM solver. The ADAM solver has an initial learning rate of 1e-3, with early stopping based on the validation loss.

[0315] Refer again Figure 16 Decision unit 1604 uses the classification score 1612 obtained from the periodic classifier 1602 to determine whether waveform 1610 contains periodic contractions and thus whether a second classifier (i.e., spontaneous contraction classifier 1606) should be invoked. Decision unit 1604 uses decision rules to determine whether to pass waveform 1610 to the spontaneous contraction classifier 1606 or not to perform any further action (such as...). Figure 16 (Indicated by the black circle in the diagram). The decision rule determines whether the classification score 1612 indicates that there is no periodic contraction within the waveform 1610 (e.g., if the classification score 1612 takes a value such as "0", it indicates that the periodic classifier 1602 has classified the waveform 1610 as not including any periodic contraction). If there is no periodic contraction, the waveform 1610 is fed into the spontaneous contraction classifier 1606.

[0316] A spontaneous contraction classifier 1606 (alternatively referred to as a local classifier, a second classifier, or a peak detector) is used to generate a classification score 1614, which indicates whether waveform 1610 includes any spontaneous contraction of the engineered tissue within a time period. It is possible that no contraction response of the tissue is observed within waveform 1610 (i.e., the tissue did not undergo induced or spontaneous contraction). Thus, classification scores 1612 and 1614 will indicate the absence of contraction (whether induced or spontaneous) within waveform 1610. Alternatively, waveform 1610 may not include induced contraction but may include one or more spontaneous contractions (such as...). Figure 15B (Example shown). The spontaneous shrinking classifier 1606 uses waveform 1610 to determine which of the two behaviors is exhibited by the organization within waveform 1610. Thus, the spontaneous shrinking classifier 1606 can be any suitable predictive model or trained machine learning model, such as a trained neural network, support vector machine, random forest, etc.

[0317] In one embodiment, the spontaneous shrinkage classifier 1606 includes a transformation process 1622 and a thresholding operation 1624. Generally, the transformation process 1622 generates a transformed waveform from waveform 1610, and then the thresholding operation 1624 is applied to the transformed waveform. The transformation process 1622 enhances peaks within waveform 1610 while reducing noise. This helps improve the performance of the thresholding operation 1624, which can then identify any peaks exceeding one or more thresholds. If any peaks exceeding one or more thresholds exist within the transformed waveform (threshold exceedance number), then waveform 1610 includes spontaneous shrinkage of the engineered tissue; otherwise, if no threshold exceedance number exists, waveform 1610 does not include any spontaneous shrinkage of the engineered tissue.

[0318] The transformation process 1622 includes any suitable signal processing operations that can enhance the peak values ​​within the waveform 1610, such as peak sharpening or peak filtering. In one embodiment, the transformation process 1622 includes the Pan-Tompkins algorithm. In another embodiment, the transformation process 1622 includes a modified Pan-Tompkins algorithm. The Pan-Tompkins algorithm was developed to detect the QRS complex of an electrocardiogram (ECG) signal (i.e., the Q wave, R wave, and S wave). The Pan-Tompkins algorithm includes a series of filters applied to the waveform to enhance the frequency content (i.e., peak values) of the waveform while also removing noise. Generally, the Pan-Tompkins algorithm includes a denoising process and a subsequent enhancement process. The denoising process applies a bandpass filter (i.e., a low-pass filter, followed by a high-pass filter) to the input waveform. The enhancement process includes derivative operations, squaring filters, and integral filters. The output of the denoising process is provided to the derivative operation, which provides slope information. The squaring operation enhances the peak values ​​of the output of the derivative operation, and the integral filter applies a moving average to the output of the squaring operation. As will be described in more detail below, the modified Pan-Tompkins algorithm of this disclosure incorporates a stationary waveform transform into the denoising process and replaces the squared filter with a rectification operation. Thus, the modified Pan-Tompkins algorithm generates a transformed waveform with better denoising characteristics and improved peak enhancement. This helps improve the identification of spontaneous contractions, which in turn helps improve the performance of downstream tasks that use this information for tasks such as drug discovery and development.

[0319] Figure 19 A step-by-step result of a modified Pan-Tompkins algorithm according to an embodiment of the present disclosure is illustrated.

[0320] Figure 19 The results of applying the steps of the modified Pan-Tompkins algorithm of this disclosure to waveform 1902 are shown. Figure 19 Waveforms 1904 (low-pass filtered), 1906 (high-pass filtered), 1908 (wavelet transform), 1910 (differentiation processed), 1912 (rectified), and 1914 (integration) are shown. Although the modified Pan-Tompkins algorithm is described as performing each of the steps described below, those skilled in the art will understand that in some implementations, steps may be combined and / or omitted. For example, the algorithm may include performing a stationary wavelet transform and rectification; or a stationary wavelet transform, differentiation, and rectification. In some embodiments, a normalization operation is applied to the output of the modified Pan-Tompkins algorithm to normalize the waveform along... The axis is scaled to a consistent range of values.

[0321] Waveform 1902 corresponds to Figure 16 The waveform 1610 shown includes the functional response (e.g., force) of the engineered organization over a time period. Figure 19 The following examples illustrate the results of applying a modified Pan-Tompkins algorithm for noise reduction: low-pass filtered waveform 1904, high-pass filtered waveform 1906, and wavelet-transformed waveform 1908. Low-pass filtered waveform 1904 corresponds to the result of applying a low-pass filter to waveform 1902. The low-pass filter allows the portion of waveform 1902 with a frequency below a predetermined cutoff frequency to pass through, and attenuates the portion of the waveform above the predetermined cutoff frequency. Therefore, the low-pass filter helps remove major wire mismatches from waveform 1902. In one implementation, the low-pass filter comprises a two-dimensional Gaussian kernel with a standard deviation of 2. High-pass filtered waveform 1906 corresponds to the result of applying a high-pass filter to low-pass filtered waveform 1904. Therefore, high-pass filtered waveform 1906 corresponds to the band-pass filtered waveform, as it is the result of both low-pass and high-pass filtering of waveform 1902. A high-pass filter allows the portion of the low-pass filtered waveform 1904 with frequencies above a predetermined cutoff frequency to pass through, and attenuates the portion of the waveform below the predetermined cutoff frequency. Therefore, the high-pass filter aids in baseline alignment while removing drift. In one implementation, the high-pass filter comprises an elliptic filter with an order of 8, a ripple of 0.5 dB, an attenuation of 40 dB, and edge frequencies of 1 and 20.

[0322] Waveform 1908, after wavelet transform, is the result of applying a stationary wavelet transform to waveform 1906, which has been high-pass filtered. As stated above, the modified Pan-Tompkins algorithm of this disclosure performs a stationary wavelet transform as an additional step in the denoising process of the Pan-Tompkins algorithm. Advantageously, the stationary wavelet transform is offset invariant and helps reduce noise in the waveform while preserving important transitional features (changes) within the waveform that may be needed for downstream tasks such as feature extraction. The stationary wavelet transform is an extension of the wavelet transform in which wavelet coefficients are not decimated at each stage. In one implementation, the wavelet transform includes a discrete stationary wavelet transform (1D) using five decomposition levels and a Daubechies 4 (db4) wavelet.

[0323] Figure 19The results of the enhancement process of the modified Pan-Tompkins algorithm are shown in waveforms 1910 (differentiated), 1912 (rectified), and 1914 (integrated). Waveform 1910 (differentiated) is the result of differentiating waveform 1908 (wavelet transform). Differentiation is used to highlight rapid changes (i.e., contractions). The numerical gradient is calculated using uniform spacing between points in all directions. In one implementation, differentiation corresponds to having a transfer function... The five-point derivative, where This is the sampling period. Instead of squaring, the modified Pan-Tompkins algorithm of this disclosure applies rectification to the differentiated waveform 1910 to generate the rectified waveform 1912. Rectification truncates the differentiated waveform 1910, removing any negative portions of it. Therefore, rectification enhances the dominant peaks within the waveform. The integrated waveform 1914 corresponds to the result of applying a moving window integration operation to the rectified waveform 1912. The moving window integration operation applies a moving average filter (i.e., a sliding window filtering operation) to the rectified waveform 1912. Therefore, the moving window integration operation removes short-duration artifacts from the rectified waveform 1912. In one implementation, the moving average filter includes elements having a length... The moving mean filter is calculated within the sliding window, where It is the length of waveform 1610.

[0324] As from Figure 19 As can be seen in the example, the output of the modified Pan-Tompkins algorithm (integrated waveform 1914) includes a denoised version of the input waveform (waveform 1902) with an enhanced representation featuring a contraction-relaxation loop. By reducing noise within the waveform and simultaneously enhancing peaks (i.e., the contraction-relaxation loop), the identification of spontaneous contractions within the transformed waveform is improved. This, in turn, helps improve the performance of downstream tasks that use this information for tasks such as drug discovery and development.

[0325] Refer again Figure 16Thresholding operation 1624 identifies any peaks or thresholds exceeding a certain number within the transformed waveform generated by transformation process 1622. The presence of peaks within the transformed waveform indicates that waveform 1610 includes spontaneous contractions. That is, if a portion of the transformed waveform satisfies the thresholding criteria defined by thresholding operation 1624, then classification score 1614 indicates that waveform 1610 includes one or more spontaneous contractions. Similarly, the absence of thresholds exceeding a certain number within the transformed waveform indicates that waveform 1610 does not contain any spontaneous contractions (or contractions induced by the hierarchical classification system employed by system 1600). That is, if no portion of the transformed waveform satisfies the thresholding criteria defined by thresholding operation 1624, then classification score 1614 indicates that waveform 1610 does not include any spontaneous contractions.

[0326] The thresholding operation 1624 involves one or more adaptive thresholds. An adaptive threshold is a threshold based on one or more characteristics or properties of the signal (waveform) to which the adaptive threshold is applied. Thus, the adaptive threshold will vary according to the statistical properties of the waveform. In one embodiment, two adaptive thresholds are applied: a uniform adaptive threshold and a dynamic adaptive threshold. The uniform adaptive threshold remains constant over the time period of waveform 1610, while the dynamic adaptive threshold varies over that time period. This is in Figure 20 Example in.

[0327] Figure 20 An example of adaptive thresholding of a waveform according to an embodiment of the present disclosure is illustrated.

[0328] Figure 20 Waveform 2002 is shown (e.g., from...) Figure 16 The transformed representation of waveform 1610 obtained by the transformation process 1622 shown, and the plots of uniform adaptive threshold 2004 and dynamic adaptive threshold 2006. Figure 20 The diagram further illustrates points 2008, 2010, 2012, 2014, and 2016, all of which are points on waveform 2002. Points 2008 and 2010 exceed both the uniform adaptive threshold 2004 and the dynamic adaptive threshold 2006. Points 2012 and 2014 exceed the uniform adaptive threshold 2004 but not the dynamic adaptive threshold 2006. Point 2016 exceeds the dynamic adaptive threshold 2006 but not the uniform adaptive threshold 2004.

[0329] Threshold operations (such as) Figure 16The thresholding operation 1624 shown determines one or both of a uniform adaptive threshold 2004 and a dynamic adaptive threshold 2006 for waveform 2002, and uses them to identify threshold exceedances (local maxima). These threshold exceedances correspond to spontaneous contractions encoded within waveform 2002. For example, if the thresholding operation requires exceedances of two thresholds to identify spontaneous contractions, then in Figure 20 In the example points highlighted in the diagram, the first point 20020 and the second point 2010 on waveform 2002 will be identified as the locations of spontaneous contraction.

[0330] The uniform adaptive threshold 2004 remains constant over the time period of waveform 2002 and represents a lenient threshold (i.e., more points within waveform 2002 will be identified as potential spontaneous contractions compared to a more conservative threshold). The uniform adaptive threshold 2004 is determined based on the statistical properties of waveform 2002. These statistical properties are calculated over the entire (or substantially the entire) time period of waveform 2002. The statistical properties include one or more of the following: the mean of waveform 2002; the median of waveform 2002; and the mean of the maximum and minimum values ​​of waveform 2002. For example, the uniform adaptive threshold may be set to the mean of the waveform. In one embodiment, the statistical properties are based on a weighting factor. Weighted. Therefore, the weighting factor is used to control the looseness of the threshold. In one implementation, the weighting factor is selected according to a manual tuning process, whereby the user selects the threshold within a range of the waveform (i.e., the waveform's "hold-back set") during the training or calibration phase. The appropriate value. Then that value is used in the system (such as...) Figure 16 The system shown in System 1600 is fixed when used.

[0331] The dynamic adaptive threshold 2006 varies over the time period of waveform 2002 and represents a conservative threshold (i.e., fewer points within waveform 2002 will be identified as potentially spontaneously contracting compared to a more lenient threshold). Unlike the uniform adaptive threshold 2004, the dynamic adaptive threshold 2006 is determined within a sub-region of waveform 2002 such that the dynamic adaptive threshold 2006 varies over that time period. In one embodiment, the dynamic adaptive threshold 2006 is calculated for each time point within waveform 2002. A window is identified around each time point of waveform 2002 (e.g., a window containing 3 points, 5 points, 10 points, etc., or a window of 0.5s, 1s, 2s, etc.), and the local threshold for that time point is determined using the portion of the waveform within that window. The local threshold is determined using the statistical properties of that portion of the waveform in the same manner as described with respect to the uniform adaptive threshold 2004 (e.g., mean, median, etc.). As with the uniform adaptive threshold 2004, in one embodiment, the statistical properties are determined based on a weighting factor. Weighted. In one implementation, weighting factors are selected based on a manual tuning process, thereby allowing the user to select weights within a range of waveforms (i.e., the waveform's "hold-back set") during the training or calibration phase. The appropriate value. Then that value is used in the system (such as...) Figure 16 The system shown in System 1600 is fixed when used.

[0332] The use of both uniform adaptive threshold 2004 and dynamic adaptive threshold 2006 helps improve the accuracy of identifying individual spontaneous contractions within waveform 2002 by reducing the number of false positives (e.g., point 4 2014 and point 5 2016). This improvement in accuracy can then lead to improvements in the performance of downstream tasks such as contraction extraction, feature extraction, and use in drug discovery and development tasks.

[0333] Refer again Figure 16 Based on the presence or absence of any spontaneous contractions within the waveform, the spontaneous contraction classifier 1606 determines a classification score 1614. For example, the classification score can be a binary value indicating the presence (e.g., +1) or absence (e.g., -1) of any spontaneous contraction within the waveform 1610. Alternatively, the classification score can be a vector of spontaneous contraction locations within the waveform 1610, such that an empty vector indicates that no spontaneous contraction is identified within the waveform 1610.

