Predicting unobserved quantitative measurements using machine learning

Abstract

To provide prediction of unobserved quantitative measurements using machine learning.SOLUTION: In a computing environment 100, for predicting unobserved quantitative measurements using machine learning, a quantitative measurements-predicting computing device 102 is configured to: receive and process de novo quantitative measurements input using a trained machine learning model so as to generate one or more predicted experimental cell quantitative measurements; based on the generated training quantitative measurements data, train a machine learning model to generate predicted experimental cell viability measurements based on the de novo quantitative measurements input, by processing the training quantitative measurements data; and store the machine learning model as a trained machine learning model in a memory of a computer.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to methods and systems for predicting unobserved quantitative measurements using machine learning, and more particularly to techniques for generating predicted experimental cellular quantitative measurements by processing novel inputs using trained machine learning models. [Background technology]

[0002] The background discussion provided herein is for the purpose of generally presenting the context of the present disclosure. The inventor's work, to the extent described in this Background section, as well as aspects of the present specification that may not otherwise qualify as prior art at the time of filing, are not admitted, expressly or impliedly, as prior art to the present disclosure.

[0003] Properties (e.g., dose-response relationships for drugs) are difficult to measure in experiments. Such relationships are often nonlinear, meaning that the relationship between drug dose and its effect is not proportional. In some cases, low doses of a drug may have little effect, while small increases may have a large effect. This nonlinearity can complicate the evaluation of synergy, because the combined effect of two drugs may not follow a simple additive pattern.

[0004] Some drugs may reach a saturation state, where increasing the dose beyond a certain level does not increase the effect proportionately. This saturation effect can obscure the determination of synergistic interactions, since the response may plateau when drugs are combined at different doses.

[0005] Certain drugs may exhibit a threshold effect, where the effect may not be observed until a certain dose is reached. Below this threshold, the drug may have little or no effect. When multiple drugs are used in combination, it may be difficult to determine the appropriate dose to achieve a synergistic effect, especially when the threshold doses of the individual drugs differ significantly.

[0006] Obtaining complete dose-response data for individual drugs could be a solution to accurately evaluate synergistic interactions. However, in practice, researchers may only have access to limited data points due to logistical constraints or ethical considerations. This data limitation can make it difficult to accurately identify dose combinations that produce synergistic effects.

[0007] Drug toxicity can be dose-dependent: lower doses may be well tolerated, but higher doses or drug combinations may result in adverse reactions or deleterious side effects. When testing for synergy, it is important to balance the desired therapeutic effect with potential toxicity.

[0008] Some drugs may exhibit time-dependent effects, increasing or decreasing their impact over time. This temporal aspect adds another layer of complexity to measuring synergy, as the timing and sequence of drug administration may affect the overall response.

[0009] People can vary widely in their response to drug doses due to genetic differences, age, underlying health conditions, and other factors. Individual variability in dose-response relationships further complicates the assessment of synergistic effects in diverse populations.

[0010] Combination therapy, which delivers two or more drugs simultaneously to treat various medical conditions, including cancer, can increase efficacy while simultaneously reducing individual drug doses, administration intervals, and side effects. Drug combinations are measured over a range of doses in preclinical in vitro or in vivo experiments at various dose levels or concentrations to identify drugs that may have synergistic effects when combined. These results generate dose-response matrices that record drug response / effects with varying combination ratios.

[0011] However, constructing a dose-response matrix is ​​more resource intensive than evaluating a single drug dose and response due to the exponential nature of evaluating two or more drugs at various concentrations (e.g., x-power propositions for a two-drug experiment, x-power propositions for a three-drug experiment, etc., where x is the numerical dose being tested).

[0012] Overall, the dose-response characteristics of drugs play an important role in understanding and measuring synergy. Nonlinearity, saturation, threshold effects, incomplete data, toxicity concerns, time-dependent effects, and individual variability all contribute to the complexities researchers encounter when studying drug combinations and their synergistic interactions.

[0013] Conventional techniques for predicting dose-response data are inadequate, lacking the ability to calculate the survivability of missing doses, lacking the ability to calculate values ​​for unmeasured time steps, lacking the ability to process image data, and lacking the ability to augment training and operational data with additional input features.

[0014] Thus, there is an opportunity for improved and sophisticated quantitative measurement and predictive platforms and techniques. Summary of the Invention

[0015] In one aspect, a computer-implemented method for predicting unmeasured experimental cellular quantitative measurements using a machine learning model trained using data to reduce laboratory experimentation burden includes (i) receiving new input via one or more processors; and (ii) processing the new input using the machine learning model to generate one or more predicted experimental cellular quantitative measurements.

[0016] In another aspect, a computer-implemented method for training a machine learning model to identify compounds that exhibit synergistic effects and reduce the burden of laboratory experimentation includes: (i) receiving, via one or more processors, training quantitative measurement data; (ii) training, via the one or more processors, a machine learning model by processing the training quantitative measurement data to generate predicted experimental cell viability measurements based on the input of new quantitative measurements; and (iii) storing, via the one or more processors, the trained machine learning model in a computer memory.

[0017] In yet another aspect, a computing system includes one or more processors and one or more memories storing computer-executable instructions, which, when executed by the one or more processors, cause the computing system to (i) receive, via the one or more processors, an input of novel quantitative measurements, and (ii) process the input of novel quantitative measurements using a machine learning model to generate one or more predicted experimental cellular quantitative measurements.

[0018] In another aspect, a computing system includes one or more processors and one or more memories storing computer-executable instructions, which, when executed by the one or more processors, cause the computing system to (i) receive, via the one or more processors, training quantitative measurement data, (ii) train, via the one or more processors, a machine learning model to generate predicted experimental cell viability measurements based on input of new quantitative measurements by processing the training quantitative measurement data, and (iii) store, via the one or more processors, the machine learning model in the memory as a trained machine learning model.

[0019] In a further aspect, the computer-readable medium includes computer-executable instructions that, when executed by one or more processors, cause the computer to (i) receive, via the one or more processors, an input of new quantitative measurements, and (ii) process the input of new quantitative measurements using a machine learning model to generate one or more predicted experimental cellular quantitative measurements.

[0020] In yet another aspect, a computer-readable medium includes computer-executable instructions that, when executed by one or more processors, cause the computer to (i) receive, via the one or more processors, training quantitative measurement data; (ii) train, via the one or more processors, a machine learning model to generate predicted experimental cell viability measurements based on new quantitative measurement inputs by processing the training quantitative measurement data; and (iii) store, via the one or more processors, the machine learning model in memory as a trained machine learning model. [Brief description of the drawings]

[0021] The figures described below depict various aspects of the systems and methods disclosed herein. It should be understood that each figure depicts an example of an aspect of the systems and methods.

[0022] [Figure 1] FIG. 1 illustrates an example computing system for training and / or operating machine learning models to predict unmeasured quantitative measurements using one or more trained machine learning models, according to some aspects. [Diagram 2] FIG. 1 shows a block flow diagram for using a trained machine learning model to predict complete or substantially complete dose-response quantitative measurements given incomplete new quantitative measurements, and further processing the complete or substantially complete dose-response quantitative measurements to generate synergy distributions and / or synergy scores for identifying synergistic dose pairs, according to some embodiments. [Diagram 3]FIG. 1 shows a block flow diagram for predicting a perfect or substantially perfect dose-response quantitative measurement given an incomplete new quantitative measurement using a trained regression machine learning model, according to some embodiments. [Figure 4A] 1 illustrates an example synergy score calculation for a synergy score map, according to some embodiments. [Figure 4B] 1 illustrates a three-dimensional synergy score map, according to some embodiments. [Figure 5A] 1 illustrates an example variational autoencoder architecture for predicting perfect or substantially perfect dose-response quantitative measurements, in accordance with some embodiments. [Figure 5B] 5B illustrates a conceptual diagram of interpolation in latent space using the trained autoencoder architecture of FIG. 5A, in accordance with some aspects. [Figure 5C] 1 illustrates an example sparse dose-response matrix, according to some embodiments. [Figure 5D] 1 illustrates an alternative autoencoder architecture in accordance with some aspects. [Figure 5E] 1 illustrates yet another alternative autoencoder architecture in accordance with some aspects. [Figure 5F] No original text [Figure 6] 1 shows a graph of exemplary experimental savings that may be realized by using the present techniques, in accordance with some embodiments. [Figure 7A] To reduce the burden of laboratory experimentation, in accordance with some embodiments, a computer implementation is shown for making predictions of unmeasured dose-response experimental cellular quantitative measurements using machine learning models trained using dose-response data. [Figure 7B] A computer-implemented method is presented for training machine learning models to make predictions of unmeasured quantitative measurements to identify compounds that exhibit synergistic effects and reduce the burden of laboratory experimentation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0023] overview The present technology relates to methods and systems for predicting unmeasured quantitative measurements using machine learning, and more particularly to techniques for processing inputs of new quantitative measurements using trained machine learning models to generate predicted experimental cellular quantitative measurements. The present technology may include using one or more machine learning models to advantageously reduce the number of experiments required to identify compounds with additive or synergistic effects.

[0024] As used herein, the term "synergistic" compounds or substances are compounds or substances that exhibit a cooperative interaction that, when used together, results in a combined effect that is greater than the sum of their individual effects. The synergistic effect between these compounds can enhance the therapeutic outcome and provide a more potent, beneficial, or desired response compared to using the compounds separately. Synergistic interactions can be positive or negative. For example, synergistic effects can include increased toxicity or adverse effects when used in combination. Thus, when searching for synergistic compounds, the potential risks and benefits associated with their combination must be carefully evaluated. In some aspects, the technology can identify antagonistic combinations to eliminate problematic combination regimens in drug development.

[0025] As used herein, the "effect" in "synergy" generally refers to the combined and enhanced ability of two or more therapeutic agents, drugs, or agents to work together in a coordinated manner to more effectively destroy or inhibit tumor cell proliferation (e.g., in a tumor or multicellular structure such as a tumor organoid) compared to the individual effects of each agent. The technology can be configured to measure effects on organoids, including three-dimensional patient-derived organoids, in addition to or in place of individual tumor cells, as described in more detail below.

[0026] Specifically, "synergy" in the context of killing tumor cells means that the combined action of therapeutic agents or treatments results in a more significant decrease in tumor cell viability, proliferation, or survival than is achieved by using each agent separately. The combined effect is greater than the sum of the individual effects, providing a more powerful and targeted approach to fighting tumors. Synergistic interactions in tumor cell killing can involve different mechanisms, such as increasing cell death (apoptosis), disrupting cell division and growth, inhibiting angiogenesis (the formation of blood vessels to support tumor growth), enhancing the immune response to the tumor, or a combination of these effects.

[0027] The present technology may include techniques for training one or more machine learning models to predict a complete dose-response matrix (or tensors for three or more drugs) from a partial dose-response matrix dataset corresponding to a limited number of experiments. Furthermore, the present technology may be extended to predict time-based quantitative measurements, such as data from unmeasured time points in the case of experiments measuring the effect of a compound over time. This advantageously allows for accelerating the rate of drug discovery by increasing experimental resolution (e.g., using empirical data to simulate additional data), thereby allowing several more drug combinations to be tested in preclinical research programs. Specifically, the present invention increases the throughput of modeling laboratory drug screening assays, allowing more experiments to be run using the same resources (in other words, using fewer experiments to generate the same size dataset of results), thereby reducing costs and increasing environmental sustainability. Furthermore, this advantageously prevents removing experimental plates from the incubator for imaging, which affects the way organoid cells grow, leading to data degradation and inaccurate experimental conclusions. By interpolating data from intermediate time points, the technique allows cells to remain undisturbed in the incubator during the unmeasured intermediate time points, such that removal from the incubator does not alter cell growth.

[0028] These improvements, in turn, will benefit patients by accelerating drug discovery while allowing the testing of more drug combinations to facilitate the development and delivery of personalized therapies.

[0029] In some aspects, the technology may be offered to other parties, e.g., via platform-as-a-service (PaaS) offerings, white label services, etc. Such other parties may include pharmaceutical companies, biotech companies, contract research organizations, start-ups (e.g., personalized medicine companies), healthcare analytics companies, software companies, pharmacogenomics companies, healthcare consultancies, non-profit research institutions, etc.

[0030] Example Computing Environment 1 illustrates an example computing environment 100 for performing the techniques, according to some aspects. The environment 100 may include computing resources for training and / or running machine learning models to predict unmeasured quantitative measurements using one or more trained machine learning models to process training and inference inputs, and for further processing / computation of such predictions.

