Medical classification using cell-free DNA and deep learning

Abstract

There is provided a method for classification of cell-free deoxyribonucleic acid (cfDNA) into a medical state of a subject, comprising: receiving raw cfDNA sequencing data comprising cfDNA reads obtained from a biological sample of a subject, the raw cfDNA sequencing data generated by a sequencing machine, sampling a subset of raw cfDNA reads from the raw cfDNA sequencing data, wherein the subset forms a sample, wherein the sampling is performed according to a sampling process, generating a multivariate input representation for the sample, comprising nucleotide sequences of the sample and / or at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to, feeding into a machine learning (ML) model, the multivariate input representation, and obtaining from the ML model, for the sample, a probability that the subject is in the medical state.

Description

[0001] MEDICAL CLASSIFICATION USING CELL-FREE DNA AND DEEP LEARNING

[0002] RELATED APPLICATION

[0003] This application claims the benefit of priority of United Kingdom Patent Application No. GB2410447.3 filed on July 18, 2024, the contents of which are incorporated herein by reference in their entirety.

[0004] BACKGROUND

[0005] The present invention, in some embodiments thereof, relates to machine learning models and, more specifically, but not exclusively, to machine learning models for classification based on cell-free DNA (cfDNA).

[0006] Cancer diagnosis often relies on invasive tissue biopsies, which are costly, uncomfortable for patients, and provide only a snapshot of tumor heterogeneity with sampling limitations. Nikanjam et al, (2022). Liquid biopsies, which analyze circulating cell-free DNA (cfDNA) in blood, can be a non-invasive alternative for early cancer detection and monitoring. As blood passes through tumors, it collects tumor-derived cfDNA released via cell death, which may enable regular, real-time screening without tissue sampling. Raez et al (2023), Perakis et al (2017). Liquid biopsies have the potential to overcome the limitations of tissue biopsies by providing a comprehensive view of tumor heterogeneity and enabling earlier detection of cancer.

[0007] SUMMARY

[0008] According to a first aspect, a system for classification of cell-free deoxyribonucleic acid (cfDNA) into a medical state of a subject, comprising: a communication interface configured for receiving raw cfDNA sequencing data comprising a plurality of cfDNA reads obtained from a biological sample of a subject, the raw cfDNA sequencing data generated by a sequencing machine, at least one processor operatively coupled to the communication interface, the at least one processor configured for executing code for: sampling a subset of raw cfDNA reads from the raw cfDNA sequencing data, wherein the subset forms a sample, wherein the sampling is performed according to a sampling process, generating a multivariate input representation for the sample, comprising nucleotide sequences of the sample and / or at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to, feeding into a machine learning (ML) model, the multivariate input representation, and obtaining from the ML model, for the sample, a probability that the subject is in the medical state. According to a second aspect, a system of training a ML model for classification of cell- free deoxyribonucleic acid (cfDNA) into a medical state of a subject, comprising: a communication interface configured for receiving raw cfDNA sequencing data obtained from respective biological samples of a plurality of individuals, the raw cfDNA sequencing data generated by a sequencing machine, at least one processor operatively coupled to the communication interface, the at least one processor configured for executing a code for: creating a training dataset comprising a plurality of records created for a plurality of individuals, by: for each individual of the plurality of individuals: sampling a subset of raw cfDNA reads from the raw cfDNA sequencing data, wherein the subset forms a sample, wherein the sampling is performed according to a sampling process, generating a multivariate input representation for the sample, comprising nucleotide sequences of the sample and at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to, creating a record including the multivariate input representation and a ground truth label indicating a classification category associated with the medical state of the individual, training the ML model on the training dataset for generating a probability that the subject is in the medical state in response to an input of a target multivariate input representation including a sample of a subset of raw cfDNA reads of raw cfDNA sequencing data generated from a biological sample of a subject, sampled according to the sampling process, and at least one of: target quality values associated with each nucleotide of the sample, and target genomic data associated with at least one region of the sample, and providing the trained ML model and an aggregation process for aggregating a plurality of probabilities obtained from the ML model for a plurality of multivariate input representations for a plurality of samples of a subject, for computing a final subject-level classification and / or an aggregated probability that the subject is in the medical state.

[0009] According to a third aspect, a computer implemented method of classification of cell-free deoxyribonucleic acid (cfDNA) into a medical state of a subject, comprising: receiving raw cfDNA sequencing data comprising a plurality of cfDNA reads obtained from a biological sample of a subject, the raw cfDNA sequencing dta generated by a sequencing machine, in a plurality of iterations: sampling a subset of raw cfDNA reads from the raw cfDNA sequencing data, wherein the subset forms a sample, wherein the sampling is according to a sampling process, wherein in each iteration a different sample is sampled according to the sampling process, generating a multivariate input representation for the sample, comprising nucleotide sequences of the sample and at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to, feeding into a machine learning model (ML) model, the multivariate input representation, and obtaining from the ML model, for the sample of each iteration probability that the subject is in the medical state, aggregating a plurality of probabilities obtained from the ML model from the plurality of iterations, and providing the aggregation of the plurality of probabilities as an indication that the subject is in the medical state.

[0010] In a further implementation form of the first, second, and third aspects, the multivariate input representation includes the nucleotide sequences of the sample, the associated quality values, and the metadata representations.

[0011] In a further implementation form of the first, second, and third aspects, the multivariate input representation includes the nucleotide sequences of the sample and the associated quality values, and excludes the metadata representations.

[0012] In a further implementation form of the first, second, and third aspects, the multivariate input representation includes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

[0013] In a further implementation form of the first, second, and third aspects, the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and includes the metadata representations.

[0014] In a further implementation form of the first, second, and third aspects, the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and excludes the metadata representations.

[0015] In a further implementation form of the first, second, and third aspects, the multivariate input representation excludes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

[0016] In a further implementation form of the first, second, and third aspects, the multivariate input representation excludes alignment-based feature engineering and / or calculation of summary genomic features, wherein the metadata representation excludes alignment-based feature engineering and / or calculation of summary genomic features.

[0017] In a further implementation form of the first, second, and third aspects, further comprising code for: iterating the sampling, the generating, the feeding, and the obtaining, for different samples of the subject, for obtaining from the ML model a plurality of probabilities for a plurality of different samples of the subject, aggregating the plurality of probabilities of the subject, and providing the aggregation of the plurality of probabilities as a final subject-level classification and / or the probability that the subject is in the medical state.

[0018] In a further implementation form of the first, second, and third aspects, a total number of reads of the plurality of different samples used to generate the final subject-level classification and / or probability that the subject is in the medical state is less than about 100,000. In a further implementation form of the first, second, and third aspects, further comprising aggregating into a vector form, the raw cfDNA reads of the sample, wherein the vector is fed into the ML model.

[0019] In a further implementation form of the first, second, and third aspects, the sampling process comprises randomly selecting the sample of raw cfDNA reads from the plurality of raw sequences.

[0020] In a further implementation form of the first, second, and third aspects, randomly selecting comprises: wherein the subset of raw cfDNA reads of the sample are mapped to random locations of a reference genome, and / or the subset of raw cfDNA reads of the sample are read in a random order.

[0021] In a further implementation form of the first, second, and third aspects, the sampling process comprises consecutively reading the subset of raw cfDNA reads of the sample.

[0022] In a further implementation form of the first, second, and third aspects, the sampling process comprises: mapping the plurality of raw sequences to a plurality of regions of a reference genome, randomly selecting a region on the reference genome, and consecutively reading the sample of the subset of raw cfDNA reads mapped to the randomly selected region on the reference genome.

[0023] In a further implementation form of the first, second, and third aspects, the sampling process comprises including about .001-5% of the plurality of raw sequences in the sample of raw cfDNA reads.

[0024] In a further implementation form of the first, second, and third aspects, the sampling process comprises selecting a number of the raw cfDNA reads included in the sample from a range of about 512-4096.

[0025] In a further implementation form of the first, second, and third aspects, the sampling process comprises selecting a size of each of the raw cfDNA reads of the sample that is common and fixed, selected from a range of about 50-350 nucleotides.

[0026] In a further implementation form of the first, second, and third aspects, further comprising: pre-processing the raw sequence data by encoding the raw sequence data into a numerical encoding by mapping each type of cfDNA nucleotide to a corresponding numerical value, wherein the encoding of each raw cfDNA of the subset is fed into the ML model.

[0027] In a further implementation form of the first, second, and third aspects, the quality values are generated by the automated sequencing machine that generates the raw cfDNA sequencing data from the plurality of cfDNA reads. In a further implementation form of the first, second, and third aspects, further comprising: mapping each raw sequence to a corresponding genomic region of a defined reference genomic sequence, and for each respective raw cfDNA read, generating corresponding metadata including empirically derived information about the corresponding genomic region, wherein the corresponding metadata is included in the multivariate input representation.

[0028] In a further implementation form of the first, second, and third aspects, the metadata for the sample indicates whether the sample maps to a coding region, a methylated region, or a region with known disease-relevant annotations.

[0029] In a further implementation form of the first, second, and third aspects, the ML model is configured for modular interchangeability between different deep learning architectures, including transformer encoders, convolutional neural networks (CNNs), and recurrent neural networks (RNNs).

[0030] In a further implementation form of the first, second, and third aspects, the ML model is implemented as a neural network encoder with a classification head.

[0031] In a further implementation form of the first, second, and third aspects, the neural network encoder with the classification head is implemented as a transformer encoder model that excludes a decoder component and excludes positional encodings.

[0032] In a further implementation form of the first, second, and third aspects, further comprising adding a classification token (CLS token) as a first token to an input sequence of a plurality of the raw cfDNA reads for feeding into the ML model, and wherein a classification output layer is attached to an embedding of the CLS token after a final encoder layer of the ML model.

[0033] In a further implementation form of the first, second, and third aspects, aggregating is selected a group consisting of: Quantile-based aggregation, Mean, Median, and Majority.

[0034] In a further implementation form of the first, second, and third aspects, the ML model is configured for receiving the multivariate input representation directly, without requiring prior manual feature extraction and / or generation of summary genomic features based on alignment to a reference genome.

[0035] In a further implementation form of the first, second, and third aspects, the ML model is configured to learn an internal representation of the cfDNA reads that enables unsupervised clustering of samples according to the medical state.

[0036] In a further implementation form of the first, second, and third aspects, the medical condition is selected from: cancer, a specific type of cancer, specific sub-type of cancer, lymphoma types including diffuse large B-cell lymphoma (DLBCL) and high-grade B-cell lymphoma (HGBCL), each of a plurality of different types of cancer, predicting the subject’s response to a therapy administered to treat a disease, predicting the subject’ s response to a treatment for a cancer, a stage of the cancer including early-stage and late-stage.

[0037] In a further implementation form of the first, second, and third aspects, the multivariate representation includes the nucleotide sequences of the sample and associated quality values, and excludes the metadata representations of genomic regions.

[0038] In a further implementation form of the first, second, and third aspects, a representation vector is obtained from the ML model, and the probability is computed from the representation vector.

[0039] In a further implementation form of the first, second, and third aspects, the ML model comprises an ensemble of two or more independently trained ML models, wherein a final probability is computed by averaging or voting on the probabilities generated by the two or more independently trained ML models.

[0040] In a further implementation form of the first and third aspects, the ML model is trained on a training dataset comprising a plurality of records created for a plurality of individuals, by: for each individual of a plurality of individuals: sampling a subset of raw cfDNA reads from raw cfDNA sequencing data of the individual obtained from a biological sample of the individual and sequenced by the sequencing machine, wherein the subset forms a sample, wherein the sampling is performed according to a sampling process, generating the multivariate input representation for the individual, comprising nucleotide sequences of the sample and at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to, and creating a record including the multivariate input representation generated for the individual and a ground truth label indicating a classification category associated with the medical state that the individual is in.

[0041] In a further implementation form of the first, second, and third aspects, the training of the ML model comprises initializing the ML model by transfer-learning from a pre-trained model trained on a large dataset of genomic data, and training the ML model on the training dataset after the initialization.

[0042] Unless otherwise defined, all technical and / or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention pertains. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of embodiments of the invention, exemplary methods and / or materials are described below. In case of conflict, the patent specification, including definitions, will control. In addition, the materials, methods, and examples are illustrative only and are not intended to be necessarily limiting. BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0043] Some embodiments of the invention are herein described, by way of example only, with reference to the accompanying drawings. With specific reference now to the drawings in detail, it is stressed that the particulars shown are by way of example and for purposes of illustrative discussion of embodiments of the invention. In this regard, the description taken with the drawings makes apparent to those skilled in the art how embodiments of the invention may be practiced.

[0044] In the drawings:

[0045] FIG. 1 is a block diagram of components of a system for classification of cfDNA sequencing data into a medical state of a subject, in accordance with some embodiments of the present invention;

[0046] FIG. 2 is a flowchart of a method of classifying cfDNA sequencing data into a medical state of a subject, in accordance with some embodiments of the invention;

[0047] FIG. 3 is a flowchart of a method of training a ML model for classifying cfDNA sequencing data into a medical state of a subject, in accordance with some embodiments of the present invention;

[0048] FIG. 4 is a schematic depicting a high-level overview of a process for determination of a medical state of a subject based on feeding raw cfDNA sequencing data into a ML model, in accordance with some embodiments of the present invention;

[0049] FIG. 5 is a dataflow diagram of an exemplary dataflow including a read representation and aggregation to the ML model and the aggregation method and analysis of the resulting scores from the ML model, in accordance with some embodiments of the present invention;

[0050] FIG. 6 is a schematic depicting an exemplary approach of the read aggregation process, demonstrating how multiple samples from a single patient are processed and their classification results combined to yield a final patient-level decision, in accordance with some embodiments of the present invention;

[0051] FIG. 7 is a table presenting a tabular summary the distribution of patient samples across various cancer types and healthy controls in an experiment conducted by the Inventors, in accordance with some embodiments of the present invention;

[0052] FIG. 8 includes plots of unsupervised clustering of datasets of subjects with different cancers indicating the ML model's ability to learn meaningful representations from raw cfDNA data that effectively separate healthy samples from cancer samples, even for data not seen during training, in accordance with some embodiments of the present invention;

[0053] FIG. 9 is a classification matrix of an ensemble ML model for pancreatic cancer classification, in accordance with some embodiments of the present invention; FIG. 10 is a box plot of “Cancer Score" for multiple cancer classification, in accordance with some embodiments of the present invention;

[0054] FIG. 11 includes AUC plots showing good classification for early and late-stage cancers, in accordance with some embodiments of the present invention;

[0055] FIG. 12 includes graphs comparing classification performance of a traditional liquid biopsy method to at least one embodiment, in accordance with some embodiments of the present invention;

[0056] FIG. 13 includes graphs depicting a comparison based on at least one embodiment with traditional liquid biopsy methods in low-coverage sequencing, in accordance with some embodiments of the present invention;

[0057] FIG. 14 includes correlations between cancer prediction of the ML model and traditional computational methods, in accordance with some embodiments of the present invention;

[0058] FIG. 15 includes a PC A plot and a t-SNE plot demonstrating the generalization performance of the approach described herein, in accordance with some embodiments of the present invention;

[0059] FIG. 16 includes a classification performance report of separating responders and nonresponders to a cancer therapy of patients with high-grade B-cell lymphoma (HGBCL), in accordance with some embodiments of the present invention; and

[0060] FIG. 17 is a box plot of cancer scores for samples of breast cancer therapy responders and non-responders, in accordance with some embodiments of the present invention.