[0334] Feature generator 1608 generates behavioral features 1616 for an engineered organization based on classification score 1614. In one embodiment, behavioral feature 1616 includes an indication (e.g., a binary identifier) ​​of whether waveform 1610 includes any spontaneous contractions. Additionally or alternatively, behavioral feature 1616 includes a summary of spontaneous contractions within waveform 1610, such as the number of spontaneous contractions, the average amplitude of spontaneous contractions, etc. Additionally or alternatively, behavioral feature 1616 includes one or more spontaneous contractions within waveform 1610 (or their location within waveform 1610). Behavioral feature 1616 can then be output for further processing. For example, behavioral feature 1616 can be assigned to waveform 1610 as a label or as the location of spontaneous contractions in waveform 1610.

[0335] In one embodiment, behavioral feature 1616 is used to extract one or more features associated with spontaneous contraction from waveform 1610. For example, features are extracted around each spontaneous contraction location within waveform 1610, such as those described above. Figure 2 These characteristics are then described. These characteristics are then formed into an eigenvector, which serves as a descriptor for the (spontaneous) contractile response of an organization. This eigenvector, or its transformations or summations, can then be used for downstream drug discovery and development tasks, such as estimating the effects of compounds or identifying disease states.

[0336] In another embodiment, behavioral characteristic 1616 is used as a feature vector to describe the conditions or state of the engineered tissue from which waveform 1610 is generated. That is, behavioral characteristic 1616, or aggregated statistics associated with it (such as the number of spontaneous contractions, average contractile force, etc.), can be used as a phenotype of the engineered tissue. Measuring changes in the behavioral characteristics of the engineered tissue (determined using system 1600) under controlled conditions and under perturbed conditions (e.g., treatment conditions involving drugs or compounds, disease states, etc.) can be used to identify effects associated with perturbed conditions.

[0337] The description will now turn to methods for determining the spontaneous behavior of an engineered organization by utilizing the components, processes, and operations of the system described above.

[0338] Figure 21 A method 2100 for determining spontaneous behavior of an engineered organization according to one aspect of this disclosure is shown.

[0339] Method 2100 includes the following steps: obtaining 2102 a first waveform; applying 2104 a first classifier; applying 2106 a second classifier; and generating 2108 behavioral features. Method 2100 further includes optional steps of assigning 2110 behavioral features and outputting 2112 behavioral features. In one embodiment, method 2100 is... Figure 1 The signal processing unit 112 or another subunit of the control unit 104 shown is used for this purpose.

[0340] At the point of acquisition in step 2102, the first waveform is obtained (e.g., Figure 16 The waveform shown is 410. The first waveform includes the functional response of the engineered tissue over a time period. In one embodiment, the first waveform is directly emitted from a bioreactor (e.g., where the engineered tissue is held) that contains the bioreactor. Figure 1 The first waveform is obtained from the bioreactor 102 shown. Alternatively, the first waveform is obtained via an intermediate unit that converts the observed measurements from the bioreactor into a functional response waveform. The functional responses of the engineered tissue measured within the first waveform are contractile forces, displacement, calcium transient responses, etc.

[0341] Engineered tissue corresponds to any myopropelling tissue or muscle tissue. This muscle tissue is used in bioreactor devices (e.g., Figure 1 Within the apparatus 106 of the bioreactor 102 shown, cells (such as induced pluripotent stem cells (iPSCs)) seeded therein grow. In one embodiment, the engineered tissue is engineered cardiac tissue. More specifically, the engineered tissue may be engineered human cardiac tissue grown from cardiomyocytes and ventricular cardiac fibroblasts derived from human iPSCs. Alternatively, the engineered tissue is engineered skeletal muscle tissue.

[0342] At step 2104, the first classifier is applied to the first waveform to generate a first classification score (e.g., applying the periodic classifier 1602). Figure 16 The waveform 1610 shown generates a classification score 1612. The first classification score indicates whether the waveform includes periodic contractions of the engineered organization over a time period. Thus, the first classifier is alternatively referred to as a periodic classifier, a frequency-based global classifier, or a global classifier.

[0343] The first classifier predicts the global characteristics of the waveform obtained at step 2102: whether the waveform includes any periodic (regular or induced) contractions, such as those mentioned above. Figure 15A The contractions described. The first classifier includes any suitable predictive model that can predict whether a time series input contains a regular or periodic response (peak or contraction).

[0344] The first classifier includes any suitable predictive model capable of predicting whether a time-series input contains a regular or periodic response (peaking or contracting). See below for more information. Figure 22 In a more detailed description, in one embodiment, the first classifier applied at step 2104 includes a spectral transformation process and a prediction model (e.g., Figure 16 The spectral transformation process 1618 and prediction model 1620 of the periodic classifier 1602 are shown.

[0345] At step 2106, a second classifier is applied to the waveform to generate a second classification score. The second classifier is applied when the first classification score indicates that there is no periodic contraction within the waveform (e.g., when classification score 412 indicates that there is no induced contraction or periodic contraction in waveform 1610, the spontaneous contraction classifier 1606 is applied). Figure 16 The waveform shown is 1610. Thus, in one embodiment, method 2100 further includes determining whether the waveform includes one or more periodic contractions based on the classification score obtained at step 2104, prior to the step of application 2106. If no periodic contraction is identified as present in the waveform, the step of applying the second classifier in step 2106 is performed (otherwise, method 2100 terminates, for example, by returning an appropriate indication or notification that the waveform is periodic). The second classification score determined by the second classifier indicates whether the waveform includes spontaneous contractions of engineered tissue over a time period. The second classifier is alternatively referred to as a spontaneous contraction classifier, a local classifier, or a peak detection.

[0346] The second classification score generated by the second classifier indicates whether the waveform obtained at step 2102 includes any spontaneous contraction of the engineered tissue within a time period. It is possible that no contraction response of the tissue is observed within the waveform (i.e., the tissue does not exhibit induced or spontaneous contraction). In this case, the first and second classification scores will indicate the absence of contraction (whether induced or spontaneous) within the waveform. Alternatively, the waveform may exclude induced contraction but include one or more spontaneous contractions (such as...). Figure 15B (Example shown). The second classifier uses the waveform obtained at step 2102 to determine which of the two behaviors is exhibited by the tissue within the waveform. The second classification score can be a binary value that indicates the presence (e.g., +1) or absence (e.g., -1) of any spontaneous contraction within the waveform. Alternatively, the second classification score can be a vector of the location of spontaneous contractions within the waveform, such that an empty vector indicates that no spontaneous contraction is identified within the waveform.

[0347] The following text is about Figure 23 In a more detailed description, in one embodiment, the second classifier includes a transformation process and a thresholding operation (e.g., Figure 16 The transformation process 1622 and thresholding operation 1624 of the spontaneous shrinking classifier 1606 are shown.

[0348] At step 2108, behavioral characteristics are generated for the engineered organization based on the second classification score during that time period. In one embodiment, the behavioral characteristics include an indication (e.g., a binary identifier) ​​of whether the waveform obtained at step 2102 includes any spontaneous contractions. Additionally or alternatively, the behavioral characteristics include one or more spontaneous contractions associated with one or more portions of the waveform that are associated with the location of the spontaneous contraction.

[0349] At an optional step of assignment 2110, a behavioral feature is assigned to the waveform obtained at step 2102. For example, the behavioral feature is assigned as a label for the waveform, making the waveform and the label available for further processing or classification (e.g., within a drug discovery or development system). Additionally or alternatively, when the behavioral feature includes the location of spontaneous contractions within the waveform, spontaneous contractions can be assigned to the waveform by identifying the location of the spontaneous contractions within the waveform (e.g., using metadata, location vectors, etc.).

[0350] At an optional step of output 2112, a behavioral characteristic is output. In one embodiment, outputting the behavioral characteristic includes storing or saving the behavioral characteristic to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the behavioral characteristic includes transmitting the behavioral characteristic via a network (e.g., a local area network, a wide area network, etc.) or displaying the behavioral characteristic for user review. In one embodiment, the behavioral characteristic is output in conjunction with a waveform.

[0351] Figure 22 A method 2200 for obtaining a classification score from a waveform is shown according to an embodiment of the present disclosure.

[0352] In one embodiment, method 2200 is performed at the step of applying 2104 of the first classifier in method 2100 to generate a first classification score. Thus, method 2200 includes a frequency-based global classifier (periodic classifier or global classifier) ​​according to one embodiment (e.g., Figure 16 The steps performed by the periodic classifier 1602 shown are as follows. Method 2200 includes the following steps: transforming 2202 the waveform to generate a transformed waveform; applying 2204 the transformed waveform to a prediction model; and obtaining 2206 a classification score from the prediction model.

[0353] At step 2202, a spectral transformation process is applied to the waveform (i.e., the waveform obtained at step 2102 of method 2100) to generate a spectral waveform (e.g., as...). Figure 16 As part of the periodic classifier 1602, waveform 1610 is applied to the spectral transformation process 418. The spectral transformation process is any suitable spectral transform, such as a Fourier transform, which produces a frequency-based representation of the waveform. Example spectral waveform in Figure 17A and Figure 17B As shown in the image.

[0354] At step 2204, the spectral waveform generated at step 2202 is input into the prediction model. The prediction model is a trained machine learning model, such as a trained neural network, support vector machine, random forest, etc. In one embodiment, the prediction model is a convolutional neural network, such as... Figure 18 The convolutional neural network shown and described in more detail above.

[0355] At step 2206, a classification score is obtained from the prediction model based on the spectral waveform. The classification score represents the contractile behavior of the tissue associated with the waveform (i.e., associated with the spectral waveform) and indicates whether the waveform contains regular or periodic contractions. In one embodiment, the classification score is a binary classification that uses one score or value (e.g., "+1") to indicate that the waveform includes periodic contractions and another score or value (e.g., "0" or "-1") to indicate that the waveform does not include periodic contractions.

[0356] Figure 23 A method 2300 for obtaining a classification score from a waveform is shown according to an embodiment of the present disclosure.

[0357] In one embodiment, method 2300 is performed at the step of applying 906 of the second classifier in method 900 to generate a second classification score. Thus, method 2300 includes a local classifier (spontaneous shrinking classifier or peak detector) according to one embodiment (e.g., Figure 16 The steps performed by the spontaneous shrinking classifier 1606 shown are as follows. Method 2300 includes the following steps: transforming 2302 a waveform to generate a transformed waveform; determining 2304 one or more adaptive thresholds; and determining 2306 a classification score based on one or more adaptive thresholds applied to the transformed waveform.

[0358] At step 2302, the waveform (i.e., Figure 21 The method shown in step 2100 (obtaining the waveform at step 2102) applies a transformation process to generate a transformed waveform (e.g., as...). Figure 16 As shown, a portion of the spontaneous shrinking classifier 1606 applies a transformation process 1622 to waveform 1610. The transformation process includes any suitable signal processing method that can enhance peaks within the waveform while reducing noise. In one embodiment, the transformation process includes the Pan-Tompkins algorithm. In another embodiment, the transformation process includes a modified Pan-Tompkins algorithm, as described below regarding... Figure 24 A more detailed description.

[0359] At step 2304, one or more adaptive thresholds are determined based on the transformed waveform. Thus, the value of one or more adaptive thresholds depends on the transformed waveform. The adaptive thresholds include one or more of a uniform adaptive threshold and a dynamic adaptive threshold. The uniform adaptive threshold is constant over the time period of the waveform, while the dynamic adaptive threshold varies over that time period. Example of uniform and dynamic adaptive thresholds in... Figure 20 It is shown in the figure and described above.

[0360] At step 2306, a classification score is determined based on an adaptive threshold applied to the transformed waveform. Thus, at step 2306, a thresholding operation involving the adaptive threshold determined at step 2304 is performed (e.g., Figure 16The thresholding operation 1624 of the spontaneous contraction classifier 1606 is shown. A second classification score determined at step 2306 indicates that the waveform includes spontaneous contraction when one or more portions of the transformed waveform exceed at least one of one or more adaptive thresholds. Alternatively, the second classification score indicates that the waveform includes spontaneous contraction when one or more portions of the transformed waveform exceed both a uniform adaptive threshold and a dynamic adaptive threshold. For example, if no portion of the transformed waveform exceeds the adaptive threshold, the classification score is determined to be "0"; but if at least one portion of the transformed waveform exceeds the adaptive threshold, the classification score is determined to be "+1" (i.e., the classification score indicates the presence of one or more spontaneous contractions within the waveform). Alternatively, the classification score includes a vector within the waveform corresponding to the location of the identified spontaneous contraction (such that an empty vector indicates that no spontaneous contraction has been identified).

[0361] Figure 24 A modified Pans-Tompkins method 2400 according to an embodiment of the present disclosure is shown.

[0362] The modified Pan-Tompkins method 2400 includes a noise reduction process 2402 and a signal enhancement process 2404. The noise reduction process 2402 includes the following steps: filtering using a low-pass filter 2406; filtering using a high-pass filter 2408; and transforming using a stationary wavelet transform 2410. The signal enhancement process 2404 includes the steps of differentiation 2412, rectification 2414, and integration 2416. The modified Pan-Tompkins method 2400 further includes the optional step of normalizing the output of the signal enhancement process 2404 2418. In one embodiment, the modified Pan-Tompkins method 2400 is performed at the waveform transformation step 2302 in method 2300.

[0363] Although the modified Pan-Tompkins algorithm is described as performing each of the steps described below, those skilled in the art will understand that in some implementations, the steps may be combined and / or omitted. For example, the algorithm may include performing stationary wavelet transform and rectification; or stationary wavelet transform, differentiation, and rectification.

[0364] The denoising process 2402 of the modified Pan-Tompkins method 2400 generates a noise-filtered representation of the waveform. The denoising process 2402 includes applying a bandpass filter to the waveform (e.g., filtering steps 2406 and 2408) and, in contrast to the standard Pan-Tompkins algorithm, transforming the output of the bandpass filter using a stationary wavelet transform 2410.

[0365] At step 2406, a low-pass filter is applied to the waveform to generate a low-pass filtered waveform. The low-pass filter helps remove major conductor mismatches from the waveform. This is corresponding to the result of applying the low-pass filter to waveform 1902. Figure 19 The low-pass filtered waveform 1904 in the figure shows an example low-pass filtered waveform. In one implementation, the low-pass filter includes a two-dimensional Gaussian kernel with a standard deviation of 2.

[0366] At step 2408, a high-pass filter is applied to the low-pass filtered waveform (generated at step 2406) to generate a high-pass filtered waveform. The high-pass filter helps with baseline alignment and removes drift. The result corresponds to the application of a high-pass filter to the low-pass filtered waveform 1904. Figure 19 The waveform 1906 in the figure shows an example high-pass filtered waveform. In one implementation, the high-pass filter includes an elliptic filter with an order of 8, a ripple of 0.5 dB, an attenuation of 40 dB, and edge frequencies of 1 and 20.

