Systems and methods for enhancing diagnostic classifier performance in targeted subpopulations via domain adaptation

Abstract

Systems and methods of the disclosure may include a computer-implemented method, the computer-implemented method may include: receiving, at a system, a source dataset comprising a first set of data samples; receiving, at the system, a target dataset comprising a second set of data samples; performing, using a processor, feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; computing, via employing a non-parametric importance-weighting technique, importance weights for each of the first set of data samples and the second set of data samples; assigning the computed importance weights to each of the first set of data samples and the second set of data samples; and training a diagnostic classifier using the importance-weighted first set of data samples and the importance-weighted second set of data samples.

Description

SYSTEMS AND METHODS FOR ENHANCING DIAGNOSTIC CLASSIFIER PERFORMANCE IN TARGETED SUBPOPULATIONS VIA DOMAIN ADAPTATIONCROSS REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to U.S. Provisional Application No. 63 / 626,854, filed on January 30, 2024, and U.S. Provisional Application No. 63,687,127, filed on August 26, 2024, which are both incorporated by reference herein in their entireties.TECHNICAL FIELD

[0002] The present disclosure relates generally to the field of machine learning and cancer diagnostics and, more specifically, to systems and methods for improving the performance of diagnostic classifiers in targeted subpopulations using domain adaptation techniques.BACKGROUND

[0003] Current methods for cancer diagnosis using blood-based methylation signatures involve training classifiers on large datasets encompassing various cancer types and tumor burdens. While these classifiers can achieve high accuracy in general applications, their performance often degrades when applied to specific subpopulations or intended use cases, such as post-diagnosis monitoring of minimal residual disease (MRD). Previous efforts to adapt classifiers to specific subpopulations have employed techniques such as synthetic data augmentation and feature optimization. These methods introduce computational and technicalcomplexities due to the large number of hyper parameters and design constraints involved.

[0004] The background description provided herein is for the purpose of generally presenting the context of the disclosure. Unless otherwise indicated herein, the materials described in this section are not prior art to the claims in this application and are not admitted to be prior art, or suggestions of the prior art, by inclusion in this section.SUMMARY OF THE DISCLOSURE

[0005] According to certain aspects of the disclosure, systems and methods are described for improving the performance of classifiers tailored to targeted subpopulations, including in the context of cancer diagnostics via blood-based methylation signatures.

[0006] In one aspect, a computer-implemented method is disclosed. The computer-implemented may include: receiving, at a system, a source dataset comprising a first set of data samples; receiving, at the system, a target dataset comprising a second set of data samples; performing, using a processor associated with the system, feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; computing, based on the performing and via employing a non-parametric importance-weighting technique, importance weights for at least the first set of data samples; assigning, using the processor, the computed importance weights to at least the first set of data samples; and training, using the processor, a diagnostic classifier using at least the importance-weighted first set of data samples.

[0007] In another aspect, a system is disclosed. The system may include: one or more processors; one or more computer readable media storing instructions that are executable by the one or more processors to perform operations to: receive a source dataset comprising a first set of data samples; receive a target dataset comprising a second set of data samples; perform feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; compute, based on the performing and via employing a non-parametric importanceweighting technique, importance weights for at least the first set of data samples; assign the computed importance weights to at least the first set of data samples; and train a diagnostic classifier using at least the importance-weighted first set of data samples.

[0008] In yet another aspect, a non-transitory computer-readable medium is disclosed. The non-transitory computer-readable medium may store computerexecutable instructions which, when executed by a system, cause the system to perform operations including: receiving, at a system, a source dataset comprising a first set of data samples; receiving, at the system, a target dataset comprising a second set of data samples; performing, using a processor associated with the system, feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; computing, based on the performing and via employing a non-parametric importance-weighting technique, importance weights for at least the first set of data samples; assigning, using the processor, the computed importance weights to at least the first set of data samples; and training, using theprocessor, a diagnostic classifier using at least the importance-weighted first set of data samples.

[0009] In yet another aspect, a computer-implemented method is disclosed. The computer-implemented may include: receiving, at a system, a source dataset comprising a first set of data samples; receiving, at the system, a target dataset comprising a second set of data samples; performing, using a processor associated with the system, feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; performing, using the processor, a first feature augmentation process on the first set of features and a second feature augmentation process on the second set of features, wherein the first feature augmentation process is different than the second feature augmentation process; generating, using the processor and based on performing the first feature augmentation process and the second feature augmentation process, a first feature- augmented data set and a second feature-augmented data set; and training, using the processor, a diagnostic classifier using the first feature-augmented data set and the second feature-augmented data set.

[0010] Additional objects and advantages of the disclosed embodiments will be set forth in part in the description that follows, and in part will be apparent from the description, or may be learned by practice of the disclosed embodiments. The objects and advantages of the disclosed embodiments will be realized and attained by means of the elements and combinations particularly pointed out in the appended claims.

[0011] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosed embodiments, as claimed.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate several embodiments and together with the description, serve to explain the principles of the disclosure.

[0013] FIG. 1 depicts a graph illustrating the challenge of addressing covariate shift in a classification task, according to one or more embodiments of the present disclosure.

[0014] FIG. 2A depicts an exemplary computer system for executing the methods described herein, according to one or more embodiments of the present disclosure.

[0015] FIG. 2B depicts an exemplary software platform for executing the methods described herein.

[0016] FIG. 3 depicts an exemplary workflow for a method for enhancing the performance of a diagnostic classifier in a targeted subpopulation, according to one or more embodiments of the present disclosure.

[0017] FIGS. 4A and 4B depict the improvements that utilizing a Kernel Mean Matching (KMM) approach has relative to a conventional approach, according to one or more embodiments of the present disclosure.

[0018] FIG. 5 depicts a diagram containing training and validation data, according to one or more embodiments of the present disclosure.

[0019] FIG. 6 depicts a graph that presents performance of two classifiers, a first that is trained without applying importance weights and a second that is trained with importance weights, according to one or more embodiments of the present disclosure.

[0020] FIGS. 7 A and 7B depict graphs illustrating performance of weighted and unweighted samples on synthetic samples, according to one or more embodiments of the present disclosure.

[0021] FIG. 8 depicts data tables that compare performance results for a weighted and unweighted classifier, according to one or more embodiments of the present disclosure.

[0022] FIG. 9 depicts data tables that compare the performance of a Heme minimum residual disease (MRD) baseline classifier and an importance-weighted classifier on real samples, according to one or more embodiments of the present disclosure.

[0023] FIG. 10 depicts data tables that compare the performance of a Heme MRD baseline classifier and an importance-weighted classifier on real non-Hodgkin’s lymphoma samples, according to one or more embodiments of the present disclosure.

[0024] FIG. 1 1 depicts data tables that compare the performance of a Heme MRD baseline classifier and an importance-weighted classifier on synthetic cross- validation samples, according to one or more embodiments of the present disclosure.

[0025] FIG. 12 depicts data tables that compare the performance of a HemeMRD baseline classifier and an importance-weighted classifier on real non-Hodgkin’slymphoma samples, according to one or more embodiments of the present disclosure.

[0026] FIG. 13 depicts data tables that compare the performance of a Heme MRD baseline classifier and an importance-weighted classifier on holdout synthetic samples, according to one or more embodiments of the present disclosure.

[0027] FIG. 14 depicts data tables that compare the performance of a Heme MRD baseline classifier and an importance-weighted classifier on synthetic nonHodgkin’s lymphoma samples, according to one or more embodiments of the present disclosure.

[0028] FIG. 15 depicts data tables that compare the performance of a Heme MRD baseline classifier and an importance-weighted classifier on synthetic nonHodgkin’s lymphoma samples, according to one or more embodiments of the present disclosure.

[0029] FIGS. 16A and 16B depict plots that display the importance weights learned by the KMM algorithm for Heme samples, according to one or more embodiments of the present disclosure.

[0030] FIG. 17 depicts a plot that displays how the classifier assigns weights to non-cancer samples, according to one or more embodiments of the present disclosure.

[0031] FIGS. 18A and 18B depict graphs that demonstrate the performance of the classifier using importance weights with synthetic samples, specifically focusing on non-Hodgkin’s samples in a target tumor fraction region, according to one or more embodiments of the present disclosure.

[0032] FIG. 19 depicts an exemplary workflow for utilizing a target dataset for learning importance weights, according to one or more embodiments of the present disclosure.

[0033] FIG. 20 depicts an exemplary workflow for utilizing a target dataset for learning importance weights and training a classifier, according to one or more embodiments of the present disclosure.

[0034] FIG. 21 depicts an exemplary workflow for leveraging domain adaptation through feature augmentation, according to one or more embodiments of the present disclosure.

[0035] FIG. 22 depicts an example computing system, according to one or more embodiments of the present disclosure.DETAILED DESCRIPTION OF EMBODIMENTS

[0036] The terminology used below may be interpreted in its broadest reasonable manner, even though it is being used in conjunction with a detailed description of certain specific examples of the present disclosure. Indeed, certain terms may even be emphasized below; however, any terminology intended to be interpreted in any restricted manner will be overtly and specifically defined as such in this Detailed Description section. Both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the features, as claimed.