[0061] DETAILED DESCRIPTION

[0062] The present invention, in some embodiments thereof, relates to machine learning models and, more specifically, but not exclusively, to machine learning models for classification based on cell-free DNA (cfDNA).

[0063] It is to be understood that the raw sequences of cfDNA (e.g., fragments) which are obtained from a biological sample of a subject refer to a digital representation of the raw sequences. The digital representation may be based on existing digital representations, for example, as automatically generated by a sequencing machine that automatically generates the digital sequence in response to an input of physical cfDNA. The term ‘digital’ is omitted for clarity and simplicity, but is implied.

[0064] As used herein the term “raw” in the context of sequences of cfDNA refers to the digital sequence as outputted by the sequencing machine, for example, A, C, T, and G. The digital representation outputted by the sequencing machine may be encoded into a corresponding 1 : 1 encoding for each nucleotide, for example, A, C, T, and G are mapped to 0, 1, 2, 3, or to 00,01,10, 11, where the encoded representation is included within the term “raw”. Other minimal and / or basic processing is meant to be included in the term “raw”, for example, aligning to a reference genome to enable obtaining metadata of the region from which the read came from. The basic processing still represent raw reads, because in the bioinformatic world, most of the analysis that is performed does not use the raw reads at all, but rather uses processed information, for example, information about mutations, or information about structural variants, or aggregate statistics like distribution of read lengths. In all of the aforementioned analysis approaches, the information about each individual read raw is lost. Therefore, based on the research world of genomics and bioinformatics, an analysis that includes individual raw reads in it is an analysis of raw reads. Moreover, the mapping to the reference genome described herein (to obtain the metadata) is for contextual annotation rather than feature engineering in the traditional sense.

[0065] An aspect of some embodiments of the present invention relates to systems, methods, computing devices, and / or code (stored on a data storage device and executable by one or more processors) for determining a medical state of a subject by classifying a combination of one or more of: raw cfDNA reads, associated quality values, and / or genomic data, i.e., using the sequences themselves rather than a feature extracted from the sequences. Raw cfDNA sequencing data of cfDNA reads obtained from a biological sample of a subject (e.g., blood, plasma) is received and / or accessed, optionally by a communication interface. The raw cfDNA sequencing data is generated by a sequencing machine that analyzes the physical cfDNA fragments and creates a digital representation of the nucleotide sequences. A subset of raw cfDNA reads is sampled from the raw cfDNA sequencing data according to a sampling process. The subset forms a sample. Examples of the sampling process include: randomly reading, wherein the subset of raw cfDNA reads of the sample are mapped to random locations of a reference genome, consecutively reading the subset of raw cfDNA reads of the sample optionally mapped to a randomly selected region on a reference genome, and / or sampling to obtain a total number of reads less than about 100,000. A multivariate input representation is generated for the sample. The multivariate input representation includes one or more of: (i) nucleotide sequences of the sample, (ii) associated quality values, and / or (iii) metadata representations of genomic regions to which the reads of the sample map to. The multivariate input representation may exclude alignment-based feature engineering and / or calculation of summary genomic features. It is noted that the metadata representation excludes alignment-based feature engineering and / or calculation of summary genomic features. The metadata representation may represent a raw reads, because, for example, in the bioinformatic worlds, most of the analysis don't use the reads at all in their analysis, but rather processed information. The multivariate input representation is fed into a machine learning model (ML) model, optionally a transformer encoder model that excludes a decoder component and excludes positional encodings (other exemplary architectures are described herein). A probability that the subject is in the medical state is obtained from the ML model.

[0066] Optionally, the multivariate input representation includes the nucleotide sequences of the sample, the associated quality values, and the metadata representations.

[0067] Alternatively, the multivariate input representation includes the nucleotide sequences of the sample and the associated quality values, and excludes the metadata representations.

[0068] Alternatively, the multivariate input representation includes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

[0069] Alternatively, the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and includes the metadata representations.

[0070] Alternatively, the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and excludes the metadata representations.

[0071] Alternatively, the multivariate input representation excludes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

[0072] Optionally, the features of the method, including the sampling, the generating, the feeding, and the obtaining, are iterated for different samples. The multiple probabilities obtained from the ML model for the multiple samples are aggregated, optionally into a single value and / or classification category providing a final subject-level classification and / or probability that the subject is in the medical state. Examples of medical states include: whether the subject has cancer, whether the subject has a specific type of cancer - optionally the probability is computed for each cancer, whether the subject has a specific type of sub-cancer, and whether the subject is responding to treatment of the cancer.

[0073] Optionally, a total number of reads of the multiple samples used to generate the final subject-level classification and / or probability that the subject is in the medical state, optionally to a target performance threshold, is less than about 250,000, or less than about 100,000, or less than about 50,000, or other values. Such number of reads is smaller than required by known prior approaches to obtain results at the target performance threshold, which improves performance of a computer performing the subject-level classification, in terms of, for example, reduced processing time, reduced utilization of processing resources, reduced memory utilization, and the like.

[0074] An aspect of some embodiments of the present invention relates to systems, methods, computing devices, and / or code (stored on a data storage device and executable by one or more processors) for training a ML model for classification of raw representation of cfDNA sequences into a medical state of a subject. Raw cfDNA sequencing data obtained from respective biological samples of multiple individuals is received and / or accessed. The raw cfDNA sequencing data is generated by a sequencing machine that sequences the physical cfDNA nucleotides. A training dataset of records is created using the cfDNA sequencing data from multiple individuals. The following is performed for each individual. A subset of raw cfDNA reads is sampled from the raw cfDNA sequencing data. The subset forms a sample. The sampling is performed according to a sampling process. Examples of the sampling process include: randomly reading, wherein the subset of raw cfDNA reads of the sample are mapped to random locations of a reference genome, consecutively reading the subset of raw cfDNA reads of the sample optionally mapped to a randomly selected region on a reference genome, and / or sampling to obtain a total number of reads less than about 100,000. A multivariate input representation is generated for the sample. The multivariate input representation includes one or more of: (i) nucleotide sequences of the sample, (ii) associated quality values, and / or (iii) metadata representations of genomic regions to which the reads of the sample map to. The multivariate input representation may exclude alignmentbased feature engineering and / or calculation of summary genomic features. A record is generated. The record includes the multivariate input representation, and a ground truth label indicating a classification category associated with the medical state of the individual. The ML model is trained on the training dataset for generating a probability that a subject is in the medical state in response to an input of a target multivariate input representation including one or more of: (i) a sample of a subset of raw cfDNA reads of raw cfDNA reads of raw cfDNA sequencing data generated from a biological sample of a subject, sampled according to the sampling process, (ii) target quality values associated with each nucleotide of the sample and / or (iii) target genomic data associated with at least one region of the sample. The ML model and optionally an aggregation process are provided. The aggregation process is for aggregating multiple probabilities obtained from the ML model for multiple samples of the subject, for computing a final subject-level classification and / or aggregated probability that the subject is in the medical state.

[0075] Optionally, the multivariate input representation includes the nucleotide sequences of the sample, the associated quality values, and the metadata representations.

[0076] Alternatively, the multivariate input representation includes the nucleotide sequences of the sample and the associated quality values, and excludes the metadata representations.

[0077] Alternatively, the multivariate input representation includes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations. Alternatively, the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and includes the metadata representations.

[0078] Alternatively, the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and excludes the metadata representations. Alternatively, the multivariate input representation excludes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations. At least one embodiment described herein addresses the technical problem of analyzing cfDNA sequences obtained from a tissue sample of a subject to determine a medical state of the subject, for example, cancer detection and / or classification, optionally early detection of cancer. Such detection, in particular early detection, is currently challenging due to the low abundance of tumor- derived cfDNA against a normal background. The signal to noise ratio of cfDNA sequences is low, implying that a large number of cfDNA nucleotides (i.e., sequences) are to be analyzed in order to determine the medical state with sufficiently high accuracy. For example, about 80,000,000 base pairs of cfDNA sequences may be obtained from the sequencing machine. It is impractical to feed the entire 80,000,000 base pairs into an ML model to obtain a result within a practical amount of processing time and / or using a practical amount of computational hardware. Another technical problem relates to analyzing a large amount of cfDNA sequencing data by a machine learning model. The ML model is unable to process all of the raw cfDNA sequencing data in a reasonable time and / or using reasonable processing resources. Another technical problem relates to feeding sufficient data into the ML model to obtain an indication of a medical state of the subject above a target performance metric threshold. At least one embodiment described herein improves the technology of ML models for analyzing cfDNA sequences to determine a medical state of a subject, by providing ML model(s) and / or approaches for using the ML model(s) that are able to analyze raw cfDNA sequencing data obtained from a tissue sample of a subject within a practical amount of time and / or using a practical amount of computational resources. At least one embodiment described herein improves upon other approaches that use ML models to analyze cfDNA sequencing data, for example, by extracting predefined hand-crafted features from the cfDNA sequences, and analyzing the extracted hand-crafted features. The improvement of at least one embodiment described herein is in the use of a sample of the raw cfDNA sequencing data, optionally in combination with associated quality values, and metadata representations of genomic regions to which the reads of the sample map to, which provide a more accurate outcome (e.g., above the target performance metric) in comparison to extracted hand-crafted features and / or other prior approaches. At least one embodiment described herein provides the practical application of being able to use a ML model to analyze raw cfDNA sequencing data within a practical amount of time and / or using a practical amount of computational resources, to determine a medical state of a subject. For example, Inventors discovered that the medical state of the subject may be predicted with a performance metric above a target threshold with as little as about 150000, or about 100000 reads of the cfDNA sampling data.

[0079] Understanding cfDNA biology is crucial for liquid biopsy analysis. cfDNA, originating from apoptotic and necrotic cells, circulates freely in the bloodstream and carries genetic information reflective of the individual's physiological state. cfDNA offers a dynamic snapshot of tumors, reflecting their genetic and epigenetic landscapes. The increased levels of tumor-derived cfDNA in circulation, attributed to high tumor cell turnover, make it a valuable source for cancer detection and monitoring. Its low concentrations in blood make cfDNA extraction and sequencing analysis complex. However, its non-invasive presence in blood plasma creates opportunities for early cancer detection without invasive procedures.

[0080] Next-generation sequencing (NGS) technologies allow for massively parallel sequencing of cfDNA fragments obtained from a simple blood draw. Reductions in NGS costs alongside improved accessibility have opened up new avenues for cfDNA analysis. Tan et al (2024). However, detecting faint early signals of cancer in cfDNA remains challenging due to the difficulty of distinguishing tiny amounts of tumor-derived cfDNA amongst the dominant normal background.

[0081] Previous computational methods for cfDNA analysis demonstrated the potential of cfDNA analysis for cancer detection and prognostication; however, they suffer from limitations such as low sensitivity, high false positive rates, and restricted genomic coverage, while often requiring reference samples or other prior knowledge. Those methods heavily relied on manual feature engineering and domain expertise to extract input representations for machine learning models. Handcrafted features, such as fragment size (Cristiano et al. (2019), Mouliere et al. (2021), methylation patterns, Liu et al. (2020), Klein et al. (2021)) and copy number aberrations (Pham et al. (2023), Zviran et al. (2020)) were used as inputs to classical machine learning techniques Nathan Wan et al. (2019). Most works employed traditional machine learning approaches, which depended on this manual feature engineering, limiting their applicability to early-stage cancer detection and analysis. Additionally, analysis was constrained to targeted sequencing of specific genomic regions with known mutations (Peneder et al. (2021)) overlooking the broader genomic landscape. For example, one prior approach describes classifying cancer by first selecting only short cfDNA fragments and then computing relative copy-number profiles at many genomic loci, which are then fed into a classifier. Similarly, another approach is based on extracting DNA, aligns reads, and generating a vectorized feature vector (e.g., "grand canyon" plots of fragment counts or distances) for an Al model. These approaches inherently discard potentially valuable raw readlevel information.

[0082] At least one embodiment described herein relates to directly analyzing raw sequencing reads across the entire genome, potentially capturing complex patterns beyond engineered features and developing inherent robustness to biological and technical noise present in the raw data. This approach could enable earlier and more comprehensive cancer detection and analysis by leveraging signals from coding and non-coding regions across the entire genome, and without the need for laborious manual feature engineering.

[0083] At least one embodiment described herein provides improved tools for cancer detection and / or classification using cfDNA, based on direct analysis of raw sequencing data which overcome the limitations of prior approaches.

[0084] At least one embodiment described herein improves performance of a computing device implementing the ML model that analyzes raw sequences of cfDNA obtained from a biological sample of a patient, by enabling obtaining an indication that the subject is in a medical state within a practical amount of processing time and / or using a practical amount of processing resources (e.g., processor(s) and / or memory).

[0085] At least one embodiment described herein provides a solution to the aforementioned technical problem, and / or improves the aforementioned technology, and / or improves upon the aforementioned existing approaches, and / or provides the aforementioned practical application, by determining a medical state of a subject by classifying raw cfDNA sequencing data, i.e., using the sequences themselves rather than a feature extracted from the sequences. Raw cfDNA sequencing data obtained from a biological sample of a subject is received and / or accessed. A subset of raw cfDNA reads is sampled from the raw cfDNA sequencing data according to a sampling process. Examples of the sampling process include: randomly reading, wherein the subset of raw cfDNA reads of the sample are mapped to random locations of a reference genome, consecutively reading the subset of raw cfDNA reads of the sample optionally mapped to a randomly selected region on a reference genome, and / or sampling to obtain a total number of reads less than about 100,000. The subset forms a sample. A multivariate input representation is generated for the sample. The multivariate input representation includes one or more of: (i) nucleotide sequences of the sample, (ii) associated quality values and / or (iii) metadata representations of genomic regions to which the reads of the sample map to. The multivariate input representation may exclude alignment-based feature engineering and / or calculation of summary genomic features. The multivariate input representation is fed into a ML model, optionally a transformer encoder model that excludes a decoder component and excludes positional encodings (other exemplary architectures are described herein). A probability that the subject is in the medical state is obtained from the ML model. Optionally, the features of the method, including the sampling, the generating, the feeding, and the obtaining, are iterated for different samples. The multiple probabilities obtained from the ML model for the multiple samples are aggregated, optionally into a single value and / or classification category providing a final subject-level classification and / or probability that the subject is in the medical state.