[0367] At step 2410, a stationary wavelet transform is applied to the high-pass filtered waveform (generated at step 2408) to generate the transformed waveform. The stationary wavelet transform helps improve the denoising performed at step 2402, thereby improving the classification performance of the spontaneous shrinking classifier. This is based on the result corresponding to the application of the stationary wavelet transform to the high-pass filtered waveform 1906. Figure 19 Waveform 1908 in the figure shows an example transformed waveform. In one implementation, the wavelet transform includes a discrete stationary wavelet transform (1D) using 5 decomposition levels and a Daubechies 4 (db4) wavelet.

[0368] The modified Pan-Tompkins method 2400 generates an enhanced representation of the noise-filtered waveform produced by the noise reduction process 2402 through the signal enhancement process 2404. In contrast to the standard Pan-Tompkins algorithm, the squaring operation is replaced by the rectification step 2414.

[0369] At step 2412, differentiation is used to differentiate the denoised waveform obtained from denoising process 2402 to generate a differentiated waveform. Differentiation is used to highlight rapid changes (i.e., contraction). This is based on the result of differentiating the waveform 1908 obtained from the wavelet transform. Figure 19 The waveform 710 shown in the figure illustrates an example of a waveform processed by differentiation. In one embodiment, the differentiation operation corresponds to a waveform with a transfer function. The five-point derivative, where It is the sampling period.

[0370] At step 2414, the differentiated waveform is rectified to generate a rectified waveform. That is, at step 2414, the differentiated waveform is truncated, thereby removing any negative portions of the differentiated waveform. Therefore, the rectification operation enhances the dominant peak within the waveform. This is based on the result of rectifying the differentiated waveform 1910. Figure 19 The rectified waveform 1912 in the figure shows an example of a rectified waveform.

[0371] At step 2416, a moving window integration operation is applied to the rectified waveform to generate an integrated waveform. The moving window integration operation removes short-duration artifacts from the rectified waveform. This is derived from the result of applying the moving window integration operation to the rectified waveform 1912. Figure 19 The integrated waveform in Figure 1914 shows an example of an integrated waveform. In one implementation, the moving average filter includes a length... The moving mean filter is calculated within the sliding window, where It is the length of waveform 1610.

[0372] At an optional step of normalization 2418, the output of signal enhancement process 2404 (e.g., the integrated waveform produced at step integration 2416) is normalized to scale the waveform values ​​to a scaled value along the curve. The range of axis settings (e.g., between 0 and 1).

[0373] A system and method for classifying the shrinkage types of engineered tissues.

[0374] This disclosure presents a system and method for classifying and processing functional response waveforms to achieve efficient and effective extraction of features across a range of contraction types.

[0375] Figure 25 A dual-contraction type contraction-relaxation cycle model according to an embodiment of the present disclosure is illustrated.

[0376] Figure 25 The single contraction model 2508 is shown (as mentioned above). Figure 3 (described in detail) and the dual contraction model 2510 (alternatively referred to as the dual contraction-relaxation cycle model, dual-type model, or dual model). Dual contraction model 2510 Includes two single contraction models and The combination of each element in a single contraction model consists of its own set of parameters. and The definition makes the overall double contraction model 2510 determined by the parameter set. Parameterization. In this model, Axis offset Both are shared across the single model, but in some embodiments, Variation across models is permitted. By using the double-contraction model 2510 to model functional responses with double-contraction types, individual components of the contraction response can be modeled and thus vary independently. This allows double-contraction type functional responses to be modeled efficiently and effectively, thereby improving the performance of downstream tasks utilizing such models.

[0377] Figure 26 A system 2600 for processing functional response waveforms with different contraction types is shown according to an embodiment of the present disclosure.

[0378] System 2600 includes a contraction type classifier 2602, a single contraction model 2604, a double contraction model 2606, and a waveform generator 2608. System 2600 receives an input waveform 2610 and generates an output waveform 2612, which includes a noise-filtered representation of the input waveform 2610. The input waveform 2610 includes at least one contraction response and at least one relaxation response of the engineered tissue. The input waveform 2610 has a length corresponding to the expected contraction-relaxation cycle of the engineered tissue and includes either a single contraction of the engineered tissue (i.e., the input waveform is of the single contraction type) or a double contraction of the engineered tissue (i.e., the input waveform is of the double contraction type). To identify the relevant contraction type, the input waveform 2610 is classified by the contraction type classifier 2602 into a predicted contraction type 2614 (alternatively referred to as a predicted contraction-relaxation cycle type, a predicted type, or a predicted cycle type). Based on the predicted contraction type 2614, a single contraction model 2604 or a double contraction model 2606 is fitted to the input waveform 2610 to obtain a fitted single model 2616 or a fitted double model 2618. A waveform generator 2608 uses the fitted model (i.e., the fitted single model 2616 or the fitted double model 2618) to generate an output waveform 2612, which includes a noise-filtered representation of the input waveform 2610. For example, the waveform generator 2608 samples multiple points (e.g., 100, 200, 500, 1000, etc.) from the fitted model (i.e., the fitted single model 2616 or the fitted double model 2618) within a time period of the input waveform 2610 to generate the output waveform 2612. Advantageously, this allows for the generation of high-resolution waveform data, which helps ensure more accurate features of the functional response of engineered tissues are extracted. This, in turn, helps improve the accuracy and power of downstream tasks involving functional response features.

[0379] In one embodiment, using, as in the following text about Figure 33 The described method is a way to extract the input waveform from a larger waveform of the functional response 2610.

[0380] The fitted single model 2616 and the fitted dual model 2618 include parameter values ​​for the relevant contraction-relaxation cycle model. In one embodiment, an optimization method is used to further refine the parameter values ​​in the fitted single model 2616 or the fitted dual model 2618. The following will discuss... Figure 29 In more detail, the optimization methods used include simplex search algorithms, such as the Nelder-Mead method.

[0381] In one embodiment, one or more feature values ​​are extracted from the output waveform 2612, and the one or more feature values ​​are output. Examples of the one or more feature values ​​extracted from the output waveform 2612 include one or more of the following: twitch or peak amplitude values ​​(i.e., Figure 2 The peak amplitude shown is 206), and the contraction time value (i.e., Figure 2 The time to reach the peak amplitude (208) and the maximum contraction slope value (i.e., Figure 2 The maximum development rate shown is 214), and the relaxation time value (i.e., Figure 2 The values ​​shown include the time to peak descent 210, the maximum relaxation slope (i.e., the maximum descent rate 216), and the twitch duration (i.e., the duration 212). For waveforms corresponding to the double contraction type, these features may include two values ​​extracted for each feature. For example, a first peak amplitude value associated with the first peak within the double contraction type waveform, and a second peak amplitude value associated with the second peak within the double contraction type waveform.

[0382] Any suitable parameter fitting technique can be used (such as least squares-based fitting or machine learning-based fitting, such as...). Figure 6 The machine learning-based fitting shown is used to fit a single contraction-relaxation model or a double contraction-relaxation model.

[0383] Figure 27 The following is an example of a classification of contraction types for three waveforms according to an embodiment of the present disclosure.

[0384] Figure 27The baseline waveform 2702, the first perturbed waveform 2704, and the second perturbed waveform 2706 are shown. These waveforms encompass the functional response (contraction force) of the artificial tissue under baseline and two perturbed conditions. Examples of perturbed conditions include treatment of the tissue with drugs or compounds, different cell lines, different disease states, physical perturbation of the artificial tissue, and alterations made to a platform or bioreactor containing the artificial tissue. The baseline waveform 2702, the first perturbed waveform 2704, and the second perturbed waveform 2706 are associated with baseline classification vector 2708, first perturbed classification vector 2710, and second perturbed classification vector 2712, respectively. Baseline classification vector 2708 includes multiple classification values ​​associated with the functional response waveforms within the baseline waveform 2702, such as a first classification value 2714, which includes a contraction type classification for the associated functional response waveform 2716 within the baseline waveform 2702. The first perturbation classification vector 2710 includes multiple classification values ​​associated with the functional response waveform within the first perturbation waveform 2704, such as a second classification value 2718 and a third classification value 2720. The second classification value 2718 and the third classification value 2720 include contraction type classifications for the associated functional response waveforms 2722 and 2724 within the first perturbation waveform 2704. The second perturbation classification vector 2712 includes multiple classification values ​​associated with the functional response waveform within the second perturbation waveform 2706.

[0385] The classification vector (e.g., baseline classification vector 2708) is obtained by extracting multiple functional response waveforms from each waveform (as described below). Figure 33 (described in more detail) and then using a predictive model (e.g., the one mentioned above). Figure 6 The contraction type is obtained by determining the contraction type for each functional response waveform using the prediction model 600 described in Figure 7. The contraction type can be a single contraction type or a double contraction type. Figure 27 As shown, the first classification value 2714 and the second classification value 2718 are single contraction types (as indicated by light gray shading) because the predictive model determines that the associated functional response waveforms 2716 and 2722 contain single contraction. In contrast, the third classification value 2720 is a double contraction type (as indicated by dark gray shading) because the predictive model determines that the associated functional response waveform 2724 contains double contraction.

[0386] As stated above, the waveform includes the functional response of the artificial tissue under baseline and two perturbation conditions. Thus, any change in the distribution of contraction types within the classification vector indicates a change in the behavior of the artificial tissue due to perturbation. Figure 27In the example shown, baseline waveform 2702 includes the functional response of the artificial tissue under reference or control conditions, first perturbation waveform 2704 includes the functional response of the artificial tissue under a first dose of isoproterenol, and second perturbation waveform 2706 includes the functional response of the artificial tissue under a second dose of isoproterenol. As can be seen by comparing the first perturbation classification vector 2710 and the second perturbation classification vector 2712 with the baseline classification vector 2708, the contractile behavior of the artificial tissue changes with increasing doses of isoproterenol. As most clearly shown in the second perturbation classification vector 2712, isoproterenol induces regular double contractions in the artificial tissue that are not present in the behavior of the artificial tissue under control conditions.

[0387] We will now turn to methods for classifying and processing functional response waveforms using the systems and models described above.

[0388] Figure 28 A method 2800 for processing a functional response waveform according to one aspect of this disclosure is shown.

[0389] Method 2800 includes the following steps: obtaining 2802 a first waveform; determining 2804 a predicted contraction type; fitting a model to the first waveform based on the predicted contraction type 2806; and generating a second waveform 2808 from the model. Method 2800 further includes the optional steps of: extracting 2810 feature values ​​from the second waveform; and outputting 2812 feature values. In one embodiment, method 2800 is... Figure 1 The model fitting unit 104-1 of the control unit 104 shown is used.

[0390] Generally, method 2800 is used to generate a noise-filtered or noise-suppressed representation of a contraction-relaxation cyclic waveform (functional response waveform), which may include a single contraction-type response or a dual contraction-type response. In an embodiment, the contraction-relaxation cyclic waveform is derived from hardware such as a bioreactor (e.g., Figure 1 The bioreactor 102 shown is used, and due to factors such as sensor variability, signal transmission, and signal conversion, the hardware can introduce noise into the waveform. Therefore, the contraction-relaxation cycle waveform includes a potentially noisy representation of the functional response of the engineered tissue during a single contraction-relaxation cycle. Advantageously, method 2800 effectively classifies the contraction type of the waveform and efficiently generates a noise-filtered representation of the contraction-relaxation cycle, thereby improving the accuracy of the extracted features, which in turn helps improve the performance of downstream tasks utilizing such features.

[0391] At step 2802, a first waveform is obtained. The first waveform includes at least one contraction response and at least one relaxation response of the engineered or artificial tissue. The first waveform has a predetermined length corresponding to the expected length of the contraction-relaxation cycle of the engineered tissue.

[0392] The first waveform (or first functional response waveform) captures the functional response of the artificial tissue. The functional response is the contractile force of the artificial tissue (e.g., the contractile force measured using data obtained from sensor assembly 108 of bioreactor 102, within which the artificial tissue grows or is maintained). Alternatively, the functional response is a contractile displacement, a transient calcium response, or a change in membrane potential.

[0393] Artificial or engineered tissues include engineered muscle tissue, such as engineered heart tissue or engineered skeletal muscle tissue. In one embodiment, a first waveform is emitted from a bioreactor containing the artificial tissue (e.g., Figure 1 The first waveform is obtained from a bioreactor 102 shown. Thus, the first waveform is obtained from a waveform comprising multiple single or double contraction-relaxation cycles of artificial tissue, obtained from the bioreactor. In one embodiment, an extraction method (such as described below) is used. Figure 33 The method described in more detail (3300) is used to obtain or extract the first waveform from the waveform.

[0394] The predetermined length of the first waveform (which corresponds to the expected length of the contraction-relaxation cycle) is from 0.05 s to 10 s, and more specifically from 0.15 s to 1 s. In one embodiment, the predetermined length is proportional to the frequency at which the first waveform is recorded when it stimulates the engineered tissue. For example, if the engineered tissue is stimulated at 1 Hz, the predetermined length is 1 s, because the expected length of the contraction-relaxation cycle of the engineered tissue will be 1 s.

[0395] At step 2804, the predicted contraction type among multiple contraction types is determined for the first waveform. These multiple contraction types include single contraction and double contraction types.

[0396] The predicted type of contraction is determined by the classifier based on the first waveform input to the classifier. (As mentioned above...) Figure 6 and Figures 7A to 7D In more detail, the classifier comprises a trained machine learning model, such as a trained neural network. In one embodiment, the classifier includes a convolutional neural network, a dilated convolutional neural network, and a long short-term memory network. Further architectural details of the classifier are described above regarding... Figure 6 and Figures 7A to 7D Provided.

[0397] At step 2806, the model is fitted to the first waveform based on the predicted contraction type. The model, independent of the growth of at least one relaxation response of the engineered tissue, parameterizes the growth of at least one contraction response of the engineered tissue. Thus, the model does not assume that the underlying functional response is symmetric. This allows the model to fit efficiently and accurately to a range of functional responses from a wide variety of engineered tissue types. More efficient and accurate model fitting helps generate more precise features, which in turn leads to better data.

[0398] The model fitted to the first waveform is selected based on the predicted contraction type. If the predicted contraction type is a single contraction type, then a single contraction type model is fitted (as mentioned above). Figure 3 (Examples and descriptions). If the predicted contraction type is a double contraction type, then fit a double contraction type model (as described above). Figure 25 Examples and descriptions).