[0031] The computing environment 100 may include a quantitative measure prediction computing device 102, a quantitative measure client computing device 104, an electronic network 106, a quantitative measure data generation system 108, and an electronic database 110. The components of the computing environment 100 may, in some aspects, be communicatively connected to each other via the electronic network 106.

[0032] The quantitative measurement prediction computing device 102 may perform, among other things, receiving / storing / pre-processing quantitative measurement data, training quantitative measurement models, operating quantitative measurement models, and related operations. In some aspects, the quantitative measurement prediction computing device 102 may be implemented as one or more computing devices (e.g., one or more servers, one or more laptops, one or more mobile computing devices, one or more tablets, one or more wearable devices, one or more cloud computing virtual instances, etc.). The quantitative measurement analysis computing device 102 may include one or more processors 120, one or more network interface controllers 122, one or more memories 124, input devices 126, and output devices 128.

[0033] In some aspects, the one or more processors 120 may include one or more central processing units, one or more graphics processing units, one or more field programmable gate arrays, one or more application specific integrated circuits, one or more tensor processing units, one or more digital signal processors, one or more neural processing units, one or more RISC-V processors, one or more co-processors, one or more specialized processors / accelerators for artificial intelligence or machine learning applications, one or more microcontrollers, etc.

[0034] The quantitative measurement prediction computing device 102 may include one or more network interface controllers 122, such as an Ethernet network interface controller, a wireless network interface controller, etc. In some embodiments, the network interface controller 122 may include advanced features such as hardware acceleration, specialized networking protocols, etc.

[0035] The memory 124 of the quantitative measurement prediction computing device 102 may include volatile and / or non-volatile storage media. For example, the memory 124 may include one or more random access memories, one or more read-only memories, one or more cache memories, one or more hard disk drives, one or more solid state drives, one or more non-volatile memory express, one or more optical drives, one or more Universal Serial Bus flash drives, one or more external hard drives, one or more network attached storage devices, one or more cloud storage instances, one or more tape drives, etc.

[0036] As described, the memory 124 may store the module 130 therein, for example, as a set of one or more computer executable instructions. In some aspects, the module 130 may include additional storage, such as one or more operating systems (e.g., Microsoft Windows, GNU / Linux, Mac OSX, etc.). The operating system may be configured to run the module 130 during operation of the quantitative measurement prediction computing system 102, for example, the module 130 may include additional modules and / or services for receiving and processing quantitative measurement data. The module 130 may be implemented using any suitable computer programming language (e.g., Python, JavaScript, C, C++, Rust, C#, Swift, Java, Go, LISP, Ruby, Fortran, etc.).

[0037] The modules 130 may include a quantitative measurement data module 152, a machine learning training module 154, a machine learning operation module 156, and a visualization module 158. In some aspects, more or fewer modules 130 may be included. The modules 130 may be configured to communicate with each other (e.g., via inter-process communication, via a bus, etc.).

[0038] The quantitative measurement data module 152 may include instructions for receiving, acquiring, pre-processing, and / or storing quantitative measurement data, where the quantitative measurement data may be incomplete or complete quantitative measurements. For example, the quantitative measurement data module 152 may include instructions for receiving incomplete quantitative measurements as described herein (e.g., via the input device 128, as described below). In general, the quantitative measurements may relate to drug dosage, immune / non-immune cell ratios, time (e.g., treatment time or time spent in a particular condition), and the like. The quantitative measurement data may include one or more types of data of non-imaging modalities, such as tabular data, text input, binary streams, and the like. The instructions may include instructions for pre-processing the quantitative measurements or other quantitative measurement data (e.g., metadata). In some aspects, the quantitative measurement data module 152 may include instructions for accessing (e.g., storing and / or acquiring) the quantitative measurement data.

[0039] For example, the quantitative measurement data module 152 may access the quantitative measurement data via the memory 124 (e.g., in a flat file, compressed file, etc.) and / or via the electronic database 110 (e.g., via insert, update, upsert, select queries). In general, the quantitative measurement data module 152 may make the data available for processing by other elements of the computing environment 100 (e.g., other modules 130). In some aspects, the quantitative measurement data module 152 may pre-process the quantitative measurement data.

[0040] For example, in some aspects, the input device 126 may be an image capture device. The quantitative measurement data module 152 may receive medical images including, for example, microscopy images, digital histopathology slides such as hematoxylin and eosin (H&E stained slide images, immunohistochemistry (IHC) stained slide images, and / or digital images of any other staining protocol from any suitable source.

[0041] In some embodiments, the quantitative measurements data module 152 may receive, for example, a physician's clinical records and / or histopathology imaging data from the electronic database 110, or via the input device 126, or via another computing environment accessible via the network 106. Any number of medical image data sources may be accessible using the computing environment 100. The histopathology images may include one or more images captured by any suitable optical histopathology slide scanner, including any dedicated digital medical image scanner, for example, a magnified (e.g., 20x and 40x resolution) scanner. Additionally, the quantitative measurements data module 152 may receive images from one or more histopathology image repositories (e.g., one or more electronic databases, non-transitory computer readable memory, etc.). In yet other examples, images may be received from partner genome sequencing systems, such as TCGA and NCI Genomic Data Commons. Additionally, the quantitative measurements analysis computing device may receive histopathology images from an organoid modeling lab.

[0042] The quantitative measurements data module 152 may receive imaging data, genomic data, patient data, treatment data, historical data, etc., according to the techniques and processes described herein. In some aspects, the quantitative measurements data module 152 may receive imaging data from multiple imaging sources. The multiple imaging sources may be different and may generate and provide different imaging data than other providers, hospitals, etc. The quantitative measurements data module 152 may receive imaging data from different sources that differ in one or more ways, resulting in biases inherent to the different data sources, such as different dyes, biospecimen fixation, embedding, staining protocols, and different pathology imaging instruments and settings.

[0043] The quantitative measurement data module 152 may perform pre-processing to improve the image data for faster processing (e.g., to train a machine learning framework). For example, the quantitative measurement data module 152 may perform normalization processing on the received image data, including one or more of color normalization, intensity normalization, imaging source normalization, etc. The quantitative measurement data module 152 may perform normalization to compensate and correct differences in the received image data. In some examples, the quantitative measurement data module 152 receives medical images, while in other examples, the subsystems may generate medical images (e.g., via the input device 126 and / or the output device 128).

[0044] The quantitative measurement data module 152 may perform further image processing to remove artifacts and other noise from the received images, for example, by performing preliminary tissue detection to identify regions of the images that correspond to histopathologically stained tissue for subsequent analysis, classification, and segmentation.

[0045] In some embodiments, such as when image data is analyzed on a tile basis, the quantitative measurement data module 152 may pre-process the image by receiving an initial histopathology image at a first image resolution, downsampling the image to a second image resolution, and then performing normalization (such as color and / or intensity normalization) on the downsampled histopathology image and removing non-tissue objects from the image. Exemplary Computer-Implemented Machine Learning Model Training Aspects

[0046] Machine learning training module 154 may include instructions for training one or more machine learning models in accordance with the techniques described herein. For example, in some aspects machine learning module 154 may include one or more sets of instructions for implementing various machine learning techniques for performing quantitative measurement analysis based on information, such as quantitative measurement data received by quantitative measurement data module 152.

[0047] For example, in some aspects, the machine learning training module 154 may include instructions for training one or more generative models (e.g., variational autoencoder machine learning models, regression and classification models, etc.). The machine learning training module 154 may train such models to predict complete or substantially complete quantitative measurements (e.g., unobserved / unmeasured / missing / corrupted dose-response measurements, imputed time points, immune cell activity, etc.), as described in more detail herein. The unobserved quantitative measurements may be unobserved due to selection, instrument failure, experimental design, etc. In general, the machine learning training module 154 may include instructions for training one or more variational autoencoder models by mapping inputs (e.g., images) into a continuous latent space, where the variational autoencoder learns to recreate the inputs via sampling of the latent space. This trained model may then be used in practice (e.g., via the machine learning operations module 156) to predict quantitative measurements at run time based on new input data (e.g., data accessible via the quantitative measurement data module 152), including input incomplete matrices or tensors of values. Here, "new" data refers to data that the model has not seen before (e.g., data that was retained while training the model, or data that became accessible to the model after the model was trained). The incomplete data may correspond to experimental measurements (e.g., sampling of dose-response data). Thus, the machine learning training module 154 may train the trained model to automatically generate experimental results in a computational manner without having to actually perform those results. Avoiding exhaustive experimentation is a particularly advantageous and compelling improvement provided to the automated quantitative measurement prediction system provided by the present technology.

[0048] As described above, many types of training data may be used in conjunction with the machine learning models of the present invention. For example, the machine learning training module 154 may train one or more machine learning models using training quantitative measurements including drug efficacy, viability, cell proliferation rate, apoptosis rate (cells in the process of dying), cell migration / invasion rate, gene expression levels, enzyme activity, cell cycle analysis, reactive oxygen species, cytokine / chemokine release, metabolic activity, neuronal activity, cell morphology, cell adhesion, binding affinity, and the like. In some embodiments, the machine learning training module 154 may train one or more models using numerical quantitative measurements. In some embodiments, the machine learning training module 154 may train one or more models using training quantitative measurements that are represented using other data modalities, such as microscopic image input, text input, tabular input, and the like. As used herein, data modality may differ from data representation / entity. For example, microscopic images may be provided directly to the model as input or may be encoded into different formats (e.g., binary format, text-based format, and the like). In yet another embodiment, the training quantitative measurements may correspond to measurements of off-target, general cytotoxicity / cytotoxicity (e.g., also killing healthy non-cancerous cells), or adverse effects.

[0049] The machine learning training module 154 may train one or more models using image data preprocessed as described above, images from H&E slides or images from IHC slides (with or without human annotation), including IHC slides targeting PD-L1, PTEN, EGFR, beta-catenin / catenin beta 1, NTRK, HRD, PIK3CA, and hormone receptors including HER2, AR, ER, and PR. In some embodiments, the training data may include DNA sequences, RNA sequences, metabolomic data, proteomic / cytokine data, epigenomic data, organoid data, raw karyotype data, transcriptional data, transcriptomics, metabolomics, microbiomics, and immunology, identification of SNPs, MNPs, InDels, MSI, TMB, CNV fusions, loss of heterozygosity, loss or gain of function. Epigenomic data includes DNA methylation, histone modifications, or other factors that inactivate genes or cause changes in gene function without changing the sequence of nucleotides within the gene. Microbiomics includes data on viral infections that may affect the treatment and diagnosis of certain diseases, as well as bacteria present in a patient's gastrointestinal tract that may affect the effectiveness of drugs taken by the patient. Protein data includes information on protein composition, structure, and activity, when and where proteins are expressed, rates of protein production, degradation, and steady-state abundance, how proteins are modified, e.g., post-translational modifications such as phosphorylation, trafficking of proteins between subcellular compartments, protein involvement in metabolic pathways, how proteins interact with each other, or post-translational modifications to proteins from RNA, such as phosphorylation, ubiquitination, methylation, acetylation, glycosylation, oxidation, or nitrosylation.

[0050] The training data may further include demographic data and tumor response data (including data regarding the reduction in tumor growth following exposure to a particular therapy, e.g., immunotherapy, DNA damaging therapy such as PARP inhibitors or platinum, or HDAC inhibitors). Demographic data includes age, sex, race, country of origin, etc. Tumor response data may include epigenomic data, examples of which include changes in chromatin morphology and histone modifications.

[0051] Tumor response data may include cellular pathways. Examples include IFNγ, EGFR, MAP kinase, mTOR, CYP, CIMP, and AKT pathways, as well as pathways downstream of HER2 and other hormone receptors. Tumor response data may include cellular status indicators including collagen composition, appearance, or refractive index (e.g., extracellular vs. fibroblastic, nodular fasciitis), stroma density or other stroma characteristics (e.g., stromal thickness, wet vs. dry), and / or general appearance of angiogenesis or vasculature (including distribution of vasculature in collagen / stroma, also described as epithelial-mesenchymal transition or EMT). Tumor response data may include tumor characteristics. Examples include the presence of tumor budding or other morphological features / characteristics indicative of tumor complexity, tumor size (including tumor bulkiness or low-profile), tumor aggressiveness (e.g., known as high-grade basal-like tumors, particularly in colorectal cancer, or high-grade dysplasia, particularly in Barrett's esophagus), and / or tumor immune status (e.g., inflamed / "hot" vs. non-inflamed / "cold" vs. immune exclusion). Trained model module 136 may load training data from database 110, for example.