[0086] Feeding the entire dataset of sequences of the cfDNA sequencing data obtained from the biological sample of the individual into the ML model is impractical, in terms of impractically long processing time to process the entire dataset and / or in terms of requiring an impractically large amount of processing resources (e.g., processors, memory) to enable processing the entire dataset. In contrast to other approaches that extract features from the dataset of sequences and analyze the extracted features (e.g., using machine learning models or other approaches), Inventors discovered that a sample of a much smaller number of raw sequences sampled from the entire dataset, optionally randomly, enables obtaining a sufficiently accurate outcome indicating the medical state of the subject.

[0087] At least one embodiment described herein provides a solution to the aforementioned technical problem, and / or improves the aforementioned technology, and / or improves upon the aforementioned existing approaches, and / or provides the aforementioned practical application, by using an end-to-end deep learning approach based on a transformer encoder model for direct analysis of raw genomic sequencing data from cfDNA, including nucleotide sequences, and optionally quality values, and / or metadata representations, without manual feature engineering. The transformer-based deep learning approach may be used for early cancer detection directly from cfDNA, requiring only a minimal amount of sequencing data. Embodiments of the disclosure therefore enable cost-effective, widespread early cancer screening and personalized treatment guidance.

[0088] At least one embodiment described herein relates to devices, systems, and methods for the classification of medical conditions using sequencing data obtained from cfDNA. In particular, it pertains to detecting the presence or absence of cancer — including early-stage cancers — and / or predicting treatment outcomes, such as response or non-response to therapy, using deep learning models trained on minimally processed cfDNA data.

[0089] In contrast to extracting hand-crafted features as done by some prior approaches (for example, copy-number variation or specific mutations extracted by conventional bioinformatic pipelines) at least one embodiment described herein is based on constructing an input representation directly from raw sequencing reads. At least one embodiment described herein does not merely use any neural network or conventional machine-learning model on cfDNA — but rather has a specific way of representing the raw sequence data (and optionally accompanying quality and / or annotation data) and / or aggregating multiple such “read-images” to leam complex, long-range patterns that are not accessible to conventional bioinformatic approaches. At least one embodiment described herein relates to an approach that may have clinical applicability in both multi-cancer screening and / or in the prediction of treatment response.

[0090] At least one embodiment described herein relates to an end-to-end deep learning methodology for the direct classification and detection of cancer (or other medical state) from cfDNA sequencing data. At least one embodiment represents a significant departure from conventional approaches by eliminating the need for manual feature engineering, instead enabling a neural network, optionally a transformer encoder model, to autonomously discern intricate patterns within the raw genomic sequences.

[0091] At least one embodiment described herein relates an end-to-end deep learning framework that operates directly on raw sequencing data, thereby eliminating the need for manual feature engineering. This allows the neural network to capture complex genomic patterns and long-range dependencies that traditional, feature engineering-based machine learning approaches might miss. The ML model jointly optimizes feature extraction and training, leading to superior prognostic predictions. This approach also substantially reduces the laborious human effort and specialized domain expertise traditionally required for feature extraction, streamlining the diagnostic process.

[0092] At least one embodiment described herein processes input data differently than other approaches. Instead of relying on aligned reads or pre-computed features, at least one embodiment processes raw sequencing reads directly. Optionally, each cfDNA nucleotide within the sequences is assigned a numerical value (e.g., 0-4) for optimized storage and efficient processing. From the patient's sequencing data, a random subset of reads, typically 4,096 reads, is extracted to construct samples. The input to the deep learning model can comprise individual or a combination of nucleotide sequences, their associated quality values, and / or metadata representations. The metadata representation may include empirically derived information about the genomic region to which a read might map (e.g., whether it's a gene region, a coding region, a region known to be methylated in certain cells, or any other relevant contextual information). This structured, yet unengineered, representation empowers the deep learning model to leam the relevant features autonomously. The inclusion of per-read quality scores and gene-region mapping / methylation is different than other approaches, some of which focus on aggregate genomic signals. The early detection of cancer signals in cell-free DNA (cfDNA) presents significant challenges due to the low signal-to-noise ratio (SNR) inherent in distinguishing tumor-derived cfDNA from normal cfDNA background. The present disclosure describes, inter alia, an end-to- end deep learning approach utilizing transformers to directly learn informative features from raw genomic sequencing reads. This approach is different than prior approaches and / or overcomes the limitations of prior approaches that rely on manual feature engineering and / or pre-computed aggregate features, by allowing the ML model to discover subtle, non-obvious patterns directly from the raw data, including nucleotide sequences, quality scores, and generalized metadata.

[0093] At least one embodiment described herein relates to a validated high-performance multicancer classifier capable of detecting signals across several types of cancer. In experiments performed by the Inventors described in the “Examples” section below, the ML model achieves a sensitivity of 79% for multi-cancer detection at 85% specificity, highlighting its potential for widespread early screening using liquid biopsies. In at least one embodiment, focusing on a single cancer type, such as breast cancer, can lead to even higher sensitivity and specificity, emphasizing the trade-off between performance and generalizability.

[0094] The reliable generalization capabilities of the ML model across multiple cancer types suggest its potential utility in clinical settings. At least one embodiment accounts for the impact of dataset biases and biological factors on performance, particularly in certain cancers. Further embodiments can employ improved sample acquisition and specialized tuning for cancers with poorer performance metrics.

[0095] To further enhance the performance and applicability of the ML model, at least one embodiment incorporates supplementary metadata, which can provide richer representations of malignant signals. For instance, annotations on DNA methylation, histone modifications, nucleosome positioning, or fragment lengths may complement genomic data and improve classification accuracy, especially in challenging cases.

[0096] Moreover, at least one embodiment couples a multi-cancer classifier with cell type tracing techniques, which can specifically locate malignancies within the body. By identifying the tissue source of highly weighted reads, the at least one embodiment may enable broad cancer detection without the need for separate per-cancer models. Additionally, at least one embodiment leverages large language models for feature extraction and classification, combined with dynamic tracking of cfDNA over time and treatment, thereby potentially enhancing the utility of liquid biopsy monitoring in precision oncology. Further, at least one embodiment incorporates clinical feedback on disease progression, which could refine staging predictions and guide optimal interventions, ultimately improving patient outcomes. Understanding stage shifting and lead time bias will be crucial as cfDNA screening advances. Real-world screening outcomes must be carefully evaluated to inform appropriate clinical responses and address open research questions. The present disclosure highlights the potential of purpose-built datasets, supplementary annotations, representation learning, dynamic tracking, and stage shifting modeling to advance blood-based cancer screening methodologies.

[0097] The present disclosure demonstrates the feasibility and effectiveness of using ML models such as deep learning approaches (e.g., including one or more neural networks) to detect cancer signals in cfDNA sequencing data. For example, a high-performance multi-cancer classifier capable of detecting various cancer types with promising sensitivity and specificity metrics is described. At least one embodiment shows potential for widespread early screening using liquid biopsies. Further embodiments address challenges related to dataset biases, biological factors, and performance variability across different cancer types. Still further embodiments incorporate supplementary metadata and leverage advanced techniques, such as cell type tracing and large language models for cfDNA “language,” thereby enhancing performance and applicability in clinical settings. The present disclosure underscores the importance of continued research and development efforts to improve blood-based cancer screening methodologies, with the ultimate goal of enabling earlier detection and better outcomes for patients.

[0098] Embodiments of the disclosure relate to one or more exemplary features:

[0099] • An end-to-end learning framework operates directly on raw sequencing data, eliminating the need for manual feature engineering. This allows one or more neural networks (e.g., one or more transformer models, transformer encoder models, and / or CFormer) to capture complex genomic patterns and long-range dependencies that may be missed by traditional bioinformatic or feature engineering based machine learning approaches.

[0100] • A unified screening test may be used, which is capable of detecting multiple cancer types simultaneously, reducing the occurrence of false positive results.

[0101] • Embodiments leverage expansive genomic context by analyzing cfDNA fragments across a broad genomic range, aggregating signals across the wider landscape across the genome.

[0102] • Embodiments require significantly fewer sequencing reads compared to existing techniques, making them more cost-effective and scalable for clinical use.

[0103] • Capability to predict patient response to cancer treatment therapies. By leveraging the power of deep learning models to capture complex patterns and long- range dependencies in sequential data, embodiments of the disclosure have the potential to unlock new insights from raw cfDNA sequencing data, paving the way for accurate and non-invasive early cancer detection and treatment response prediction from liquid biopsies.

[0104] One or more differences of at least one embodiment described herein over prior approaches include:

[0105] • Reliance on raw reads (e.g., random or consecutive fragments) as the core input — in contrast to engineered summary statistics (e.g. bin counts) as done by some prior approaches.

[0106] • Inclusion of read-level metadata (e.g., quality scores, mapping position, epigenetic or methylation marks, gene-region annotations) in a single, end-to-end deep learning framework.

[0107] • Potential to detect multiple cancer types simultaneously, and / or predict therapy response, using a unifying end-to-end architecture.

[0108] • Elimination of the need for handcrafted features (e.g., copy-number profiles, fragmentation metrics, etc.), potentially resulting in superior accuracy at low coverage and / or enabling detection across broader genomic landscapes.

[0109] One or more potential advantages of at least one embodiment described herein, which are not provided by prior approaches (which suffer from low sensitivity, high false positive rates, restricted genomic coverage, and / or reliance on manual feature engineering) include:

[0110] • Reduction or avoiding Manual Feature Engineering: The end-to-end learning framework bypasses labor-intensive and expertise-dependent traditional bioinformatics.

[0111] • Unified Screening Test: Providing a unified test capable of simultaneously detecting multiple cancer types, reducing false positives.

[0112] • Expansive Genomic Context: Leveraging broad genomic context by analyzing cfDNA fragments across a wide genomic range.

[0113] • Reduced Sequencing Requirements: At least one embodiment requires significantly fewer sequencing reads compared to existing techniques, making it more cost-effective and scalable for clinical deployment.

[0114] • High Accuracy: For multi-cancer classification, the ML model achieved a sensitivity of 79% at 85% specificity, improving to 86% for breast cancer in experiments conducted by Inventors, as described below in the “Examples” section. • Treatment Response Prediction: Ability to detect patients who will respond or not respond to a cancer treatment therapy.

[0115] • Robustness to Sequencing Depth: Performance of the ML model remained consistent even when reducing sequencing depth from 2,000,000 reads to 80,000 reads per sample, indicating robustness and cost-effectiveness for cancer screening.

[0116] Some improvements over existing approaches, and / or alternatives over existing approaches, provided by one or more embodiments include:

[0117] • One prior approach appears to classify cancer using size-selected cfDNA reads (<160 nt) to compute genome-wide copy-number at loci, and applies a classifier to these copy-number features to predict cancer class, using aggregated CNV features. In contrast, at least one embodiment uses raw read sequences and optional quality as input rather than copy-number. It does not filter by fragment length in the same way, and does not rely on copy-number profiles. The other approach does not teach or relate to using neural nets on raw data.

[0118] • Another approach appears to perform Al-based cancer diagnosis: extract DNA, align reads, generate vectorized data from aligned fragments (e.g. counts per bin), then feed to trained Al model. Specifically the approach vectorizes aligned reads (hand- engineered features). In contrast, at least one embodiment uses the unaltered raw reads themselves. The other approach performs traditional pre-processing (alignment, counting). . In contrast, at least one embodiment feeds bases, and optional qualities, etc. directly. The other approach does not teach or relate to per-base info like quality scores or methylation.

[0119] • Yet another approach detects chromosomal abnormalities via Al: align cfDNA reads, generate grand canyon plot (GC plot) from read counts or fragment distances per genomic bin, and input to neural net to decide aneuploidy. It uses Al on sequencing data but for aneuploidy (prenatal) not cancer class. Relies entirely on binned aggregate features (GC plots). Does not use raw sequence content or quality. In contrast, at least one embodiment uses neural nets on raw reads.

[0120] • Yet another approach performs cancer detection / monitoring from cfDNA: sequence DNA, then use ML / AI to assign reads as ctDNA based on patient- specific mutations and general ctDNA features (e.g. coverage, fragment ends, epigenetic markers). This approach employ ML, but its methodology is feature-driven: identifying known somatic mutations or methylation patterns to flag ctDNA reads. In contrast, at least one embodiment relies on letting the network learn patterns from raw nucleotides and metadata. The other approach is based on supervised by known mutations; while in contrast, at least one embodiment does not require prior knowledge of a tumor mutation.

[0121] • Yet another approach applies ML to sequencing for copy-number analysis: align reads and count per genomic segment to form input for ML classifier; variations include ML- based alignment / binning. This other approach uses ML on sequence data, but in context of prenatal CNV detection. It mainly still uses aggregate counts per bin (alignment and counting). In contrast, at least one embodiment does not perform explicit counting / bias correction; instead, it inputs raw-sequence features.

[0122] Before explaining at least one embodiment of the invention in detail, it is to be understood that the invention is not necessarily limited in its application to the details of construction and the arrangement of the components and / or methods set forth in the following description and / or illustrated in the drawings and / or the Examples. The invention is capable of other embodiments or of being practiced or carried out in various ways.

[0123] The present invention may be a system, a method, and / or a computer program product. The computer program product may include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present invention.

[0124] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium may be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A non-exhaustive list of more specific examples of the computer readable storage medium includes the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire. Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network may comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0125] Computer readable program instructions for carrying out operations of the present invention may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate arrays (FPGA), or programmable logic arrays (PLA) may execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present invention.

[0126] Aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer readable program instructions.

[0127] These computer readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0128] The computer readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process, such that the instructions which execute on the computer, other programmable apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0129] The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustration, and combinations of blocks in the block diagrams and / or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts or carry out combinations of special purpose hardware and computer instructions.