[0399] The single contraction type model includes the first contraction function. and the first relaxation function A double-contraction type model includes a first contraction function and a first relaxation function, as well as a second contraction function and a second relaxation function. In one embodiment, the contraction function is an ascending logic function with a positive growth rate, and the relaxation function is a descending logic function with a negative growth rate. Therefore, the contraction response of a single or double contraction-relaxation cycle is modeled by the ascending logic function, and the relaxation response of a single or double contraction-relaxation cycle is modeled by the descending logic function. A single-contraction type model includes the product of a first ascending logic function and a first descending logic function. A double-contraction type model includes a combination (i.e., an additive combination) of two single-contraction type models.

[0400] The contraction-relaxation cycle model includes multiple parameters associated with at least one contraction response and at least one relaxation response. These multiple parameters include at least one maximum parameter. At least one rate of ascent parameter At least one descent rate parameter , Offset parameter At least one increase Offset parameter and at least one decline Offset parameter The above text is about Figure 3 and Figure 25 The relationship between each of these parameters and the overall single contraction-relaxation cycle model or double contraction-relaxation cycle model is shown and described in more detail.

[0401] In one embodiment, data from a machine learning model (e.g., Figure 6The predicted parameter values ​​obtained by the prediction model 600 shown are used to fit the model to the first waveform, as described below. Figure 29 Method 2900 is described in more detail.

[0402] At step 2808, a second waveform is generated from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform. For example, within the time period of the first waveform, multiple points (e.g., 100, 200, 500, 1000, etc.) are sampled from the model fitted at step 2806. Advantageously, this allows for the generation of high-resolution waveform data from the contraction-relaxation cycle, which helps ensure more accurate features of the functional response of engineered tissues are extracted. This, in turn, helps improve the accuracy and power of downstream tasks involving functional response features.

[0403] Optionally, a second waveform may be output. In one embodiment, outputting the second waveform includes storing or saving the second waveform to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the second waveform includes transmitting the second waveform via a network (e.g., a local area network, a wide area network, etc.) or displaying the waveform for user review.

[0404] At the optional step of extraction 2810, one or more feature values ​​are extracted from the second waveform.

[0405] The second waveform is a noise-filtered representation of the first waveform (i.e., a noise-filtered representation of a single or double contraction-relaxation cycle), thus achieving more accurate values ​​of the features of the underlying functional response to be extracted. One or more feature values ​​include one or more of the following: at least one contraction-relaxation cycle or peak amplitude value (i.e., Figure 2 The peak amplitude shown is 206), and the contraction time value (i.e., Figure 2 The time to reach the peak amplitude (208) and the maximum contraction slope value (i.e., Figure 2 The maximum development rate shown is 214), and the relaxation time value (i.e., Figure 2 The time to reach the peak descent (210) and the maximum relaxation slope value (i.e., Figure 2 The maximum rate of decline (216) and the duration of the contraction-relaxation cycle (i.e., Figure 2 The duration shown is 212).

[0406] Once extracted, one or more eigenvalues ​​can be used as quantitative descriptors of single or double contraction-relaxation cycles, and thus provide a numerical representation of the functional response of artificial tissues. (See also: Regarding...) Figure 35 In more detail, such features are used for a variety of downstream processing tasks, such as effect identification in drug discovery and development.

[0407] At an optional step of output 2812, one or more feature values ​​extracted from the second waveform are output. In one embodiment, outputting one or more feature values ​​includes storing or saving one or more feature values ​​to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting one or more feature values ​​includes transmitting one or more feature values ​​via a network (e.g., a local area network, a wide area network, etc.) or displaying one or more feature values ​​for user review.

[0408] Figure 29 A method 2900 for fitting a model to a functional response waveform according to an embodiment of the present disclosure is shown.

[0409] In one embodiment, method 2900 is performed as part of the fitting 906 step of method 900. Method 2900 includes a step of predicting 2902 multiple values ​​for multiple parameters, and further includes an optional step of optimizing 2904 multiple values.

[0410] The steps in method 2900 are used to predict parameter values ​​for the model from the first waveform. This model is a single-shrinkage type model (…). Figure 3 ) or double contraction type model ( Figure 25 Which model is fitted by method 2900 depends on the type of contraction of the prediction determined at step 2804 of method 2800. The trained machine learning model (e.g., Figure 6 The parameter estimation model shown is used to predict parameter values ​​so that the fitted model closely approximates the first waveform.

[0411] At step 2902, multiple values ​​are predicted for multiple parameters of the model, such that the model fitted to the first waveform includes multiple values ​​for multiple parameters. These multiple values ​​are predicted by either a single-contraction type model or a double-contraction type model.

[0412] At an optional step of optimization 2904, the multiple values ​​determined at the prediction 2902 step are optimized by minimizing the error between the first waveform and the model fitted to the first waveform (i.e., using multiple values ​​for multiple parameters of the model).

[0413] Given multiple values ​​for the model determined at step 2902. In this case, at step 2904, multiple updated values ​​are sought. , making .here, It is a cost or loss function that measures the first waveform. According to the parameter value set Model fitted to the first waveform The error between them. A lower value indicates a better fit of the model to the first waveform. In one embodiment, the cost function It is the root mean square error:

[0414]

[0415] In one embodiment, it corresponds to:

[0416]

[0417] The optimization of this is a multidimensional problem because the model involves multiple parameters. Therefore, making Minimization requires fitting multiple parameters simultaneously. To perform this minimization, in one embodiment, a simplex search algorithm (such as the Nelder-Mead method) is used to optimize multiple values. Advantageously, performing optimization after obtaining initial predictions of the parameters helps to achieve an accurate model fit, which helps avoid local minima and makes better use of processing resources, as the optimization process starts from a solution expected to be close to the optimal solution.

[0418] Figure 30 A method 3000 for training a classifier to predict tissue contraction type from functional response waveforms is shown.

[0419] Method 3000 includes the following steps: obtaining 3002 multiple waveforms; extracting 3004 a first set of multiple parameters; extracting 3006 a second set of multiple parameters; determining 3008 the distribution of the multiple parameter sets; generating 3010 a synthetic training dataset; and training 3012 a classifier on the synthetic training dataset. In one embodiment, method 3000 consists of... Figure 1 The control unit 104 or its subunit shown is used for this purpose.

[0420] In many cases, the effectiveness of predictive models (e.g., machine learning models such as deep neural networks) is limited by the quantity and quality of available training data. Without a large volume of high-quality training data, predictive models will often fail to produce sufficient output. In this disclosure, this problem leads to inaccurate identification of single or double-type shrinkage, which subsequently reduces the effectiveness and applicability of using such identifiers in practice for tasks such as drug discovery and development. Method 3000 aims to address this problem by generating high-fidelity synthetic training data, thereby allowing for the generation of a virtually unlimited amount of data. This helps improve the performance of the trained predictive model, which in turn improves the accuracy of the model's predictions. This improvement in accuracy helps drive improvements in downstream tasks that utilize classifications derived from such models.

[0421] At step 3002, multiple waveforms are obtained. These waveforms represent the functional responses of one or more engineered tissues. Each of the multiple waveforms has a predetermined length corresponding to the expected length of the contraction-relaxation cycle of the engineered tissue.

[0422] At step 3004, a first set of multiple parameters is extracted from a first subset of multiple waveforms associated with a single contraction type. The first set of parameters in the first set of multiple parameters represents a first waveform in the first subset of the waveforms.

[0423] At step 3006, a second set of multiple parameters is extracted from a second subset of multiple waveforms associated with the double contraction type. The second set of parameters in the second set of multiple parameters characterizes a second waveform within the second subset of the waveforms.

[0424] At step 3008, multiple parameter set distributions are determined. These multiple parameter set distributions include a first parameter set distribution determined from a first set of multiple parameters and a second parameter set distribution determined from a second set of multiple parameters.

[0425] At step 3010, a synthetic training dataset is generated. Each element of the synthetic training dataset includes a synthetic waveform and the corresponding tissue contraction type associated with that waveform. The synthetic waveform is generated using a parameter set distribution from multiple parameter set distributions associated with the corresponding tissue contraction type.

[0426] At step 3012 of the training algorithm, the classifier is trained using a synthetic training dataset. The classifier trained using the synthetic training dataset determines the predicted type of tissue contraction for the input waveform.

[0427] Figure 31 A method 3100 for predicting the type of tissue contraction for a waveform using a synthetically trained classifier is shown according to an embodiment of the present disclosure.

[0428] Method 3100 includes the steps of obtaining 3102 a first waveform and predicting 3104 the type of tissue contraction. Method 3100 also includes the optional step of outputting 3106 the type of tissue contraction. In one embodiment, method 3100 is... Figure 1 The control unit 104 shown or its sub-units (such as model fitting unit 104-1) are used for this purpose.

[0429] At step 3102, a first waveform is obtained. The first waveform includes the functional response of the first artificial tissue. The first waveform has a length corresponding to the expected length of the contraction-relaxation cycle of the first artificial tissue.

[0430] At the step of predicting 3104, a classifier trained on the synthetically generated training dataset is used (as mentioned above). Figure 30The type of tissue contraction is predicted from the first waveform.

[0431] At an optional step of output 3106, the tissue shrinkage type is output. In one embodiment, outputting the tissue shrinkage type includes storing or saving the tissue shrinkage type to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the tissue shrinkage type includes transmitting the tissue shrinkage type via a network (e.g., a local area network, a wide area network, etc.) or displaying the tissue shrinkage type for user review.

[0432] Figure 32 A method 3200 for training a parameter estimation model using synthetic training data is shown according to one aspect of this disclosure.

[0433] Method 3200 includes the following steps: obtaining 3202 multiple waveforms; extracting 3204 multiple parameter sets; determining 3206 the distribution of the parameter sets; generating 3208 a synthetic training dataset; and training 3210 a prediction model on the synthetic training dataset. In one embodiment, method 3200 consists of... Figure 1 The control unit 104 or its subunit shown is used for this purpose.

[0434] In many cases, the effectiveness of predictive models (e.g., machine learning models such as deep neural networks) is limited by the quantity and quality of available training data. Without a large volume of high-quality training data, predictive models will often fail to produce sufficient output. In this disclosure, this problem leads to inaccurate shrink-relaxation models being fitted, which subsequently reduces the effectiveness and applicability of such models in practice for tasks such as drug discovery and development. Method 3200 aims to address this problem by generating high-fidelity synthetic training data, thereby allowing for the generation of a virtually unlimited amount of data. This helps improve the performance of the trained predictive model, which in turn improves the accuracy of models fitted using that predictive model. This improvement in accuracy helps drive improvements to downstream tasks that utilize features extracted from such models.

[0435] At step 3202, multiple waveforms are obtained. These waveforms represent the functional responses of one or more artificial tissues. All of the waveforms include a common contraction type from a predetermined set of contraction types. For example, all of these waveforms may be either single-contraction type or double-contraction type waveforms. Therefore, the model trained by method 3200 is trained to predict the parameters of a single-contraction type model or a double-contraction type model based on the common contraction type among the multiple waveforms.

[0436] Multiple waveforms correspond to real or guided data from which synthetic training datasets will be generated. Each waveform corresponds to a time series representing the values ​​of the functional response (e.g., contractile force, calcium transient, etc.) of an artificial or engineered tissue within a single contraction-relaxation cycle or a double contraction-relaxation cycle. Therefore, each waveform contains at least one contraction cycle and at least one relaxation cycle, and can be determined according to the above regarding... Figure 3 and Figure 25 The described single or double contraction-relaxation models are parameterized.

[0437] To help ensure that the synthetic training data represents the broadest possible population of contraction-relaxation cycles, multiple waveforms are preferably obtained from a variety of artificial tissues under a series of different conditions. Artificial tissues include one or more engineered muscle tissues, such as engineered cardiac tissue and / or engineered skeletal muscle tissue. Diversity is achieved by obtaining waveforms from artificial tissues across a range of different cell lines, disease states, and treatment methods. Alternatively, the range of conditions within the multiple waveforms is restricted, allowing the synthetic data and subsequent parameter estimation models to be fine-tuned for specific applications. For example, the multiple waveforms may be restricted to vector-processed waveforms to generate control parameter estimation models, or they may be restricted to specific tissue types (e.g., engineered cardiac tissue) to generate tissue-specific parameter estimation models. Advantageously, this helps improve the performance of the parameter estimation models when it is known which category of waveforms the parameter estimation model will be used for.

[0438] At step 3204, multiple parameter sets are extracted from multiple waveforms. The parameter sets in the multiple parameter sets represent the corresponding waveforms in the multiple waveforms.

[0439] The parameter set includes parameters for single or double contraction-relaxation cycle models (e.g., as mentioned above regarding...). Figure 3 and Figure 25 As stated above, the model used depends on the common contraction type of the multiple waveforms obtained at step 3202 described above. In one embodiment, the set of parameters associated with the waveform includes at least one maximum parameter value (as described above). ), offset parameter value ( ), at least one contraction midpoint parameter value ( ), at least one contraction rate parameter value ( ), at least one relaxation midpoint parameter value ( ) and at least one relaxation rate parameter value ( ).

[0440] Multiple parameter sets are extracted using supervised, unsupervised, or semi-supervised methods. With supervised methods, the parameter sets are manually fitted to each waveform. For example, a first waveform is presented to the user, who adjusts the values ​​of the parameters within the parameter set until a second waveform, generated by a contraction-relaxation loop model fitted based on that parameter set, closely matches the first waveform. The final parameter set producing the closely matched second waveform is then used as the parameter set in one of the multiple parameter sets associated with the first waveform. With unsupervised methods, the parameter set is automatically fitted to the waveform (e.g., using the parameters described above). Figure 9 and Figure 10 (The method described). In one embodiment, the unsupervised method utilizes a trained machine learning model to predict parameter set values ​​for a waveform. According to the semi-supervised method, the automatically determined parameter set obtained from the unsupervised method is manually reviewed and refined by one or more users.

[0441] At step 3206, the parameter set distribution is determined from multiple parameter sets.

[0442] The multiple parameter sets extracted at the extraction step 3204 of method 3200 include multiple values ​​for each parameter of the (single or double) contraction-relaxation cycle model. For example, if 100 parameter sets are extracted, there will be 100 parameter values ​​extracted for each parameter of the contraction-relaxation cycle model. At the determination step 3206, a distribution or value distribution is determined for each parameter of the model. In one embodiment, the distribution is determined independently for each parameter. Alternatively, a multivariate distribution is determined for the multiple parameters forming the parameter sets.

[0443] Any suitable method can be used to determine the parameter set distribution, such as histogram-based methods, density estimation methods, and clustering methods. In one embodiment, a kernel density estimation (KDE) method is used to determine the parameter set distribution, which utilizes a kernel and bandwidth parameters to estimate the parameter set distribution. In one implementation, a normal (Gaussian) kernel is used in conjunction with bandwidth selected using cross-validation or a bandwidth selection method such as the Scott rule or Silverman rule.

[0444] Once the distribution of the parameter set has been determined, the parameter set can be obtained by sampling from that distribution (i.e., sampling from each individual distribution or from the joint distribution).