[0037] In the field of cancer diagnostics, machine learning classifiers play an important role in interpreting complex biological data to detect and monitor cancer.These classifiers are typically trained on large, diverse datasets that include a multitude of cancer types and various tumor burdens (e.g., a dataset may containcomprehensive methylation signatures from numerous cancer subjects). However, while classifiers trained on such extensive datasets can achieve high overall accuracy, their performance may diminish when applied to specific subpopulations or specialized clinical scenarios. This performance drop is particularly problematic in sensitive applications like post-diagnosis monitoring of minimal residual disease (MRD), where the accurate detection of low tumor fractions is paramount to downstream decision-making.

[0038] Attempts to address this issue have primarily focused on synthetic data augmentation and feature optimization techniques. Synthetic data augmentation involves generating additional artificial samples that mimic the characteristics of the target subpopulation, which are then used to retrain or fine-tune the classifier. Feature optimization, on the other hand, aims to select or transform features to improve the classifier's performance for the specific subpopulation. While these methods can enhance performance, they come with drawbacks. Synthetic data augmentation is computationally expensive, requiring extensive resources to generate and validate artificial data. It also involves a large number of hyperparameters, making the process complex and difficult to optimize. Feature optimization similarly demands substantial computational power and introduces a high degree of technical complexity due to the intricate design constraints and numerous hyperparameters involved.

[0039] The concepts described herein propose a novel solution to these challenges by utilizing domain adaptation techniques, specifically importance weighting methodologies, that address the issue of covariate shift, where the distribution of features differs between the training (source) and deployment (target) datasets, but the conditional distribution of labels given to the features remains thesame. For instance, FIG. 1 presents a graph 100 that illustrates the challenge of covariate shift in a classification task, using a hypothetical example. In this context, the task may be to classify images (e.g., distinguishing between cats and dogs) based on pixel representations. The training set (source domain) consists of real images (blue dots), and the classifier is trained on these samples of real images. However, the deployment scenario (target domain) involves classifying cartoon images of cats and dogs, so that the underlying features (pixel representations) differ significantly from those in the training set. The red line in graph 100 represents the true function that ideally maps input features to correct classifications. The blue dots denote the training samples used to learn this function. When the classifier is trained solely on the blue training samples, it learns a function (green line) that fits the training data well but does not generalize to the test samples (black cross marks). This is because the test samples follow a different distribution than the training samples due to the covariate shift. The discrepancy between the learned function and the true function in the test samples highlights the problem: the classifier trained on the source domain fails to perform accurately in the target domain. This mismatch results in prediction errors when the classifier is deployed on data it was not optimized for, demonstrating the need for domain adaptation techniques.

[0040] To mitigate the impact of covariate shift, a method based on the concept of likelihood functions may be employed to provide a statistical model of the data. One way to correct for covariate shift is to reweight the likelihood function using importance weights. These weights are computed based on the ratio of the probability densities of the target and source data. This ratio reflects how much more likely a particular sample is to appear in the target distribution relative to the source distribution. In a parametric setting, where the likelihood function may be explicitlydefined, such as in a linear regression problem, the process involves adjusting each term of the likelihood by the importance weight. This adjustment helps align the training data distribution with the test data distribution. The optimization problem then becomes finding the maximum weighted likelihood estimate (MWLE). This MWLE approach ensures that the model is not biased towards the training data distribution but is instead optimized to perform well on the target data distribution.

[0041] As an example of the foregoing, and in consideration of a regression problem in which the aim is to fit a model to an underlying cubic function, if the probability distributions of the source (training data) and target (test data) may be analytically derived, the importance weights can be computed for each sample. Without any weighting, fitting the model to the source data may result in a prediction that deviates significantly from the true target distribution, represented by a misaligned regression line. By applying the importance weights, the influence of each training sample may be adjusted, resulting in a weighted likelihood function that better represents the target distribution. Consequently, the fitted model aligns more closely with the true function, thereby reducing prediction error.

[0042] Although effective, the importance weighting process described above becomes considerably more complex in high-dimensional spaces, such as those involving mixed model features representing underlying biological states, like when analyzing amino acids, such as RNA or DNA, including methylation patterns in circulating free DNA (cfDNA). In these scenarios, the dimensionality can reach tens of thousands, making it infeasible to derive an analytical likelihood function. The challenge may be further compounded by the intricate nature of the data, necessitating advanced techniques to estimate the probability densities and compute the importance weights accurately.

[0043] To overcome the limitations described above, the concepts described herein employ a method for addressing covariate shift by leveraging an importance weighting approach known as Kernel Mean Matching (KMM). KMM provides a nonparametric approach to estimate the importance weights without requiring explicit density estimation. Specifically, KMM leverages the concept of kernel functions to project data into a different feature space, where the task of matching the distributions of source and target datasets becomes more manageable. Kernels, such as the Gaussian or Radial Basis Function (RBF) kernels, transform the data into a higher or lower-dimensional space where linear separability or other advantageous properties may be exploited. Given the appropriate kernels, Kernel Mean Matching can be reduced to a quadratic programming problem that can be solved via an interior point method or projected gradient approach. This transformation allows the method to bypass the full computation of kernel estimates, thereby significantly improving computational efficiency.

[0044] By incorporating KMM into the classifier adaptation process, the proposed method offers several key advantages over conventional approaches. First, it simplifies the adaptation process by reducing the number of hyperparameters to just one (e.g., the kernel bandwidth hyperparameter, which determines the scale of the kernel function, influencing how much influence each data point has over its neighbors in the feature space), significantly lowering the technical complexity. Second, it is computationally efficient, avoiding the need for costly synthetic data generation and complex feature transformations. Instead, KMM adjusts the distribution of training samples in the feature space to match the target samples, facilitating a more straightforward and faster adaptation process. Third, the method iseasy to implement within existing classifier frameworks, facilitating smooth integration and deployment.

[0045] In view of the foregoing, the concepts described herein may enhance the performance of classifiers for specific subpopulations and maintain computational efficiency and simplicity while doing so. In practical applications, such as MRD detection and monitoring, the method has demonstrated improved diagnostic accuracy, particularly in low tumor fraction samples. By focusing on the specific subpopulation of interest, the adapted classifier achieves better performance in the intended use cases, thus addressing a critical need in cancer diagnostics and improving clinical outcomes (as further illustrated and described herein). Overall, this approach offers a robust and scalable solution to the limitations of conventional methods, paving the way for more precise and effective cancer diagnostic tools.

[0046] The subject matter of the present disclosure will now be described more fully hereinafter with reference to the accompanying drawings, which form a part hereof, and which show, by way of illustration, specific exemplary embodiments. An embodiment or implementation described herein as “exemplary” is not to be construed as preferred or advantageous, for example, over other embodiments or implementations; rather, it is intended to reflect or indicate that the embodiment(s) is / are “example” embodiment(s). Subject matter may be embodied in a variety of different forms and, therefore, covered or claimed subject matter is intended to be construed as not being limited to any exemplary embodiments set forth herein; exemplary embodiments are provided merely to be illustrative. Likewise, a reasonably broad scope for claimed or covered subject matter is intended. Among other things, for example, subject matter may be embodied as methods, devices, components, or systems. Accordingly, embodiments may, for example, take the formof hardware, software, firmware, or any combination thereof. The following detailed description is, therefore, not intended to be taken in a limiting sense.

[0047] Throughout the specification and claims, terms may have nuanced meanings suggested or implied in context beyond an explicitly stated meaning. Likewise, the phrase “in one embodiment” or “in some embodiments,” or “in one aspect” or “in some aspects” as used herein does not necessarily refer to the same embodiment or aspect, and the phrase “in another embodiment” or “in another aspect” as used herein does not necessarily refer to a different embodiment or aspect. It is intended, for example, that claimed subject matter include combinations of exemplary embodiments in whole or in part.

[0048] Diseases referred to herein may include cancer. Non-limiting cancer types that the concepts described herein may be applied to include, for example, breast cancer, lung cancer (e.g., non-small cell lung cancer (NSCLC)), prostate cancer, colorectal cancer, renal cancer, uterine cancer, pancreatic cancer, cancer of the esophagus, a lymphoma, head and neck cancer, ovarian cancer, a hepatobiliary cancer, a melanoma, cervical cancer, multiple myeloma, leukemia, thyroid cancer, bladder cancer, and gastric cancer. Additionally, it is also important to note that although the concepts described throughout this disclosure are made in reference to cancer, these designations are for exemplary purposes only and are not intended to be limiting. Specifically, the concepts described herein may be applicable to other disease types and other disease-detecting machine-learning classifiers.