[0052] The machine learning training module 154 may be configured to train any number of machine learning models using any suitable machine learning technique including, for example, variational autoencoders, regression models, fully connected networks, deep learning techniques, probabilistic encoders / decoders, embedding vectors, artificial neural networks, convolutional neural networks, recurrent neural networks, etc. The machine learning training module 154 may store one or more trained models, including model weights / hyperparameters and / or other model metadata, in a suitable location for later retrieval, such as in the memory 124 and / or in the electronic database 110.

[0053] With the trained classifier model, a machine learning framework can be used to analyze and diagnose quantitative measurement analysis in subsequent images collected from the patient over time. In this manner, images and other data of previously treated and analyzed patients are utilized through the trained model to provide analysis and diagnosis capabilities for future patients.

[0054] The quantitative measurements data module 152 may be a standalone system that interfaces with external (e.g., third party) network accessible systems. In some examples, the quantitative measurements data module 152 may be integrated with one or more of these systems, including as part of a distributed cloud-based platform. Any of the functionality described in the technology herein may be distributed across one or more network accessible devices, including cloud-based devices.

[0055] In some examples, the quantitative measurement data module 152 may be part of a comprehensive quantitative measurement analysis prediction, diagnosis, and treatment system. For example, the quantitative measurement data module 152 and the quantitative measurement analysis computing device 104 may be configured to receive the generated quantitative measurement information and communicate the quantitative measurement analysis computing information to an external system, including a computer-based pathology lab / oncology system, which may use it to further diagnose the patient's cancer condition and identify matching therapies for use in treating the patient. The quantitative measurement analysis computing device 104 may also send the generated report to the patient's primary care provider's computer system and / or the physician's clinical record system to store the patient's report along with a database of previously generated reports for the patient and / or generated reports for other patients for use in future patient analyses, including machine learning based analyses as described herein.

[0056] Example Computer Implemented Machine Learning Model Operation The machine learning operations module 156 may, in some aspects, run one or more models trained by the machine learning training module 154. Once trained, the machine learning operations module 156 may load one or more trained models (e.g., from the database 110). The machine learning operations module 156 may include instructions to prepare the loaded models for use, such as installing dependencies, downloading one or more pre-trained models, downloading the models themselves, loading one or more model architectures (e.g., having one or more layers, connections and / or parameters), loading pre-trained weights into corresponding layers of the models, and performing one or more inference / prediction sub-operations, such as performing forward inference. In some aspects, the machine learning operations module 156 may be configured to perform post-processing and / or fine-tuning based on the model output.

[0057] The machine learning operation module 156 and / or another component may load the trained model (e.g., by accessing the trained model database 110). In some aspects, the trained model may be stored and accessed via memory of the computing device 102 and / or via distributed memory (e.g., via the network 106).

[0058] The quantitative measurement data module 152 may be a standalone system that interfaces with an external (e.g., third party) network-accessible system. In some examples, the quantitative measurement data module 152 may be integrated with one or more of these systems, including as part of a distributed cloud-based platform. For example, the quantitative measurement data module 152 may be integrated with a histopathology imaging system, such as a digital H&E stain imaging system, to enable rapid quantitative measurement analysis and reporting at the imaging station. Indeed, any of the functions described in the technology herein may be distributed across one or more network-accessible devices, including cloud-based devices. Histopathology slides may be used, for example, to confirm that an organoid is a particular cancer type or is not healthy tissue.

[0059] In some examples, the quantitative measurement data module 152 may be part of a comprehensive quantitative measurement analysis prediction, diagnosis, and treatment system. For example, the quantitative measurement data module 152 and the quantitative measurements of the analysis computing device 104 may be configured to communicate the quantitative measurement analysis computing information to an external system, including a computer-based pathology lab / oncology system, which may receive the generated quantitative measurement information and use it to further diagnose the patient's cancer condition and identify matching therapies for use in treating the patient. The quantitative measurement analysis computing device 104 may also send the generated report to the patient's primary care provider's computer system and / or the physician's clinical record system for storing the patient's report along with a database of previously generated reports for the patient and / or generated reports for other patients for use in future patient analysis, including machine learning based analysis as described herein.

[0060] The quantitative measurement client computing device 104 may be configured to access the quantitative measurement data and receive results (e.g., results generated by the machine learning operation module 156). For example, the quantitative measurement client computing device 104 may include one or more processors, memory, and / or network interface controllers (not shown) having functionality similar to that of the processor of the quantitative measurement prediction computing device 102. The memory of the quantitative measurement client computing device 104 may include one or more sets of instructions that perform functions different from those of the quantitative measurement prediction computing device 102. For example, the quantitative measurement client computing device 104 may include a viewer application 170 that includes instructions for accessing and displaying the results generated by the machine learning operation module 156. Thus, the viewer application 170 may enable a user (e.g., an employee, contractor, third party, physician, diagnostician, software engineer, system administrator, etc.) to review the results generated via the present technology.

[0061] In some aspects, the viewer application 170 may include instructions that enable a user to operate one or more trained models operated by the machine learning operation module 156 in a different manner (e.g., by parameterizing the inputs in a different manner) and / or modify the training parameters of one or more networks trained by the machine learning training module 154. The viewer application 170 may include instructions that enable a user to download results, share results (e.g., via email), and / or visualize results (e.g., through displaying one or more charts or graphs via the quantitative measurement client computing device 104). The viewer application 170 may include instructions that enable the viewer application 170 to access components of the computing environment 100 (e.g., the quantitative measurement prediction computing device 102, the electronic network 110, etc.) via the network 106, and components for displaying visualizations generated by the quantitative measurement prediction computing device 102.

[0062] Example Visualization Generation The visualization module 158 may include a set of computer-executable instructions for generating visualizations of quantitative measurement data and outputs generated by training one or more machine learning models using the machine learning training module 154 and / or running one or more machine learning models using the machine learning operation module 156. For example, the visualization module 158 may include a set of computer-executable instructions for generating drug-drug synergy score calculations and associated synergy score maps according to some embodiments. In some embodiments, for example, the visualization module 158 may generate a synergy score map such as in FIG. 4A. In some embodiments, the visualization module 158 may generate a three-dimensional phase synergy score map such as in FIG. 4B. In some embodiments, the visualization module 158 may generate an incomplete matrix visualization such as in FIG. 5C, or a complete matrix visualization.

[0063] The visualization module 158 may store the generated visualization (e.g., in memory 124 or electronic database 110) and / or display it directly (in some aspects, e.g., via an output device 128 which is or includes a display device) and / or send it over the network 106 (e.g., for display via a viewer application 170).

[0064] An electronic network 106 may communicatively couple elements of the environment 100. The network 106 may include a public network, such as the Internet, a private network, such as a research institute or corporate private network, and / or any combination thereof. The network 106 may include a local area network (LAN), a wide area network (WAN), a cellular network, a satellite network, and / or other network infrastructure, whether wireless or wired. In some aspects, the network 106 may be communicatively coupled to and / or part of a cloud-based platform (e.g., a cloud computing infrastructure). The network 106 may utilize communication protocols including packet-based and / or datagram-based protocols, such as the Internet Protocol, Transmission Control Protocol, User Datagram Protocol, and / or other types of protocols. The network 106 may include one or more devices that facilitate network communications and / or form a hardware base for the network, such as one or more switches, one or more routers, one or more gateways, one or more access points (e.g., wireless access points), one or more firewalls, one or more base stations, one or more repeaters, one or more backbone devices, etc.

[0065] The quantitative measurement generating system 108 may include any computing system, machine, or software that generates quantitative measurement data that may be used to train the machine learning models described herein and / or for prediction / classification purposes using the trained machine learning models described herein. For example, the quantitative measurement data generating system 108 may generate the quantitative measurement data received by the quantitative measurement data module 152 described above from a repository of microscope or digital microscope images, a histopathology slide generator or repository thereof, an immunohistochemistry slide image generator or repository thereof, any source of digital data (including images, text data, tabular data, binary data, etc.), etc.

[0066] The electronic database 110 may include one or more suitable electronic databases for storing and retrieving data, such as a relational database (e.g., a MySQL® database, an Oracle database, a Microsoft SQL Server database, a PostgreSQL database, etc.). The electronic database 110 may be a NoSQL database, such as a key-value store, a graph database, a document store, etc. The electronic database 110 may be an object-oriented database, a hierarchical database, a spatial database, a time series database, an in-memory database, etc. In some aspects, some or all of the electronic database 110 may be distributed.

[0067] In some aspects, some or all of the components of environment 100 may be implemented as cloud computing instances (e.g., electronic database 110 may be a cloud computing database instance, quantitative measurement prediction computing device 102 may be implemented as one or more cloud computing instances, etc.). Any suitable cloud computing infrastructure may be used to implement the cloud-based components, including Amazon Web Services, Microsoft Azure, Google Cloud Platform, etc. Additionally, in some aspects, some of the components of environment 100 may be implemented using a private cloud infrastructure and / or a hybrid cloud infrastructure.

[0068] In operation, a user may configure one or more training data sets (e.g., via quantitative measurements data module 152) and one or more untrained models (e.g., via machine learning training module 154) by accessing environment 100 (e.g., via quantitative measurements client computing environment 104). A user may cause the configuration of the training data sets and untrained models to occur by accessing elements of viewer application 170 (e.g., via a graphical user interface). A user may perform a training procedure by accessing machine learning operations module 156.

[0069] In some aspects, the data preparation, model configuration, and training processes may be done automatically, for example, using a scheduler (e.g., Cron), a timer, a task scheduler, PowerShell scripts, or another automation tool / framework (e.g., Ansible, Chef, Puppet, etc.). A user and / or an automation tool may generate the training procedure and then trigger an action (i.e., inference) operation once the training procedure is finished. A user may view the output of the inference and share the results (e.g., with a third party via email). In some aspects, the quantitative measurement prediction computing device 102 may include instructions for further processing of the machine learning results, such as generating visualizations as described above, via the visualization module 158.

[0070] In some aspects, a user may use environment 100 to configure the training and / or operation of one or more autoencoder architectures, such as autoencoder architecture 500 depicted in FIG. 5A below. A user may further configure the training and / or operation of one or more regressors, such as regressor 524 in FIG. 5B. A user may select a dataset of incomplete quantitative measurements (e.g., via a graphical user interface displayed in viewer application 170 in FIG. 1) and have environment 100 train one or more machine learning models using the dataset. A user may then select one or more new inputs (i.e., inputs that were not used to train the model) and provide them to environment 100 to have the trained model process the new inputs and provide predictions based thereon. For example, a user may provide an incomplete matrix, such as the sparse matrix of dose-response in FIG. 5C, as a new input at run-time for inference / prediction.

[0071] Additionally, more complex arrangements of models can be constructed, trained, and run by a user (or automatically as described above), such as the model of Figure 5D, where the decoder head 554 is parameterized using latent space interpolation and additional features for richer predictability. It should be appreciated that the ability of the present technique to further enhance the variational autocoder predictions with additional features is an improvement over any prior techniques that may apply a generative approach, leading to improved predictions.

[0072] Exemplary Regression or Classification-Based Dose-Response Prediction FIG. 2 illustrates a block flow diagram 200 for predicting complete or substantially complete dose-response quantitative measurements provided with incomplete novel quantitative measurements using a trained machine learning model, according to some embodiments, and further processing the complete or substantially complete dose-response quantitative measurements to generate a synergy distribution and / or synergy score to identify synergistic dose pairs. In general, the quantitative measurements may include drug efficacy, viability, cell proliferation rate, apoptosis rate, cell migration / invasion rate, gene expression levels, enzyme activity, cell cycle analysis, reactive oxygen species, cytokine / chemokine release, metabolic activity, neuronal activity, cell morphology, cell adhesion, binding affinity, and the like. In some embodiments, the quantitative measurements may be numerical and / or time-based. In other embodiments, the measurements may be expressed using other data modalities, such as microscopic image input, text input (e.g., SMILES encoded data), tabular input, and the like. In yet other embodiments, the quantitative measurements may correspond to measurements of off-target, general cytotoxicity / cytotoxicity (e.g., also killing healthy non-cancer cells), or adverse effects.