[0445] At step 3208, a synthetic training dataset is generated. Each element of the synthetic training dataset includes a synthetic waveform and a corresponding parameter set used to generate the synthetic waveform. The corresponding parameter set is obtained from a parameter set distribution.

[0446] The synthetic training dataset is generated by repeatedly sampling the parameter set for the contraction-relaxation loop model from the parameter set distribution (as described above) and generating a corresponding waveform for each of the sampled parameters. In this way, training datasets of any size can be generated efficiently (e.g., 1000, 10000, 100000 training data elements, etc.). The synthetic data will also closely approximate real waveform data because the parameter set distribution is modeled on real-world data. The synthetic waveform will be a single contraction type waveform (e.g., regarding...). Figure 3 (Examples and descriptions) or double-contraction type waveforms (such as those related to...) Figure 25 Examples and descriptions).

[0447] Optionally, a noise component is added to each of the waveforms in the synthetic training dataset. The noise component is determined via a uniform distribution derived from multiple waveforms.

[0448] At step 3210, the prediction model is trained using a synthetic training dataset. The prediction model is trained to estimate the output parameter set from the input waveform. In one embodiment, the prediction model corresponds to the above regarding... Figure 6 The parameter estimation model described is (i.e., prediction model 600). As stated above, in one embodiment, training the parameter estimation model includes using mini-batch gradient descent with a batch size of 128 and an ADAM solver. The ADAM solver has an initial learning rate of 1e-3, where early stopping is performed based on validation loss. Further details regarding the training of the parameter estimation model at step 3210 are provided above regarding... Figure 6 The description is given.

[0449] Figure 33 A method 3300 for extracting single or double contraction-relaxation cycle waveforms according to one aspect of this disclosure is shown.

[0450] Method 3300 includes the following steps: obtaining 3302 a first waveform; convolving the first waveform with a pulse train 3304; identifying 3306 a first position; and extracting 3308 a second waveform from the first waveform at the first position. Method 3300 also includes the optional step of outputting 3310 the second waveform. In one embodiment, method 3300 is... Figure 1 The signal processing unit 104-2 of the control unit 104 shown performs this function.

[0451] Generally, method 3300 extracts single or double contraction-relaxation cycle waveforms from larger waveforms that include multiple contraction-relaxation cycles. The larger waveforms can be derived from hardware devices such as bioreactors (i.e., Figure 1The waveform obtained from the bioreactor 102 shown typically corresponds to the functional response of an artificial tissue under specific conditions. For example, a larger waveform could include the contractile response of engineered cardiac tissue stimulated at a frequency of 1 Hz over a 30-second time period. In this example, the larger waveform would include approximately 30 peaks or 30 contraction-relaxation cycles, corresponding to the contractions of the engineered cardiac tissue in response to electrical stimulation. Method 3300 provides an efficient and accurate mechanism for extracting each (single or double) contraction-relaxation cycle from the larger waveform, such that these sub-waveforms can then be used for further processing and analysis (e.g., fitting models to these waveforms and extracting relevant features, as described above regarding...). Figure 28 (as described).

[0452] At step 3302, a first waveform is obtained. The first waveform includes multiple functional responses of the artificial tissue stimulated at a first frequency.

[0453] In one embodiment, the first waveform originates from a bioreactor (e.g., Figure 1 The bioreactor 102 shown provides an environment in which artificial or engineered tissues are grown / maintained. As previously stated, artificial tissues include engineered muscle tissues, such as engineered heart tissues or engineered skeletal muscle tissues.

[0454] As mentioned above Figure 1 The procedure involves applying electrical stimulation to the cell culture during maturation and, once matured, to the artificial tissue to mimic the natural physiological environment of the artificial tissue, thereby allowing measurement of the artificial tissue's functional response to the stimulation. The first waveform obtained at step 3302 includes the artificial tissue's functional response to stimulation at a first frequency. In one embodiment, the first frequency or pacing frequency for stimulating the artificial tissue is 0.1 Hz to 20 Hz. In another embodiment, the first frequency is 1 Hz to 6 Hz.

[0455] Optionally, method 3300 includes a step of stimulating (not shown) an artificial tissue at a first frequency prior to the step of obtaining the first waveform 3302. For example, stimulating the artificial tissue into a bioreactor containing the artificial tissue (e.g., Figure 1 The bioreactor 102 in the middle sends instructions (e.g., Figure 1 Instruction 126) or command in order to cause the bioreactor to stimulate the artificial tissue at a first frequency.

[0456] At step 3304 of convolution, the first waveform is convolved with the pulse train to generate the convolved waveform. The pulse train is generated at a first frequency.

[0457] A pulse train corresponds to an idealized representation of the functional response of an artificial tissue at a first frequency. It is well known that a pulse train, or pulse wave, comprises... Duration of frequency The waveform of a non-sinusoidal (rectangular) pulse or wave, in which This is the period of the pulse train. Therefore, the duty cycle of the pulse train is... Therefore, the pulse train convolved with the first waveform includes pulses with periodicity. and duration A rectangular pulse sequence, in which It is the first frequency.

[0458] Optionally, method 3300 further includes a step of generating (not shown) a pulse train at a first frequency prior to the step of performing convolution 3304.

[0459] Given a first waveform and pulse train In the case of convolution at step 3304, the convolution is performed. for:

[0460]

[0461] At step 3306, a first position is identified that is associated with the maximum value of the convolved waveform. This first position corresponds to the expected position of the first (single or double) contraction-relaxation cycle.

[0462] The maximum value of the convolved waveform corresponds to the optimal alignment point between the first waveform and the pulse train. Thus, the location of the maximum value of the convolved waveform is used to identify the most likely location of a single contraction-relaxation cycle within the first waveform. The location of the maximum value of the convolved waveform also provides an anchor point from which other contraction-relaxation cycles can be extracted from the first waveform.

[0463] At step 3308, a second waveform is extracted from a first position of the first waveform. The second waveform includes a first contraction-relaxation cycle and has a first duration proportional to the first frequency. The second waveform is a single contraction type (e.g., ...). Figure 3 (as shown in the example) or double contraction type (such as...) Figure 25 example in).

[0464] The first position identified at step 3306 corresponds to the most likely location of a (single or double) contraction-relaxation cycle within the first waveform. Therefore, the second waveform extracted from the first position includes this contraction-relaxation cycle. Since the first waveform corresponds to the functional response of the artificial tissue when stimulated at a predetermined frequency, the duration or length of the second waveform is proportional to that frequency. For example, if the artificial tissue is stimulated at a frequency of 1 Hz, the first duration will be 1 s; if the artificial tissue is stimulated at a frequency of 2 Hz, the first duration will be 0.5 s, and so on.

[0465] Therefore, the second waveform corresponds to a window or subframe within the first waveform, the window or subframe having a length corresponding to the first duration. In one embodiment, the second waveform is centered at a first position such that the midpoint of the second waveform is aligned or substantially aligned with the first position of the first waveform.

[0466] At an optional step of output 3310, a second waveform is output. In one embodiment, outputting the second waveform includes storing or saving the second waveform to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the second waveform includes transmitting the second waveform via a network (e.g., a local area network, a wide area network, etc.) or displaying the waveform for user review.

[0467] In one embodiment, outputting the second waveform includes outputting the second waveform to another process or method of this disclosure. For example, the second waveform may be output to the method 2800 described above, such that the step of obtaining 2802 includes obtaining the second waveform from method 3300.

[0468] Figure 34 A method 3400 for extracting additional single or double contraction-relaxation cycles from a waveform is shown according to an embodiment of the present disclosure.

[0469] Method 3400 includes the steps of identifying a second position 3402, extracting a third waveform 3404 from a first waveform at the second position, and optionally outputting a third waveform 3406. Method 3400 is performed after method 3300. Specifically, method 3400 may be performed after the step of identifying the first position 3306, and may be performed in parallel with the step of extracting the second waveform 3308. In one embodiment, method 3400 is... Figure 1 The signal processing unit 104-2 of the control unit 104 shown performs this function.

[0470] Method 3400 is used to extract additional waveforms from a first waveform, including single or double contraction-relaxation cycles. Advantageously, the extraction performed in method 3400 is efficient and highly parallel because method 3400 utilizes prior information about the expected location of single or double contraction-relaxation cycles within the first waveform, thereby enabling the independent extraction of contraction-relaxation cycles.

[0471] At step 3402, the second position is identified based on the first position and the first frequency. The second position corresponds to the expected position of the second contraction-relaxation cycle.

[0472] The first position (identified at step 3306 of method 3300) corresponds to the optimal alignment between the first waveform and the pulse train. Therefore, the first position can be understood as the most likely location of a contraction-relaxation cycle within the first waveform. Since the first waveform includes the functional response of the artificial tissue at a predetermined frequency (i.e., the first frequency), other contraction-relaxation cycles related to the functional response of the artificial tissue are likely located at positions spaced apart from the first position. Therefore, the first position can serve as an anchor point within the first waveform from which other contraction-relaxation cycle waveforms can be extracted.

[0473] The second position corresponds to the intended position of the second contraction-relaxation cycle (single or double) and will be spaced from the first position by a distance proportional to the first frequency. Specifically, the first position within a given first waveform In the case of the second position It will be ,in It is the step size factor, and It is the period of the pulse train. Therefore, the scaling factor can be set to... To identify the next single or double contraction-relaxation cycle, and also by setting the scaling factor to... This is used to identify the previous contraction-relaxation cycle.

[0474] At step 3404, a third waveform is extracted from a second position of the first waveform. The third waveform includes a second (single or double) contraction-relaxation cycle and has a second duration proportional to the first frequency.

[0475] The third waveform corresponds to the functional response of the artificial tissue when stimulated at a predetermined frequency (i.e., a single or double contraction-relaxation cycle). Therefore, the duration or length of the third waveform is proportional to the frequency. For example, if the artificial tissue is stimulated at a frequency of 1 Hz, the second duration will be 1 s; if the artificial tissue is stimulated at a frequency of 2 Hz, the second duration will be 0.5 s, and so on. In one embodiment, the first and second durations are the same.

[0476] In one embodiment, the third waveform is centered at the second position such that the midpoint of the third waveform is aligned or substantially aligned with the second position of the first waveform.

[0477] At an optional step of output 3406, a third waveform is output. In one embodiment, outputting the third waveform includes storing or saving the third waveform to a persistent storage device, such as non-volatile memory, non-transitory media, etc. Additionally or alternatively, outputting the third waveform includes transmitting the third waveform via a network (e.g., a local area network, a wide area network, etc.) or displaying the waveform for user review.

[0478] In one embodiment, outputting a third waveform includes outputting the third waveform to another process or method of this disclosure. For example, the third waveform may be output to the method 2900 described above, such that the step of obtaining 2802 includes obtaining the third waveform from method 3400.

[0479] As stated above, method 3200 for extracting additional single-twitch-relaxation cycle waveforms or double-twitch-relaxation cycle waveforms from a first waveform can be repeated for all twist-relaxation cycles within the first waveform. Since the extraction performed by method 3400 depends only on the first position and the first frequency, no additional signal processing or analysis is required to identify the positions of the additional twist-relaxation cycles. Therefore, method 3400 provides a fast and efficient method for extracting single-twitch-relaxation cycles or double-twitch-relaxation cycles from a waveform. These waveforms can then be further processed, for example, by fitting a model to the waveform to generate a noise-filtered representation of the waveform.

[0480] Figure 35 A method 3500 for predicting the effects of a disturbance is shown according to one aspect of this disclosure.

[0481] Method 3500 includes the following steps: obtaining 3502 multiple signals; splitting the multiple signals 3504 into a first plurality of waveforms; determining 3506 the predicted type of contraction; fitting a model 3508 to each of the first plurality of waveforms; generating 3510 a second plurality of waveforms from the model; extracting 3512 a first feature value; extracting 3512 a second feature value; and determining 3516 the effect. In one embodiment, method 3500 is composed of... Figure 1 The control unit 104 or its subunit shown is used for this purpose.

[0482] Generally, Method 3500 describes the application of a single contraction-relaxation cycle model, or a single contraction-relaxation cycle model, to downstream drug discovery / development tasks. Specifically, the contraction-relaxation cycle model is used to efficiently generate accurate feature values ​​from baseline and perturbation signals of engineered tissues. Accurate feature extraction from these signals allows for efficient and accurate identification of effects associated with perturbations.

[0483] At step 3502, multiple signals are obtained. These signals include a baseline signal and a perturbation signal. The baseline signal includes a first set of multiple functional responses of the engineered organization under reference or baseline conditions. The perturbation signal includes a second set of multiple functional responses of the engineered organization under a first set of perturbation conditions.

[0484] Baseline and perturbation signals include multiple functional responses (i.e., multiple contraction-relaxation cycles or peaks) of the engineered tissue under a reference condition and a first set of perturbation conditions. Generally, a reference condition refers to a condition that provides a baseline comparison to the perturbation conditions. In embodiments, a reference condition is a condition associated with a control setting or environment. A reference condition may correspond to an engineered or artificial tissue in its default, natural, or unaltered state (i.e., without a dose of drug or agent). Alternatively, a reference condition may correspond to engineered tissue treated with a carrier. A perturbation condition refers to a condition in which the engineered tissue has been perturbed in some way. Examples of perturbation conditions include administration of a drug or compound (i.e., a perturber), disease state, different cell lines, physical perturbation applied to the engineered tissue, or alteration of the environment of the engineered tissue. Thus, perturbation conditions may alternatively be referred to as therapeutic conditions. In the case of perturbation involving a drug or compound, these conditions are further associated with effects related to the drug or compound, such as mechanism of action or toxicity. Given a range of different perturbation conditions, engineered tissues can be associated with more than one perturbation condition (e.g., engineered tissues that have been treated with a specific compound).

[0485] In one embodiment, the baseline signal and the perturbation signal are emitted from the bioreactor (e.g., Figure 1 The engineered or artificial tissue is obtained in a bioreactor 102, in which it is grown / maintained. As previously stated, the artificial tissue includes engineered muscle tissue, such as engineered cardiac tissue or engineered skeletal muscle tissue. The baseline signal and perturbation signal are obtained at two distinct time points. For example, the baseline signal is obtained at a first time point, and then the engineered tissue is perturbed according to a first perturbation (e.g., by applying a first dose of the compound to the engineered tissue), and the perturbation signal is obtained at a second time point after the first time point. The baseline signal and perturbation signal include the functional response of the engineered tissue when stimulated at a predetermined pacing frequency (e.g., 0.1 Hz, 0.5 Hz, 1 Hz, 2 Hz, etc.). Alternatively, the baseline signal and perturbation signal include the spontaneous functional response of the engineered tissue in the absence of external stimulation.