[0049] FIG. 2A depicts an exemplary system for enhancing the performance of a diagnostic classifier for targeted subpopulations. Exemplary system 200 includes a data collection component 10, a database 20, and device data intelligence component 30, operably connected to each other via network 40. Alternatively, oradditionally, one or more of the components may be connected with another component locally without reliance on network connection; e.g., through a wired connection. In many aspects described herein, sequencing data of cell-free nucleic acids are used to illustrate the concepts. However, one of skill in the art would understand that the current method may be applied to sequencing data of DNA, RNA, or other materials, as well from a variety of sample types, e.g., a blood sample (e.g., a serum sample, a plasma sample, a whole blood sample), a urine sample, a saliva sample, a tissue sample, a bone marrow sample, etc.

[0050] As disclosed herein, data collection component 10 may include a device or machine with which sequencing data may be generated. In some embodiments, data collection component 10 may include one or more sequencing devices or a facility that uses one or more sequencing devices to generate nucleic acid (e.g., DNA or RNA) sequence data of biological samples. In some aspects, data collection component 10 may be a database that receives sequencing information generated from one or more sequencing devices. Any suitable liquid or solid biological samples may be used for sequencing. In some embodiments, a biological sample may be cell-based, for example, one or more types of tissue. In some embodiments, a biological sample may be a sample that includes cell-free nucleic acid fragments. Examples of biological samples include, but are not limited to, a blood sample (e.g., a cell-free DNA (cfDNA) sample, a cell-free RNA (cfRNA) sample, a serum sample, a plasma sample, a whole blood sample), a urine sample, a saliva sample, a tissue sample, a bone marrow sample, etc. Further, although sequencing of DNA from these samples is discussed herein, RNA from these samples may alternatively or additionally be sequenced.

[0051] Examples of sequencing data may include, but are not limited to, sequence read data of targeted genomic locations, partial or whole genome sequencing data of the genome represented by nucleic acid fragments in cell-free or cell-based samples, partial or whole genome sequencing data including one or more types of epigenetic modifications (e.g., methylation), or combinations thereof.

[0052] Data acquired by the data collection component 10 may be transferred to database 20 via network 40 or a local or network connection. In some embodiments, data collection component 10 may alternatively receive data from one or more sequencing devices. In some embodiments, the collected data may be analyzed by data intelligence component 30, via network 40 or a local or network connection. FIG. 1 B depicts exemplary functional modules that may be implemented to perform tasks of data intelligence component 30.

[0053] FIG. 2B depicts an exemplary computer system 210 for enhancing the performance of a diagnostic classifier for targeted subpopulations. Exemplary system 210 achieves such functionalities by implementing, on one or more computer devices, user input and output (I / O) module 220, memory or database 230, data processing module 240, data analysis module 250, classification module 260, network communication module 270, and any other functional modules that may be needed for carrying out a particular task (e.g., an error correction or compensation module, a data compression module, etc.). As disclosed herein, user I / O module 220 may further include an input sub-module, such as a keyboard, and an output submodule, such as a display (e.g., a printer, a monitor, or a touchpad). In some embodiments, all functionalities may be performed by one computer system. In some embodiments, the functionalities are performed by more than one computer system.

[0054] Also disclosed herein, a particular task may be performed by implementing one or more functional modules. In particular, each of the enumerated modules itself may, in turn, include multiple sub-modules. For example, data processing module 240 may include a sub-module for data quality evaluation (e.g., for discarding very short sequence reads or sequence reads including obvious errors), a sub-module for normalizing numbers of sequence reads that align to different regions of a reference genome, a sub-module to compensate / correct guanine-cytosine (GC) biases, a sub-module for matching data associated with a cancer sample with other data associated with one or more non-cancer samples, etc.

[0055] In some embodiments, a user may use I / O module 220 to manipulate data that is available either on a local device or can be obtained via a network connection from a remote service device or another user device. For example, I / O module 220 may allow a user, e.g., via a keyboard, a mouse, or a touchpad, to initiate or perform data analysis via a graphical user interface (GUI). In some embodiments, a user may manipulate data via voice control. In some embodiments, user authentication may be required before a user is granted access to the data being requested. In some embodiments, user I / O module 220 may be used to manage various functional modules. For example, a user may request via user I / O module 220 input data while an existing data processing session is in process. A user may do so by selecting a menu option or type in a command discretely without interrupting the existing process. In another example, a user may utilize user I / O module 220 to set various thresholds, configure sample matching settings, and / or provide other instructions to computer system 210 that dictate how data may be analyzed. As disclosed herein, a user may use any type of input to direct and control data processing and analysis via I / O module 220.

[0056] In some embodiments, system 210 further comprises a memory or database 230. In some embodiments, database 230 comprises a local database that may be accessed via user I / O module 220. In some embodiments, database 230 comprises a remote database that may be accessed by user I / O module 220 via network connection. In some embodiments, database 230 is a local database that stores data retrieved from another device (e.g., a user device or a server). In some embodiments, memory or database 230 may store data retrieved in real-time from internet searches. In some embodiments, database 230 may send data to and receive data from one or more of the other functional modules, including, but not limited to, a data collection module (not shown), data processing module 240, data analysis module 250, classification module 260, network communication module 270, and etc.

[0057] In some embodiments, database 230 may be a database local to the other functional modules. In some embodiments, database 230 may be a remote database that may be accessed by the other functional modules via wired or wireless network connection (e.g., via network communication module 270). In some embodiments, database 230 may include a local portion and a remote portion.

[0058] In some embodiments, system 210 comprises a data processing module 240. Data processing module 240 may receive data from I / O module 220 or database 230. In some embodiments, data processing module 240 may perform standard data processing algorithms, such as one or more of noise reduction, signal enhancement, normalization of counts of sequence reads, correction of GC bias, etc.In some embodiments, data processing module 240 may be configured to detect and measure methylation signatures, and specifically abnormal and / or differentially methylated features.

[0059] In some embodiments, system 210 comprises a data analysis module 250. In some embodiments, data analysis module 250 includes identifying and treating systematic errors in sequencing data, as described in connection with data processing module 240.

[0060] In some embodiments, system 210 comprises a classification module 260, which may embody a “machine-learning model” or “trained classifier.” As used herein, a “machine-learning model” or “trained classifier” generally encompasses instructions, data, and / or a model configured to receive input, and apply one or more of a weight, bias, classification, or analysis on the input to generate an output. The output may include, for example, a classification of the input, an analysis based on the input, a design, process, prediction, or recommendation associated with the input, or any other suitable type of output. A machine-learning model is generally trained using training data, e.g., experiential data and / or samples of input data, which are fed into the model in order to establish, tune, or modify one or more aspects of the model, e.g., the weights, biases, criteria for forming classifications or clusters, or the like. Aspects of a machine-learning model may operate on an input linearly, in parallel, via a network (e.g., a neural network), or via any suitable configuration. In some aspects, the machine-learning model may be trained on a combination of real and synthetic sample data.

[0061] The execution of the machine-learning model may include deployment of one or more machine-learning techniques, such as k-nearest neighbors, linear regression, logistic regression, random forest, gradient boosted machine (GBM), deep learning, a deep neural network, and / or any other suitable machine-learning technique that solves problems in the field of Natural Language Processing (NLP). Supervised, semi-supervised, and / or unsupervised training maybe employed. For example, supervised learning may include providing training data and labels corresponding to the training data, e.g., as ground truth. Unsupervised approaches may include clustering, classification or the like. K-means clustering or K-Nearest Neighbors may also be used, which may be supervised or unsupervised. Combinations of K-Nearest Neighbors and an unsupervised cluster technique may also be used. Any suitable type of training may be used, e.g., stochastic, gradient boosted, random seeded, recursive, epoch or batch-based, etc.

[0062] In an exemplary use case, a machine-learning model may be trained to analyze data from a test sample from a test subject whose status with respect to a medical condition is unknown and subsequently classifies the unknown test sample from the test subject based on the likelihood of the subject fitting into a particular category. In some embodiments, the one or more parameters may include a binomial probability score that is calculated based on logistic regression analysis. As disclosed herein, the binomial probability score may correspond to the likelihood of a subject having a certain medical condition, such as cancer. For example, a score of over a predefined threshold may indicate that the subject associated with a test sample is more likely to have cancer than not have cancer. In some embodiments, the one or more parameters may include a sequencing or methylation data distribution pattern correlating with the presence of cancer. A subject associated with a test sample having sequencing or methylation data with a pattern resembling the cancer pattern to a sufficient degree may be predicted as having cancer. In some embodiments, a sequencing or methylation data distribution pattern may be identified in connection with a specific type of cancer, determining a tissue of origin or cancer signal origin, thus allowing a test sample to be classified as indicative of a certain cancer type.