[0073] Block flow diagram 200 includes receiving incomplete quantitative measurements (depicted as a number of sparse matrices) (block 202). As used herein, "sparse" may have its strict mathematical definition (e.g., the number of zero elements is much greater than the number of non-zero elements in a given matrix) or a more relaxed definition, where a matrix is ​​technically incomplete as long as at least one element / entry of the matrix has a null, undefined, or zero value. The quantitative measurements may be represented using any suitable data structure and may include numerical information or other information (e.g., image-based information).

[0074] Block flow diagram 200 includes processing the quantitative measurements using a trained machine learning model (block 204) to predict one or more quantitative measurements and complete a sparse matrix (block 206). In the depicted example, the machine learning model processes a sparse matrix, where the sparseness of the matrix refers to unmeasured (i.e., incomplete, or intermediate) measurements for one or more drug-drug combinations. As shown, the matrix represents a Cartesian product of the concentrations of two drugs (Drug 1 along the X-axis, and Drug 2 along the Y-axis) measured in nanomoles. Of course, the drugs can be transposed (e.g., arranged in row or column order). The values ​​of the matrix represent clinical outcome values, such as relative inhibition rates. The values ​​can also correspond to the percentage of total cells that are not dead or dying, referred to as "viability." In various embodiments, tumor organoids can be grown using non-cancer cells, including immune cells or any animal cells. For example, the subject technology may be used in conjunction with that described in U.S. patent application Ser. No. 17 / 816,395, entitled "Platform For Co-Culture Imaging To Characterize In Vitro Efficacy Of Heterotypic Effector Cellular Therapies In Cancer," filed July 29, 2022, and incorporated by reference in its entirety into this specification for all purposes.

[0075] For example, see USSN 17 / 816,395, the method of characterizing cancer organoid response in response to immune cell-based therapy can be expanded using the present technology. In one aspect, one or more incomplete quantitative measurements corresponding to cancer organoid response can be provided to the present technology, which can process those measurements using the trained machine learning model described herein to generate complete / substantially complete quantitative measurements. In particular, the incomplete quantitative measurements can correspond to various administration schedules of cancer organoids and / or immune cells in conjunction with cancer organoids, and measurements corresponding to their response (e.g., morphological changes). For example, organoids can be combined with immune cells in an experimental environment. The experimental design can determine the ratio of immune cells (effector cells) to organoid cells (target cells) (e.g., 5:1, 4:1, 3:1, 2:1, 1:1, 1:2, 1:3, 1:4, 1:5, etc.). One or more agents can then be administered to this environment. The drug may not directly kill the organoid, but rather increase immune cell response to kill the cells. In some examples, immune cells may be engineered to directly kill organoid cells, with or without the addition of drug. Thus, organoid death may still be observed, and data related to dilute drug administration may be collected and processed using the present technology to predict data for intermediate doses and / or effector-to-target cell ratios that have not been experimentally tested in wet labs. In some embodiments, the data may measure the effect of immune cells on organoid cells without the addition of drug. The present technology may measure and predict different concentrations and / or distances between immune cells and / or organoid / tissue cells.

[0076] Advantageously, this may allow the present technology to significantly reduce the complexity (e.g., number, time, etc.) of the experiments required to characterize cancer organoid responses. It should be understood that many other experimental techniques may be accelerated in this manner.

[0077] Agents / drug types that may be used in dose-response experiments as described herein may include immune cell-based therapeutics, such as cellular therapeutics, as well as large molecule (biologics) and small molecule (pharmaceutical) agents that modulate immune cell activity. Examples of adoptive cell therapy include, for example, tumor infiltrating lymphocyte therapy and immune cells (such as T cells and NK cells) engineered to express chimeric antigen receptors (CAR T cells, CAR NK cells). Autologous or allogeneic effector cells, as well as autologous or allogeneic engineered or synthetic effector cell populations (such as CART cells), are contemplated. Examples of CAR T cell therapy include, but are not limited to, Axicabtagene ciloleucel, Brexucabtagene autoleucel, Idecabtagene vicleucel, Lisocabtagene maraleucel, and Tisagenlecleucel. Examples of biologics include recombinant proteins, such as antibodies and antibody-like constructs, that modulate immune cell activity and / or recruit immune effector cells to tumor cells. In various aspects, the immune cell-based therapeutic is a recombinant protein (e.g., an antibody or antibody-like construct) that engages / activates effector immune cells when in the presence of tumor organoid cells. For example, in some aspects, the immune cell-based therapeutic is a bispecific antibody or bispecific antibody-like construct (e.g., scFv-based construct, bispecific Fab, diabody, etc.) that binds to immune effector cells and tumor cell antigens. In various aspects, the antibody or antibody-like construct is a checkpoint inhibitor. The immune cell-based therapeutic can also be DNA or RNA-based, such as RNA-based nanoparticles that upregulate T cell activation. Alternatively, the immune cell-based therapeutic can be a small molecule drug, such as a drug that inhibits PI3K or Btk, or a drug that activates T cells, for example, via TLR.

[0078] In various embodiments, experiments may involve exposing non-cancerous cells to treatments and / or combinations of treatments and then measuring the viability of the cells, which measures off-target effects(y / ies) of the treatments (e.g., to see if a cancer therapeutic also adversely affects healthy non-cancerous cells).

[0079] Here, in general, the term "viability" refers to the ability of tumor cells to survive, grow, and proliferate. It is a measure of how healthy and functioning tumor cells are, and indicates their likelihood of continuing to divide and form new cells. A decrease in cell viability indicates that the treatment has a cytotoxic effect on the tumor cells, resulting in a decrease in survival and growth. On the other hand, if the cells remain viable, it suggests that the treatment may not be effectively targeting the tumor cells or that the tumor is resistant to the treatment. In various embodiments, the treatment may have a cytostatic effect on the tumor cells, preventing existing cells from dividing and increasing the total number of cells. In this case, a high percentage of cells are expected to remain viable, but the number of cells will be lower than expected for the time the cells have been proliferating and / or significantly lower than a population of cells that have proliferated for a similar time without treatment. Thus, the technology can be configured to predict different aspects of viability. In some aspects, one or more models can be trained to quantify cytotoxic and / or cytostatic effects.

[0080] The technique may complete a sparse matrix by destructive modification of an existing data structure (e.g., by inserting model predicted values ​​into the sparse matrix). Doing so may be more computationally efficient than making a copy of the sparse matrix. In either case, predicting one or more quantitative measurements and completing the matrix may result in one or more fully dense or substantially dense matrices corresponding to the incomplete / sparse matrix. Here, "fully dense" means all elements / items in the completed matrix that have predicted dose values ​​(e.g., relative percent inhibition values), and "substantially dense" means nearly all elements / items in the completed matrix that have predicted dose values. As used herein, "dense matrix" refers to a fully dense or substantially dense matrix.

[0081] Block flow diagram 200 includes generating a synergy distribution map that includes a plurality of synergy scores (i.e., synergy distribution) for the dense matrix predicted in block 204 (block 208). In the depicted example, drug-drug combinations in the middle and upper right quadrants of the synergy distribution map generally exhibit high synergy scores, indicating drug-drug dose combinations that may have increased relative inhibition.

[0082] In general, one way in which the present technology significantly and advantageously improves the performance of a computer or another technology or technical field (e.g., computerized drug screening assays) is by reducing the number of experiments (e.g., testing intermediate doses) required to identify combinations of compounds that exhibit synergistic effects. Further advantages are described in connection with the following figures.

[0083] 3 shows a block flow diagram 300 for predicting complete or substantially complete dose-response quantitative measurements (block 306) by processing incomplete new quantitative measurements (block 302) using a trained regression machine learning model (block 304). In general, block flow diagram 300 may correspond to techniques including in vitro analysis, experimental and image protocols to quantify the effect of each compound and dose combination on cell death, in some examples using intensity measurements of key dye staining (e.g., TOPRO3 and caspase 3 / 7) as endpoints.

[0084] The machine learning model in block 304 may in some embodiments correspond to the machine learning model in block 204 of Figure 2. The dose-response matrix in block 302 is sparse or incomplete insofar as it includes many intermediate, unmeasured values. The machine learning model in block 304 may be, for example, a regression model trained to interpolate survival probabilities for unmeasured doses from the dose-response matrix in block 302. The machine learning model may be any suitable model that performs regression or classification.

[0085] The machine learning model of block 304 may be a parametric or non-parametric machine learning regression or statistical model (e.g., ensembles of bagged and / or boosted trees such as Random Forest or XG Boost, k-nearest neighbors, artificial neural networks, support vector machines, Gaussian regression, polynomial regression, ridge regression, linear and lasso regression, non-negative matrix decomposition, etc.) that takes as input one or more of the following input quantitative measurement features: (1) Cell viability at a measured combined dose of two or more drugs (e.g., mean and standard deviation of TOPRO3 or caspase 3 / 7, and / or other statistical measurements, and / or other vital dye stains in whole cells) (2) Type of cancer (3) Dose combinations and, for multiple time point experiments (e.g., when multiple measurements are taken after treatment initiation), time points (e.g., duration of treatment at the time of measurement), and, for experiments involving cell therapy, effector to target cell ratios. (4) DNA / RNA sequencing data of cells / tumor organoids (5) Image-derived (e.g., visually detectable by a human or computer) morphological (human-interpretable) features, including cell size, shape, circularity, texture, pixel intensity, fractal dimension, etc. (i) at the image level (e.g., can be calculated for each image, for a collection of two or more organoids in an image) using standard computer vision methods, (ii) at the organoid level (e.g., for each organoid) using segmentation techniques that can segment the organoids, and / or (iii) organoid-level morphological features calculated using computer vision after segmentation has been performed. (6) Image embeddings obtained by passing unlabeled brightfield and / or fluorescent and / or cell-painted images through a pre-trained deep learning autoencoder, convolutional neural network, transformer, or other deep learning architecture trained using unsupervised or self-supervised learning; and / or (7) Drug molecule structural features (e.g., Simplified Molecular Input Line Entry System or SMILES). The machine learning model in block 304 may impute intermediate time points (e.g., impute unmeasured data corresponding to time points between two times the data was actually measured).

[0086] The machine learning model in block 304 may output predicted cell viability for an unmeasured combination dose and / or time point and / or effector to target cell ratio (e.g., in the case of a treatment involving immune cells). For example, activity may be measured at multiple different time points. In general, the technology may be used to predict a defect time point within an experiment in addition to (or as an alternative to) a defect dose. Some experiments may have time points, immune doses, and immune ratios as parameters. Any combination of these may be predicted using the technology. Morphological features may be cellular (e.g., associated with biological cells or many biological cells / organoids).

[0087] For example, the machine learning model in block 304 may be a random forest model, an XG boost model, an artificial neural network, a Gaussian regression model, a support vector machine, or a ridge and lasso regression model. In some aspects, the machine learning model may be an ensemble of multiple different machine learning models (e.g., two or more regression models).

[0088] The machine learning model may be trained (e.g., by a training module stored in memory 160 of FIG. 1B) by receiving and processing (e.g., from database 110) a training dataset of dose-response experiments that includes complete or substantially complete / dense matrices (or tensors if more than two compounds are used) of one or more dose-responses for several experiments. This training dataset may further include a respective cell viability value for each measured dose. The training module may then apply a masking procedure such that a set of values ​​from the dose-response matrix is ​​removed or obscured. The masking procedure may include a modified training dataset or an incomplete new training dataset in which only a subset of the dose-responses (experiments) are measured. This masked data may be used as an input training dataset by the training module. The output (target) dataset used for training may include an unmasked / unmasked dose-response matrix. The training module may train the machine learning model to predict these masked (i.e., unmeasured or missing) dose-responses from the input set (and one or more other additional features as described herein). The training procedure may be considered a supervised process when the target value (e.g., cell viability) is known. The training module may train a machine learning model to predict more than one unmeasured dose at a time (i.e., training a multi-output model), or to predict a single dose (e.g., training a single output model). Furthermore, as described herein, the technology may include predicting the viability (or different outputs) at different time points, or different treatment variables such as the effector-to-target cell ratio. In this case, the masked values ​​correspond to experiments at different time steps / points (or immune-tumor cell ratios), and the model may be trained to predict the viability for the unmeasured time steps (and optionally the unmeasured dose or the unmeasured immune-tumor cell ratio).