[0130] Reference is now made to FIG. 1, which is a block diagram of components of a system 100 for classification of cfDNA sequencing data 124 into a medical state of a subject, in accordance with some embodiments of the present invention. Reference is also made to FIG. 2, which is a flowchart of a method of classifying cfDNA sequencing data into a medical state of a subject, in accordance with some embodiments of the invention. Reference is also made to FIG. 3, which is a flowchart of a method of training a ML model for classifying cfDNA sequencing data into a medical state of a subject, in accordance with some embodiments of the present invention. Reference is also made to FIG. 4, which is a schematic 402 depicting a high-level overview of a process for determination of a medical state of a subject based on feeding raw cfDNA sequencing data into a ML model, in accordance with some embodiments of the present invention. Reference is also made to FIG. 5, which is a dataflow diagram 502 of an exemplary dataflow including a read representation and aggregation to the ML model and the aggregation method and analysis of the resulting scores from the ML model, in accordance with some embodiments of the present invention. Reference is also made to FIG. 6, which is a schematic 602 depicting an exemplary approach of the read aggregation process, demonstrating how multiple samples from a single patient are processed and their classification results combined to yield a final patient- level decision, in accordance with some embodiments of the present invention. Reference is also made to FIG. 7, which is a table 702 presenting a tabular summary the distribution of patient samples across various cancer types and healthy controls in an experiment conducted by the Inventors, in accordance with some embodiments of the present invention. Reference is also made to FIG. 8, which includes plots 802A-B of unsupervised clustering of datasets of subjects with different cancers indicating the ML model's ability to learn meaningful representations from raw cfDNA data that effectively separate healthy samples from cancer samples, even for data not seen during training, in accordance with some embodiments of the present invention. Reference is also made to FIG. 9, which is a classification matrix 902 of an ensemble ML model for pancreatic cancer classification, in accordance with some embodiments of the present invention. Reference is also made to FIG. 10, which is a box plot 1002 of “Cancer Score" for multiple cancer classification, in accordance with some embodiments of the present invention. Reference is also made to FIG. 11, which includes AUC plots 1102 showing good classification for early and late-stage cancers, in accordance with some embodiments of the present invention. Reference is also made to FIG. 12, which includes graphs 1202 comparing classification performance of a traditional liquid biopsy method to at least one embodiment, in accordance with some embodiments of the present invention. Reference is also made to FIG. 13, which includes graphs 1302 depicting a comparison based on at least one embodiment with traditional liquid biopsy methods in low-coverage sequencing, in accordance with some embodiments of the present invention. Reference is also made to FIG. 14, which includes correlations 1402 between cancer prediction of the ML model and traditional computational methods, in accordance with some embodiments of the present invention. Reference is also made to FIG. 15, which includes a PCA plot 1502 and a t-SNE plot 1504 demonstrating the generalization performance of the approach described herein, in accordance with some embodiments of the present invention. Reference is also made to FIG. 16, which includes a classification performance report 1602 of separating responders and non- responders to a cancer therapy of patients with high-grade B-cell lymphoma (HGBCL), in accordance with some embodiments of the present invention. Reference is also made to FIG. 17, which is a box plot 1702 of cancer scores for samples of breast cancer therapy responders and nonresponders, in accordance with some embodiments of the present invention.

[0131] Referring now back to FIG. 1, system 100 may implement the acts of the methods described herein, by processor(s) 102 of a computing environment 104 executing code instructions 106A stored in a data storage device 106 (also referred to as a memory and / or program store).

[0132] Computing environment 104 may be implemented as, for example, a client terminal, a server, a single computer, a group of computers, a computing cloud, a virtual server, a virtual machine, a mobile device, a desktop computer, a thin client, a Smartphone, a Tablet computer, a laptop computer, a wearable computer, glasses computer, and a watch computer.

[0133] Multiple architectures of system 100 based on computing environment 104 may be implemented.

[0134] In an exemplary implementation of a centralized architecture, computing environment 104 storing code 106A, may be implemented as one or more servers (e.g., network server, web server, a computing cloud, a virtual server) that provides services (e.g., one or more of the acts of the methods described herein) to one or more client terminals 112 and / or to one or more sequencing devices 122 and / or to one or more servers 120 over a network 114, for example, providing software as a service (SaaS), providing software services accessible using a software interface (e.g., application programming interface (API), software development kit (SDK)), providing an application for local download, and / or providing functions using a remote access session such as through a web browser. For example, computing environment 104 receives cfDNA sequencing data 124 generate by sequencing device 122 from a sample of tissue of a subject. cfDNA sequencing data may be received, optionally over a network 114, for example, from sequencing device 122, from client terminal 112 and / or from server 120. cfDNA sequencing datal24 may be used to generate a ML model training dataset 116B for generating a trained ML model 116C, and / or fed into trained ML model 116C, as described herein. Multiple users may use their respective client terminals 112 to access computing environment 104, which may be remotely located. Optionally, each client terminal 112 and / or server 120 and / or sequencing device 122 may provide its own cfDNA sequencing 124 (e.g., from different subjects) for feeding into the trained ML model 116C running on computing environment 104, for example, via the API, and / or via an application locally installed on client terminal 112, and / or by another file transfer protocol. Computing environment 104 may centrally pre-processes cfDNA sequencing data 124 of different subjects. Computing environment 104 may centrally sample raw cfDNA reads from the sequencing data 124. Alternatively, the sampling is done by respective client terminal(s) 112 and provided to computing environment 104. Computing environment 104 may centrally compute a multivariate input representation from the samples, associated quality values 126, and / or metadata (e.g., as described herein). Alternatively, the computation of the multivariate input representation is done by respective client terminal(s) 112 and provided to computing environment 104. Computing environment 104 may centrally feed the multivariate input representation into trained ML model 116C to generate an outcome. Multiple outcomes obtained for different samples may be aggregated to obtain an indication of a medical state of the subject, as described herein. Computing environment 104 may provide the determined medical state to the corresponding client terminal 112 and / or server 120 (based on respective cfDNA sequencing data 124) for example, for presentation on a display associated with client terminal 112 and / or for storage in an electronic health record.

[0135] In another example of a localized architecture, computing environment 104 may include locally stored software (e.g., code 106A) that performs one or more of the acts described with reference to methods described herein, for example, as a self-contained system such as a laboratory server (e.g., 120) in communication with sequencing device 122. Code 106A may be implemented as a plug-in and / or additional feature set for integration with existing software that controls sequencing device 122. For example, cfDNA obtained in a tissue sample (e.g., blood) is sequenced by sequencing device 122, and cfDNA sequencing data 124 is obtained. The cfDNA genetic sequences may be locally pre-processed and / or locally sampled. A multivariate input representation may be centrally computed from the samples, associated quality values 126, and / or metadata (e.g., as described herein). The cfDNA multivariate input representation is fed into ML model 116C for obtaining one or more outcomes for one or more samples, which may be aggregated to obtain an indication of a medical state of the subject, as described herein. The medical state of the subject may be presented on a display (e.g., 108) of computing environment 104, stored in data repository 116, forwarded to another device (e.g., server 120), and the like.

[0136] ML model training dataset 116B may be centrally and / or locally created based on multivariate input representations including one or more of: samples of cfDNA sequencing data obtained from one or more sequencing devices 122, quality scores 126 associated with the sequenced nucleotides and / or metadata (e.g., as described herein). ML model 116C may be centrally and / or locally trained based on the centrally and / or locally obtained ML model training dataset 116B. For example, a central ML model training dataset 116B is created from different tissue samples obtained from different subjects which are sequenced at different sequencing devices 122, for example, in different cities, countries, and the like. A general trained ML model 116C may be created by training on the ML model training dataset 116B. In another example, specialized and / or personalized ML model training datasets 116B are created, for example, per anatomical location of the sample (e.g., blood, plasma), and / or per patient population (e.g., elderly, children, patients with specific cancer), and / or per geographical location (e.g., healthcare facility, city). Respective specialized and / or personalized ML models 116C may be created by training on respective specialized and / or personalized ML models training datasets 116B.

[0137] Processor(s) 102 of computing environment 104 may be implemented, for example, as a central processing unit(s) (CPU), a graphics processing unit(s) (GPU), field programmable gate array(s) (FPGA), digital signal processor(s) (DSP), and application specific integrated circuit(s) (ASIC). Processor(s) 102 may include multiple processors (homogenous or heterogeneous) arranged for parallel processing, as clusters and / or as one or more multi core processing devices. Processor(s) 102 may be arranged as a distributed processing architecture, for example, in a computing cloud, and / or using multiple computing devices. Processor(s) 102 may include a single processor, where optionally, the single processor may be virtualized into multiple virtual processors for parallel processing, as described herein.

[0138] Data storage device 106 stores code instructions executable by processor(s) 102, for example, a random access memory (RAM), read-only memory (ROM), and / or a storage device, for example, non-volatile memory, magnetic media, semiconductor memory devices, hard drive, removable storage, and optical media (e.g., DVD, CD-ROM). Storage device 106 stores code 106 A that implements one or more features and / or acts of the method described with reference to methods described herein when executed by processor(s) 102.

[0139] Computing environment 104 may include a data repository 116 for storing data, for example, storing one or more of a repository set for storing cfDNA sequencing data 116A, ML model training dataset 116B created as described herein, trained ML model 116C created as described herein and / or used as described herein, and / or other repositories of data described herein (not shown). Data repository 116 may be implemented as, for example, a memory, a local harddrive, virtual storage, a removable storage unit, an optical disk, a storage device, and / or as a remote server and / or computing cloud (e.g., accessed using a network connection).

[0140] Computing environment 104 may include a network interface 118 for connecting to network 114, for example, one or more of, a network interface card, a wireless interface to connect to a wireless network, a physical interface for connecting to a cable for network connectivity, a virtual interface implemented in software, network communication software providing higher layers of network connectivity, and / or other implementations. Network 114 may be implemented as, for example, the internet, a local area network, a virtual private network, a wireless network, a cellular network, a local bus, a point to point link (e.g., wired), and / or combinations of the aforementioned.

[0141] Computing environment 104 and / or client terminal(s) 112 and / or server(s) 120 include and / or are in communication with one or more physical user interfaces 108 that include a mechanism for a user to enter data (e.g., provide the cfDNA sequencing data 124 for preprocessing and / or as input into trained ML model 116C) and / or view the determined medical state of the subject, optionally within a GUI. Exemplary user interfaces 108 include, for example, one or more of, a touchscreen, a display, a keyboard, a mouse, and voice activated software using speakers and microphone.

[0142] Referring now back to FIG. 2, at 202, a ML model is trained and / or accessed. Additional exemplary details of training the ML model are described with reference to FIG. 3.

[0143] The ML model may be implemented as, for example, a neural network encoder, optionally a transformer encoder-based classification model, with a classification head. The transformer encoder model may exclude a decoder component and / or may exclude positional encodings.

[0144] The ML model may include an ensemble of two or more independently trained ML models, where the final classification (e.g., probability) is computed by averaging or voting on the outputs of the two or more independently trained ML models.

[0145] The encoder is designed to receive the raw reads of the cfDNA and compute a latent representation, optionally a vector.

[0146] The ML model is designed for receiving an input of the raw cfDNA read directly, excluding manual extracted features and / or excluding an alignment to a reference genome.

[0147] The transformer encoder architectures of the ML may represent an exemplary implementation due to their proficiency in handling sequences and capturing relationships within data, and represent an exemplary implementation due to their scalability and / or contextual sensitivity. However, the deep learning model may alternatively be implemented other architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or hybrid combinations thereof. It is noted that other architectures may be used, for example, a deep learning classifier, one or a combination of convolutional neural networks (CNNs), recurrent neural networks (RNNs) and / or transformers, optionally with a classification head, and / or other transformer, convolutional, or recurrent architectures.

[0148] Optionally, the ML model is designed for modular interchangeability between different deep learning architectures, including for example, transformer encoders, CNNs, and RNNs.

[0149] The ML model may be trained using an end-to-end deep learning method. The ML model is designed to accept as input raw cfDNA sequencing data for direct analysis of thereof.

[0150] Optionally, the ML model is designed for receiving the multivariate input representation directly, without requiring prior manual feature extraction and / or generation of summary genomic features based on alignment to a reference genome.

[0151] Optionally, the ML model is designed to learn an internal representation of the cfDNA reads that enables unsupervised clustering of samples according to the medical state.

[0152] Neural networks, including transformers, excel at capturing intricate patterns and long- range dependencies in sequential data like genomics. One or more neural networks as described herein may classify a patient's full set of sampled reads, optionally simultaneously, enabling the ML model to integrate signals from distinct genomic regions. By learning directly from raw sequencing data, including nucleotide sequences, optionally including quality values, and / or metadata representations, a neural network optimizes feature extraction and model training jointly, offering superior prognostic predictions compared to traditional methods.

[0153] The neural network encoder-based classification model may be designed to accept, as an input, a multivariate sequence data (as described herein), including nucleotide sequences, and / or quality values, and / or metadata representations. The metadata may be generated based on the mapping location of the read. However, in some embodiments, nucleotides and / or quality values could be used without mapping of the reads to the genome.

[0154] Optionally, the ML model, optionally the neural network encoder-based classification model, lacks (i.e., excludes) a decoder component.

[0155] The ML model, optionally the neural network encoder-based classification model, may include a classification token (CLS token) added, as a first token, to an input sequence of raw cfDNA reads for feeding into the ML model. The CLS token may aggregate information across all other tokens, providing an effective mechanism for sequence-level outputs. This token representation may aggregates information across all other tokens in the sequence through the selfattention mechanism described herein. The CLS token classification head may be used for sequence classification tasks, for providing an effective mechanism for aggregating sequence representations and / or producing informative sequence-level outputs.

[0156] For sequence classification tasks, a classification output layer is attached to the embedding of the CLS token after a final encoding layer of the ML model, optionally the final Transformer encoder layer. Mathematically, if the input representation of the CLS token after the final Transformer block is hCLS 6 RA(dmodel), then the classification logits z are generated by: z = WhCLS Here, W 6 RA(Ncxdmodel) is a trainable linear classification layer with Nc being the number of target classes. This offers an effective architecture for sequence classification by designating the CLS token as an aggregate sequence representative. The self-attention layers allow incorporating context from the entire input via this single CLS token. The training loss over z then optimizes hCLS as well as the Transformer encoder blocks to produce an informative sequencelevel representation. During inference, the predicted class y is obtained using the softmax probability over z. The CLS token head thus provides an end-to-end mechanism for sequence classification in Transformers.