[0063] As disclosed herein, network communication module 270 may be used to facilitate communications between a user device, one or more databases, and any other suitable system or device through a wired or wireless network connection. Any communication protocol / device may be used, including, without limitation, a modem, an Ethernet connection, a network card (wireless or wired), an infrared communication device, a wireless communication device, and / or a chipset (such as a Bluetooth™ device, an 802.11 device, a WiFi device, a WiMax device, cellular communication facilities, etc.), a near-field communication (NFC), a Zigbee communication, a radio frequency (RF) or radio-frequency identification (RFID) communication, a PLC protocol, a 3G / 4G / 5G / LTE based communication, and / or the like. For example, a user device having a user interface platform for processing / analyzing CHIP-related methylation signature data may communicate with another user device with the same platform, a regular user device without the same platform (e.g., a regular smartphone), a remote server, a physical device of a remote loT local network, a wearable device, a user device communicably connected to a remote server, and etc.

[0064] The functional modules described herein are provided by way of example. It will be understood that different functional modules may be combined to create different utilities. It will also be understood that additional functional modules or sub-modules may be created to implement a certain utility.

[0065] Referring now to FIG. 3, an exemplary workflow 300 is provided for enhancing the performance of a diagnostic classifier in targeted subpopulations using domain adaptation techniques, specifically through importance weighting methodologies such as KMM. Aspects of the exemplary workflow 300 may be performed in accordance with some or all components described in FIG. 2A and 2B.

[0066] At step 305, system 200 may receive data associated with source and target datasets. In an aspect, the source dataset may be composed of labeled data samples used to train a model. These samples are typically collected from a specific domain and exhibit a particular distribution of features and labels. A non-limiting example of a source dataset may be the Circulating Cell-free Genome Atlas 2 (CCGA2) dataset, which encompasses a wide array of cancer types and tumor burdens. This dataset may serve as the source domain and provide the initial training samples for the classifier.

[0067] Alongside the source dataset, system 200 may also receive a target dataset. In an aspect, the target dataset, or target subpopulation dataset, may be a subset of data that the machine learning model is ultimately deployed on. In some cases, the target dataset may differ significantly from the source dataset used for training. In an aspect, the target dataset may be characterized by features that may not be fully represented in the source dataset. This can include differences in feature distributions, data collection conditions, or even the presence of new patterns or anomalies that were not seen during training. For instance, in the context of MRD detection, the target dataset may consist of samples with low tumor burdens or specific types of cancer such as non-Hodgkin’s lymphoma. Additionally to the foregoing, training a classifier directly on the target dataset may not be feasible or practical in certain contexts. For instance, in an aspect, the target dataset may contain limited data. More particularly, the target dataset may be much smaller compared to the source dataset, and there may not be enough target samples to train a robust classifier. In some instances, medical data, especially for specific conditions or subpopulations, may be rare and expensive to collect, thereby making it challenging to gather a sufficient number of target samples for training. In anotheraspect, an existing classifier may already be trained on a large, diverse source dataset, which may include a broad spectrum of cases, providing a robust foundation for the classifier. Retraining a classifier from scratch on a smaller target dataset would not leverage the valuable information and patterns learned from the extensive source dataset.

[0068] Once the data is collected, one or more preprocessing steps may be taken to promote consistency and comparability. For instance, the data values may be normalized or scaled to a common scale without distorting differences in the ranges of values (e.g., often transforming them to a range of [0,1] or [-1 ,1]). This may help in mitigating the impact of varying data scales on the classifier’s performance. Additionally or alternatively, preprocessing may involve identifying and addressing missing data in the datasets. Techniques such as imputation (e.g., where missing values are filled in using statistical methods) or exclusion (e.g., where missing values are removed) may be employed based on the extent and / or nature of the missing data. Additionally or alternatively, for classification tasks, the class distribution may be highly imbalanced (e.g., a large number of non-cancer samples compared to cancer samples). In these situations, the dataset may be balanced through techniques such as oversampling, undersampling, or synthetic data generation. Balancing the dataset may help in preventing the classifier from being biased towards the majority class. Additionally or alternatively, in some cases, the dataset may be augmented with additional synthetic samples (e.g., generating synthetic data for rare and / or unavailable cancer types in the target subpopulation may enhance the classifier’s ability to generalize). Additionally or alternatively, one or more dimensionality reduction techniques may be employed (e.g., principal component analysis (PCA)) to reduce the data dimensionality while still retaining importantinformation. Additionally or alternatively, the datasets may be split into training, validation, and test sets. The training set may be used to train the classifier, the validation set may be used to tune hyperparameters and prevent overfitting, and the test set may be used to evaluate the final performance of the adapted classifier.

[0069] At step 310, a feature selection process may be employed to identify and select the most relevant features from the source and target datasets. In the context of sequencing data, this could mean identifying specific methylation sites, gene expression levels, or other biomarkers indicative of cancer presence. These features capture the underlying patterns and structures within the data, making it easier to apply machine learning algorithms. Ultimately, eliminating redundant or irrelevant features may reduce the amount of noise in the data and prevent overfitting. In an aspect, various feature selection techniques may be employed to choose the most relevant features (e.g., this may involve using PCA or other dimensionality reduction methods to focus on the most informative features.).

[0070] At step 315, system 200 may receive an indication of a selected kernel function based on the selected features in step 310. In machine learning contexts, a kernel is a mathematical function used to compute the similarity or relationship between two data points in a transformed feature space, without explicitly performing the transformation. Accordingly, kernels enable the modeling of complex, non-linear relationships by transforming the data into a higher-dimensional space where linear methods may be applied. For instance, given a set of data points that are not linearly separable in the original space, by applying a kernel function, these points are mapped to a higher-dimensional space where they become linearly separable. This enables the mapping to be done implicitly, meaning the actual coordinates in the higher-dimensional space are never explicitly computed. Instead,the kernel function directly computes the similarity (dot product) in this highdimensional space. Common types of kernels may include a linear kernel, polynomial kernel, Gaussian Radial Basis Function (RBF) Kernel, Sigmoid Kernel, and the like. In an aspect, each kernel has its own characteristics and suitability depending on the nature of the data. For high-dimensional data like methylation signatures in cfDNA, the Gaussian kernel may be preferred due to its ability to handle non-linear relationships and its flexibility in adjusting the kernel width parameter. Accordingly, the remainder of the discussion is facilitated with the kernel being a Gaussian RBF kernel. However, such a designation is non-limiting, and other types of kernels may be utilized.

[0071] At step 320, system 200 may utilize the selected kernel function to project the feature data into a new feature space where the distributions of the source and target domains may be more effectively aligned. In an aspect, once the kernel is selected, the next step is to project the data into the kernel-induced feature space. This transformation may involve computing the kernel matrix, which contains the pairwise similarities between all samples in the source and target datasets. The goal is to represent the data in a space where the distributions of the source and target samples may be aligned more effectively. In the kernel-induced feature space, the KMM technique aims to adjust the distribution of the source samples in a feature space so that the mean of the distribution of the source samples matches the mean of the distribution of the target samples. Matching the distributions of the source samples and the target samples is a convex optimization problem. This alignment facilitates correcting of the covariate shift and ensuring that the classifier trained on the source data generalizes well to the target data.

[0072] At step 325, system 200 may calculate importance weights to rebalance the source dataset to better reflect the target dataset. In an aspect, importance weights quantify how much more likely a given sample from the source domain is to appear in the target domain. Mathematically, the importance weight for a sample is the ratio of the probability density of the target distribution to the source distribution at that sample’s location. These weights are used to adjust the influence of each training sample during the learning process, effectively aligning the source data distribution with the target data distribution and thereby mitigating the effects of covariate shift and improving the classifier’s performance on the target population. In an aspect, directly estimating these densities in high-dimensional space may be challenging and error-prone. Therefore, KMM offers a non-parametric approach to estimate these weights without explicit density estimation.

[0073] In view of Equation 1 below, KMM frames the problem of computing importance weights as an optimization problem, where the objective is to find weights 3 for each source sample x, such that the weighted mean of the source samples matches the mean of the target samples in the kernel-induced feature space. This may be formulated as a convex quadratic programming problem:Equation 1

[0074] Subject to the constraints *== i , whereO(x) represents the feature mapping induced by the kernel, x, are the source samples, Xjteare the target samples, ntris the number of source samples, nteis the number of target samples, and is a regularization parameter to prevent overfitting.In an aspect, to solve the optimization problem, an efficient numerical solver, such as a quadratic programming solver, may be employed. These solvers iteratively adjust the weights to minimize the different in the weighted means. The result is a set of importance weights that may be applied to the source samples during the training of the classifier.

[0075] At step 330, once the importance weights are computed, the classifier is then trained using the reweighted samples. More particularly, these weights are used to adjust the loss function, giving higher priority to samples that better represent the target distribution. More particularly, the classifier is trained by iteratively updating the model parameters to minimize the weighted loss function. Techniques such as stochastic gradient descent (SGD) may be adapted to incorporate the importance weights, thereby ensuring that each update step takes into account the relevance of each sample to the target domain. Accordingly, given the foregoing, the classifier is caused to pay more attention to the source samples that are more representative of the target domain. More specifically, as a result of the reweighted training, the classifier is better equipped to handle new, unseen data by focusing on samples more representative of the target domain. Additionally, this process reduces bias by adjusting the influence of training samples, thereby mitigating biases introduced by the source domain, and leading to more accurate predictions in the target domain. Furthermore, this process enhances robustness, making the newly trained model more resilient to variations between the training and target datasets, improving its real-world applicability.