[0089] The machine learning model in block 304, once trained, may receive input features such as measured mean cell viability (e.g., from TOPRO3 staining). The machine learning model may process the input features to generate a model output including an unmeasured predicted viability dose. The machine learning model may be configured to have a cost function (e.g., a single cost function, a multi-target mean squared error function, etc.). In some aspects, the input features may include respective means, means and / or standard deviations of the viability measurements (e.g., TOPRO3, caspase, etc.), one or more doses, one or more cancer types, one or more image embeddings, one or more morphological features, etc. In some aspects, the machine learning model in block 304 may be a single output / single dose regression model.

[0090] Predicting the dose-response matrix at block 306 may include storing the predicted values ​​in a non-transitory computer memory, such as memory 160 of Figure 1B or database 110 of Figure 1A. In some embodiments, predicting the dose-response may also include sending one or more predicted values ​​and / or matrices over communication network 106 of Figures 1A and 1B. Example Computer-Implemented Synergy Score and Summary Synergy Score Calculations

[0091] FIG. 4A illustrates an example synergy score calculation 402 relative to a synergy score map 404 .

[0092] In general, the synergy score measures the amount of synergy or antagonism of a combination between the compounds (e.g., drugs) being tested. The synergy score may be calculated by comparing the observed drug combination response to the expected response, calculated using a reference model that does not assume interactions between the drugs. Multiple synergy modeling algorithms may be used, including Bliss, Highest Single Agent (HSA), Loewe, and / or Zero Interaction Potency (ZIP). Furthermore, in some embodiments, multiple of these models may be combined to provide a consensus scoring that can advantageously further reduce the risk of false positive synergy results, as described in Ianevski, Aleksandr, Anil K. Giri, and Tero Aittokallio. "SynergyFinder 3.0: Interactive Analysis and Consensus Interpretation of Multiple Drug Synergy Across Multiple Samples." Nucleic Acids Research 50.W1(2022):W739-W743, incorporated herein by reference in its entirety for all purposes.

[0093] For example, Bliss synergy assumes that the drugs independently induce their effects. Bliss synergy is calculated as follows: Sbliss=EA,B(d1,d2)-(1-(1-EA(d1))(1-EB(d2))) where EA(d1) and EB(d2) are the individual effects of drugs A and B at doses 1 and 2, respectively, and EA,B(d1,d2) is the experimentally measured combined effect. Here, the effect can refer to cell inhibition, i.e., the percentage of cells killed by the compound. For example, it can be seen that the effect on drug A is 30% (i.e., drug A kills 30% of the cells), and the effect on drug B is 60%. The experimentally measured combined effect EA,B(d1,d2) is 80%. The Bliss model predicts that this combination will kill 72% of the cells (the second term in the formula). The difference between the measured effect EA,B(d1,d2) and the expected effect is the synergistic effect Sbliss=0.08 (8%).

[0094] In some embodiments, a summary synergy score may be calculated as the average excess response across doses. An example is shown below. Less than -10. The interaction between the two drugs is likely to be antagonistic. -10 to 10. The interaction between the two drugs is likely to be additive. greater than 10. The interaction between the two drugs is likely to be synergistic.

[0095] In some embodiments, open source packages can be used to calculate synergy scores. One example is SynergyFinder, available in the R programming language. Another is the Python package Synergy (Wooten, David J., and Reka Albert. "synergy: a Python library for calculating, analyzing and visualizing drug combination synergy" Bioinformatics 37.10 (2021): 1473-1474).

[0096] FIG. 4B illustrates a three-dimensional synergy score map 450, according to some embodiments.

[0097] Illustrative Imaging-Based Generation Dose-Response Predictions The techniques may include training a deep neural network encoding model (convolutional neural network, variational autoencoder, generative adversarial network, transformer, RNN and / or LSTM) trained on unlabeled images (e.g., brightfield and / or fluorescent and / or cell images) to represent images in a compressed (latent) space. In some aspects, the training may be performed using unsupervised learning or self-supervised learning. For example, the training may be performed via the computing environment 100 of FIG. 1A.

[0098] 5A illustrates an example variational autoencoder architecture 500 for predicting perfect or substantially perfect dose-response quantitative measurements. The variational autoencoder is configured to map one or more inputs 502-A (e.g., brightfield and / or fluorescent images 504-A) into a continuous latent space 508 from which samples may be taken by training a probabilistic encoder 506-A and a probabilistic decoder 506-B of the autoencoder architecture 500 until one or more reconstructed inputs 502-B resemble inputs 502-A.

[0099] Once the autoencoder architecture 500 is trained, the latent space 508 can be sampled to generate an unmeasured dose image 504-B that includes a quantitative measurement (e.g., the survivability of the unmeasured dose).

[0100] Figure 5B shows a conceptual diagram 520 of interpolating in a latent space 522 using the trained autoencoder architecture 500 of Figure 5A. Interpolation in the latent space between doses can be performed to reconstruct embeddings of images of unmeasured doses and / or time points / steps. Interpolation can include identifying extreme points in the embedding space and interpolating between them.

[0101] For example, as shown in Figure 5B, if we consider two points in the latent space, Z a and Z b The values ​​of Z are known because they are part of the reduced dimensional representation of the input image 504-A. a may be the first extreme point, which is the centroid of the image treated with the maximum dose of the drug, and Z b Another point Z in the latent space can be the second extreme point, which is the centroid of the image treated with the minimum dose of the drug. c The value of is unknown, but it can be determined by using a trained probabilistic decoder 506-B to obtain the vector Z b -Z aThe latent space can be interpolated by sampling along the direction of the unmeasured dose image 504-B to generate the unmeasured dose image 504-B. The vector connecting these two points estimates the dose-response curve. Sampling along this vector gives an embedding of intermediate interpolated samples. The embedding vector is decoded to generate data points (images) corresponding to the unmeasured doses and / or time points. Decoding these samples with a generative model produces an image similar to the raw image, making the model explainable. This approach is extensible to multiple extreme points, so that evaluation of multiple drug combinations is envisioned (e.g., untreated to a first drug, untreated to a second drug, and first drug to a second drug, and back). In general, the latent space can be sampled along two (or more) directions of a hyperplane defined by three (or more) treatment centroids, which allows for the generation of data for intermediate conditions defined by the treatment centroids. In other words, the technique can be used to perform synthetic experiments of drug combinations, each with a configurable dosage, by selecting appropriate points on the hyperplane / latent space and decoding them using decoder 506-B.

[0102] The embedding vector may in some embodiments be augmented with a set of morphologically human-interpretable features at the image and / or organoid level extracted from the input image (including cell size, shape, texture, and pixel intensity). As shown below (e.g., as shown in FIG. 5D), the embedding vector may be augmented using modalities / data other than image information, such as cell sequencing features, drug molecular structure features, etc.

[0103] As described below, a second neural network (e.g., an image-based classifier or regression head) can input the generated image or its embedding vector and output: (1) predicted cell viability for an unmeasured combination dose, (2) predicted cell viability at an unmeasured time point, or (3) a predicted synergy map or synergy score for the compound.

[0104] The sampled values ​​may be passed to an image-based regressor 524. In some embodiments, the regressor 524 may include a fully connected network 526 capable of image embedding, for example as described below. U.S. patent application Ser. No. 17 / 301,975, entitled "Artificial Fluorescent Image Systems and Methods," filed April 20, 2021, U.S. patent application Ser. No. 17 / 857,901, entitled "Large Scale Organoid Analysis," filed July 5, 2022, and U.S. patent application Ser. No. 17 / 114,386, entitled "Large Scale Organoid Analysis," filed December 7, 2020, each of which is incorporated herein by reference in its respective entirety for all purposes. For example, hardware and software such as those described in USSN 17 / 301,975 may be used to automate image capture such as brightfield microscopy and use additional optical accessories or capabilities (e.g., darkfield, phase contrast, differential interference contrast, polarizing, confocal, light sheet, etc.). The technology may process brightfield images and / or fluorescence both during inference, training, and during inference. Raw brightfield images may be stored in database 110, for example, along with respective three-channel brightfield images. The raw images and / or three-channel brightfield images may be used to train one or more machine learning models. For example, input 502-A and / or input 504-A may include raw and / or channel fluorescent and / or brightfield images. For example, the three channels of data in a three-channel brightfield (or fluorescent) image may include a blue / total nuclei channel, a green / apoptotic channel, and a red / pink / dead channel. Each channel may be used to form a fluorescent image. In some embodiments, the machine learning model can be trained to output a combined three-channel fluorescent image including a blue / total nuclei channel, a green / apoptotic channel, and a red / pink / dead channel.

[0105] The regressor 524 may be trained to predict information 528 decoded from the images (e.g., survivability, synergy score, or other information). It should be appreciated that the method of interpolating across time points can be extended for experiments using images acquired at multiple times.

[0106] In general, training of the autoencoder architecture 500 may be performed as follows: First, the autoencoder architecture 500 is trained on a dataset of brightfield and / or fluorescent and / or cell paint images. These images may correspond to various test conditions, including different compound doses (combined or alone), cancer strains, time points (in case of multi-timepoint experiments), and any other conditions that define the experiment. The autoencoder 500 is trained to reconstruct only the input images such that image reconstruction error is minimized. Thus, the network does not require labels associated with the input images at this stage, and this training stage may be characterized as unsupervised. As mentioned above, the autoencoder may also be trained on other data modalities other than images, such as DNA / RNA cell sequencing data, drug molecular structural features, and human-interpretable image morphological features.

[0107] Once the autoencoder 500 is trained, a second network (e.g., network 526, a regression head, etc.) can be trained to predict endpoints (e.g., predicted viability for a missed dose, synergy score, etc.) from the autoencoder image embeddings. The endpoints can be, for example, the viability of a compound, or a synergy score, or a classification of whether a compound is synergistic based on the current image embeddings (i.e., the experiment corresponding to the input image). This stage of training can be supervised, such that each input with an embedded label is provided with a label (a binary value for viability, synergy score, or synergy). The regression head can be trained to use these labels to predict synergy scores, viability scores, etc. In general, the autoencoder 500 is trained in a label-free manner (i.e., unsupervised training), whereas the network 526 is trained using the labels (i.e., supervised training).

[0108] During inference, the technique may sample embeddings from the latent space corresponding to intermediate doses or time points, and the resulting sampled embeddings are passed to a second network to predict the endpoint.

[0109] 5C shows an example sparse dose-response matrix 540. Sparse dose-response matrix 540 includes many intermediate / incomplete values, such as value 542 (coordinates: d1c, d2c). Values ​​544 (coordinates: d1b, d2b) and 546 (coordinates: d1a, d2a) are known and may be used to predict, impute, or interpolate value 542 using any of the techniques described above.

[0110] As will be explained, the technology involves machine learning / deep learning models that predict quantitative measurements (e.g., cell viability, response to a drug, etc.) for unmeasured dose-response measurements. It will be appreciated that variations of the above autoencoder architecture 500 are envisioned, as described with respect to the following figures.

[0111] FIG. 5D illustrates an alternative autoencoder architecture 550 according to some embodiments. In some embodiments, the autoencoder architecture 550 may be identical in some / all respects to the architecture 500 of FIG. 5A, except that the autoencoder architecture 550 may include one or more additional input features 552 and one or more corresponding embeddings 554. The embeddings 554 may be referred to as embedding vectors. In some embodiments, the additional input features 552 may represent additional raw input features (e.g., RNA / DNA sequencing data), and the corresponding embedding vector 554 may be a concatenation of the additional input features 552 (e.g., feature vectors) and the embeddings 554 (e.g., embedding vectors). Furthermore, the regression head 526 may be trained to predict the viability of the current input image from the image latent space embeddings augmented by the additional feature embeddings 554. The embedding vectors 554 may be similarly augmented with cell sequencing data (e.g., DNA / RNA) and / or drug molecular structure features (e.g., SMILES notation). In some embodiments, the technique may include sampling the embedding vectors with additional features (e.g., cell sequencing data such as RNA / DNA, drug molecular structure features such as those represented in SMILES notation, etc.). In some embodiments, the autoencoder architecture 550 may be trained on tabular data (e.g., morphological image features, cell sequencing data, drug molecular structure data) rather than the images themselves. The output of the regression head may be either a viability or synergy score map, indicating whether the compounds are synergistic, additive, or antagonistic.