[0486] At step 3504, the multiple signals are split into a first plurality of waveforms. Each of the first plurality of waveforms includes at least one contraction response and at least one relaxation response of the engineered tissue during a single contraction-relaxation cycle.

[0487] The method 3300 described above for extracting contraction-relaxation cycle waveforms is used to split multiple signals into a first plurality of waveforms. Alternatively, the multiple signals can be split by manually annotating or extracting regions within the first plurality of waveforms that correspond to individual contraction-relaxation cycles.

[0488] The first subset of waveforms generated at step 3504 includes a first subset of waveforms associated with waveforms extracted from the baseline signal and a second subset of waveforms associated with waveforms extracted from the perturbation signal. Subsequent model fitting (described below) is independent of the source of the waveforms (i.e., independent of whether the waveform is a reference waveform or a perturbation waveform), but uses an identifier indicating whether the waveform corresponds to a reference condition or a perturbation condition in a subsequent feature extraction step.

[0489] At step 3506, for each of the first plurality of waveforms, a predicted contraction type among multiple contraction types is determined. These multiple contraction types include single contraction types and double contraction types.

[0490] The predicted type of contraction is determined by the classifier based on the waveform input to the classifier. (As mentioned above...) Figure 6 and Figures 7A to 7D In more detail, the classifier comprises a trained machine learning model, such as a trained neural network. In one embodiment, the classifier includes a convolutional neural network, a dilated convolutional neural network, and a long short-term memory network. Further architectural details of the classifier are described above regarding... Figure 6 and Figures 7A to 7D Provided.

[0491] At the step of fitting 3508, the model is fitted to each of the first plurality of waveforms based on the corresponding predicted contraction type. For example, if the waveform is a single contraction type, a single contraction type model is fitted at the step of fitting 3508. Alternatively, if the waveform is a double contraction type, a double contraction type model is fitted at the step of fitting 3508. This model independently parameterizes the growth of at least one contraction response and at least one relaxation response of the engineered tissue.

[0492] The step of fitting the model to 3508 for each waveform corresponds to the above section on... Figure 28 The steps of fitting 2806 are described in more detail. Therefore, at step 3508, step 2806 is repeated for each of the first plurality of waveforms. The process of fitting the model to the waveform is described above regarding... Figure 28 and Figure 29 To describe in more detail.

[0493] At step 3510, a second plurality of waveforms is generated from the model fitted to each of the first plurality of waveforms. The second plurality of waveforms includes a plurality of filtered baseline waveforms associated with a baseline signal and a plurality of filtered perturbation waveforms associated with a perturbation signal.

[0494] The step of generating 3510 corresponds to the step of generating 908 (as mentioned above). Figure 28(To be described in more detail) The model is repeatedly applied to each waveform fitted to the first plurality of waveforms. The process of generating the second waveform from the model fitted to the first waveform is described above regarding... Figure 28 and Figure 29 To describe in more detail.

[0495] At step 3512, the first feature value of the first feature is extracted from multiple filtered baseline waveforms.

[0496] Multiple filtered baseline waveforms include noise-filtered or noise-suppressed representations of the contraction-relaxation cycle in the baseline signal. Because the signal has been filtered to remove noise, features can be accurately extracted from these waveforms.

[0497] The step of extracting the first feature value (3512) includes extracting a plurality of feature values ​​from a plurality of filtered baseline waveforms, such that the first feature value includes a plurality of feature values ​​or a representation of a plurality of feature values. Thus, a value of the first feature is extracted from each of the plurality of filtered baseline waveforms to determine the first feature value. In one embodiment, the first feature value includes the average (mean, median, etc.) of the first feature determined from the plurality of filtered baseline waveforms. In another embodiment, the first feature value includes the maximum value, minimum value, or value distribution determined from the plurality of filtered baseline waveforms.

[0498] The primary characteristic is one of the following: convulsions or peak amplitude (i.e., Figure 2 The peak amplitude shown is 206), and the contraction time (i.e., Figure 2 The time to reach peak amplitude (208) and the maximum contraction slope (i.e., Figure 2 The maximum development rate shown is 214), and the relaxation time (i.e., Figure 2 The time to reach the peak descent (210) and the maximum relaxation slope (i.e., Figure 2 The maximum rate of decline shown is 216) or the duration of the convulsion (i.e., Figure 2 The duration shown is 212).

[0499] In one embodiment, the first feature is the total number of single-contraction type waveforms or the total number of double-contraction type waveforms (as mentioned above). Figure 27 (as described above). Thus, the first characteristic value corresponds to the number or distribution of single-contraction type waveforms and / or double-contraction type waveforms within multiple filtered baseline waveforms.

[0500] At step 3514, the second feature value of the first feature is extracted from multiple filtered perturbation waveforms.

[0501] The step of extracting the second feature value 3514 includes extracting a plurality of feature values ​​from a plurality of filtered perturbation waveforms, such that the second feature value includes the plurality of feature values ​​or a representation of the plurality of feature values. Thus, a value of a first feature is extracted from each of the plurality of filtered perturbation waveforms to determine the second feature value. In one embodiment, the second feature value includes the average (mean, median, etc.) of the first feature determined from the plurality of filtered perturbation waveforms. In another embodiment, the second feature value includes the maximum value, minimum value, or value distribution determined from the plurality of filtered perturbation waveforms.

[0502] As stated above, in one embodiment, the first feature is the total number of single-shrinkage type waveforms or the total number of double-shrinkage type waveforms, such that the second feature value corresponds to the number or distribution of single-shrinkage type waveforms and / or double-shrinkage type waveforms within a plurality of filtered perturbation baseline waveforms.

[0503] At step 3516, the effect associated with the first perturbation is determined based on a comparison of the first eigenvalue and the second eigenvalue.

[0504] The first eigenvalue is a quantitative descriptor of the functional response of the engineered organization under reference conditions. The second eigenvalue is a quantitative descriptor of the functional response of the engineered organization under perturbation conditions involving the first perturbation. Thus, the descriptions of the first and second eigenvalues ​​reveal any changes in the functional response or effect of the engineered organization due to the first perturbation. (The above text is about...) Figure 27 Here is an example of such a comparison.

[0505] As another example, the first perturbation may correspond to the application of a compound with unknown physiological effects. A comparison of a first eigenvalue (in this example, corresponding to the peak amplitude of the contractile force waveform of the engineered tissue under reference conditions) and a second eigenvalue (corresponding to the peak amplitude of the contractile force waveform of the engineered tissue under the perturbation conditions involving the application of the compound) reveals an increase in the average peak amplitude. Therefore, it can be inferred that the compound has an effect associated with increased contractile force of the engineered tissue during the contraction-relaxation cycle. Advantageously, since the eigenvalues ​​are determined from noise-filtered waveforms, the effects resulting from differences between the eigenvalues ​​can be more accurately identified, leading to improved treatment and potentially improved patient outcomes.

[0506] Systems and methods for tracking tissue scaffolds.

[0507] Figure 36 A bioreactor (such as) according to an embodiment of the present disclosure is shown. Figure 1 Well 3602 of the bioreactor 102 shown.

[0508] Well 3602 includes artificial tissue 3604 (or engineered tissue) attached to a first flexible support 3606 and a second flexible support 3608. Figure 36 Further illustrated is an imaging device 3610 configured to acquire one or more images of the artificial tissue 3604, the first flexible scaffold 3606, and / or the second flexible scaffold 3608. In one embodiment, the well 3602, the artificial tissue 3604, the first flexible scaffold 3606, the second flexible scaffold 3608, and the imaging device 3610 respectively correspond to Figure 1 The imaging device or optical sensor shown includes well 114, engineered tissue 124, first support 120-1, second support 120-2, and sensor assembly 108.

[0509] As mentioned above Figure 1 As stated, the artificial tissue 3604 (or engineered tissue) is grown within well 3602 by cells seeded in the well. During maturation, the artificial tissue 3604 attaches to a first flexible scaffold 3606 and / or a second flexible scaffold 3608. Both the first flexible scaffold 3606 and the second flexible scaffold 3608 are disposed across well 3602, thereby allowing the artificial tissue 3604 to attach thereto. The first flexible scaffold 3606 and the second flexible scaffold 3608 include flexible elements arranged to deflect or deform in response to forces applied thereto. For example, the first flexible scaffold 3606 is arranged in response to contractile forces exerted on the first flexible scaffold 3604 by the artificial tissue 3604. Deformation occurs in the first direction. Similarly, the second flexible scaffold 3608 is arranged to respond to the contractile force exerted on the second flexible scaffold 3608 by the artificial tissue 3604. Deformation occurs in the second direction. In another example, the first flexible support 3606 and the second flexible support 3608 are arranged to deflect in response to a predetermined force applied thereto by a probe or other instrument. In one embodiment, the first flexible support 3606 and the second flexible support 3608 are made of flexible polymer threads, such as poly(octamethylene maleate (anhydride) citrate) (POMaC) threads. In such embodiments, the flexible support may be referred to as a thread or flexible wire.

[0510] Imaging device 3610 is configured to detect deformation or deflection of the first flexible scaffold 3606 and / or the second flexible scaffold 3608 (e.g., due to contractile forces applied to the flexible scaffold by the artificial tissue 3604 or probe). Imaging device 3610 is configured to obtain multiple image-based representations of one or more deflections of the first flexible scaffold 3606 and / or the second flexible scaffold 3608 within a time frame or a time period. For example, imaging device 3610 may be configured to detect deformation or deflection of the first flexible scaffold 3606 and / or the second flexible scaffold 3608 within each time frame or time period. Captures images or frames of tissue or a region of tissue attached to a tissue scaffold in seconds. Here, Associated with a predetermined rate at which the image or frame will be captured. For example, when At that time, one image is captured per second. (Optional) Any suitable value, such as In one embodiment, the frame rate is determined based on the frequency at which the tissue within the device is being stimulated. Therefore, an image or frame sequence of the tissue within a time frame captures one or more deflections of the first flexible scaffold 3606 and / or the second flexible scaffold 3608 within that time frame. Images captured by the imaging device 3610 can be output from the device or bioreactor to a control unit or processing unit (e.g., images captured by the sensor assembly 108 can be output to...). Figure 1 (The control unit 104 shown).

[0511] Figure 37 Example images of tissue scaffolds according to embodiments of the present disclosure under different contractile forces are shown.

[0512] Figure 37 An imaging device from a bioreactor (e.g., is shown) Figure 36 The imaging device 210 shown obtains a first image 3702 and a second image 3704. The first image 3702 captures the tissue scaffold 3706 at a first time point. The first deflection at the location. The second image 3704 captures the tissue scaffold 3708 at the second time point. The second deflection at that point. To understand the difference in deflection, the deflection is shown relative to a common reference line (the vertical dashed line in both images). and Changes in the deflection can be used to encode or otherwise characterize the contractile force exerted on the tissue scaffold (e.g., by tissue attached to the scaffold or by a probe or other instrument).

[0513] This disclosure relates to flexible scaffold tracking, which allows for the accurate and efficient extraction and measurement of tissue scaffold deflection, thereby enabling the generation of models characterizing the contractile forces that produce the deflection. Advantageously, such models can encode the contractile response of engineered tissues, thus providing a quantification of the functional response of engineered tissues. Alternatively, such models can be used to encode or calibrate the relationship between contractile forces and measured displacements, thereby improving the accuracy of contractile force measurements obtained from such models. Improvements to such models provide improvements to downstream tasks utilizing such models while also improving the efficiency and performance of computational systems used to generate and deploy such models.

[0514] Generally speaking, flexible scaffold tracking can be achieved by transferring the scaffold from a device or bioreactor (e.g., from...). Figure 1The method begins by obtaining an image of the flexible scaffold at a first time point using a bioreactor 102. This image captures the deflection of the flexible scaffold along a first dimension at the first time point due to the (contraction) force applied to it. A curve is fitted to the image such that the curve extends within the image along or approximately along the centerline of the flexible scaffold. Displacement values ​​are determined between the curve and a reference line extending along a second dimension perpendicular to the first dimension. A model is generated based on these displacement values, such that the model characterizes the (contraction) force applied to the flexible scaffold at the first time point. This method... Figure 38 Examples are shown below, and further details are described below.

[0515] Figure 38 A method for tracking a flexible stent according to one aspect of this disclosure is shown.

[0516] Figure 38 Image region 3802, including an image of the flexible support 3804, taken at a first time point, is shown. Image region 3802 captures the deflection of the flexible support 3804 along the first dimension 3806 due to the contractile force applied to the flexible support 3804 at the first time point. Also... Figure 38 The diagram shows a second dimension 3808 perpendicular to the first dimension 3806. In one embodiment, the first dimension 3806 corresponds to the vertical axis of the image region 3802, and the second dimension 3808 corresponds to the horizontal axis of the image region 3802. Curve 3810 is shown extending along the centerline of the flexible support 3804. Curve 3810 intersects the first edge 3812 of the image region 3802 at a first intersection point 3814, and intersects the second edge 3816 of the image region 3802 at a second intersection point 3818. Reference line 3820 extends along the second dimension 3808 between the first intersection point 3814 and the second intersection point 3818. The measurement result 3822 performed between curve 3810 and reference line 3820 corresponds to the displacement value of the flexible support 3804 at a first time point. Measurement result 3822 is determined at reference point 3824 along the second dimension 3808. Reference point 3824 is determined such that the orthogonal distance between curve 3810 and reference line 3820 along the first dimension 3806 is maximized at reference point 3824 (i.e., the distance between the first point 3826 on curve 3810 at reference point 3824 along the second dimension 3808 and the second point 3828 on reference line 3820 at reference point 3824 along the second dimension 3808 corresponds to the maximum orthogonal distance between curve 3810 and reference line 3820 along the first dimension 3806). Figure 38Further illustrated is a predetermined expected range 3830, within which the maximum orthogonal distance between curve 3810 and reference line 3820 is expected to lie. The predetermined expected range 3830 includes a portion of the length of image region 3802 along the second dimension 3808 (e.g., 60% of the height of image region 3802). As will be described in more detail below, the displacement values ​​determined from measurement result 3822 are used to generate a model characterizing the contractile force applied to the flexible stent 3804 (e.g., due to the contraction of engineered tissue attached to the flexible stent 3804 or a predetermined force applied to the flexible stent 3804 by a probe or other instrument).

[0517] Image region 3802 includes images obtained from a device or bioreactor (e.g., from an imaging device of a bioreactor such as...). Figure 36 The image or a portion of an image obtained by the imaging device 210 shown. Image region 3802 includes grayscale bright-field or fluorescence images. Figure 36 As shown more clearly, image region 3802 captures one of the tissue scaffolds of the device or bioreactor. Image region 3802 can be captured such that only the portion of the device containing the relevant tissue scaffold is captured within image region 3802. Alternatively, image region 3802 can be cropped from a larger image of the entire device / well or tissue. In such cases, image region 3802 can be cropped manually or automatically (e.g., by cropping the larger image to a predetermined bounding box having known coordinates corresponding to the portion of the larger image containing the tissue scaffold).