[0076] In optional step 335, the adapted classifier’s performance may be evaluated using cross-validation techniques on one or both of synthetic and real test samples. In an aspect, the original available training dataset may be divided intodistinct subsets: a training subset, a validation subset, and a test subset. The training subset may be used to fit the model, while the validation subset may provide an unbiased evaluation during the training process, helping to select the best model parameters and prevent overfitting. The final evaluation may be conducted on the test subset, which may be held out from the training and validation phases, to assess the modefs performance on completely unseen data.

[0077] One or more known cross-validation techniques may be employed to enhance the reliability of the evaluation results. For instance, k-fold cross validation may be utilized, in which the dataset is divided into k equally-sized folds, and the model is trained and validated k times, each time using a different fold as the validation set and the remaining k-1 folds as the training set. The average performance across all folds may be considered, providing a more comprehensive assessment of the model’s generalization capabilities. This method may help in detecting overfitting and underfitting issues, thereby ensuring that the model’s performance is consistent across different subsets of the data.

[0078] In the context of high-dimensional sequencing data and importance weighting, the evaluation phase may also involve comparing the performance of the classifier trained with reweighted samples against one trained without reweighting. This comparison demonstrates the effectiveness of importance weighting in improving the model’s ability to generalize to the target domain. Metrics such as the reduction in prediction error and improved AUC-ROC scores may highlight the benefits of the reweighted training process.

[0079] In an aspect, validation techniques may include the use of holdout sets from different domains or conditions, such as synthetic data resembling potential future datasets. This approach may test the model’s adaptability to varyingconditions and ensures its robustness in diverse real-world scenarios. For example, in the context of a heme collaboration for classifying cancer types, the model may be evaluated on holdout sets composed of real and synthetic samples representing different tumor fractions. This may help to verify the model’s performance across a range of conditions it may encounter in actual deployment.

[0080] FIGS. 4A and 4B presents graph 400 and boxplot 405, respectively, that collectively illustrate the improvements that utilizing the KMM approach has relative to the conventional density ratio method.

[0081] Graph 400 illustrates how different models fit polynomial data of degree 1 using various methods: Ordinary Least Squares (OLS) and Weighted Ordinary Least Squares (WOLS). Stated differently, graph 400 demonstrates that different methods of weighting the training data influence the model’s fit to better align with the true function for the subpopulation of interest and the target data points. The blue circles represent the training samples used to fit the models. The red line shows the true underlying function that the models aim to approximate. The blue dashed line represents the model fit using OLS on the training samples. The red crosses represent the target data points where the model will eventually be deployed. For the WOLS fitting lines, the yellow dashed line shows the model fit using traditional ratio estimates where analytical densities are calculated (e.g., this method computes the ratio of test to training density but requires knowledge of the full likelihood), the green dashed line represents the model fit using KMM, and the pink dashed line represents a minimum importance criterion (IC) approach.

[0082] Graph 405 presents a box plot that shows the average performance of the three WOLS methods compared to OLS. Each box plot summarizes the distribution of the loss values. The yellow box represents the performance of theWOLS method that uses the ratio of test to training density. The green box represents the performance of the KMM approach. The purple box represents the performance of the WOLS method using the minimum importance criterion. The red box depicts the performance of the OLS method trained directly on the labeled test points. Graph 405 illustrates that the KMM approach exhibits better performance than the conventional importance weighting method.

[0083] In an aspect, a classifier trained using the concepts described above may be utilized to distinguish between cancerous and non-cancerous samples and to identify a cancer signal origin (CSO). More particularly, importance weights may be learned from currently available datasets that are expected to be similar to currently unavailable datasets that may be encountered in the future.

[0084] In an aspect, a classifier based on the concepts described above may be developed and validated specifically for heme cancers (such as those involving the blood, bone marrow, or lymph nodes) by using importance weighting to improve the classifier’s performance on a dataset that may have different characteristics than the training data. This classifier may be configured to distinguish between cancer and non-cancer cases, and identify the cancer signal origin (CSO). To address the differences between the training data and the target deployment data, importance weights may be calculated. These weights help adjust the model to perform better on data that may have different properties from the training data. In an aspect, the training data may include data associated with approximately 1200 heme cancers. Specifically, the data may be derived from non-Hodgkin’s lymphoma samples with low tumor fractions. The training data may additionally include approximately 10,000 non-cancer samples to help the classifier learn to differentiate between cancerous and non-cancerous cases. After training the classifier and applying importanceweighting, the classifier’s performance may be tested using additional data. This includes both real samples from one or more existing datasets and synthetic samples that are generated to simulate different conditions (e.g., such as tumor fraction ratios). FIG. 5 presents a diagram containing additional parameters associated with the training and validation datasets.

[0085] Referring now to FIG. 6, graph 600 is provided that presents a receiver operating characteristic (ROC) curve that illustrates the performance of two classifiers, a first classifier that is trained without applying importance weights and a second classifier that is trained with importance weights computed via KMM on an existing dataset. Each of these classifiers was trained on a mix of real and synthetic samples and was validated against an existing dataset. The true positive rate, which measures the proportion of actual positives correctly identified by the classifier, is plotted on the y-axis whereas the false positive rate, which measures the proportion of actual negatives incorrectly identified as positives by the classifier, is plotted on the x-axis. The dashed blue diagonal line represents the performance of a random classifier. Any point above this line indicates better performance than random guessing. The orange curve 62 is associated with the first classifier (which is trained without importance weights), and the green curve 64 is associated with the second classifier (which is trained with importance weights). Examination of graph 600 reveals that the area under the curve (AUG) for both the unweighted and weighted classifiers are almost identical, with each having an ROC of 0.78. This indicates that applying importance weights via KMM does not significantly change the classifier’s performance.

[0086] Referring now to FIG. 7A, graph 700 is provided that presents theROC curves for an unweighted 72 and weighted 74 classifier when tested onsynthetic samples. The ROC curves for each classifier are almost identical, indicating that the importance weighting did not significantly affect the classifier’s performance on synthetic samples. Both curves show a high true positive rate across various false positive rates, suggesting good performance. Further enforcing the foregoing, the AUC for the weighted and unweighted classifier is 0.78.

[0087] Referring now to FIG. 7B, graph 705 is provided that presents the ROC curves for an unweighted 76 and weighted 78 classifier when tested on real samples. Here, the unweighted classifier performs better than the weighted classifier on real samples, as indicated by the higher AUC (i.e., unweighted AUC is 0.8 where weighted AUC is 0.77). Ultimately, in consideration of the foregoing, the classifier improves performance slightly on synthetic samples with importance weighting, but exhibits performance decline on real samples. This trade-off suggests that the classifier is more focused on the synthetic samples due to importance weighting, which might be beneficial for specific tasks or datasets.

[0088] Referring now to FIG. 8, data table 800 is provided that shows performance results for an unweighted classifier trained on the CCGA2 dataset, and data table 805 is provided that shows performance results for a weighted classifier trained on the CCGA2 dataset but with applied importance weights. Each data table provides details on the actual labels of the samples, the sum of true positives, the count of total samples, and the percentage of correctly classified samples for each category. The importance-weighted classifier shows improvements in specific areas where it has been instructed to focus, such as non-Hodgkin lymphoma samples, at the cost of performance in other areas, such as myeloid samples. This trade-off highlights the effect of importance weighting, which shifts the classifier’s attention towards more critical or representative cases of the target domain.

[0089] Referring now to FIG. 9, data tables 900 and 905 are provided. Data tables 905 and 910 compare the performance of the Heme MRD baseline classifier and the importance-weighted classifier on real samples from the CCGA2 dataset, which are broken down by tumor methylated fraction (TMeF) categories. The importance-weighted classifier shows a slight improvement in the target tumor fraction category (i.e. , 1e-4in row 1), which was designated as the focus. However, this improvement comes at the cost of reduced performance for samples with unknown tumor fractions (row 5). The performance in other categories remained relatively stable, indicating that the importance weighted classifier effectively targeted the desired tumor fraction without drastically altering the performance in other categories.

[0090] Referring now to FIG. 10, data tables 1000 and 1005 are provided. Data tables 1000 and 1005 compare the performance of the Heme MRD baseline classifier and the importance-weighted classifier on real non-Hodgkin’s lymphoma samples from the CCGA2 dataset, broken down by TMeF categories. The importance-weighted classifier demonstrates a significant improvement in the target tumor fraction category (i.e., 1 e'4in row 1 ) for non-Hodgkin’s lymphoma samples, showing a higher accuracy compared to the baseline. Performance in other tumor fraction categories remained stable, indicating that the importance-weighted classifier generally did not lose accuracy in those categories, except for a small decrease in the smallest tumor fraction category (i.e., as shown in row 0).