[0112] FIG. 5E illustrates a further alternative autoencoder architecture 570 having a first autoencoder 572-A and a second autoencoder 572-B in accordance with some aspects of the present technology. The autoencoder 572-A and the autoencoder 572-B may generate image embeddings 574-A and feature embeddings 574-B, respectively. The embeddings 574 may be concatenated to generate embeddings 554, which may correspond to the embeddings 554 of FIG. 5D. In FIG. 5E, the second autoencoder 572-B may encode the input features using an architecture similar to that of the autoencoder 572-A. The architecture 570 may sample embeddings from the autoencoders 572-A and 572-B and concatenate the embeddings to form embeddings 554, which may be further processed by the regression head 526 to generate the decoded information 528.

[0113] 5F illustrates yet another alternative autoencoder architecture 590 in accordance with some aspects. The autoencoder architecture 590 may, for example, use two loss functions (image reconstruction loss, or MSE 画像 and survival probability predicted loss MSE 生存可能性 5A and the autoencoder architecture 550 of FIG. 5D by including a structure that jointly optimizes the regression head 526 with the autoencoder 590 by combining the decoder head and the regression head. The architecture 590 is expected to generate a latent space 508 that represents images whose neighboring points have similar survivability values. Arrows 594 indicate that the loss values ​​are backpropagated through the decoder head and the regression head, respectively.

[0114] 1B, in some aspects, the machine learning operational module 154 may include, for example, a set of computer-executable instructions that, when executed, cause a computer (e.g., computing device 104) to select one or more trained models (e.g., regression model 304 of FIG. 3, autoencoder 500 of FIG. 5A, etc.) based on inputs provided to the controller. For example, the instructions may retrieve a model from database 110 based on inputs provided to the machine learning operational module 154.

[0115] The benefits of examples The present technique advantageously reduces the scale of experiments (as measured by reagents, plate wells, cells / organoids, etc.) that need to be performed for a given modeling task. For example, Figure 6 shows a graph 600 of exemplary experimental savings that can be realized by using the present technique. Demonstration experiments have demonstrated a 66% or greater reduction in the number of experiments required.

[0116] Exemplary Computer-Implemented Method 7A shows a computer-implemented method 700 for predicting experimental cellular quantitative measurements using a machine learning model trained using quantitative measurement data to reduce the burden of laboratory experiments, according to some embodiments. The method 700 may include receiving an input of new quantitative measurements via one or more processors (block 702). The method 700 may include processing the input of new quantitative measurements using the machine learning model to generate one or more predicted experimental cellular quantitative measurements (block 704).

[0117] The machine learning model of method 700 may be a regression machine learning model, a classification machine learning model, etc. The input of new quantitative measurements may include a sparse matrix of quantitative measurements. For example, the input of new quantitative measurements may correspond to the incomplete dose matrix block 202 of FIG. 2, and the machine learning model may correspond to the machine learning model block 204 of FIG. 2. In some embodiments, the regression machine learning model may be at least one of: (i) a random forest model, (ii) an XG boost model, (iii) an artificial neural network, (iv) a Gaussian regression model, (v) a support vector machine, or (vi) a ridge and lasso regression model.

[0118] The regression machine learning model may include a mean squared error cost function for single or multiple targets.

[0119] In some embodiments, the predicted experimental cellular quantitative measurement includes data corresponding to at least one microscopic image associated with the experiment or experimental result. For example, the data source used to measure viability may be a microscopic fluorescent image. This may be independent of the underlying model used to predict the quantitative information. In some embodiments, the machine learning model is trained with at least some masked training quantitative measurement data.

[0120] In some embodiments, the machine learning model is an artificial neural network capable of encoding and decoding input images and trained using quantitative measurement data including one or more brightfield training images and / or one or more fluorescent training images. In that case, the input of the new quantitative measurement may correspond to at least one point of the continuous latent space embedding, for example, as described with respect to FIG. 5D. The artificial neural network may be a variational autoencoder network, a transformer network, a recurrent neural network, or a long-term short-term network. For example, the machine learning model may include an encoder trained to map one or more portions of the brightfield or fluorescent training images to a continuous latent space embedding representing a compressed, low-dimensional representation of the one or more training images, and a decoder trained via one or more processors to map the one or more continuous latent space embeddings to one or more portions of a reconstructed image by sampling from the continuous latent space embedding to generate a reconstructed input corresponding to one or more portions of the training image, and update the continuous latent space embedding based on a comparison between the reconstructed input and one or more portions of the training input. An example of this architecture is described with respect to FIG. 5D, FIG. 5E, and FIG. 5F. The method 700 may include training one or more parallel autoencoders, as shown in FIG. 5E.

[0121] In some aspects, the method 700 may include predicting a viability or synergy score or other quantitative measurement metric from the image using a machine learning model. The method 700 may include extending the continuous latent space embedding using an encoder by processing one or more human-interpretable morphological features, or other additional features. The human-interpretable morphological features may be generated by computer vision techniques.

[0122] Method 700 may further include generating one or both of: (i) a predicted survival probability responsive to the input of the new quantitative measurements; and (ii) a synergy map or score responsive to the input of the new quantitative measurements.

[0123] 7B illustrates a computer-implemented method 720 for training a machine learning model to make predictions of unmeasured quantitative measurements to identify compounds that exhibit synergistic effects and reduce the burden of laboratory experimentation, according to some embodiments. The method 720 may include receiving training quantitative measurement data via one or more processors (block 722). The method 720 may include training a machine learning model via one or more processors to generate predicted experimental cell viability measurements based on input of new quantitative measurements by processing the training quantitative measurement data (block 724). The method may include storing the machine learning model via one or more processors as a trained machine learning model in a memory of the computer (block 726).

[0124] Method 720 may be a regression machine learning model. In some embodiments, the training quantitative measurement data of method 720 includes a substantially dense or fully dense matrix of one or more training quantitative measurements, each of the matrices including a respective experimental cell viability measurement. The new quantitative measurement input may include a sparse matrix of quantitative measurements. The new quantitative measurement input may correspond to an inference time (i.e., runtime) input that the model has not seen before. The regression machine learning model may be at least one of: (i) a random forest model, (ii) an XG boost model, (iii) an artificial neural network, (iv) a Gaussian regression model, (v) a support vector machine, or (vi) a ridge and lasso regression model. The regression machine learning model may include a mean squared error cost function for single or multiple targets. In some embodiments, other cost functions may be used.

[0125] Method 720 may include training a machine learning model by receiving one or more additional input features via one or more processors and processing the additional input features with one or more substantially dense or fully dense training quantitative measurement matrices. For example, the input modality of the additional input features may be binary data, vectors of embeddings, text data, tabular data, etc. The material of the additional input features may correspond to at least one of: (i) fluorescent nuclear and chromosomal counterstain values, (ii) indications of cell death, (iii) indications of cell apoptosis, (iv) cancer type, (v) one or more image embeddings, (vi) one or more morphological features, (vii) one or more compound doses, (viii) one or more features encoding a molecular structure of a compound, (ix) cellular RNA sequencing data, or (x) cellular DNA sequencing data. Training the machine learning model may include masking at least a portion of the training quantitative measurement data.

[0126] In some embodiments, the machine learning model may be a variational autoencoder, for example as shown in Figures 5D-5F. The training quantitative measurement data may include one or more bright field and / or fluorescent training images, and the new quantitative measurement input may correspond to at least one point of the continuous latent space embedding. Training the machine learning model may include an encoder, via one or more processors, for mapping one or more portions of the training images to a continuous latent space embedding representing a compressed, low-dimensional representation of the one or more training images, and a decoder, trained via one or more processors, for mapping the one or more continuous latent space embeddings to one or more portions of a reconstructed image by sampling from the continuous latent space embedding to generate, via one or more processors, a reconstructed input corresponding to the one or more portions of the training image, and updating the continuous latent space embedding based on a comparison between the reconstructed input and the one or more portions of the training input. Training the machine learning model may include training an imaging-based regression model to predict a viability or synergy score from the reconstructed image or from the continuous latent space embedding. The imaging-based regression model may be trained using supervised learning. On the other hand, an autoencoder may be trained using unsupervised learning. Training the imaging-based regression model may include training a fully connected network to generate decoded survivability based on input from the decoder.

[0127] The method 720 includes receiving human-interpretable morphological features, where the human-interpretable morphological features may include at least one of size, shape, circularity, fractal dimension, texture, and pixel intensity. In some aspects, the human-interpretable morphological features may be generated by computer vision techniques. Training the encoder may include extending the latent space embedding by processing one or more human-interpretable morphological features. The method 720 may include training a regression head to generate one or both of (i) a predicted viability corresponding to an input of a novel quantitative measurement, and (ii) a synergy map or score corresponding to an input of a novel quantitative measurement. Training the machine learning model may include optimizing the regression head and the variational autoencoder by combining two or more loss functions.

[0128] The various embodiments described above can be combined to provide further embodiments. All U.S. patents, U.S. patent application publications, U.S. patent applications, foreign patents, foreign patent applications and non-patent publications mentioned herein and / or listed in the Application Data Sheet are incorporated herein by reference in their entirety. Aspects of the embodiments can be modified as necessary to employ concepts from various patents, applications, and publications to provide still further embodiments.

[0129] These and other variations can be made to the embodiments in light of the above detailed description. In general, in the following claims, the terms used should not be construed to limit the claims to the specific embodiments disclosed in the specification and claims, but should be construed to include all possible embodiments, along with the full scope of equivalents to which such claims are entitled. Thus, the claims are not limited by this disclosure.

[0130] Aspects of the technology described in this disclosure may include any of the following aspects, either alone or in combination.

[0131] 1. A computer-implemented method for predicting unmeasured quantitative measurements (experimental cellular quantitative measurements) using a machine learning model trained using quantitative measurement data to reduce the burden of laboratory experiments, the computer-implemented method comprising: receiving, via one or more processors, an input of new quantitative measurements; and processing the input of new quantitative measurements using the machine learning model to generate one or more predicted experimental cellular quantitative measurements.

[0132] 2. The computer-implemented method of aspect 1, wherein the machine learning model is a regression machine learning model or a classification machine learning model and the input of new quantitative measurements comprises a sparse matrix of quantitative measurements.

[0133] 3. The computer-implemented method of aspect 1 or 2, wherein the regression machine learning model is at least one of: (i) a random forest model; (ii) an XG Boost model; (iii) an artificial neural network; (iv) a Gaussian regression model; (v) a support vector machine; or (vi) a ridge and lasso regression model.

[0134] 4. The computer-implemented method of any one of aspects 1-3, wherein the regression machine learning model comprises a mean squared error cost function for a single or multiple targets.

[0135] 5. The computer-implemented method of embodiment 1, wherein the predicted experimental cellular quantitative measurements include at least one microscopic image associated with the experiment or experimental outcome.

[0136] 6. The computer-implemented method of aspect 1 or 2, wherein the machine learning model is trained with at least some masked training quantitative measurement data.

[0137] 7. The computer-implemented method of aspect 1, wherein the machine learning model is an artificial neural network capable of encoding and decoding input images and trained using quantitative measurement data including one or more brightfield training images and / or one or more fluorescent training images, and wherein a new quantitative measurement input corresponds to at least one point of the continuous latent space embedding.

[0138] 8. The computer-implemented method of embodiment 7, wherein the artificial neural network is a variational autoencoder network, a transformer network, a recurrent neural network, or a long-term short-term network.

[0139] 9. The computer-implemented method of any one of aspects 1-7, wherein the machine learning model includes: an encoder trained to map one or more portions of a bright field or fluorescent training image to a continuous latent space embedding that represents a compressed, low-dimensional representation of the one or more training images; and a decoder trained via one or more processors to map the one or more continuous latent space embeddings to one or more portions of a reconstructed image by sampling from the continuous latent space embedding to generate reconstructed inputs that correspond to the one or more portions of the training images, and to update the continuous latent space embedding based on a comparison between the reconstructed inputs and the one or more portions of the training inputs.

[0140] 10. The computer-implemented method of any one of aspects 1-9, further comprising using a machine learning model to predict a viability or synergy score or other quantitative measurement metric from the images.

[0141] 11. The computer-implemented method of any one of aspects 1-9, further comprising using an encoder to extend the continuous latent space embedding by processing one or more human-interpretable morphological features.

[0142] 12. The computer-implemented method of aspect 11, wherein the human-interpretable morphological features are generated by computer vision techniques.

[0143] 13. The computer-implemented method of any one of aspects 1-10, further comprising generating one or both of: (i) a predicted survival probability responsive to the input of new quantitative measurements; and (ii) a synergy map or score responsive to the input of new quantitative measurements.