[0157] Optionally, layer normalization is applied between each block of the ML model, optionally the neural network encoder-based classification model, to stabilize hidden state dynamics and / or and facilitate gradient flow. Layer normalization may be performed by normalizing activations at each layer, thereby assisting gradient flow during optimization, particularly in deep models like transformers.

[0158] Optionally, positional encodings is excluded from the ML model, optionally from the neural network transformer encoder-based classification model. The positional encodings may be excluded from the transformer architecture based on the understanding that the input sequence that includes randomly sampled and shuffled cfDNA reads, includes independent data instances without inherent ordering and / or temporal significance, and that positional bias could negatively impact model convergence.

[0159] Various positional encoding mechanisms have been developed to provide order information to at least one neural network (e.g., the Transformer model). However, according to embodiments of the disclosure, the model further omits positional encodings from the Transformer architecture. As a large set of independent data instances are classified without any inherent ordering, the positional information may be unnecessary. The input sequence consists of shuffled, discrete items with no temporal or positional significance. The goal is to learn representations of each item based on co-occurrence statistics. Additional positional bias can negatively impact convergence. Hence, the ML model does not necessarily rely on positional embeddings. This allows the ML model to internally learn useful data relationships without explicit positional encoding schemes.

[0160] To manage the significant computational demands of processing long input sequences (e.g., 4,096 reads per sample), FlashAttention or other corresponding process for transformer selfattention focused on speed and memory efficiency may be implemented. The FlashAttention or other corresponding process may reduce memory and / or reduce computational complexity from O(nA2) to O(kn). The ML model may include self-attention layer(s). One advantage of self-attention compared to other layer types can be as follows: (1) It can model long-range dependencies between items in a sequence equally well, using fewer parameters than fully connected layers, which is critical for application of a long input sequence of 512 reads. (2) Convolutional layers have a limited context window. Stacking many convolutional layers is needed to understand whole sequence of reads. Self-attention directly models relationships between all reads in the sequence, regardless of distance, with a small constant number of layers. (3) It performs parallel computations for all read pairs in the sequence. This computational efficiency and full context modeling allows it to capture long-range dependencies better than sequential recurrent layers like LSTMs. The order of input sequence elements doesn’t matter in self-attention, which works with embodiments of the disclosure that are agnostic on the order of sequences, but prefer the classification of a big amount of random reads at once.

[0161] Attention mechanisms are at the core of transformer-based models, but the default implementation can be computationally expensive. The standard transformer’s attention layer consumes memory proportional to the square of the input sequence size, represented as O(nA2), where ‘n’ is the size of the input sequence. This quadratic scaling can become a bottleneck when dealing with large datasets or input sequences, as it significantly increases memory requirements and computation time. To address this challenge, some embodiments may employ FlashAttention, an efficient drop-in replacement for standard attention introduced in PyTorch 2.0.

[0162] The efficiency gains of FlashAttention arise from its high-performance mixed precision and kernel fusion optimizations for faster attention computation. FlashAttention utilizes a few techniques to optimize the standard Transformer attention. Firstly, it employs low-bit- width matrix multiplication kernels down to FP8 precision for faster computation. Additionally, it reduces peak memory consumption via gradient checkpointing, recomputing activations instead of storing them all during the backward pass. Together these accelerate the feedforward propagation during training. At inference time, further speedup is obtained by kernel fusion, specifically, combining multiple matrix operations including attention, feedforward networks, and layernorm into one consolidated kernel.

[0163] Mathematically, FlashAttention uses a low-rank decomposition to approximate the standard dot product attention. For an input sequence of length n, the attention matrix A G RA(nxn) is approximated by: A ~ PQT P = XWP G RA(nxk) Q = XWQ G RA(nxk) Here, X refers to the combined query-key input after projection, and WP, WQ are projection parameter matrices to a k- dimensional latent space where k « n. This formulation reduces both memory and compute complexity from O(nA2) to O(kn). The outputs are then rescaled by the factors in WP and WQ appropriately to restore the magnitude for the downstream network. Overall, FlashAttention offers an optimized drop-in replacement for standard Transformer attention to improve efficiency.

[0164] At 204, raw cfDNA sequencing data obtained from a biological sample of a subject (blood or plasma) is received and / or accessed.

[0165] The raw cfDNA sequencing data may be represented using digital representations for each nucleotide in the sequence, for example, A, T, C, and G, for the four different types of base nucleotide types.

[0166] The raw cfDNA sequencing data may be generated by an automated sequencing device that analyzes the physical sequences of nucleotides on the physical cfDNA sequences.

[0167] The total number of nucleotides of the raw cfDNA sequencing data may be, for example, about 2,000,000 - 80,000,000. This large number is impractical for directly feeding into the ML model, as described herein.

[0168] At 206, the raw cfDNA sequencing data may be pre-processed.

[0169] The raw cfDNA sequencing data may be standardized to facilitate uniform data handling and preprocessing.

[0170] The pre-processing may be done by assigning each cfDNA nucleotide in the cfDNA sequencing data a value indicating its type, where the representation of the value improves computational efficiency and / or improves storage efficiency and / or facilitates simultaneous classification of a larger amount of cfDNA reads.. For example, each one of the four different types of nucleotides (i.e., A, T, C, and G) is assigned a corresponding unique value, for example, a number between zero and three (0, 1, 2, 3), between one and four (1, 2, 3, 4) or other values. The number may be represented in binary notation, for example, 00, 01, 01, and 11.

[0171] The initial raw cfDNA sequencing data may be pre-processed to transform and / or standardize for downstream analysis. Raw cfDNA sequencing data may be processed into BAM files, which may facilitate integration and comparison across multiple sources. These BAM files may be subsequently converted into optimized binary files, for reducing memory requirements and / or enabling rapid data loading.

[0172] At 208, a subset of raw cfDNA reads is sample from the raw cfDNA sequencing data according to a sampling process.

[0173] The subset forms a sample.

[0174] The sampling may improve handling the vast amount of cfDNA data efficiently while ensuring optimal ML model performance. The sampling may promote efficient utilization of processing resources (e.g., processor, memory, processor utilization, processing time) without significant loss of information. Optionally, the sample of the subset represents raw sequencing data of cfDNA without additional biological contextualization.

[0175] The sampling process defines how the sampling is performed. Examples of the sampling process include: randomly reading, wherein the subset of raw cfDNA reads of the sample are mapped to random locations of a reference genome, consecutively reading the subset of raw cfDNA reads of the sample optionally mapped to a randomly selected region on a reference genome, and / or sampling to obtain a total number of reads less than about 100,000.0ptionally, the sampling process may define reading the subset of raw cfDNA reads of the sample in a random order.

[0176] Alternatively or additionally, the sampling process may define: mapping the raw sequences to multiple regions of a reference genome, randomly selecting a region on the reference genome, and consecutively reading the sample of the subset of raw cfDNA reads mapped to the randomly selected region on the reference genome.

[0177] Optionally, the sampling process defines an amount of reads to include in the sample. The amount may be an absolute value, for example, a number of the raw cfDNA reads of the subject to include in the same is selected from a range of about 500-5000, or about 100-1000, or about 2000-5000, or other ranges, or for example, about 512, or about 1024, or about 4096, or other values. Alternatively or additionally, the sampling process defines the number of reads to include in the sample as a percentage of the raw sequences, for example, about 0.001-5%, or about 0.001- 1%, or other values.

[0178] Optionally, the sampling process defines a total number of reads to include in multiple iterations of sampling, for computing the final subject-level classification and / or probability at a performance of the ML above a threshold of one or more performance metrics. Inventors discovered that the number of total reads may be significantly less than in other known approaches, for example, less than about 250000, or 200000, or 150000, or 100000, or 75000, or 50000, or other values.

[0179] Optionally, the sampling process defines a fixed size of each of the raw cfDNA reads that is common to all samples. The fixed size is selected from a range of, for example, about 50-500, or about 50-350, or other values. The number may be based on the technical specification of the sequencing machine.

[0180] Optionally, the raw cfDNA sequences of the sample may be arranged in parallel, for example, into one or more larger data structures, for example, a single long sequence, a set of long sequences, a two dimension (2D) array (e.g., “image”), a three dimensional (3D) array (e.g., “volume), and the like. Optionally, the sampling process defines selecting the subset as consecutive groups aligned and / or mapped to a region in the reference genome. Alternatively, the sampling process defines selecting a random genetic region of the reference genome, and extracting closely mapped reads from the random region.

[0181] Optionally, the sampling process defines mapping each raw sequence to a corresponding genomic region of a defined reference genomic sequence. A random genomic region may be selected from the multiple genomic regions of the genomic sequence. The sampled subset of raw cfDNA reads may be mapped to the random genomic region and sequential according to a natural genomic order.

[0182] Optionally, the sampling process defines a sampling factor, denoted Ns, to (optionally randomly) select a predefined number (e.g., about 100-10000, or about 2000-5000, or about 4096) of reads from each sector of the dataset of cfDNA sequences, simulating the total reads from sequencing.

[0183] At 210, the quality score (also referred to herein as quality value) and / or other confidence score associated with the automated sequencing is received (e.g., accessed).

[0184] Optionally, the automated sequencer device that generates the raw sequences from the physical nucleotides of the cfDNA fragments generates a quality value (or score) indicating the accuracy of a base call, i.e., how confident the sequence is that a specific nucleotide - A, T, C, or G, is correct, for example, a Phred quality score (Q).

[0185] The quality score associated with each sequenced nucleotide may be fed into the ML model in combination with the raw cfDNA sequences of the subset, as described herein.

[0186] At 212, one or more metadata may be received (e.g., accessed). The metadata may be a metadata representations of genomic regions to which the reads of the sample map to.

[0187] Metadata may be associated with respective raw cfDNA reads of the subset.

[0188] The corresponding metadata may be fed into the ML model in combination with each respective raw cfDNA read of the subset.

[0189] Metadata values may be based on a mapping of the subset of reads to a reference genome prior to inference.

[0190] The metadata representations may be derived from genome alignment or other external annotations, for example, gene regions or other gene-region mappings, methylation patterns and / or status, and the like.

[0191] Optionally, each raw sequence is mapped to a corresponding genomic region of a defined reference genomic sequence. Corresponding metadata, including empirically derived information about the corresponding genomic region, is generated for each respective raw cfDNA read. The metadata for each respective raw cfDNA may indicate whether the respective raw cfDNA maps to a coding region, a methylated region, or a region with known disease-relevant annotations.

[0192] The metadata may be obtained (e.g., extracted, received from) external annotation sets, for example, known methylation sites and / or gene boundaries from public databases like UCSC Genome Browser or methylation atlas.

[0193] At 214, a multivariate (multi-dimensional) input representation may be generated, and fed into the ML model. The multivariate input representation includes one or more of: (i) the nucleotide sequences of the sample of the subset of raw cfDNA reads or encoding thereof (e.g., into the numbers 0-3 as described herein), (ii) the quality values associated with the nucleotide sequences of the sample, and / or (iii) the metadata, as described herein.

[0194] In some embodiments, the multivariate input representation includes the nucleotide sequences of the sample and associated quality scores, and excludes the metadata. Such embodiment may enable inference directly from minimally processed raw reads. In other embodiments, the multivariate input representation includes the nucleotide sequences of the sample and the metadata and excludes the quality scores. The selected implementation may be based on, for example, available preprocessing resources, available data, desired information, and the like.

[0195] Different combinations of components of the multivariate input representation are possible, including:

[0196] • The multivariate input representation includes the nucleotide sequences of the sample, the associated quality values, and the metadata representations.

[0197] • The multivariate input representation includes the nucleotide sequences of the sample and the associated quality values, and excludes the metadata representations.

[0198] • The multivariate input representation includes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

[0199] • The multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and includes the metadata representations.

[0200] • The multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and excludes the metadata representations.

[0201] • The multivariate input representation excludes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations. The multivariate input representation may exclude alignment-based feature engineering and / or calculation of summary genomic features and / or other hand-crafted features directly computed from the raw cfDNA sequencing data.

[0202] At 216, the multivariate input representation of the sample is fed into the ML model.

[0203] Optionally, the raw cfDNA reads of the sample are fed as a single (or two or more large) data structure that may be a concatenation of multiple individual raw cfDNA reads, for example, the single long sequence, the “image”, or the “volume”, as described herein. The raw cfDNA reads may be aggregated into a vector form, and the vector is fed into the ML model.

[0204] The ML model may be designed to accept the multivariate input representation directly. The ML model (e.g., neural network encoder-based classification model) may implement FlashAttention of other corresponding process for efficient processing of long input sequences, as described herein.

[0205] At 218, an indication of the medical state of the subject is obtained from the ML model. The indication may include a classification category of the medical state selected from multiple candidate classification models. Alternatively or additionally, a probability (or other indication of likelihood) of the medical state of the subject identified by the ML model (e.g., classification category) is generated.

[0206] Optionally, a representation vector is obtained from the ML model, and the probability is computed from the representation vector.

[0207] The indication of the medical state is generated for the respective sample of the subset of raw cfDNA fed into the ML model.

[0208] Examples of the medical condition include:

[0209] • Cancer, for example, cancerous or non-cancerous.

[0210] • Different types of cancer, for example, lung, colon, breast, prostate, and the like. The indication and / or probability may be generated for each type.

[0211] • Distinguishing between different cancer sub-types, for example, between different types of lymphoma such as diffuse large B-cell lymphoma (DLBCL) and high-grade B-cell lymphoma (HGBCL).

[0212] • A stage of the cancer, for example, early-stage or late-stage, or stages I, II, III, or IV.

[0213] • Predicting a patient's response to a therapy administered to treat a disease, optionally to treat cancer.

[0214] The subject may be treated with treatments known to be effective for specific cancers. For example: • For cancer - chemotherapy, targeted therapy immunotherapy, hormone therapy, monoclonal antibodies, radiopharmaceuticals, surgery, radiation therapy, ablation therapy, embolization, stem cell (bone marrow) transplant, photodynamic therapy, and nanoparticle therapy.

[0215] • For Diffuse Large B-Cell Lymphoma (DLBCL):

[0216] First-Line Treatments:

[0217] R-CHOP (Rituximab + Cyclophosphamide + Doxorubicin + Vincristine + Prednisone)

[0218] R-EPOCH (Rituximab + Etoposide + Prednisone + Vincristine + Cyclophosphamide + Doxorubicin).