[0518] As stated above, curve 3810 extends along or approximately along the centerline of the flexible scaffold 3804 captured within image region 3802. To determine the centerline of the flexible scaffold 3804 within image region 3802, one or more image processing operations are used to enhance the structure and / or appearance of the flexible scaffold 3804 within image region 3802. (As from...) Figure 37 As can be seen in the example images, the flexible scaffold (e.g., tissue scaffold 3706 or tissue scaffold 3708) appears as a vascular structure within the image. Thus, image processing and / or filtering operations can be used to enhance the regions of the image containing the vascular structure.

[0519] In one embodiment, a filtering operation is used to transform image region 3802 to generate a transformed image region. The filtering operation determines the probability that a region or portion (i.e., a pixel or pixel neighborhood) of image region 3802 contains the flexible stent 3804. Therefore, the transformed image region includes a transformed representation of image region 3802, wherein each pixel value in the transformed image region corresponds to the probability that that pixel value is located on the flexible stent. The filtering operation includes a vascular enhancement filter, such as a multi-scale vascular enhancement filter (e.g., a Frangi vascular saliency filter based on the Hessian matrix) or a diffusion filter (e.g., a coherence enhancement diffusion filter). In one embodiment, the filtering operation includes convolving a one-dimensional kernel across one dimension of image region 3802, such as... Figure 39 exemplified in .

[0520] Figure 39 A one-dimensional vascular enhancement filter according to an embodiment of the present disclosure is shown.

[0521] Figure 39 It shows the path along the line between points A and A' Figure 38 The plot 3902 shows the pixel intensity values ​​captured by a cross-section of the image region 3802. Figure 39 The diagram further illustrates a one-dimensional kernel 3904, a convolution operation 3906, and a plot of the response values ​​3908. The one-dimensional kernel 3904 has a shape corresponding to the approximate cross-sectional shape of the flexible scaffold. The shape of the one-dimensional kernel 3904 is square or approximately square. The response values ​​shown in plot 3908 are produced by convolving the one-dimensional kernel 3904 across the pixel intensity values ​​shown in plot 3902 via the convolution operation 3906. The response values ​​shown in Figure 3908 correspond to the pixel intensity values ​​for pixel rows A to A' in the transformed representation of the image region.

[0522] As can be seen from the response value plot 3908, the pixel position corresponding to the approximate center point of the flexible scaffold includes the maximum response value as a result of the convolution operation 3906. Therefore, across Figure 38 All rows of the image region 3802 shown are subjected to repeated one-dimensional convolution to produce a transformed image that approximates the pixel intensity with the maximum along the centerline of the flexible tissue scaffold. Therefore, the pixel with the maximum intensity value within each row of the transformed image is selected as the most likely estimate of the pixel along the centerline of the tissue scaffold. This operation is repeated for each row of the transformed image to generate multiple data points that correspond to possible locations along the centerline of the flexible scaffold 3804 within the image region 3802. Figure 40 As shown and as described below, the curve can be fitted to multiple data points to determine the approximate centerline of the flexible stent 3804.

[0523] Figure 40 A plot 4002 of data points obtained from a transformed image of a tissue scaffold according to an embodiment of the present disclosure is shown.

[0524] Plotting 4002 includes multiple data points (represented as black dots within plotting 4002) corresponding to the pixel location with the maximum local intensity (i.e., the maximum intensity within a row of the pixel). Plotting 4002 further includes a curve 4004 fitted to the multiple data points, a first neighborhood 4006 surrounding the first outlier 4008, and multiple additional neighborhoods 4010-4014.

[0525] Curve 4004 corresponds to the location of the maximum pixel value along the first dimension within the transformed image, as described above regarding multiple data points (i.e., the location of the maximum pixel value along the first dimension within the transformed image). Figure 39 The curve 4004 is a quadratic curve. In one embodiment, the least squares method is used to fit the curve 4004 to multiple data points, such that the sum of squares of the residuals (the deviations between the multiple data points and the corresponding set of data points on the curve) are minimized. Once fitted, the curve 4004 corresponds to the approximate centerline of the flexible stent.

[0526] In one embodiment, one or more outliers are removed from a plurality of data points before curve 4004 is fitted. Advantageously, this helps improve the accuracy of the curve fitting process, resulting in a centerline that more closely corresponds to the centerline of the tissue scaffold. An outlier is considered an outlier if a data point among the plurality of data points meets or exceeds a predetermined outlier threshold.

[0527] In one embodiment, the predetermined anomaly threshold includes a neighborhood distance threshold. In the example shown in Figure 4002, the first anomaly 4008 is identified as an anomaly because the distance between the first anomaly 4008 and each of the plurality of data points exceeds the predetermined distance threshold. In other words, no other data points from multiple data points are located within the first neighborhood 4006 surrounding the first outlier 4008. A predetermined distance threshold is used. Related to the expected width of the flexible stent. For example, if the flexible stent has Pixels or The expected width (e.g., , , (etc.), then the distance threshold is set to ,in This is a parameter that can take a value between 1 and 2. In one implementation, Additionally or alternatively, a data point is considered an outlier if fewer than a predetermined number of other data points lie within its neighborhood. In the example shown in plot 4002, data points within each of several additional neighborhoods 4010-4014 are visually identifiable as outliers, but these neighborhoods still overlap. For example, data points associated with neighborhoods 4012 and 4014 are located within neighborhood 4010. Thus, if fewer than... Other data points are located within the neighborhood of the data point (e.g., determined by a distance threshold). If the definition is correct, then the data point can be considered an outlier. Here, It is a predetermined value, such that or In one implementation, .

[0528] Additionally or alternatively, the predetermined anomaly threshold includes a pixel intensity threshold such that the pixel value at the image location corresponding to the anomaly exceeds the predetermined intensity threshold. In one implementation, the predetermined intensity threshold is three times the median absolute deviation (MAD) of the intensity of all pixels placed along the tissue scaffold.

[0529] Refer again Figure 38 The curve 3810 fitted to the flexible support 3804, as described above, is used to determine the measurement result 3822 corresponding to the displacement value of the flexible support 3804 at the first time point.

[0530] The displacement value encodes the deflection of the flexible support 3804 at a first time point relative to its expected positioning when at rest. That is, the displacement or deflection of the flexible support 3804 at a given time point corresponds to the difference between the positioning of the flexible support 3804 at that time point and its expected positioning when at rest. The expected positioning of the flexible support 3804 when at rest is modeled by a reference line 3820, such that the distance between curve 3810 (extending along the centerline of the flexible support 3804) and reference line 3820 is used to determine the displacement value for the flexible support 3804 at the first time point.

[0531] Reference line 3820 (optionally referred to as a stationary line, baseline, or expected stationary location) is determined using the intersection of curve 3810 and the boundary or edge of image region 3802, such that reference line 3820 extends between the intersections. Figure 38In this configuration, the first intersection point 3814 is determined by the intersection of the marker curve 3810 and the first edge 3812 of the image region 3802. The second intersection point 3818 is determined by the intersection of the marker curve 3810 and the second edge 3816 of the image region 3802. Here, the second edge 3816 of the image region 3802 is opposite to the first edge 3812 of the image region 3802. A reference line 3820 is then fitted such that it extends between the first intersection point 3814 and the second intersection point 3818. In one embodiment, the reference line 3820 defines a second dimension 3808 (i.e., defines the angle / direction of the second dimension 3808 relative to one or more axes of the image region 3802). In some cases, the reference line 3820 may be aligned or substantially aligned with an image axis (e.g., a vertical axis) such that the second dimension 3808 is aligned or substantially aligned with the image axis. In some other cases, reference line 3820 is not aligned with any of the image axes, causing the second dimension 3808 to be substantially misaligned with any of the image axes.

[0532] Then, a measurement result 3822 for determining the displacement value of the flexible support 3804 at the first time point is calculated between curve 3810 and reference line 3820. Measurement result 3822 is determined at reference point 3824 along the second dimension 3808. As stated above, reference point 3824 is determined such that the orthogonal distance between curve 3810 and reference line 3820 along the first dimension 3806 (which is perpendicular to the second dimension 3808) is maximized at reference point 3824 (i.e., the distance between the first point 3826 on curve 3810 at reference point 3824 along the second dimension 3808 and the second point 3828 on reference line 3820 at reference point 3824 along the second dimension 3808 corresponds to the maximum orthogonal distance between curve 3810 and reference line 3820 along the first dimension 3806). Alternatively, the displacement value is determined using a measurement corresponding to the estimated area of ​​the shape defined by curve 3810 and reference line 3820 (i.e., the shape defined by the path from the first intersection point 3814 to the second intersection point 3818 along curve 3810 and the shape closed by the path from the second intersection point 3818 to the first intersection point 3814 along reference line 3820).

[0533] In one embodiment, if the reference point is outside the predetermined expected range 3830, the predetermined reference point is used to replace the reference point determined using the process described above. As previously stated, the predetermined expected range 3830 corresponds to the portion of the image region 3802 where the reference point 3824 is expected to be located; that is, the expected portion of the image region 3802 includes the maximum orthogonal distance between curve 3810 and reference line 3820. Therefore, the predetermined expected range 3830 includes a proportion of the length of the image region 3802 along the second dimension 3808, centered along the second dimension 3808. This proportion is 20% to 80% of the length of the image region 3802 along the second dimension 3808. In one implementation, this proportion is 60% of the length of the image region 3802 along the second dimension 3808. If the reference point is outside the predetermined expected range 3830, the predetermined reference point used to replace ...

Claims

1. A method for processing a functional response waveform, the method comprising: A first waveform is obtained by one or more processors, the first waveform including the contraction response and relaxation response of the artificial tissue during a single contraction-relaxation cycle; The model is fitted to the first waveform by the one or more processors, wherein the model independently parameterizes the growth of the contraction response and the relaxation response; as well as The one or more processors generate a second waveform from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform.

2. The method according to claim 1, further comprising: One or more feature values ​​are extracted from the second waveform by the one or more processors.

3. The method according to claim 2, further comprising: The one or more processors output the one or more feature values ​​extracted from the second waveform.

4. The method according to claim 2, wherein the one or more feature values ​​include one or more of the following: twitching amplitude value, contraction time value, maximum contraction slope value, relaxation time value, maximum relaxation slope value, and twitching duration value.

5. The method of claim 1, further comprising: The second waveform is output by the one or more processors.

6. The method of claim 1, wherein the model comprises an ascending logic function with a positive growth rate and a descending logic function with a negative growth rate.

7. The method of claim 6, wherein the contraction response is modeled by the rising logic function.

8. The method of claim 6, wherein the relaxation response is modeled by the descent logic function.

9. The method of claim 6, wherein the model comprises the product of the rising logic function and the falling logic function.

10. The method of claim 1, wherein the model comprises a plurality of parameters associated with the contraction response and the relaxation response.

11. The method of claim 10, wherein the plurality of parameters includes a maximum value parameter, an ascent rate parameter, a descent rate parameter, a y-offset parameter, an ascent x-offset parameter, and a descent x-offset parameter.

12. The method of claim 10, wherein the step of fitting the model to the first waveform comprises: The one or more processors predict multiple values ​​for the multiple parameters of the model, such that the model fitted to the first waveform includes the multiple values ​​for the multiple parameters.

13. The method of claim 12, wherein the plurality of values ​​are predicted using a machine learning model.

14. The method of claim 13, wherein the machine learning model comprises a trained initial model.

15. The method of claim 13, wherein the machine learning model comprises a trained convolutional neural network with dilated convolutions.

16. The method of claim 13, wherein the machine learning model comprises a trained long short-term memory neural network.

17. The method of claim 12, wherein the step of fitting the model to the first waveform further comprises: The multiple values ​​are optimized by the one or more processors by using multiple values ​​of the multiple parameters for the model to minimize the error between the first waveform and the model fitted to the first waveform.

18. The method of claim 17, wherein the plurality of values ​​are optimized using a simplex search algorithm.

19. The method of claim 1, wherein the first waveform is obtained from a waveform comprising a plurality of contraction-relaxation cycles of the artificial tissue.

20. The method of claim 1, wherein the contraction response comprises the contractile force of the artificial tissue during the contraction cycle of the single contraction-relaxation cycle.

21. The method of claim 1, wherein the relaxation response comprises the contractile force of the artificial tissue during the relaxation cycle of the single contraction-relaxation cycle.

22. The method of claim 1, wherein the first waveform is obtained from a bioreactor comprising the artificial tissue.

23. The method of claim 1, wherein the artificial tissue comprises engineered muscle tissue.

24. The method of claim 23, wherein the engineered muscle tissue is engineered heart tissue.

25. The method of claim 23, wherein the engineered muscle tissue is engineered skeletal muscle tissue.

26. A non-transitory computer-readable medium storing instructions, which, when executed by one or more processors of a device, cause the one or more processors of the device to: A first waveform is obtained, which includes the contraction and relaxation responses of the artificial tissue during a single contraction-relaxation cycle; The model is fitted to the first waveform, wherein the model independently parameterizes the growth of the contraction response and the relaxation response; and A second waveform is generated from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform.

27. A system comprising: One or more processors; and Memory, which stores instructions. The instructions, when executed by the one or more processors, cause the one or more processors to: A first waveform is obtained, which includes the contraction and relaxation responses of the artificial tissue during a single contraction-relaxation cycle; The model is fitted to the first waveform, wherein the model independently parameterizes the growth of the contraction response and the relaxation response; as well as A second waveform is generated from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform.

28. A method for training a model using synthetic training data, the method comprising: Multiple waveforms are obtained by one or more processors, the multiple waveforms including the functional response of one or more artificial tissues during a single contraction-relaxation cycle; The one or more processors extract multiple parameter sets from the multiple waveforms, wherein the parameter sets in the multiple parameter sets characterize the corresponding waveforms in the multiple waveforms; The parameter set distribution is determined by the one or more processors from the plurality of parameter sets; A synthetic training dataset is generated by the one or more processors, each element of the synthetic training dataset including a synthetic waveform and a corresponding parameter set for generating the synthetic waveform, wherein the corresponding parameter set is obtained from the parameter set distribution; as well as The one or more processors use the synthetic training dataset to train a prediction model, wherein the prediction model is trained to estimate an output parameter set from an input waveform.

29. The method of claim 28, wherein the prediction model includes an initial model.

30. The method of claim 28, wherein the prediction model comprises a convolutional neural network with dilated convolutions.