[0144] 14. The computer-implemented method of claim 1, wherein the predicted cellular quantitative measurements correspond to at least one of a defective time point, a defective dose response, or a defective immune ratio.

[0145] 15. A computer-implemented method for training a machine learning model to make predictions of unmeasured quantitative measurements to identify compounds that exhibit synergistic effects and reduce the burden of laboratory experimentation, the method comprising: receiving, via one or more processors, training the machine learning model, via the one or more processors, to generate predicted experimental cell viability measurements based on input of new quantitative measurements by processing the training quantitative measurement data; and storing, via the one or more processors, the machine learning model in a memory of the computer as a trained machine learning model.

[0146] 16. The computer-implemented method of aspect 14, wherein the machine learning model is a regression machine learning model, the training quantitative measurement data comprises one or more substantially dense or fully dense matrices of training quantitative measurements, each of the matrices comprising a respective experimental cell viability measurement, and the input of new quantitative measurements comprises a sparse matrix of quantitative measurements.

[0147] 17. The computer-implemented method of aspect 15, wherein the regression machine learning model is at least one of: (i) a random forest model; (ii) an XG Boost model; (iii) an artificial neural network; (iv) a Gaussian regression model; (v) a support vector machine; or (vi) a ridge and lasso regression model.

[0148] 18. The computer-implemented method of aspect 15 or 16, wherein the regression machine learning model comprises a mean squared error cost function for a single or multiple targets.

[0149] 19. The computer-implemented method of any one of aspects 14-17, wherein training the machine learning model includes receiving, via one or more processors, one or more additional input features and processing the additional input features together with one or more substantially dense or fully dense training quantitative measurement matrices.

[0150] 20. The computer-implemented method of embodiment 19, wherein the additional input features comprise at least one of: (i) fluorescent nuclear and chromosomal counterstain values; (ii) an indication of cell death; (iii) an indication of cell apoptosis; (iv) cancer type; (v) one or more image embeddings; (vi) one or more morphological features; (vii) one or more compound doses; (viii) one or more features encoding the molecular structure of the compound; (ix) RNA sequencing data of the cell; or (x) DNA sequencing data of the cell.

[0151] 21. The computer-implemented method of any one of aspects 14-19, wherein training the machine learning model includes masking at least a portion of the training quantitative measurement data.

[0152] 22. The computer-implemented method of any one of aspects 14-20, wherein the machine learning model is a variational autoencoder, the training quantitative measurement data includes one or more bright field and / or fluorescent training images, and the new quantitative measurement input corresponds to at least one point of the continuous latent space embedding.

[0153] 23. The computer-implemented method of aspect 21, wherein training the machine learning model includes training a trained encoder, via one or more processors, to map one or more portions of a training image to a continuous latent space embedding that represents a compressed, low-dimensional representation of the one or more training images; and training a decoder, via one or more processors, to map the one or more continuous latent space embeddings to one or more portions of a reconstructed image by sampling from the continuous latent space embedding to generate reconstructed inputs that correspond to the one or more portions of the training images, and update the continuous latent space embedding based on a comparison between the reconstructed inputs and the one or more portions of the training inputs.

[0154] 24. The computer-implemented method of aspect 21 or 22, wherein training the machine learning model comprises training an imaging-based regression model to predict the viability or synergy score from the reconstructed images or from a continuous latent space embedding.

[0155] 25. The computer-implemented method of any one of aspects 21-23, wherein training the imaging-based regression model includes training a fully connected network to generate a decoded survivability based on input from the decoder.

[0156] 26. The computer-implemented method of any one of aspects 14 to 24, further comprising receiving human-interpretable morphological features, the human-interpretable morphological features comprising at least one of size, shape, circularity, fractal dimension, texture, and pixel intensity.

[0157] 27. The computer-implemented method of aspect 25, wherein the human-interpretable morphological features are generated by computer vision techniques.

[0158] 28. The computer-implemented method of any one of aspects 21-24, wherein training the encoder includes extending the latent space embedding by processing one or more human-interpretable morphological features.

[0159] 29. The computer-implemented method of any one of aspects 14 to 23, further comprising training a regression head to generate one or both of: (i) a predicted survival probability corresponding to the input of new quantitative measurements; and (ii) a synergy map or score corresponding to the input of new quantitative measurements.

[0160] 30. The computer-implemented method of embodiment 28, wherein training the machine learning model includes optimizing a regression head and a variational autoencoder by combining two loss functions.

[0161] 31. A computing system comprising one or more processors and one or more memories storing computer-executable instructions, wherein when the computer-executable instructions are executed by the one or more processors, the computing system receives an input of novel quantitative measurements via the one or more processors, and processes the input of novel quantitative measurements using a machine learning model to generate one or more predicted experimental cellular quantitative measurements.

[0162] 32. A computing system comprising one or more processors and one or more memories storing computer-executable instructions, wherein when the computer-executable instructions are executed by the one or more processors, the computing system receives, via the one or more processors, training quantitative measurement data; processes the training quantitative measurement data to train, via the one or more processors, a machine learning model to generate predicted experimental cell viability measurements based on input of new quantitative measurements; and stores, via the one or more processors, the machine learning model in the memory as a trained machine learning model.

[0163] 33. A computer-readable medium storing computer-executable instructions that, when executed by one or more processors, cause a computer, via the one or more processors, to receive an input of new quantitative measurements and process the input of new quantitative measurements using a machine learning model to generate one or more predicted experimental cellular quantitative measurements.

[0164] 33. A computer-readable medium storing computer-executable instructions, which when executed by one or more processors, cause the computer to receive, via the one or more processors, training quantitative measurement data, train, via the one or more processors, a machine learning model to generate predicted experimental cell viability measurements based on input of new quantitative measurements by processing the training quantitative measurement data, and store, via the one or more processors, the machine learning model in memory as a trained machine learning model.

[0165] Additional Considerations The computer-readable medium may include executable computer-readable code stored thereon to program a computer (including, for example, a processor and a GPU) to the techniques of the present specification. Examples of such computer-readable storage media include hard disks, CD-ROMs, digital versatile disks (DVDs), optical storage devices, magnetic storage devices, ROMs (read-only memories), PROMs (programmable read-only memories), EPROMs (erasable programmable read-only memories), EEPROMs (electrically erasable programmable read-only memories), and flash memories. More generally, the processing unit of the computing device 1300 may represent a CPU-type processing unit, a GPU-type processing unit, a TPU-type processing unit, a field programmable gate array (FPGA), another class of digital signal processors (DSPs), or other hardware logic components that may be driven by a CPU.

[0166] Although the example deep learning framework herein is described as being comprised of an example machine learning architecture (FCN configuration), it should be noted that any number of suitable artificial neural network architectures may be used. In general, the deep learning framework herein may implement any suitable statistical model (e.g., a neural network or other model implemented through a machine learning process) that is applied to each of the received images. As described herein, the statistical model may be implemented in various manners. In some examples, machine learning is used to evaluate the training images. In some examples, image features may be identified as training classifiers using a learning algorithm such as a neural network, a support vector machine (SVM), or other machine learning process. Once the classifiers in the statistical model are properly trained on a series of training images, the statistical model may be used in real time to analyze subsequent images that are provided as input to the statistical model to predict the likelihood of survival. In some examples, when the statistical model is implemented using a neural network, the neural network may be configured in various ways. In some examples, the neural network may be a deep neural network and / or a convolutional neural network. In some examples, the neural network may be a distributed and scalable neural network. The neural network may be customized in various ways, including but not limited to providing a specific top layer, such as a logistic regression top layer. A convolutional neural network may be considered as a neural network that includes a set of nodes with associated parameters. A deep convolutional neural network may be considered as having a stacked structure with multiple layers. A neural network or other machine learning process may include many different sizes, numbers of layers, and levels of connectivity. Some layers may correspond to stacked convolutional layers (optionally followed by contrast normalization and max pooling), followed by one or more fully connected layers.For neural networks trained by large datasets, the number of layers can be increased and layers can be expanded by using dropout to address the potential problem of overfitting. In some instances, neural networks can be designed to abandon the use of fully connected upper layers at the top of the network. By forcing the network to reduce the dimensionality of the intermediate layers, very deep neural network models can be designed while dramatically reducing the number of learned parameters.

[0167] A system for performing the methods described herein may include a computing device, and more specifically may be implemented in one or more processing units, such as a central processing unit (CPU) and / or a graphical processing unit (GPU) including one or more or a cluster of CPUs and / or GPUs. The described features and functions may be stored and implemented in one or more non-transitory computer-readable media of the computing device. The computer-readable medium may include, for example, an operating system and software modules, or "engines," that perform the methods described herein. More generally, the computer-readable medium may store batch normalization processing instructions for an engine to perform the techniques herein. The computing device may be a distributed computing system, such as Amazon Web Services or Google Cloud Platform cloud computing solutions.

[0168] The computing device includes a network interface communicatively coupled to a network for communicating with a portable personal computer, a smart phone, an electronic document reader, a tablet, and / or a desktop personal computer, or other computing devices. The computing device further includes an I / O interface connected to devices such as a digital display, a user input device, and the like.

[0169] The engine's functionality may be implemented across distributed computing devices interconnected via communication links, and the like. In other examples, the system's functionality may be distributed across any number of devices, including the illustrated portable personal computers, smartphones, electronic documents, tablets, and desktop personal computing devices. The computing devices may be communicatively coupled to a network and another network. The network may be a public network, such as the Internet, a private network of an academic institution or enterprise, or any combination thereof. The network may include a local area network (LAN), a wide area network (WAN), a cellular network, a satellite network, or other network infrastructure, whether wireless or wired. The network may utilize communication protocols, including packet-based and / or datagram-based protocols, such as the Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), or other types of protocols. Additionally, the network may include numerous devices, such as switches, routers, gateways, access points (such as wireless access points as illustrated), firewalls, base stations, repeaters, backbone devices, and the like, that facilitate network communication and / or form the hardware base of the network.

[0170] A computer-readable medium may include executable computer-readable code stored thereon for programming a computer (including, for example, a processor and a GPU) to the techniques herein. Examples of such computer-readable storage media include hard disks, CD-ROMs, digital versatile disks (DVDs), optical storage devices, magnetic storage devices, ROMs (read-only memories), PROMs (programmable read-only memories), EPROMs (erasable programmable read-only memories), EEPROMs (electrically erasable programmable read-only memories), and flash memories. More generally, the processing unit of a computing device may represent a CPU-type processing unit, a GPU-type processing unit, a field programmable gate array (FPGA), another class of digital signal processor (DSP), or other hardware logic components that may be driven by a CPU.

[0171] Throughout this specification, multiple instances may implement components, operations, or structures described as a single instance. Although individual operations of one or more methods have been illustrated and described as separate operations, one or more of the individual operations may be performed simultaneously, and the operations need not be performed in the order illustrated. Structures and functions presented as separate components in an illustrated configuration may be implemented as a combined structure or component. Similarly, structures and functions presented as a single component may be implemented as a separate component or multiple components.

[0172] Furthermore, certain aspects are described herein as including logic or some routines, subroutines, applications, or instructions. These may constitute either software (e.g., code embodied on a machine-readable medium or in a transmission signal) or hardware. In hardware, routines, etc. are tangible units capable of performing certain operations and may be configured or arranged in a particular way. In an exemplary aspect, one or more computer systems (e.g., standalone, client, or server computer systems) or one or more hardware modules (e.g., processors or groups of processors) of a computer system may be configured by software (e.g., applications or application portions) as hardware modules that operate to perform certain operations as described herein.

[0173] In various aspects, a hardware module may be implemented mechanically or electronically. For example, a hardware module may include dedicated circuitry or logic that is permanently configured (e.g., as a microcontroller, a field programmable gate array (FPGA), or a special purpose processor such as an application specific integrated circuit (ASIC)) to perform certain operations. A hardware module may also include programmable logic or circuitry that is temporarily configured by software (e.g., contained within a general purpose processor or other programmable processor) to perform certain operations. It will be appreciated that whether a hardware module is implemented mechanically, in dedicated permanently configured circuitry, or in temporarily configured circuitry (e.g., configured by software) is determined by cost and time considerations.