[0219] R-ACVBP (Rituximab + Doxorubicin + Cyclophosphamide + Vindesine + Bleomycin + Prednisone)

[0220] • For Relapsed / Refractory DLBCL:

[0221] CAR-T Cell Therapy:

[0222] Axicabtagene ciloleucel (Yescarta)

[0223] Tisagenlecleucel (Kymriah)

[0224] Lisocabtagene maraleucel (Breyanzi)

[0225] Salvage Chemotherapy + Autologous Stem Cell Transplant:

[0226] R-ICE (Rituximab + Ifosfamide + Carboplatin + Etoposide)

[0227] R-DHAP (Rituximab + Dexamethasone + Cytarabine + Cisplatin)

[0228] R-GDP (Rituximab + Gemcitabine + Dexamethasone + Cisplatin)

[0229] Targeted Therapies:

[0230] Polatuzumab vedotin (anti-CD79b antibody-drug conjugate)

[0231] Tafasitamab (anti-CD19 monoclonal antibody)

[0232] Selinexor (XPO1 inhibitor)

[0233] • For High-Grade B-Cell Lymphoma (HGBCL):

[0234] R-CODOX-M / IVAC (alternating cycles):

[0235] Cycle A: Rituximab + Cyclophosphamide + Vincristine + Doxorubicin + Methotrexate

[0236] Cycle B: Ifosfamide + Etoposide + Cytarabine

[0237] R-EPOCH

[0238] R-Hyper-CVAD - Alternating with high-dose methotrexate and cytarabine

[0239] • For Relapsed / Refractory HGBCL:

[0240] CAR-T Cell Therapy (same as DLBCL) High-dose chemotherapy + Autologous SCT

[0241] Allogeneic stem cell transplant (in selected cases)

[0242] • Examples of Emerging / Investigational Therapies:

[0243] Bispecific antibodies: Mosunetuzumab, Glofitamab

[0244] BTK inhibitors: Ibrutinib, Acalabrutinib

[0245] BCL-2 inhibitors: Venetoclax

[0246] Checkpoint inhibitors: Pembrolizumab, Nivolumab Antibody-drug conjugates: Loncastuximab tesirine

[0247] At 220, features described with reference to 208-218 are iterated. In each iteration a different sample of different raw cfDNA reads of the subset are obtained and fed into the ML model. The ML model generates the indication of the medical state for each feeding.

[0248] At 222, the multiple indications generated by the ML model for multiple samples of subsets of raw cfDNA sample reads (e.g., generated during the iterations) may be aggregated into a single indication, optionally a single subject-level classification category and / or single probability (or other single indication of likelihood).

[0249] The multiple probabilities generated by the ML model may be aggregated.

[0250] Exemplary aggregating approaches include: Quantile-based aggregation, Mean, Median, and Majority.

[0251] The aggregation may be selected for increasing the signal of the final aggregated classification and / or probability, which improves classification outcomes, such as increasing the accuracy of the classification outcome.

[0252] At 224, the aggregated indication and / or probability is provided as the indication of that the subject is in the medical state.

[0253] The aggregated indication and / or probability, optionally the subject-level classification category, may be presented on a display, fed into another executing process, forwarded to a remote device, stored on a data storage device (e.g., within an electronic medical record of the subject), and the like.

[0254] Referring now back to FIG. 3, the ML model is trained. The training of the ML model is performed to correspond to the inference features described with reference to FIG. 2.

[0255] At 304, raw cfDNA sequencing data obtained from a biological sample of a respective individual (of multiple individuals) is received. Multiple biological samples are obtained from multiple individuals. The raw cfDNA sequencing data are obtained from one individuals (e.g., per iteration). As described herein, during iterations another set of raw sequences are obtained from a different individual.

[0256] Exemplary details of the raw cfDNA sequencing data are described, for example, with reference to 202 of FIG. 2.

[0257] At 306, the raw cfDNA sequencing data may be pre-processed, for example, as described with reference to 206 of FIG. 2.

[0258] At 308, a sample of a subset of raw cfDNA reads is obtained by sampling from the raw cfDNA sequencing data, for example, as described with reference to 208 of FIG. 2.

[0259] At 310, quality values associated with the raw cfDNA sequencing data of the sample are received, for example, as described with reference to 210 of FIG. 2.

[0260] At 312, metadata associated with the sample of the subset of raw cfDNA reads is obtained, for example, as described with reference to 212 of FIG. 2.

[0261] At 314, a multivariate input representation may be generated for the sample. The multivariate input representation includes one or more of: (i) the nucleotide sequences of the sample, (ii) associated quality values, and / or (iii) metadata representations of genomic regions to which the reads of the sample map to, for example, as described with reference to 214 of FIG. 2.

[0262] Different combinations of components of the multivariate input representation are possible, including:

[0263] • The multivariate input representation includes the nucleotide sequences of the sample, the associated quality values, and the metadata representations.

[0264] • The multivariate input representation includes the nucleotide sequences of the sample and the associated quality values, and excludes the metadata representations.

[0265] • The multivariate input representation includes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

[0266] • The multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and includes the metadata representations.

[0267] • The multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and excludes the metadata representations.

[0268] • The multivariate input representation excludes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations. At 316, an indication of the medical state of the subject is received, for example, manually entered by a user, extracted from an electronic medical record of the subject, and the like. Examples of medical states are described, for example with reference to 216 of FIG. 2.

[0269] At 318, a record is generated. The includes a respective multivariate input representation, and a ground truth label indicating the medical state of the subject, for example, a classification category selected from multiple classification categories.

[0270] The data included in the record may correspond to one or more data structures described with reference to FIG. 2, for example, the data may be represented in the multivariate input representation, in a single long sequence, the “image”, and / or the “volume”.

[0271] At 320, one or more features described with reference 304-318 may be iterated. Each iteration may be for a different sample of a different subset of raw cfDNA reads, and / or may be for a different individual.

[0272] A respective record may be created during each iteration.

[0273] At 322, a training dataset is generated from the multiple records.

[0274] At 324, the ME model is trained on the training dataset.

[0275] The ML model is trained on the training dataset for generating an indication (e.g., classification category, a probability and / or other indication of likelihood) that the subject is in the medical state, in response to an input of a target multivariate input representation including a sample of a subset of raw cfDNA reads of raw cfDNA sequencing data generated from a biological sample of a subject, sampled according to the sampling process, and target quality values associated with each nucleotide of the sample and / or target genomic data associated with at least one region of the sample.

[0276] Optionally, the trained ML model is provided with an associated aggregation process for aggregating multiple indications (e.g., classification categories, probabilities, other indications of likelihood) that the subject is in the medical state, obtained from the ML model for multiple multivariate input representations for multiple different samples of the subject, for computing a final subject-level classification and / or an aggregated probability that the subject is in the medical state.

[0277] Additional details of the aggregation is described, for example, with reference to 222 of FIG. 2.

[0278] Optionally, the training of the ML model includes a first stage of initializing the ML model by transfer-learning from a pre-trained model trained on a large dataset of genomic data, and a second state of training the ML model on the training dataset after the initialization during the first stage. The ML model training may be performed for a number of epochs (e.g., 15 in the experiments described herein) with a defined batch size (e.g., 8 in the experiments described herein), selected for balancing performance and / or utilization of computational resources. The training may be performed using an adaptive momentum optimizer, for example NAdam. The NAdam optimizer, is an adaptive momentum optimizer based on Adam with Nesterov accelerated gradients. Various optimizers may be used to prevent vanishing and / or exploding gradients, which may be significant for the self-attention layers in modeling long-range dependencies.

[0279] Exemplary evaluation metrics are now described. Robust evaluation metrics may be selected essential for developing and / or assessing predictive modeling performance, for example, for applications like cancer diagnosis. At least one embodiment of the disclosure utilizes a suite of complementary metrics to evaluate various aspects of classification performance during model development, selection, and final testing.

[0280] Maximizing recall (true positive rate) at a high level of specificity may be significant for cancer classification due to the risks associated with false negative and false positive predictions. A false negative, where the model fails to detect cancer in a patient who truly has it, can lead to delayed treatment and severe consequences. Conversely, a high false positive rate can result in unnecessary treatment or procedures for healthy patients incorrectly diagnosed with cancer. By setting a specificity threshold, for example, of 85%, at least one embodiment constrains false positives to an acceptable rate of, for example, 15%. At this fixed level of false positives, a ML model is selected that maximizes sensitivity or recall. This approach may optimize cancer detection rates without exceeding a reasonable false positive tolerance, which may align with critical objectives of accurate detection and / or responsible resource management in clinical settings. This metric can be used in various cancer classification methods. Achieving high recall at high specificity may be significant in healthcare diagnostics to accurately identify cancer cases while minimizing false alarms. This metric may optimize the trade-off between reliable cancer detection and acceptable false positive rates, ensuring comprehensive identification of genuine cases for timely interventions while mitigating misdiagnosis and unnecessary patient burden.

[0281] The Area Under The Curve of the Receiver Operating Characteristic (AUC-ROC) may be used to evaluate the ML model’s overall performance on the test set. AUC quantifies the ML model’s ability to discriminate between positive and negative classes across all possible classification thresholds. AUC scores range from 0 to 1, with a score of 1 representing perfect discrimination and a score of 0.5 representing random chance. For each classification threshold, the false positive rate (FPR) and true positive rate (TPR) are calculated as: All those values represent an ROC curve, and to summarize the results for all thresholds, at least one embodiment of the disclosure calculate the AUC: Z: Area Under Curve (AUC) = TPR x (FPR) x dFPR (1) The ROC curve is constructed by plotting the TPR against the FPR for all possible thresholds. The AUC is then computed as the integral of the TPR with respect to the FPR:Z: Area Under the Curve (AUC) = TPR x dFPR (2) The AUC may provide a comprehensive evaluation of the ML model’s discriminative ability across the entire range of classification thresholds, complementing the specific operating point chosen for recall at, for example, 85% specificity. By employing both recall at, for example, 85% specificity and AUC, a comprehensive understanding of the ML model’s performance may be gained, balancing the need for high cancer detection rates with acceptable false positive rates, while also assessing its overall discriminative ability. This rigorous evaluation approach is crucial for developing reliable and clinically applicable cancer detection models.

[0282] The Matthews correlation coefficient (MCC) is a metric that can be used in cancer detection to provide a balanced measure of performance, robust to class imbalance. MCC values range from -1 to +1, with +1 indicating perfect prediction, 0 indicating random prediction, and -1 indicating complete disagreement. Unlike accuracy, MCC uses all four categories of the confusion matrix (true positives, true negatives, false positives, and false negatives), making it suitable for imbalanced datasets. This metric was used for model selection during training.

[0283] The confusion matrix breaks down the ML model’s predictions into true positives, true negatives, false positives, and false negatives for each class, providing insights into areas of strong or weak performance. This analysis may be useful for debugging and / or understanding the ML model’s behavior across different cancer types.

[0284] The Fl-score is a standard classification metric that combines precision and recall into a single measure, calculated as the harmonic mean of the two. It provides a balanced evaluation of the ML model’s performance, complementing the other metrics used. By employing this comprehensive suite of evaluation metrics, a thorough understanding of the ML model’s performance may be gained, enabling optimization for high cancer detection rates while maintaining acceptable false positive rates.

[0285] Model selection was driven by tracking validation set MCC and loss during training. Checkpointing may be implemented to save ML model weights each time a new highest validation MCC and loss were achieved, facilitating recovery of optimal weights. The final ML model for testing may be selected based on the checkpoint with the highest validation set metrics.

[0286] Referring now back to FIG. 4, schematic 402 provides a high-level schematic overview of the end-to-end deep learning method for classification of a medical state of a subject, including forward pass visualization. Schematic 402 illustrates the process from obtaining raw sequencing data (e.g., FASTA / FASTQ files) from multiple patients to generating a final cancer (or other medical state) classification probability. Element 404 indicates raw sequencing reads (e.g., "read 1: ATGC...") obtained from various patients, for example, a sequencing file. These raw reads may undergo a Preprocessing step 406, transforming them into an optimized "binary format" for efficient handling. The preprocessed data may be fed 408 into a Deep learning model which is trained using patient information (e.g., Patient 1 - Lung cancer, Patient 3 - Healthy control). The ML model, optionally based on transformer encoder layer(s) and a classification head, may learn an internal representation for each patient. During the testing phase 410, new sample data is input into the trained ML model, which then outputs a Probability of cancer classification (e.g., Lung cancer 0.12, Colorectal cancer 0.81, Healthy 0.07). FIG. 4 depicts the direct use of raw sequencing data, bypassing traditional alignment and / or explicit feature engineering, and indicates the end-to- end nature of the deep learning pipeline.

[0287] Referring now back to FIG. 5, dataflow diagram 502 depicts the exemplary dataflow including the read representation and aggregation to the model and the aggregation method and analysis of the resulting scores from the ML model. Raw sequencing data 504, including individual reads (e.g., 300 nucleotides each), is shown as the initial input. From the raw sequence data, a random subset of N reads is selected 506 may be selected to form a Sample (represented as N x 300, #reads x read size). This sample is directly fed into a "Classifier Model" 508 which may include a "Transformer Encoder" 510 and a "Classification Head" 512. The Transformer Encoder 510 processes the raw read data (e.g., "CCTTTCTGCC," "TCCGGGGGCA") to learn intricate patterns. The Classification Head 512 then generates "Multiple classification scores (per sample)" 514. These per-sample scores are then subjected to an "Aggregation Method" 516 (e.g., Quantilebased aggregation, Mean, Median, Majority) to produce a "Final Classification Per-Patient" 518 (e.g., "Patient 1: Cancer 0.1 Healthy 0.9"). FIG. 4 underscores the direct processing of raw, unaligned reads, the use of a transformer encoder for feature learning, and the subsequent aggregation of sample-level predictions to derive a robust patient-level diagnosis.

[0288] Referring now back to FIG. 6, schematic 602 visually elaborates on the read aggregation process, demonstrating how multiple samples from a single patient are processed and their classification results combined to yield a final patient-level decision. For a given "Patient," 604 multiple "Sample" inputs 606 (e.g., each 4096 x 300 reads) are generated. Each of these samples is independently processed by the "Classifier Model" 608 which outputs a "Probability" score 610 for that specific sample. These individual sample probabilities are then fed into an "Aggregation Method" 612. Examples of aggregation methods are provided, including "Quantile-based aggregation," "Mean," "Median," and "Majority." The application of this aggregation method results in a "Patient Level Class" 614 providing a final, robust classification for the patient (e.g., "Cancer" or "Healthy"). FIG. 6 emphasizes the robust patient-level classification achieved by leveraging multiple sample-level predictions, thereby increasing the signal-to-noise ratio and improving diagnostic accuracy.