[0091] Referring now to FIG. 11 , data tables 1100 and 1105 are provided.Data tables 1100 and 1105 compare the performance of the Heme MRD baseline classifier and the importance-weighted classifier on synthetic cross-validation samples from the CCGA2 dataset, broken down by TMeF categories. Theimportance-weighted classifier demonstrates a significant improvement in the target tumor fraction category (i.e., 1e-4in row 1) for synthetic samples, showing higher accuracy compared to the baseline. Performance in other tumor fraction categories also shows slight improvements or remains stable, indicating that the importance weighting effectively targeted the desired category without sacrificing accuracy in other categories.

[0092] Referring now to FIG. 12, data tables 1200 and 1205 are provided. Data tables 1200 and 1205 compare the performance of the Heme MRD baseline classifier and the importance-weighted classifier on synthetic cross-validation samples from the CCGA2 dataset, broken down by TMeF categories. The importance-weighted classifier gains approximately 5% sensitivity in the target tumor fraction category (i.e., 1 e-4in row 1) on synthetic cross-validation samples. This indicates that the importance-weighted classifier is better at detecting cancer in these difficult-to-detect samples compared to the baseline classifier. Performance in other tumor fraction categories also shows slight improvements or remains stable, indicating that the importance weighting effectively targeted the desired category without sacrificing accuracy in other categories.

[0093] Referring now to FIG. 13, data tables 1300 and 1305 are provided. Data tables 1305 and 1305 compare the performance of the Heme MRD baseline classifier and the importance-weighted classifier on CCGA3 holdout synthetic samples, broken down by TMeF categories. The importance-weighted classifier demonstrates a substantial sensitivity boost (-17%) in the target TMeF (e.g., 1e-4) on CCGA3 holdout synthetic samples. This improvement highlights the effectiveness of the importance weighting approach in enhancing the classifier’s ability to detect low-tumor-fraction samples in the specific TMeF range targeted for improvement.Performance in other tumor fraction categories also shows slight improvements or remains stable, indicating that the importance weighting effectively targeted the desired category without sacrificing accuracy in other categories.

[0094] Referring now to FIG. 14, data tables 1400 and 1405 are provided. Data tables 1400 and 1405 compare the performance of the Heme MRD baseline classifier and the importance-weighted classifier specifically on non-Hodgkin’s lymphoma synthetic samples, broken down by TMeF categories. The importance- weighted classifier demonstrates substantial improvements in performance, especially in the target TMeF category (e.g., 1 e-4) for non-Hodgkin’s lymphoma synthetic samples. This -30% increase in sensitivity highlights the classifier’s enhanced ability to detect low-tumor-fraction samples, an important improvement for effective cancer detection.

[0095] Referring now to FIG. 15, data tables 1500 and 1505 are provided. Data tables 1500 and 1505 compare the performance of the Heme MRD baseline classifier and the importance-weighted classifier specifically on non-Hodgkin’s lymphoma synthetic samples, broken down by TMeF categories. These tables illustrate a trade-off where the importance-weighted classifier experiences a loss in performance for non-target CSOs in the target MVAF regime.

[0096] Referring now to FIGS. 16A and 16B, plots 1600 and 1605 are presented that display the importance weights learned by the KMM algorithm for heme cancer samples. These weights indicate how much emphasis the classifier assigns to different samples, particularly focusing on non-Hodgkin’s lymphoma samples within a specific molecular variant allele-fraction (MVAF) range. Plot 1600 inFIG. 16A presents importance weights by real or synthetic samples. The blue dots represents synthetic samples. These are primarily non-Hodgkin’s lymphomasamples in the target MVAF regime (between 2e-5 and 1 e-4). These samples are assigned higher weights, indicating the classifier’s focus on them. The orange dots represent real samples. The real non-Hodgkin’s lymphoma samples also receive higher weights, reflecting the classifier’s focus on samples similar to the synthetic target dataset.

[0097] Plot 1605 in FIG. 16B presents importance weights by sample type. The green dots represent non-Hodgkin’s lymphoma samples. These samples, both synthetic and real, are assigned higher weights, particularly in the target MVAF regime (between 2e-5 and 1e-4). The upweighting of these samples demonstrates the classifier’s focus on the desired target profile. The other colors represent other sample types such as myeloid samples. These samples generally receive lower weights, indicating the classifier’s reduced focus on these non-target sample types. Collectively, the plots demonstrates how the importance weights learned by KMM influence the classifier’s focus. Non-Hodgkin’s lymphoma samples, both synthetic and real, within the target MVAF regime, receive higher weights, resulting in better performance for these samples. Conversely, non-target samples are downweighted, leading to decreased performance for other cancer types.

[0098] Referring now to FIG. 17, plot 1700 illustrates how the classifier assigns weights to non-cancer samples, focusing on those that are closer to the target MVAF regime. The blue dots represent non-cancer samples. These samples are plotted based on their Iog10 MVAF values and the weights assigned by the classifier. The classifier upweights non-cancer samples with intermediate tumor fractions and those closer to the target MVAF regime, ensuring that it focuses on samples most likely to resemble non-Hodgkin’s lymphoma. In effect, this helps tomaintain overall performance by emphasizing relevant negative examples, explaining the observed balance in sensitivity and specificity.

[0099] Referring now to FIGs. 18A and 18B, graphs 1800 and 1805 are presented that demonstrate the performance of the classifier using importance weighting with synthetic samples, specifically focusing on non-Hodgkin’s samples in the 1e-4 MVAF regime. Graph 1800 illustrates how the classifier performs in cross- validation samples, and graph 1805 illustrates how the classifier performs in hold-out samples. Turning first to graph 1800 in FIG. 18A, which focuses on the CCGA2 cross-validation samples, the orange line 182 represents the performance of the classifier without applying importance weights. The AUG is 0.66, indicating the classifier’s ability to distinguish between true positives and false positives without weighting adjustments. The green line 184 represents the performance of the classifier with importance weights applied. The AUG is 0.65, which is slightly lower than the unweight AUC, showing a marginal trade-off in overall performance due to the application of weights. Turning next to graph 1805 in FIG. 18B, which focuses on the CCGA3 synthetic hold-out samples, the orange line 186 represents the performance of the classifier without applying importance weights. The AUC is 0.67, reflecting its baseline capability to differentiate between true positives and false positives. The green line 188 represents the performance of the classifier with importance weights applied. The AUC is 0.67, indicating that the importance weighting has maintained the classifier’s performance in this holdout sample.

[0100] The experimental results described above indicate that importance weighted classifiers may be used to selectively increase performance in a given target regime. The techniques described herein may be used to adapt general classifiers trained on large datasets to specific subpopulations given examples of thespecific subpopulation. These methods may further be generalized to multi-omics datasets. KMM and other importance weighting methodologies may be leveraged for adapting classifiers trained on a general set of data to more specific subpopulations of interest.

[0101] Referring now to FIG. 19, an exemplary workflow 1900 is provided for utilizing a target dataset for learning importance weights. Aspects of the exemplary workflow 1900 may be performed in accordance with some or all components described in FIG. 2A and 2B.

[0102] At steps 1905 and 1910, raw input from source samples (e.g., CCGA fragment files) and raw input from target samples (e.g., MRD fragment files), respectively, may be received at system 200. The source samples provide the initial data that may be used to extract features and form the basis for training the classifier. The target samples provide understanding of the specific characteristics and features of the data that the classifier will eventually need to work with in the real-world application.

[0103] At step 1915, feature extraction may occur that involves identifying relevant features from the collected data. In the context of genetic sequencing data, this could mean identifying specific methylation sites, gene expression levels, or other biomarkers indicative of disease, e.g., cancer, presence in the target samples population. After extracting the features, feature selection techniques may be employed to choose the most relevant features that contribute significantly to the classifier’s performance. This may involve using PCA or other dimensionality reduction methods to focus on the most informative subset of features. Accordingly, the feature extraction process at step 1915 may be utilized to identify source set features 1920 (which are features that represent the various aspects of the sourcedata that will be used to train the classifier) and target data set features 1925 (which are utilized to understand the specific characteristics of the target data, which the classifier needs to accurately predict or classify samples in the target group).

[0104] At step 1930, importance weights may be computed for samples in the source set. In an aspect, the importance weights may be computed using KMM (e.g., in which a selected kernel function, such as an RBF Gaussian kernel, is applied to transform the original feature space into a higher-dimensional space). Specifically, the importance weights may be computed by comparing the source data distribution with the target data distribution. The weights are then applied, at step 1935, to each data sample of the source training samples, with samples more relevant to the target distribution receiving higher weights, while less relevant samples receive lower weights.

[0105] At step 1940, a classifier may then be trained using the reweighted samples from the source dataset, incorporating the computed importance weights into the training process. In an aspect, once the classifier is trained with the importance weights, the trained classifier may be applied to sequencing data of unknown samples to identify whether or not subjects associated with those unknown samples have a target disease (e.g., cancer, or a specific type of cancer). This approach emphasizes samples that are more representative of the target deployment scenario, enhancing the classifier’s performance in the desired context. In an aspect, the performance of the diagnostic classifier may be evaluated (e.g., by utilizing a validation dataset). The classifier performance may be evaluated subsequent to classifier training or substantially during the training process.