31. The method of claim 28, wherein the prediction model comprises a long short-term memory neural network.

32. The method of claim 28, further comprising: A first waveform is obtained by the one or more processors, the first waveform including the functional response of a first artificial tissue; as well as The first set of parameter values ​​is predicted from the first waveform by the one or more processors and using the prediction model trained on the synthetic training dataset.

33. The method of claim 32, further comprising: The first set of parameter values ​​is output by the one or more processors.

34. The method of claim 28, wherein the step of generating the synthetic training dataset further comprises: The one or more processors add noise components to the synthetic waveform of each element of the synthetic training dataset according to a noise model.

35. The method of claim 34, wherein the noise model is determined from the plurality of waveforms.

36. The method of claim 28, wherein the parameter set comprises parameter values ​​of a model, the model comprising an ascending logic function with a positive growth rate and a descending logic function with a negative growth rate.

37. The method of claim 36, wherein the parameter values ​​include a maximum value parameter value, an ascent rate parameter value, a descent rate parameter value, a y-offset parameter value, an ascent x-offset parameter value, and a descent x-offset parameter value.

38. The method of claim 28, wherein the one or more artificial tissues comprise one or more engineered muscle tissues.

39. The method of claim 38, wherein the one or more engineered muscle tissues are engineered heart tissues.

40. The method of claim 38, wherein the one or more engineered muscle tissues are engineered skeletal muscle tissues.

41. A non-transitory computer-readable medium storing instructions, which, when executed by one or more processors of a device, cause the one or more processors of the device to: Multiple waveforms are obtained, which include the functional response of one or more artificial tissues during a single contraction-relaxation cycle; Multiple parameter sets are extracted from the multiple waveforms, wherein the parameter sets in the multiple parameter sets represent the corresponding waveforms in the multiple waveforms; Determine the parameter set distribution from the multiple parameter sets; Generate a synthetic training dataset, wherein each element of the synthetic training dataset includes a synthetic waveform and a corresponding parameter set for generating the synthetic waveform, wherein the corresponding parameter set is obtained from the parameter set distribution; as well as The synthetic training dataset is used to train a prediction model, wherein the prediction model is trained to estimate the output parameter set from the input waveform.

42. A system comprising: One or more processors; and Memory, which stores instructions. The instructions, when executed by the one or more processors, cause the one or more processors to: Multiple waveforms are obtained, which include the functional response of one or more artificial tissues during a single contraction-relaxation cycle; Multiple parameter sets are extracted from the multiple waveforms, wherein the parameter sets in the multiple parameter sets represent the corresponding waveforms in the multiple waveforms; Determine the parameter set distribution from the multiple parameter sets; Generate a synthetic training dataset, wherein each element of the synthetic training dataset includes a synthetic waveform and a corresponding parameter set for generating the synthetic waveform, wherein the corresponding parameter set is obtained from the parameter set distribution; as well as The synthetic training dataset is used to train a prediction model, wherein the prediction model is trained to estimate the output parameter set from the input waveform.

43. A method for extracting a contraction-relaxation cyclic waveform, the method comprising: A first waveform is obtained by one or more processors, the first waveform including multiple functional responses of an artificial tissue stimulated at a first frequency; The first waveform is convolved with a pulse train by the one or more processors to generate a convolved waveform, wherein the pulse train is generated at the first frequency; The one or more processors identify a first position associated with the maximum value of the convolved waveform, wherein the first position corresponds to the expected position of the first contraction-relaxation cycle; as well as The one or more processors extract a second waveform comprising the first contraction-relaxation cycle from the first position of the first waveform, wherein the second waveform has a first duration proportional to the first frequency.

44. The method of claim 43, further comprising: The second waveform is output by the one or more processors.

45. The method of claim 43, further comprising: The second position is identified by the one or more processors based on the first position and the first frequency, wherein the second position corresponds to the expected position of the second contraction-relaxation cycle; as well as The third waveform, comprising the second contraction-relaxation cycle, is extracted from the second position of the first waveform by the one or more processors, wherein the second waveform has a second duration proportional to the first frequency.

46. ​​The method of claim 45, further comprising: The third waveform is output by the one or more processors.

47. The method of claim 45, wherein the first duration and the second duration are the same.

48. The method of claim 43, further comprising: The pulse train is generated by the one or more processors at the first frequency.

49. The method of claim 43, wherein the midpoint of the second waveform is substantially aligned with the first position of the first waveform.

50. The method of claim 43, wherein the first frequency is from 0.1 Hz to 20 Hz.

51. The method of claim 50, wherein the first frequency is from 1 Hz to 6 Hz.

52. The method of claim 43, wherein the first waveform is obtained from a bioreactor comprising the artificial tissue.

53. The method of claim 43, wherein the artificial tissue comprises engineered muscle tissue.

54. The method of claim 53, wherein the engineered muscle tissue is engineered heart tissue.

55. The method of claim 53, wherein the engineered muscle tissue is engineered skeletal muscle tissue.

56. A non-transitory computer-readable medium storing instructions, which, when executed by one or more processors of a device, cause the one or more processors of the device to: A first waveform is obtained, the first waveform including multiple functional responses of an artificial tissue stimulated at a first frequency; The first waveform is convolved with a pulse train to generate a convolved waveform, wherein the pulse train is generated at the first frequency; Identify a first position associated with the maximum value of the convolved waveform, wherein the first position corresponds to the expected position of the first contraction-relaxation cycle; as well as Extract a second waveform comprising the first contraction-relaxation cycle from the first position of the first waveform, wherein the second waveform has a first duration proportional to the first frequency.

57. A system comprising: One or more processors; and Memory, which stores instructions. The instructions, when executed by the one or more processors, cause the one or more processors to: A first waveform is obtained, the first waveform including multiple functional responses of an artificial tissue stimulated at a first frequency; The first waveform is convolved with a pulse train to generate a convolved waveform, wherein the pulse train is generated at the first frequency; Identify a first position associated with the maximum value of the convolved waveform, wherein the first position corresponds to the expected position of the first contraction-relaxation cycle; as well as Extract a second waveform comprising the first contraction-relaxation cycle from the first position of the first waveform, wherein the second waveform has a first duration proportional to the first frequency.

58. A method for predicting treatment effects, the method being performed at a device including one or more processors and a memory, the method comprising: Multiple signals are obtained by one or more processors, the multiple signals including a baseline signal and a perturbation signal, wherein the baseline signal includes a first plurality of functional responses of the engineered organization under reference conditions, and the perturbation signal includes a second plurality of functional responses of the engineered organization under perturbation conditions involving the first perturbation; The plurality of signals are split into a first plurality of waveforms by the one or more processors, each of the first plurality of waveforms including the contraction response and relaxation response of the engineered tissue during a single contraction-relaxation cycle; The model is fitted to each of the first plurality of waveforms by the one or more processors, wherein the model independently parameterizes the growth of the contraction response and the relaxation response of the engineered tissue during the single contraction-relaxation cycle of each waveform; The one or more processors generate a second plurality of waveforms from the model fitted to each of the first plurality of waveforms, wherein the second plurality of waveforms includes a plurality of filtered baseline waveforms associated with the baseline signal and a plurality of filtered perturbation waveforms associated with the perturbation signal; The processor extracts a first feature value of a first feature from the plurality of filtered baseline waveforms; The processor extracts a second feature value of the first feature from the plurality of filtered perturbation waveforms; as well as The effect associated with the first perturbation is determined by the one or more processors based on a comparison of the first feature value and the second feature value.

59. The method of claim 58, wherein the first feature is one of the following: twitching amplitude, contraction time, maximum contraction slope, relaxation time, maximum relaxation slope, and twitching duration.

60. The method of claim 58, wherein the first feature value comprises the average value of the first feature determined from the plurality of filtered baseline waveforms.

61. The method of claim 58, wherein the second feature value comprises the average of the first feature determined from a plurality of filtered treatment waveforms.

62. The method of claim 58, further comprising: The effect associated with the first perturbation is output by the one or more processors.

63. The method of claim 58, wherein the model comprises an ascending logic function with a positive growth rate and a descending logic function with a negative growth rate.

64. The method of claim 63, wherein the contraction response is modeled by the rising logic function.

65. The method of claim 63, wherein the relaxation response is modeled by the descent logic function.

66. The method of claim 63, wherein the model comprises the product of the rising logic function and the falling logic function.

67. The method of claim 58, wherein the model includes a plurality of parameters associated with the contraction response and the relaxation response.

68. The method of claim 67, wherein the plurality of parameters includes a maximum value parameter, an ascent rate parameter, a descent rate parameter, a y-offset parameter, an ascent x-offset parameter, and a descent x-offset parameter.

69. The method of claim 67, wherein the step of fitting the model to each of the first plurality of waveforms includes, for a first waveform of the first plurality of waveforms: The one or more processors predict multiple values ​​for the multiple parameters of the model, such that the model fitted to the first waveform includes the multiple values ​​for the multiple parameters.

70. The method of claim 69, wherein the plurality of values ​​are predicted using a machine learning model.

71. The method of claim 70, wherein the machine learning model includes an initial model.

72. The method of claim 70, wherein the machine learning model comprises a trained convolutional neural network with dilated convolutions.

73. The method of claim 70, wherein the machine learning model comprises a trained long short-term memory neural network.

74. The method of claim 69, wherein the step of fitting the model to each of the first plurality of waveforms includes, for the first waveform among the first plurality of waveforms: The multiple values ​​are optimized by the one or more processors by using multiple values ​​of the multiple parameters for the model to minimize the error between the first waveform and the model fitted to the first waveform.

75. The method of claim 74, wherein the plurality of values ​​are optimized using a simplex search algorithm.

76. The method of claim 58, wherein the plurality of signals are obtained from a bioreactor including the engineered tissue.

77. The method of claim 58, wherein the engineered tissue comprises engineered muscle tissue.

78. The method of claim 77, wherein the engineered muscle tissue is engineered heart tissue.

79. The method of claim 77, wherein the engineered muscle tissue is engineered skeletal muscle tissue.

80. A non-transitory computer-readable medium storing instructions, said instructions causing said one or more processors of the device, when executed, to: Multiple signals are obtained, including a baseline signal and a perturbation signal, wherein the baseline signal includes a first plurality of functional responses of the engineered tissue under reference conditions, and the perturbation signal includes a second plurality of functional responses of the engineered tissue under perturbation conditions involving the first perturbation; The plurality of signals are split into a first plurality of waveforms, each of the first plurality of waveforms including the contraction response and relaxation response of the engineered tissue during a single contraction-relaxation cycle; The model is fitted to each of the first plurality of waveforms, wherein the model independently parameterizes the growth of the contraction response and the relaxation response of the engineered tissue during the single contraction-relaxation cycle of each waveform; A second plurality of waveforms are generated from the model fitted to each of the first plurality of waveforms, wherein the second plurality of waveforms includes a plurality of filtered baseline waveforms associated with the baseline signal and a plurality of filtered perturbation waveforms associated with the perturbation signal; Extract the first feature value of the first feature from the plurality of filtered baseline waveforms; Extract the second feature value of the first feature from the plurality of filtered perturbation waveforms; as well as The effect associated with the first perturbation is determined based on a comparison of the first eigenvalue and the second eigenvalue.

81. A system comprising: One or more processors; and Memory, which stores instructions. The instructions, when executed by the one or more processors, cause the one or more processors to: Multiple signals are obtained, including a baseline signal and a perturbation signal, wherein the baseline signal includes a first plurality of functional responses of the engineered tissue under reference conditions, and the perturbation signal includes a second plurality of functional responses of the engineered tissue under perturbation conditions involving the first perturbation; The plurality of signals are split into a first plurality of waveforms, each of the first plurality of waveforms including the contraction response and relaxation response of the engineered tissue during a single contraction-relaxation cycle; The model is fitted to each of the first plurality of waveforms, wherein the model independently parameterizes the growth of the contraction response and the relaxation response of the engineered tissue during the single contraction-relaxation cycle of each waveform; A second plurality of waveforms are generated from the model fitted to each of the first plurality of waveforms, wherein the second plurality of waveforms includes a plurality of filtered baseline waveforms associated with the baseline signal and a plurality of filtered perturbation waveforms associated with the perturbation signal; Extract the first feature value of the first feature from the plurality of filtered baseline waveforms; Extract the second feature value of the first feature from the plurality of filtered perturbation waveforms; as well as The effect associated with the first perturbation is determined based on a comparison of the first eigenvalue and the second eigenvalue.

82. A method for processing a functional response waveform, the method comprising: A first waveform is obtained by one or more processors, the first waveform including the contraction response and relaxation response of an artificial tissue during a single contraction-relaxation cycle, wherein the first waveform is obtained from a bioreactor including the artificial tissue; The model is fitted to the first waveform by the one or more processors, wherein the model independently parameterizes the growth of the contraction response and the relaxation response; The one or more processors generate a second waveform from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform; One or more feature values ​​are extracted from the second waveform by the one or more processors; Train a machine learning model on the one or more feature values ​​from the second waveform; as well as By inputting a dataset of human tissue into the machine learning model, predictions defining one or more characteristics of the human tissue are generated, the predictions corresponding to at least one of the one or more feature values ​​of the second waveform.

83. A system configured to process a functional response waveform, the system comprising: One or more bioreactors configured to grow artificial tissues therein; and A control unit, communicatively coupled to the bioreactor, and including one or more processors configured to execute instructions that cause the one or more processors to: A first waveform is obtained, comprising the contraction and relaxation responses of the artificial tissue during a single contraction-relaxation cycle, wherein the first waveform is obtained from the bioreactor comprising the artificial tissue. The model is fitted to the first waveform, wherein the model independently parameterizes the growth of the contraction response and the relaxation response. A second waveform is generated from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform. Extract one or more feature values ​​from the second waveform. Train a machine learning model on the one or more feature values ​​from the second waveform, and By inputting a dataset of human tissue into the machine learning model, predictions defining one or more characteristics of the human tissue are generated, the predictions corresponding to at least one of the one or more feature values ​​of the second waveform.

84. A non-transitory computer-readable medium storing instructions, which, when executed by one or more processors of a device, cause the one or more processors of the device to: A first waveform is obtained, the first waveform including the contraction response and relaxation response of the artificial tissue during a single contraction-relaxation cycle, wherein the first waveform is obtained from a bioreactor including the artificial tissue; The model is fitted to the first waveform, wherein the model independently parameterizes the growth of the contraction response and the relaxation response; A second waveform is generated from the model fitted to the first waveform, such that the second waveform includes a noise-filtered representation of the first waveform; Extract one or more feature values ​​from the second waveform; Train a machine learning model on the one or more feature values ​​from the second waveform; as well as By inputting a dataset of human tissue into the machine learning model, predictions defining one or more characteristics of the human tissue are generated, the predictions corresponding to at least one of the one or more feature values ​​of the second waveform.

Images