[0289] Various embodiments and aspects of the present disclosure as delineated hereinabove and as claimed in the claims section below find experimental and / or calculated support in the following examples.

[0290] EXAMPLES

[0291] Reference is now made to the following examples, which together with the above descriptions illustrate some embodiments of the invention in a not necessarily limiting fashion.

[0292] Technical Implementation. All experiments and model training were conducted using, e.g., a single Nvidia RTX 3090 GPU, with PyTorch and PyTorch Lightning library employed as the machine learning framework. The Hydra configuration manager abstracted away boilerplate code, streamlining model building for the research tasks.

[0293] Experimental results from multi-cancer and single cancer models, demonstrating the effectiveness of embodiments in detecting various cancer types from cfDNA sequencing data of previously published clinical cancer datasets as well as other samples, are described. Results demonstrating the ML model's ability to generalize to unseen datasets and predict patient response to therapy are also presented.

[0294] At least one embodiment described herein was applied to multiple clinical datasets containing sequencing data from cancer patients and healthy individuals, including down- sampled low-coverage samples simulating ultra-low-cost sequencing. For multi-cancer classification, the model achieved 79% sensitivity at 85% specificity. Using a breast cancer- specific model improved the sensitivity to 86% at 85% specificity for detecting breast cancer. Remarkably, the ML model’s performance remained consistent across various sequencing depths, unlike traditional methods that suffer from reduced accuracy with lower coverage. At least one embodiment has also demonstrated the capability to predict patient response to cancer treatment therapies and to distinguish between different cancer subtypes.

[0295] Results

[0296] The performance of various embodiments were evaluated using a large dataset combining several publicly available sequencing datasets from published studies along with a dataset from the Shomron Lab containing breast cancer and healthy patient samples (referred to as Dataset 1). K- Fold Cross-Validation was employed to assess the model’s performance, wherein the dataset was partitioned into K folds, and for each iteration, one fold was held out for testing while the remaining K-l folds were used for training.

[0297] Various embodiments make predictions at the individual sample level based on a context of 512 sequencing reads of size 300 nucleotides, corresponding to 153600 nucleotides per sample. These sample-level predictions are then aggregated to produce a patient-level diagnosis classification.

[0298] Referring now back to FIG. 7, table 702 (also referred to herein as Table 4.1) details the distribution of patient samples across various cancer types and healthy controls used in the experiment conducted by the Inventors. Table 702 lists five distinct datasets (Dataset 1 through Dataset 5) and the number of samples for "Healthy" individuals and patients diagnosed with "Breast Cancer," "Colorectal Cancer," "Lung Cancer," "Ovarian Cancer," "Pancreatic Cancer," and "Renal Cancer." A "Total" row provides the aggregate count for each category across all datasets. FIG. 7 demonstrates the comprehensive and diverse nature of the clinical datasets utilized for training and evaluating the deep learning model, underscoring the broad applicability and validation of the described approach across multiple cancer types.

[0299] Referring now back to FIG. 8, plots 802A-B of unsupervised clustering of datasets of subjects with different cancers indicate the ML model's ability to learn meaningful representations from raw cfDNA data that effectively separate healthy samples from cancer samples, even for data not seen during training. Two plots are presented: a Principal Component Analysis (PCA) plot 802A and a t-Distributed Stochastic Neighbor Embedding (t-SNE) plot 802B, both colored by phenotype (Ovarian Cancer, Healthy, Pancreatic Cancer, Breast Cancer, Colorectal Cancer, Lung Cancer). In both plots, distinct clusters are observable, demonstrating that the ML model, through its end-to-end learning from raw data, can inherently discern underlying biological differences between healthy and various cancer states. The clear separation of healthy samples from cancer samples in an unsupervised manner, on unseen data, provides strong evidence of the model's capacity to extract robust and generalizable biological signals.

[0300] Referring now back to FIG. 9, classification matrix 902 (confusion matrix) and associated performance metrics for an ensemble ML model applied to pancreatic cancer classification is depicted. Classification matrix 902 includes "precision," "recall," "fl-score," and "support" for both "Pancreatic Cancer" (labeled as 1) and "Healthy" (labeled as 0) classes. Overall "accuracy," "macro avg," and "weighted avg" are also provided. For instance, the model achieved a precision of 0.95 and recall of 0.89 for healthy samples, and a precision of 0.59 and recall of 0.76 for pancreatic cancer samples, with an overall accuracy of 0.87. FIG. 9 demonstrates the ML model's performance in classifying a particularly challenging cancer type, highlighting its ability to contribute to the diagnosis of difficult-to-detect malignancies.

[0301] Referring back to FIG. 10, box plot 1002 illustrates the distribution of "Prediction Probabilities" (referred to as "cancer scores") produced by an ensemble ML model trained on nucleotide, quality, and metadata inputs for multiple cancer classification, as described herein. Box plot 1002 shows the distribution of scores for "Ovarian Cancer," "Healthy," "Pancreatic Cancer," "Breast Cancer," "Colorectal Cancer," and "Lung Cancer" phenotypes. Box plot 1002 visually represents the median, quartiles, and outliers of the cancer scores for each group. Notably, harder-to-detect cancers like pancreatic and breast cancers exhibit a larger spread of scores, indicating the inherent variability and challenge in their classification, while healthy samples show a tight distribution near 0.0 and easily detectable cancers show distributions closer to 1.0. FIG. 10 provides insight into the ML model's confidence and discriminative power across different cancer types.

[0302] Referring back to FIG. 11, a series of Receiver Operating Characteristic (ROC) curves 1102 and their corresponding Area Under the Curve (AUC) values, demonstrates the ML model's robust classification performance for both early and late-stage cancers across various types. Separate ROC curves 1102 are shown for "Breast Cancer Early," "Breast Cancer Late," "Colorectal Cancer Early," "Colorectal Cancer Late," "Lung Cancer Early," "Lung Cancer Late," "Ovarian Cancer Early," "Ovarian Cancer Late," "Pancreatic Cancer Early," and "Pancreatic Cancer Late." Each plot includes the mean recall at 85% specificity and the AUC value. For example, Breast Cancer Early achieved 91% Mean Recall at 85% Specificity (AUC=0.87), and Breast Cancer Late achieved 96% Mean Recall at 85% Specificity (AUC=0.90). FIG. 11 provides evidence that the approach described herein is effective across different stages of cancer progression, which is crucial for early detection and intervention.

[0303] Referring back to FIG. 12, graphs 1202 (e.g., bar charts) compare the classification accuracy of at least one embodiment against a traditional liquid biopsy method based on mutational analysis, as well as a "simple" Al model and a more sophisticated transformer ML model (representing the approach described herein). The accuracy is shown for "Healthy controls," "Breast Cancer," "Colorectal Cancer," "Lung Cancer," and "Ovarian Cancer." The chart illustrates the "Prediction improvement with newer models," showing that both the "Simple" Al model and, more significantly, the "New transformer model" (based on at least one embodiment) achieve substantially higher accuracy compared to "Random Prediction" and the "Traditional mutational model" across all cancer types. FIG. 12 empirically demonstrates the superior performance and technical advantage of the deep learning approach described herein, particularly the transformerbased model, over conventional methods.

[0304] Referring now back to FIG. 13, line graphs 1302 comparing the Fl Score of the approach based on at least one embodiment described herein ("Embodiment(s) ") against traditional liquid biopsy methods, specifically "ichorCNA" (based on copy number variations (CNV) analysis) and "LIQUORICE" (quantification of changes in sequencing coverage around regions of interest in cfDNA, showing methylation patterns), in a setting of low-coverage sequencing. The comparison is made across different amounts of sequencing reads: 50 Million (Mil), 5 Mil, and 0.1 Mil. Graphs 1302 clearly demonstrate that "Embodiment(s)" consistently maintains a higher Fl Score, particularly at lower sequencing depths (e.g., 0.1 Mil reads), compared to ichorCNA and LIQUORICE. FIG. 13 provides strong evidence of the embodiment's superior robustness and effectiveness in low-coverage scenarios, making it more cost-effective and scalable for widespread clinical deployment.

[0305] Referring now back to FIG. 14, correlation 1402 between the cancer prediction scores generated by the ML model described herein and scores from traditional computational methods, are presented. The traditional computational methods include tMAD (trimmed median absolute deviation from copy number neutrality), which is a score related to copy number variation, ichorCNA which is related to copy number variation, and liquorice which is related to methylation. The observed correlation 1402 indicate that at least one embodiment based on the ML model, while operating on raw, un-engineered data, is capable of learning and reflecting biological signals that are also captured by established, feature-engineered methods like tMAD and others. This demonstrates that the ML model described herein is extracting biologically relevant information, validating its ability to identify complex patterns indicative of cancer.

[0306] Referring now back to FIG. 15, PCA plots 1502 and t-SNE plot 1504 demonstrate the generalization performance of at least one embodiment. ML models initially developed for general cancer detection were applied to a completely different, unseen, out-of-distribution dataset from a different laboratory, comprising patients with diffuse large B-cell lymphoma (DLBCL) and Highgrade B-cell lymphoma (HGBCL). The plots show that the model, using cancer classification scores derived from its early cancer detection training, was able to distinguish between patients from these two different lymphoma types using unsupervised clustering methods. This figure provides compelling evidence of the method's capability to learn fundamental biological signals related to cancer that are highly generalizable and can be utilized for purposes beyond its initial training, such as distinguishing between different lymphoma subtypes, without requiring additional training. Referring now back to FIG. 16, classification performance report 1602 for the task of separating responders from non-responders to a cancer therapy in patients with High-Grade B-cell Lymphoma (HGBCL), is presented. An ensemble ML model, previously trained on early cancer detection datasets, was applied to a completely unseen, out-of-distribution (OOD) dataset for this new task. Report 1602 includes overall "Accuracy" (0.75), a "Confusion Matrix" ([

[0347] ,

[0914] ]), and a "Classification Report" detailing "precision," "recall," "fl-score," and "support" for "False" (non-responders) and "True" (responders) classes. Report 1602 also provides "Sensitivity at specified specificity thresholds" (e.g., 56.52% sensitivity at 95.00% specificity). FIG. 16 demonstrates the significant capability of at least one embodiment to classify patients into groups based on their response to therapy, using (optionally only) the cancer classification scores derived from a ML model trained for early cancer detection, highlighting its versatility and clinical utility in guiding personalized treatment.

[0307] Referring now back to FIG. 17, box plot 1702 is of "cancer scores" for breast cancer patient samples obtained from a local collaborating hospital, categorized into two groups: those "responding to a breast cancer therapy" and "non responders to a breast cancer therapy." FIG. 17 visually demonstrates that at least one embodiment described herein, even though trained for the early detection of multiple cancers, can be applied without additional training to distinguish between breast cancer patients who will or will not respond to a specific cancer therapy. The distinct distributions of cancer scores between the two groups indicate the ML model's ability to capture subtle biological differences related to treatment response, further underscoring its broad applicability and potential for clinical decision support in personalized medicine.

[0308] Performance Evaluation. Evaluating performance is crucial to estimate its generalization ability and predictive performance on unseen data. To perform model selection (hyperparameter optimization) and performance estimation on the dataset, 5-Fold Cross Validation was employed. This technique may involve partitioning the dataset into K folds, where one fold is held out for testing, and the remaining K-l folds are used for training and validation in each iteration.

[0309] Performance evaluation is illustrated in FIG. 5 (data flow illustration) and FIG. 6 (read aggregation illustration). Specifically, the data from each of the five dataset sources for each cancer class is divided into five folds. For each fold, the fit set (training data) was randomly split into 80% for training and 20% for validation. These train, validation, and test sets were aggregated across all folds. The validation set, comprising 20% of the training data, was used to monitor the training process and determine when to stop to avoid overfitting, while the best-performing checkpoint was saved for evaluation on the test set. This stratification strategy aimed to maintain consistent class probabilities across different folds and groups. One advantage of this cross-validation approach is that each patient appears once in the test set, allowing the method to be evaluated on all patients in the dataset. The test set was used solely for estimating model performance for each cross-validation fold, ensuring an unbiased evaluation of the model’s generalization capabilities. 5 is a good tradeoff between bias-variance computations. By employing 5-Fold Cross Validation and a robust data splitting strategy, the model’s performance and generalization ability could be reliably assessed, providing valuable insights into its potential for clinical application in cancer diagnosis.

[0310] Sample-Level Classification Analysis. At the sample level, the ML model exhibited reliable classification capabilities, even with reduced sequencing depth. Subsampling the sequencing files to 80,000 reads per sample resulted in clear separation between cancer and healthy classes, with most samples classified with high confidence. The model achieved a mean recall of 67% across all folds, while constraining the false positive rate to 15%.

[0311] Patient-Level Classification Analysis. Aggregating sample classifications to the patient level significantly improved recall at 85% specificity, demonstrating an 18% absolute improvement over single-sample classifications. Patient- level ROC curves showcased the model’s ability to achieve higher recall rates with reasonable specificity thresholds. FIG. 8 shows output probabilities of all samples per cancer type, demonstrating unsupervised clustering and separation of healthy from cancer samples. FIG. 10 further illustrates the distribution of "cancer scores" for multiple cancer types, showing the model's discriminative power.

[0312] Per-Cancer Analysis. The ML model’s performance varied across different cancer types, with colorectal, lung, ovarian, and renal cancers exhibiting strong outcomes, while breast and pancreatic cancers showed comparatively lower effectiveness. These differences may stem from dataset biases or clinical challenges associated with diagnosing certain cancer types. FIG. 9 provides a detailed classification matrix for pancreatic cancer, demonstrating the model's performance on this challenging cancer type. Per-cancer ROC curves are shown in FIG. 11, illustrating good classification for both early and late-stage cancers across various types.

[0313] Breast Cancer Experiment. A specialized breast cancer detection ML model achieved a recall of 86% at 85% specificity, outperforming the multi-cancer model by 27% in breast cancer samples. However, this specialized ML model’s inability to generalize across other cancer types highlights the trade-off between specialization and flexibility.

[0314] Effect of Sampling Factor. Reducing sequencing depth to simulate lower-cost testing scenarios did not significantly impact patient-level classification metrics. The ML model demonstrated robustness and effectiveness for cancer screening, even with substantially fewer reads. The demonstration proves the method is suitably performed with coverage as low as 0.1 million reads. This is further illustrated in FIG. 13, which compares at least one embodiment’s Fl score with traditional liquid biopsy methods (ichorCNA and LIQUORICE) at low sequencing coverage, demonstrating superior performance.