[0106] Referring now to FIG. 20, an exemplary workflow 2000 is provided for utilizing a target dataset for learning importance weights and training a classifier.Aspects of the exemplary workflow 2000 may be performed in accordance with some or all components described in FIG. 2A and 2B.

[0107] At steps 2005 and 2010, raw input from source samples (e.g., CCGA fragment files) and raw input from target samples (e.g., MRD fragment files), respectively, may be received at system 200. The source samples provide the initial data that may be used to extract features and form the basis for training the classifier. The target samples provide understanding of the specific characteristics and features of the data that the classifier will eventually need to work with in the real-world application.

[0108] At step 2015, feature extraction may occur that involves identifying relevant features from the collected data. In the context of genetic sequencing data, this could mean identifying specific methylation sites, gene expression levels, or other biomarkers indicative of disease, e.g., cancer, presence. After extracting the features, feature selection techniques may be employed to choose the most relevant features that contribute significantly to the classifier’s performance. This may involve using PCA or other dimensionality reduction methods to focus on the most informative subset of features. Accordingly, the feature extraction process at step 2015 may be utilized to identify source set features 2020 (which are features that represent the various aspects of the source data that will be used to train the classifier) and target data set features 2025 (which are utilized to understand the specific characteristics of the target data, which the classifier needs to accurately predict or classify samples in the target group).

[0109] At step 2030, importance weights may be computed for samples in both the source set and the target dataset. This involves determining how representative each sample is relative to the target scenario. The importanceweights help to adjust the influence of each sample during the training process to better align with the target distribution. In an aspect, the importance weights may be computed using KMM (e.g., in which a selected kernel function, such as an RBF Gaussian kernel, is applied to transform the original feature space into a higherdimensional space). For each sample in the source set, an importance weight is assigned, at step 2035, which indicates the relevance of each source sample to the target scenario, ensuring that more relevant samples have a greater impact during training. Similarly, for each sample in the target dataset, an importance weight is assigned, at step 2040, to ensure that the classifier gives appropriate consideration to the characteristics of the target data during training.

[0110] At step 2045, a classifier may then be trained using both the source and target samples, with the loss function weighted by the computed importance weights. This means that each sample’s contribution to the training process is adjusted based on its importance weight, helping the classifier to learn more effectively and accurately for the specific target application. In an aspect, the performance of the diagnostic classifier may be evaluated (e.g., by utilizing a validation dataset). The classifier performance may be evaluated subsequent to classifier training or substantially during the training process.

[0111] Referring now to FIG. 21 , an exemplary workflow 2100 is provided for leveraging domain adaptation through feature augmentation. More particularly, workflow 2100 describes a domain adaptation method where the feature matrices are augmented differently based on whether they originate from source or target datasets. This approach is distinct from importance weighting, offering a flexible way for the classifier to adapt to domain-specific characteristics without reweightingindividual samples. Aspects of the exemplary workflow 2100 may be performed in accordance with some or all components described in FIG. 2A and 2B.

[0112] At steps 2105 and 2110, raw input from source samples (e.g., CCGA fragment files) and raw input from target samples (e.g., MRD fragment files), respectively, may be received at system 200. The source samples provide the initial data that may be used to extract features and form the basis for training the classifier. The target samples provide understanding of the specific characteristics and features of the data that the classifier will eventually need to work with in the real-world application.

[0113] At step 21 15, feature extraction may occur that involves identifying relevant features from the collected data. In the context of genetic sequencing data, this could mean identifying specific methylation sites, gene expression levels, or other biomarkers indicative of the disease, e.g., cancer, presence. After extracting the features, feature selection techniques may be employed to choose the most relevant features that contribute significantly to the classifier’s performance. This may involve using PCA or other dimensionality reduction methods to focus on the most informative features. Accordingly, the feature extraction process at step 2115 may be utilized to identify source set features 2120 (which are features that represent the various aspects of the source data that will be used to train the classifier) and target data set features 2125 (which are utilized to understand the specific characteristics of the target data, which the classifier needs to accurately predict or classify samples in the target group).

[0114] At step 2130, system 200 may perform feature augmentation based on source and target labels. More particularly, instead of reweighting samples, this method augments the feature matrix to allow the classifier to perform domainadaptation independently. In an aspect, the augmentation process may vary for the source and target features. With respect to the former, the original feature matrix of the source dataset may be expanded to three times its original size (e.g., the original feature matrix, which may have approximately 20,000 features, is tripled in size to 60,000 features). Thereafter, the original source features may be copied into the first third of the expanded matrix. This part remains unchanged and represents the source features as they are. The original source features may also be copied into the second third of the expanded matrix. The last third of the expanded matrix may be filled with zeros, which indicate that these features do not correspond to the target dataset. A similar process may be employed for the target features. More particularly, the original feature matrix of the target dataset may be expanded to three times its original size. The original target features may be copied into the first third of the expanded matrix. The second third of the expanded matrix may be filled with zeros (e.g., indicating that these features do not correspond to the source dataset). The final third of the expanded matrix may include a copy of the original target features. This augmentation strategy effectively triples the size of the feature matrix, with different sections representing the source and target features differently.

[0115] At step 2135, the augmented feature set for the source samples may be created, incorporating the modified structure explained in step 2130. This augmented set now contains both the original and augmented features tailored for the source data. At step 2140, the target set features may be augmented according to the same structured augmentation process. This augmented set includes both the original and augmented features specific to the target samples.

[0116] At step 2145, a classifier may be trained using the feature-augmented input matrices from both the source and target datasets. The classifier utilizes theseaugmented features to learn how to adjust the importance of different features based on the domain (source or target), thereby improving its ability to adapt and perform well across different datasets. In an aspect, the performance of the diagnostic classifier may be evaluated (e.g., by utilizing a validation dataset). The classifier performance may be evaluated subsequent to classifier training or substantially during the training process.

[0117] In general, any process discussed in this disclosure that is understood to be computer-implementable may be performed by one or more processors of a computer system, such as system environment 210, as described above. A process or process step performed by one or more processors may also be referred to as an operation. The one or more processors may be configured to perform such processes by having access to instructions (e.g., software or computer-readable code) that, when executed by the one or more processors, cause the one or more processors to perform the processes. The instructions may be stored in a memory of the computer server. A processor may be a central processing unit (CPU), a graphics processing unit (GPU), or any suitable types of processing unit.

[0118] A computer system, such as system environment 210, may include one or more computing devices. If the one or more processors of the computer system are implemented as a plurality of processors, the plurality of processors may be included in a single computing device or distributed among a plurality of computing devices. If a system environment comprises a plurality of computing devices, the memory of the computer system may include the respective memory of each computing device of the plurality of computing devices.

[0119] FIG. 22 is a simplified functional block diagram of a computer system 2200 that may be configured as a computing device for executing the processes described herein, according to exemplary embodiments of the present disclosure. FIG. 22 is a simplified functional block diagram of a computer that may be configured according to exemplary embodiments of the present disclosure. In various embodiments, any of the systems herein may be an assembly of hardware including, for example, a data communication interface 2220 for packet data communication. The platform also may include a central processing unit (“CPU”) 2202, in the form of one or more processors, for executing program instructions. The platform may include an internal communication bus 2208, and a storage unit 2206 (such as ROM, HDD, SDD, etc.) that may store data on a computer readable medium 2222, although the system 2200 may receive programming and data via network communications via electronic network 2225 (e.g., voice, video, audio, images, or any other data over the electronic network 2225). The system 2200 may also have a memory 2204 (such as RAM) storing instructions 2224 for executing techniques presented herein, although the instructions 2224 may be stored temporarily or permanently within other modules of system 2200 (e.g., processor 2202 and / or computer readable medium 2222). The system 2200 also may include input and output ports 2212 and / or a display 2210 to connect with input and output devices such as keyboards, mice, touchscreens, monitors, displays, etc. The various system functions may be implemented in a distributed fashion on a number of similar platforms, to distribute the processing load. Alternatively, the systems may be implemented by appropriate programming of one computer hardware platform.

[0120] In this disclosure, the term “based on” means “based at least in part on.” The singular forms “a,” “an,” and “the” include plural referents unless thecontext dictates otherwise. The term “exemplary” is used in the sense of “example” rather than “ideal.” The terms “comprises,” “comprising,” “includes,” “including,” or other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, or product that comprises a list of elements does not necessarily include only those elements, but may include other elements not expressly listed or inherent to such a process, method, article, or apparatus. Relative terms, such as “about,” “approximately,” “substantially,” and “generally,” are used to indicate a possible variation of ±10% of a stated or understood value. In addition, the term “between” used in describing ranges of values is intended to include the minimum and maximum values described herein. The use of the term “or” in the claims and specification is used to mean “and / or” unless explicitly indicated to refer to alternatives only if the alternatives are mutually exclusive, although the disclosure supports a definition that refers to only alternatives and “and / or.” As used herein “another” may mean at least a second or more.