[0174] Thus, the term "hardware module" should be understood to encompass a tangible entity that is physically configured, permanently configured (e.g., wired), or temporarily configured (e.g., programmed) to operate in a particular manner or perform particular operations described herein. Given the temporarily configured (e.g., programmed) aspect of the hardware modules, each of the hardware modules need not be configured or instantiated at any one instance in time. For example, if the hardware modules include a general-purpose processor configured using software, the general-purpose processor may be configured as different hardware modules at different times. Thus, the software may configure the processor, for example, to configure a particular hardware module at one instance of time and a different hardware module at a different instance of time.

[0175] Hardware modules can provide information to and receive information from other hardware modules. Thus, the described hardware modules can be considered to be communicatively coupled. When multiple such hardware modules are present simultaneously, communication can be achieved through signal transmission (e.g., via appropriate circuits and buses) connecting the hardware modules. In aspects where multiple hardware modules are configured or instantiated at different times, communication between such hardware modules can be achieved, for example, through storage and retrieval of information in memory structures accessed by the multiple hardware modules. For example, one hardware module can perform an operation and the output of that operation can be stored in a memory device to which the hardware module is communicatively coupled. An additional hardware module can then access the memory device at a later time to retrieve and process the stored output. Hardware modules can also initiate communication with input or output devices and operate on resources (e.g., collections of information).

[0176] Various operations of the example methods described herein may be performed, at least in part, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the associated operations. Whether temporarily or permanently configured, such processors may constitute processor-implemented modules that operate to perform one or more operations or functions. Modules referred to herein may, in some example aspects, include processor-implemented modules.

[0177] Similarly, the methods or routines described herein may be at least partially processor-implemented. For example, at least some of the operations of the methods may be performed by one or more processors or processor-implemented hardware modules. Performance of certain of the operations may be distributed among one or more processors that are not only present in a single machine, but are also spread across multiple machines. In some example aspects, the processor may be located in a single location (e.g., in a home environment, in an office environment, or as a server farm), while in other aspects the processor may be distributed across multiple locations.

[0178] Performance of certain of the operations may be distributed among one or more processors that are not only present within a single machine, but are also spread across multiple machines. In some example aspects, one or more processors or processor-implemented modules may be located in a single geographic location (e.g., in a home environment, an office environment, or a server farm). In other example aspects, one or more processors or processor-implemented modules may be distributed across several geographic locations.

[0179] Unless specifically stated otherwise, descriptions herein using words such as "processing," "computing," "calculating," "determining," "presenting," "displaying," and the like may refer to machine (e.g., computer) actions or processes that manipulate or transform data represented as physical (e.g., electronic, magnetic, or optical) quantities in one or more memories (e.g., volatile memory, non-volatile memory, or a combination thereof), registers, or other machine components that receive, store, transmit, or display information.

[0180] As used herein, any reference to "one embodiment" or "an embodiment" means that a particular element, feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. The appearances of the phrase "in one embodiment" in various places in the specification do not necessarily all refer to the same embodiment.

[0181] Some embodiments may be described using the terms "coupled" and "connected," along with their derivatives. For example, some embodiments may be described using the term "coupled" to indicate that two or more elements are in direct physical or electrical contact. However, the term "coupled" may mean that two or more elements are not in direct contact with each other, but still cooperate or interact with each other. The embodiments are not limited in this context.

[0182] As used herein, the terms "comprises," "comprising," "including," "has," "having," or any other variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, article, or apparatus that includes a list of elements is not necessarily limited to only those elements, but may include other elements not expressly listed in or inherent to such process, method, article, or apparatus. Furthermore, unless expressly stated to the contrary, "or" refers to an inclusive or, not an exclusive or. For example, a condition A or B is satisfied by any one of the following: A is true (or exists) and B is false (or does not exist), A is false (or does not exist) and B is true (or exists), or both A and B are true (or exist).

[0183] Additionally, the use of "a" or "an" is used to describe elements and components of aspects herein. This is done merely for convenience and to give the description a general meaning. This description and the claims that follow should be read to include one or at least one, and the singular also includes the plural unless it is clear that it is meant otherwise.

[0184] This detailed description should be construed as merely exemplary and does not describe every possible embodiment, as describing every possible embodiment would be impractical, if not impossible. One can implement many alternative embodiments using either current technology or technology developed after the filing date of this application.

Claims

1. 1. A computer-implemented method for predicting unmeasured quantitative measurements using a machine learning model trained using quantitative measurement data to reduce the burden of laboratory experimentation, the method comprising: receiving an input of new quantitative measurements via one or more processors; and processing the input of the novel quantitative measurements using the machine learning model to generate one or more predicted experimental cellular quantitative measurements.

2. The machine learning model is a regression machine learning model or a classification machine learning model, The computer-implemented method of claim 1 , wherein the input of new quantitative measurements comprises a sparse matrix of quantitative measurements.

3. 3. The computer-implemented method of claim 2, wherein the regression machine learning model is at least one of: (i) a random forest model, (ii) an XG boost model, (iii) an artificial neural network, (iv) a Gaussian regression model, (v) a support vector machine, or (vi) a ridge and lasso regression model.

4. The computer-implemented method of claim 2 , wherein the regression machine learning model comprises a mean squared error cost function for a single or multiple targets.

5. The computer-implemented method of claim 1 , wherein the predicted experimental cellular quantitative measurements include at least one microscopic image associated with an experiment or an experimental outcome.

6. The computer-implemented method of claim 2 , wherein the machine learning model is trained with at least some masked training quantitative measurement data.

7. the machine learning model is an artificial neural network capable of encoding and decoding input images and trained using quantitative measurement data including one or more brightfield training images and / or one or more fluorescent training images; The computer-implemented method of claim 1 , wherein the novel quantitative measurement input corresponds to at least one point of a continuous latent space embedding.

8. 8. The computer-implemented method of claim 7, wherein the artificial neural network is a variational autoencoder network, a transformer network, a recurrent neural network, or a long-term short-term network.

9. The machine learning model, an encoder trained to map one or more portions of the brightfield or fluorescent training images to the continuous latent space embedding that represents a compressed, low-dimensional representation of the one or more training images; 8. The computer-implemented method of claim 7, further comprising: a decoder trained via one or more processors to map one or more continuous latent space embeddings to one or more portions of a reconstructed image by sampling from the continuous latent space embedding to generate reconstructed inputs corresponding to one or more portions of the training image, and to update the continuous latent space embeddings based on a comparison between the reconstructed inputs and the one or more portions of the training inputs.

10. 10. The computer-implemented method of claim 9, further comprising using the machine learning model to predict a viability or synergy score or other quantitative measurement metric from an image.

11. 10. The computer-implemented method of claim 9, further comprising using the encoder to extend the continuous latent space embedding by processing one or more human-interpretable morphological features.

12. The computer-implemented method of claim 11 , wherein the human-interpretable morphological features are generated by computer vision techniques.

13. 11. The computer-implemented method of claim 10, further comprising generating one or both of: (i) a predicted survival probability responsive to the input of the novel quantitative measurements; and (ii) a synergy map or score responsive to the input of the novel quantitative measurements.

14. 2. The computer-implemented method of claim 1, wherein the predicted cellular quantitative measurements correspond to at least one of a defective time point, a defective dose response, or a defective immune ratio.

15. 1. A computer-implemented method for training a machine learning model to make predictions of unmeasured quantitative measurements to identify compounds that exhibit synergistic effects and reduce the burden of laboratory experimentation, the method comprising: receiving, via one or more processors, training quantitative measurement data; training a machine learning model via one or more processors to process the training quantitative measurement data to generate predicted experimental cell viability measurements based on input of new quantitative measurements; and storing, via one or more processors, the machine learning model in a computer memory as a trained machine learning model.

16. the machine learning model is a regression machine learning model; the training quantitative measurement data comprises one or more substantially dense or fully dense matrices of training quantitative measurements, each of said matrices comprising a respective experimental cell viability measurement; The computer-implemented method of claim 15 , wherein the input of new quantitative measurements comprises a sparse matrix of quantitative measurements.

17. 17. The computer-implemented method of claim 16, wherein the regression machine learning model is at least one of: (i) a random forest model, (ii) an XG boost model, (iii) an artificial neural network, (iv) a Gaussian regression model, (v) a support vector machine, or (vi) a ridge and lasso regression model.

18. 20. The computer-implemented method of claim 16, wherein the regression machine learning model includes a mean squared error cost function for a single or multiple targets.

19. Training the machine learning model receiving, via one or more processors, one or more additional input features; and processing the additional input features together with the one or more substantially dense or fully dense training quantitative measurement matrices.

20. 20. The computer-implemented method of claim 19, wherein the additional input features comprise at least one of: (i) fluorescent nuclear and chromosomal counterstain values; (ii) an indication of cell death; (iii) an indication of cellular apoptosis; (iv) cancer type; (v) one or more image embeddings; (vi) one or more morphological features; (vii) one or more compound doses; (viii) one or more features encoding a molecular structure of a compound; (ix) cellular RNA sequencing data; or (x) cellular DNA sequencing data.

21. 20. The computer-implemented method of claim 16, wherein training the machine learning model comprises masking at least a portion of the training quantitative measurement data.

22. the machine learning model is a variational autoencoder; the training quantitative measurement data includes one or more brightfield and / or fluorescent training images; 16. The computer-implemented method of claim 15, wherein the novel quantitative measurement input corresponds to at least one point of a continuous latent space embedding.

23. Training the machine learning model training, via one or more processors, an encoder for mapping one or more portions of the training images into the continuous latent space embedding that represents a compressed, reduced dimensional representation of the one or more training images; 23. The computer-implemented method of claim 22, comprising: mapping one or more continuous latent space embeddings to one or more portions of a reconstructed image by sampling from the continuous latent space embedding to generate reconstructed inputs corresponding to one or more portions of the training image; training a decoder via one or more processors; and updating the continuous latent space embeddings via one or more processors based on a comparison between the reconstructed inputs and the one or more portions of the training inputs.

24. Training the machine learning model 24. The computer-implemented method of claim 23, comprising training an imaging-based regression model to predict a viability or synergy score from reconstructed images or from the continuous latent space embedding.

25. 25. The computer-implemented method of claim 24, wherein training the imaging-based regression model comprises training a fully connected network to generate a decoded survivability based on input from a decoder.

26. 16. The computer-implemented method of claim 15, further comprising receiving human-interpretable morphological features, the human-interpretable morphological features comprising at least one of size, shape, circularity, fractal dimension, texture, and pixel intensity.

27. 27. The computer-implemented method of claim 26, wherein the human-interpretable morphological features are generated by computer vision techniques.

28. 26. The computer-implemented method of claim 25, wherein training the encoder comprises extending the latent space embedding by processing one or more human-interpretable morphological features.

29. 25. The computer-implemented method of claim 24, further comprising training a regression head to generate one or both of: (i) a predicted survival probability responsive to the input of the novel quantitative measurements, and (ii) a synergy map or score responsive to the input of the novel quantitative measurements.

30. 30. The computer-implemented method of claim 29, wherein training the machine learning model comprises optimizing the regression head and the variational autoencoder by combining two loss functions.

31. 1. A computing system comprising: one or more processors; and one or more memories storing computer-executable instructions that, when executed by the one or more processors, cause the computing system to: receiving, via one or more processors, an input of new quantitative measurements; A computing system that processes the input of the novel quantitative measurements using a machine learning model to generate one or more predicted experimental cellular quantitative measurements.

32. 1. A computing system comprising: one or more processors; and one or more memories storing computer-executable instructions that, when executed by the one or more processors, cause the computing system to: receiving, via said one or more processors, training quantitative measurement data; training a machine learning model via the one or more processors to process the training quantitative measurement data to generate predicted experimental cell viability measurements based on input of new quantitative measurements; A computing system that, via one or more processors, stores the machine learning model in a memory as a trained machine learning model.

33. A computer-readable medium having stored thereon computer-executable instructions that, when executed by one or more processors, cause a computer to: receiving, via one or more processors, an input of new quantitative measurements; A computer readable medium that processes the input of said novel quantitative measurements using a machine learning model to generate one or more predicted experimental cellular quantitative measurements.

34. A computer-readable medium having stored thereon computer-executable instructions that, when executed by one or more processors, cause a computer to: receiving, via said one or more processors, training quantitative measurement data; training a machine learning model via the one or more processors to process the training quantitative measurement data to generate predicted experimental cell viability measurements based on input of new quantitative measurements; A computer-readable medium that, via one or more processors, stores the machine learning model in a memory as a trained machine learning model.

Images