[0315] Effect of Checkpointing and Aggregation. Checkpointing based on validation loss improved patient-level performance, indicating its importance in achieving generalized diagnosis capabilities. Aggregation techniques such as median probability achieved the highest recall at 85% specificity, emphasizing the importance of careful selection in translating sample-level insights into robust patient-level outcomes.

[0316] Comparison to Baseline Methods. Comparisons with baseline methods demonstrated the superiority of embodiments of the disclosure in achieving high recall rates, particularly at the patient level. FIG. 12 illustrates the significant prediction improvement achieved by the new transformer model compared to random prediction and traditional mutational models across various cancer types. FIG. 14 further shows a correlation between the cancer prediction of the deep learning ML model described herein and traditional computational methods (tMAD), validating the biological relevance of the learned features. Despite challenges in dataset availability and reproducibility, embodiments of the disclosure showcased a novel capability for accurate classification with low input sample counts, addressing a critical need in cancer diagnosis.

[0317] Generalization and Treatment Response Prediction. At least one embodiment has demonstrated remarkable generalization capabilities. As shown in FIG. 15, models trained for general cancer detection were successfully applied to an entirely different, unseen dataset of patients with diffuse large B-cell lymphoma (DLBCL) and High-grade B-cell lymphoma (HGBCL). The ML model was able to distinguish between these two lymphoma types using unsupervised clustering methods based on its learned cancer scores, indicating its ability to extract fundamental biological signals relevant to cancer beyond its initial training task.

[0318] Furthermore, at least one embodiment described herein has shown significant utility in predicting patient response to cancer therapy. FIG. 16 presents a classification performance report for separating responders from non-responders to cancer therapy in HGBCL patients, using an ensemble model trained for early cancer detection on an out-of-distribution (OOD) dataset. The cancer classification scores alone were sufficient to classify patients based on their treatment response, without requiring additional model training or fine-tuning for this specific task. Similarly, FIG. 17 illustrates the model's ability to distinguish between breast cancer patients who will or will not respond to therapy, using samples from a local collaborating hospital, again without additional training. These results highlight the versatility and clinical utility of the inventive method in guiding personalized medicine and treatment decisions. The descriptions of the various embodiments of the present invention have been presented for purposes of illustration, but are not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terminology used herein was chosen to best explain the principles of the embodiments, the practical application or technical improvement over technologies found in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

[0319] It is expected that during the life of a patent maturing from this application many relevant ML models will be developed and the scope of the term ML model is intended to include all such new technologies a priori.

[0320] As used herein the term “about” refers to ± 10 %.

[0321] The terms "comprises", "comprising", "includes", "including", “having” and their conjugates mean "including but not limited to". This term encompasses the terms "consisting of" and "consisting essentially of".

[0322] The phrase "consisting essentially of" means that the composition or method may include additional ingredients and / or steps, but only if the additional ingredients and / or steps do not materially alter the basic and novel characteristics of the claimed composition or method.

[0323] As used herein, the singular form "a", "an" and "the" include plural references unless the context clearly dictates otherwise. For example, the term "a compound" or "at least one compound" may include a plurality of compounds, including mixtures thereof.

[0324] The word “exemplary” is used herein to mean “serving as an example, instance or illustration”. Any embodiment described as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments and / or to exclude the incorporation of features from other embodiments.

[0325] The word “optionally” is used herein to mean “is provided in some embodiments and not provided in other embodiments”. Any particular embodiment of the invention may include a plurality of “optional” features unless such features conflict.

[0326] Throughout this application, various embodiments of this invention may be presented in a range format. It should be understood that the description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the invention. Accordingly, the description of a range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range. For example, description of a range such as from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6 etc., as well as individual numbers within that range, for example, 1, 2, 3, 4, 5, and 6. This applies regardless of the breadth of the range.

[0327] Whenever a numerical range is indicated herein, it is meant to include any cited numeral (fractional or integral) within the indicated range. The phrases “ranging / ranges between” a first indicate number and a second indicate number and “ranging / ranges from” a first indicate number “to” a second indicate number are used herein interchangeably and are meant to include the first and second indicated numbers and all the fractional and integral numerals therebetween.

[0328] It is appreciated that certain features of the invention, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable subcombination or as suitable in any other described embodiment of the invention. Certain features described in the context of various embodiments are not to be considered essential features of those embodiments, unless the embodiment is inoperative without those elements.

[0329] Although the invention has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, it is intended to embrace all such alternatives, modifications and variations that fall within the spirit and broad scope of the appended claims.

[0330] It is the intent of the applicant(s) that all publications, patents and patent applications referred to in this specification are to be incorporated in their entirety by reference into the specification, as if each individual publication, patent or patent application was specifically and individually noted when referenced that it is to be incorporated herein by reference. In addition, citation or identification of any reference in this application shall not be construed as an admission that such reference is available as prior art to the present invention. To the extent that section headings are used, they should not be construed as necessarily limiting. In addition, any priority document(s) of this application is / are hereby incorporated herein by reference in its / their entirety.

Claims

1. WHAT IS CLAIMED IS:

1. A system for classification of cell-free deoxyribonucleic acid (cfDNA) into a medical state of a subject, comprising: a communication interface configured for receiving raw cfDNA sequencing data comprising a plurality of cfDNA reads obtained from a biological sample of a subject, the raw cfDNA sequencing data generated by a sequencing machine; at least one processor operatively coupled to the communication interface, the at least one processor configured for executing code for: sampling a subset of raw cfDNA reads from the raw cfDNA sequencing data, wherein the subset forms a sample, wherein the sampling is performed according to a sampling process; generating a multivariate input representation for the sample, comprising nucleotide sequences of the sample and / or at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to; feeding into a machine learning (ML) model, the multivariate input representation; and obtaining from the ML model, for the sample, a probability that the subject is in the medical state.

2. The system of claim 1, wherein the multivariate input representation includes the nucleotide sequences of the sample, the associated quality values, and the metadata representations.

3. The system of claim 1, wherein the multivariate input representation includes the nucleotide sequences of the sample and the associated quality values, and excludes the metadata representations.

4. The system of claim 1, wherein the multivariate input representation includes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

5. The system of claim 1, wherein the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and includes the metadata representations.

6. The system of claim 1, wherein the multivariate input representation excludes the nucleotide sequences of the sample, includes the associated quality values, and excludes the metadata representations.

7. The system of claim 1, wherein the multivariate input representation excludes the nucleotide sequences of the sample, excludes the associated quality values, and includes the metadata representations.

8. The system of claim 1, wherein the multivariate input representation excludes alignmentbased feature engineering and / or calculation of summary genomic features, wherein the metadata representation excludes alignment-based feature engineering and / or calculation of summary genomic features.

9. The system of claim 1, further comprising code for: iterating the sampling, the generating, the feeding, and the obtaining, for different samples of the subject, for obtaining from the ML model a plurality of probabilities for a plurality of different samples of the subject; aggregating the plurality of probabilities of the subject ; and providing the aggregation of the plurality of probabilities as a final subject-level classification and / or the probability that the subject is in the medical state.

10. The system of claim 9, wherein a total number of reads of the plurality of different samples used to generate the final subject-level classification and / or probability that the subject is in the medical state is less than about 100,000.

11. The system of claim 1, further comprising aggregating into a vector form, the raw cfDNA reads of the sample, wherein the vector is fed into the ML model.

12. The system of claim 1, wherein the sampling process comprises randomly selecting the sample of raw cfDNA reads from the plurality of raw sequences.

13. The system of claim 12, wherein randomly selecting comprises: wherein the subset of raw cfDNA reads of the sample are mapped to random locations of a reference genome, and / or the subset of raw cfDNA reads of the sample are read in a random order.

14. The system of claim 1, wherein the sampling process comprises consecutively reading the subset of raw cfDNA reads of the sample.

15. The system of claim 1, wherein the sampling process comprises: mapping the plurality of raw sequences to a plurality of regions of a reference genome; randomly selecting a region on the reference genome; and consecutively reading the sample of the subset of raw cfDNA reads mapped to the randomly selected region on the reference genome.

16. The system of claim 1, wherein the sampling process comprises including about .001-5% of the plurality of raw sequences in the sample of raw cfDNA reads.

17. The system of claim 1, wherein the sampling process comprises selecting a number of the raw cfDNA reads included in the sample from a range of about 512-4096.

18. The system of claim 1, wherein the sampling process comprises selecting a size of each of the raw cfDNA reads of the sample that is common and fixed, selected from a range of about 50- 350 nucleotides.

19. The system of claim 1, further comprising: pre-processing the raw sequence data by encoding the raw sequence data into a numerical encoding by mapping each type of cfDNA nucleotide to a corresponding numerical value, wherein the encoding of each raw cfDNA of the subset is fed into the ML model.

20. The system of claim 1, wherein the quality values are generated by the automated sequencing machine that generates the raw cfDNA sequencing data from the plurality of cfDNA reads.

21. The system of claim 1, further comprising:mapping each raw sequence to a corresponding genomic region of a defined reference genomic sequence; and for each respective raw cfDNA read, generating corresponding metadata including empirically derived information about the corresponding genomic region, wherein the corresponding metadata is included in the multivariate input representation.

22. The system of claim 1, wherein the metadata for the sample indicates whether the sample maps to a coding region, a methylated region, or a region with known disease-relevant annotations.

23. The system of claim 1 , wherein the ML model is configured for modular interchangeability between different deep learning architectures, including transformer encoders, convolutional neural networks (CNNs), and recurrent neural networks (RNNs).

24. The system of claim 1, wherein the ML model is implemented as a neural network encoder with a classification head.

25. The system of claim 24, wherein the neural network encoder with the classification head is implemented as a transformer encoder model that excludes a decoder component and excludes positional encodings.

26. The system of claim 24, further comprising adding a classification token (CLS token) as a first token to an input sequence of a plurality of the raw cfDNA reads for feeding into the ML model, and wherein a classification output layer is attached to an embedding of the CLS token after a final encoder layer of the ML model.

27. The system of claim 1, wherein aggregating is selected a group consisting of: Quantilebased aggregation, Mean, Median, and Majority.

28. The system of claim 1, wherein the ML model is configured for receiving the multivariate input representation directly, without requiring prior manual feature extraction and / or generation of summary genomic features based on alignment to a reference genome.

29. The system of claim 1, wherein the ML model is configured to learn an internal representation of the cfDNA reads that enables unsupervised clustering of samples according to the medical state.

30. The system of claim 1, wherein the medical condition is selected from: cancer, a specific type of cancer, specific sub-type of cancer, lymphoma types including diffuse large B-cell lymphoma (DLBCL) and high-grade B-cell lymphoma (HGBCL), each of a plurality of different types of cancer, predicting the subject’s response to a therapy administered to treat a disease, predicting the subject’s response to a treatment for a cancer, a stage of the cancer including early - stage and late- stage.

31. The system of claim 1, wherein the multivariate representation includes the nucleotide sequences of the sample and associated quality values, and excludes the metadata representations of genomic regions.

32. The system of claim 1, wherein a representation vector is obtained from the ML model, and the probability is computed from the representation vector.

33. The system of claim 1, wherein the ML model comprises an ensemble of two or more independently trained ML models, wherein a final probability is computed by averaging or voting on the probabilities generated by the two or more independently trained ML models.

34. The system of claim 1, wherein the ML model is trained on a training dataset comprising a plurality of records created for a plurality of individuals, by: for each individual of a plurality of individuals: sampling a subset of raw cfDNA reads from raw cfDNA sequencing data of the individual obtained from a biological sample of the individual and sequenced by the sequencing machine, wherein the subset forms a sample, wherein the sampling is performed according to a sampling process; generating the multivariate input representation for the individual, comprising nucleotide sequences of the sample and at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to; andcreating a record including the multivariate input representation generated for the individual and a ground truth label indicating a classification category associated with the medical state that the individual is in.

35. The system of claim 34, wherein the training of the ML model comprises initializing the ML model by transfer-learning from a pre-trained model trained on a large dataset of genomic data, and training the ML model on the training dataset after the initialization.

36. A system of training a ML model for classification of cell-free deoxyribonucleic acid (cfDNA) into a medical state of a subject, comprising: a communication interface configured for receiving raw cfDNA sequencing data obtained from respective biological samples of a plurality of individuals, the raw cfDNA sequencing data generated by a sequencing machine; at least one processor operatively coupled to the communication interface, the at least one processor configured for executing a code for: creating a training dataset comprising a plurality of records created for a plurality of individuals, by: for each individual of the plurality of individuals: sampling a subset of raw cfDNA reads from the raw cfDNA sequencing data, wherein the subset forms a sample, wherein the sampling is performed according to a sampling process; generating a multivariate input representation for the sample, comprising nucleotide sequences of the sample and at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to; creating a record including the multivariate input representation and a ground truth label indicating a classification category associated with the medical state of the individual; training the ML model on the training dataset for generating a probability that the subject is in the medical state in response to an input of a target multivariate input representation including a sample of a subset of raw cfDNA reads of raw cfDNA sequencing data generated from a biological sample of a subject, sampled according to the sampling process, and at least one of: target quality values associated with each nucleotide of the sample, and target genomic data associated with at least one region of the sample; andproviding the trained ML model and an aggregation process for aggregating a plurality of probabilities obtained from the ML model for a plurality of multivariate input representations for a plurality of samples of a subject, for computing a final subject-level classification and / or an aggregated probability that the subject is in the medical state.

37. A computer implemented method of classification of cell-free deoxyribonucleic acid (cfDNA) into a medical state of a subject, comprising: receiving raw cfDNA sequencing data comprising a plurality of cfDNA reads obtained from a biological sample of a subject, the raw cfDNA sequencing dta generated by a sequencing machine; in a plurality of iterations: sampling a subset of raw cfDNA reads from the raw cfDNA sequencing data, wherein the subset forms a sample, wherein the sampling is according to a sampling process, wherein in each iteration a different sample is sampled according to the sampling process; generating a multivariate input representation for the sample, comprising nucleotide sequences of the sample and at least one of: associated quality values, and metadata representations of genomic regions to which the reads of the sample map to; feeding into a machine learning model (ML) model, the multivariate input representation; and obtaining from the ML model, for the sample of each iteration probability that the subject is in the medical state; aggregating a plurality of probabilities obtained from the ML model from the plurality of iterations; and providing the aggregation of the plurality of probabilities as an indication that the subject is in the medical state.

Images