[0121] As used herein, the term “user” generally encompasses any person or entity, such as a researcher and / or a care provider (e.g., a doctor, etc.), that may desire information, resolution of an issue, or engage in any other type of interaction with a provider of the systems and methods described herein (e.g., via an application interface resident on their electronic device, etc.). The term “electronic application” or “application” may be used interchangeably with other terms like “program,” or the like, and generally encompasses software that is configured to interact with, modify, override, supplement, or operate in conjunction with other software.

[0122] Program aspects of the technology may be thought of as “products” or “articles of manufacture” typically in the form of executable code and / or associated data that is carried on or embodied in a type of machine-readablemedium. “Storage” type media include any or all of the tangible memory of the computers, processors or the like, or associated modules thereof, such as various semiconductor memories, tape drives, disk drives and the like, which may provide non-transitory storage at any time for the software programming. All or portions of the software may at times be communicated through the Internet or various other telecommunication networks. Such communications, for example, may enable loading of the software from one computer or processor into another, for example, from a management server or host computer of the mobile communication network into the computer platform of a server and / or from a server to the mobile device. Thus, another type of media that may bear the software elements includes optical, electrical and electromagnetic waves, such as used across physical interfaces between local devices, through wired and optical landline networks and over various air-links. The physical elements that carry such waves, such as wired or wireless links, optical links, or the like, also may be considered as media bearing the software. As used herein, unless restricted to non-transitory, tangible “storage” media, terms such as computer or machine “readable medium” refer to any medium that participates in providing instructions to a processor for execution.

[0123] Furthermore, while some embodiments described herein include some but not other features included in other embodiments, combinations of features of different embodiments are meant to be within the scope of the invention, and form different embodiments, as would be understood by those skilled in the art. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0124] Thus, while certain embodiments have been described, those skilled in the art will recognize that other and further modifications may be made theretowithout departing from the spirit of the invention, and it is intended to claim all such changes and modifications as falling within the scope of the invention. For example, functionality may be added or deleted from the block diagrams and operations may be interchanged among functional blocks. Steps may be added or deleted to methods described within the scope of the present invention.

[0125] The above disclosed subject matter is to be considered illustrative, and not restrictive, and the appended claims are intended to cover all such modifications, enhancements, and other implementations, which fall within the true spirit and scope of the present disclosure. Thus, to the maximum extent allowed by law, the scope of the present disclosure is to be determined by the broadest permissible interpretation of the following claims and their equivalents, and shall not be restricted or limited by the foregoing detailed description. While various implementations of the disclosure have been described, it will be apparent to those of ordinary skill in the art that many more implementations are possible within the scope of the disclosure. Accordingly, the disclosure is not to be restricted except in light of the attached claims and their equivalents.

Claims

WHAT IS CLAIMED IS:

1. A computer-implemented method, the computer-implemented method comprising: receiving, at a system, a source dataset comprising a first set of data samples; receiving, at the system, a target dataset comprising a second set of data samples; performing, using a processor associated with the system, feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; computing, based on the performing and via employing a non-parametric importance-weighting technique, importance weights for at least the first set of data samples; assigning, using the processor, the computed importance weights to at least the first set of data samples; and training, using the processor, a diagnostic classifier using at least the importance-weighted first set of data samples.

2. The computer-implemented method of claim 1 , wherein the first set of data samples encompasses a plurality of disease types.

3. The computer-implemented method of claim 2, wherein the plurality of disease types are associated with cancer.

4. The computer-implemented method of claim 1 , wherein the target dataset represents a targeted subpopulation of disease types.

5. The computer-implemented method of claim 4, wherein the targeted subpopulation of disease types is at least one of: non-Hodgkin lymphoma, Hodgkin lymphoma, myeloid cancer, plasma cell disorders, circulating lymphoma conditions, unspecified conditions, and non-cancer conditions.

6. The computer-implemented method of claim 1 , wherein the computing the importance weights via employing the non-parametric importance-weighting technique comprises employing a Kernel Mean Matching (KMM) technique.

7. The computer-implemented method of claim 6, wherein the KMM technique comprises: selecting a kernel function to project the first set of features and the second set of features into a high-dimensional feature space; adjusting a first distribution of the first set of features in the highdimensional feature space to match a second distribution of the second set of features; and determining the importance weights that minimize a difference between means of the first distribution of the first set of features and the second distribution of the second set of features.

8. The computer-implemented method of claim 7, wherein the kernel function is a Gaussian radial basis function (RBF).

9. The computer-implemented method of claim 1 , wherein the training comprises reweighting a loss function of the classifier utilizing at least the importance-weighted first set of data samples.

10. The computer-implemented method of claim 1 , further comprising evaluating a performance of the diagnostic classifier on a validation dataset.11 . The computer-implemented method of claim 10, wherein the validation dataset comprises a combination of real data and synthetic data from a target subpopulation.

12. The computer-implemented method of claim 1 , further comprising deploying the trained classifier on blood-based methylation test data.

13. The computer-implemented method of claim 1 , wherein: the computing comprises computing the importance weights for the second set of data samples; the assigning comprises assigning the importance weights to the second set of data samples; and the training comprises using the importance-weighted second set of data samples.

14. A system, comprising: one or more processors; and one or more computer readable media storing instructions that are executable by the one or more processors to perform operations to: receive a source dataset comprising a first set of data samples; receive a target dataset comprising a second set of data samples; perform feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; compute, based on the performing and via employing a non-parametric importance-weighting technique, importance weights for at least the first set of data samples; assign the computed importance weights to at least the first set of data samples; and train a diagnostic classifier using at least the importance-weighted first set of data samples.

15. The system of claim 14, wherein the first set of data samples encompasses a plurality of disease types.

16. The system of claim 15, wherein the plurality of disease types are associated with cancer.

17. The system of claim 14, wherein the target dataset represents a targeted subpopulation of disease types.

18. The system of claim 17, wherein the targeted subpopulation of disease types is at least one of: non-Hodgkin lymphoma, Hodgkin lymphoma, myeloid cancer, plasma cell disorders, circulating lymphoma conditions, unspecified conditions, and non-cancer conditions.

19. The system of claim 14, wherein the operations to compute the importance weights via employing the non-parametric importance-weighting technique comprises operations to employ a Kernel Mean Matching (KMM) technique.

20. The system of claim 19, wherein the KMM technique comprises operations to: select a kernel function to project the first set of features and the second set of features into a high-dimensional feature space; adjust a first distribution of the first set of features in the highdimensional feature space to match a second distribution of the second set of features; and determine the importance weights that minimize a difference between means of the first distribution of the first set of features and the second distribution of the second set of features.21 . The system of claim 20, wherein the kernel function is a Gaussian radial basis function (RBF).

22. The system of claim 14, wherein the instructions are further executable by the one or more processors to perform operations to: evaluate a performance of the diagnostic classifier on a validation dataset.

23. The system of claim 22, wherein the validation dataset comprises a combination of real data and synthetic data from a target subpopulation.

24. The system of claim 14, wherein the instructions are further executable by the one or more processors to perform operations to: deploy the trained classifier on blood-based methylation test data.

25. The system of claim 14, wherein: the instructions executable by the processor to compute further comprise instructions executable by the processor to compute the importance weights for the second set of data samples; the instructions executable by the processor to assign further comprise instructions executable by the processor to assign the importance weights to the second set of data samples; and the instructions executable by the processor to train further comprise instructions executable by the processor to use the importance-weighted second set of data samples.

26. A non-transitory computer-readable medium storing computer-executable instructions which, when executed by a system, cause the system to perform operations comprising:receiving, at a system, a source dataset comprising a first set of data samples; receiving, at the system, a target dataset comprising a second set of data samples; performing, using a processor associated with the system, feature extraction on the source dataset and the target dataset to obtain a first set of features associated with the source dataset and a second set of features associated with the target dataset; computing, based on the performing and via employing a non-parametric importance-weighting technique, importance weights for at least the first set of data samples; assigning, using the processor, the computed importance weights to at least the first set of data samples; and training, using the processor, a diagnostic classifier using at least the importance-weighted first set of data samples.

27. A computer-implemented method, the computer-implemented method comprising: receiving, at a system, a source dataset comprising a first set of data samples; receiving, at the system, a target dataset comprising a second set of data samples; performing, using a processor associated with the system, feature extraction on the source dataset and the target dataset to obtain a first set of featuresassociated with the source dataset and a second set of features associated with the target dataset; performing, using the processor, a first feature augmentation process on the first set of features and a second feature augmentation process on the second set of features, wherein the first feature augmentation process is different than the second feature augmentation process; generating, using the processor and based on performing the first feature augmentation process and the second feature augmentation process, a first feature- augmented data set and a second feature-augmented data set; and training, using the processor, a diagnostic classifier using the first feature- augmented data set and the second feature-augmented data set.

Images