Model-based characterization and classification

Abstract

In various embodiments, an analysis system uses a model to determine features and classifications of a disease state. The disease state may indicate the presence or absence of cancer, the type of cancer, or the tissue of origin of cancer. The model may include a binary classifier and a source organization classifier. The analysis system may process sequence reads from a test biological sample to produce data for training the classifier. The analysis system may also train the model using a combination of machine learning techniques, which may include a multi-layer perceptron. In some embodiments, the analysis system uses methylation information to train the model to determine a prediction regarding a disease state.

Description

[0001] This application is a divisional application of the patent application with application number 202080036172.8, application date May 13, 2020, and name “Model-based Characterization and Classification”. Background Art 1. Technical Field

[0002] The present disclosure generally relates to model-based characterization and classifiers for predicting disease states from nucleic acid samples.

[0003] 2. Description of Related Technology

[0004] DNA methylation plays a role in regulating gene expression. Abnormal DNA methylation is associated with many disease processes including cancer. DNA methylation analysis using methylation sequencing (e.g., whole genome bisulfite sequencing (WGBS)) is increasingly being recognized as a valuable diagnostic tool for detecting, diagnosing and / or monitoring cancer. For example, specific patterns of differentially methylated regions can be used as molecular markers for various disease states. Summary of the invention

[0005] Disclosed herein are methods for training and applying models using nucleic acid samples to generate features and / or classifications of disease states (e.g., the presence or absence of cancer, the type of cancer, and / or the tissue of origin of the cancer). In one aspect, the present disclosure provides a method for analyzing sequence reads to generate multiple features, comprising: generating a first plurality of reference sequence reads from a first reference sample, the first sample being from a subject having a first disease state; generating a second plurality of reference sequence reads from a second reference sample, the second sample being from a subject having a second disease state, training a first probability model using the first plurality of reference sequence reads, the first probability model being associated with the first disease state; training a second probability model using the second plurality of reference sequence reads, the second probability model being associated with the second disease state; generating a plurality of training sequence reads from the training sample, and for each sequence read of the plurality of training sequence reads: applying the sequence reads to the first probability model to determine a first probability value, the first probability value being the probability of the sequence reads being derived from a sample associated with the first disease state, and applying the sequence reads to the second probability model to determine a second probability value, the second probability value being the probability of the sequence reads being derived from a sample associated with the second disease state; and identifying one or more features by comparing the first probability value and the second probability value for each sequence read.

[0006] In another aspect, the present disclosure provides a system comprising a computer processor and a memory storing computer program instructions which, when executed by the computer processor, cause the processor to perform steps comprising: accessing a first plurality of reference sequence reads from a first reference sample, the first sample being from a subject having a first disease state; accessing a second plurality of reference sequence reads from a second reference sample, the second sample being from a subject having a second disease state, training a first probability model using the first plurality of reference sequence reads, the first probability model being associated with the first disease state; training a second probability model using the second plurality of reference sequence reads, the second probability model being associated with the second disease state; accessing a plurality of training sequence reads from the training sample, and for each sequence read of the plurality of training sequence reads: applying the sequence read to the first probability model to determine a first probability value, the first probability value being a probability of the sequence read being derived from the sample associated with the first disease state, and applying the sequence read to the second probability model to determine a second probability value, the second probability value being a probability of the sequence read being derived from the sample associated with the second disease state; and identifying one or more features by comparing the first probability value and the second probability value for each sequence read.

[0007] In another aspect, the present disclosure provides a non-transitory computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to implement steps comprising: accessing a first plurality of reference sequence reads from a first reference sample, the first sample being from a subject having a first disease state; accessing a second plurality of reference sequence reads from a second reference sample, the second sample being from a subject having a second disease state, training a first probability model using the first plurality of reference sequence reads, the first probability model being associated with the first disease state; training a second probability model using the second plurality of reference sequence reads, the second probability model being associated with the second disease state; accessing a plurality of training sequence reads from the training sample, and for each sequence read of the plurality of training sequence reads: applying the sequence read to the first probability model to determine a first probability value, the first probability value being a probability of the sequence read being derived from a sample associated with the first disease state, and applying the sequence read to the second probability model to determine a second probability value, the second probability value being a probability of the sequence read being derived from a sample associated with the second disease state; and identifying one or more features by comparing the first probability value and the second probability value for each sequence read.

[0008] In some embodiments, the first disease state is cancer and the second disease state is non-cancer. In some embodiments, the first disease state is a first type of cancer and the second disease state is a second type of cancer, and wherein the first type of cancer and the second type of cancer are different.

[0009] In some embodiments, the method, system or non-transitory computer-readable medium also includes generating multiple reference sequence reads from a third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference sample, each of the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference samples having a different disease state, and wherein each of the different disease states is a different type of cancer; and training a third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability model using the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth multiple reference sequence reads, wherein each of the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models is respectively associated with a different type of cancer.

[0010] In some embodiments, the cancer or type of cancer is selected from the following, including: breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, renal pelvis ureter urothelial carcinoma, renal cancer other than urothelial carcinoma, prostate cancer, anorectal cancer, colorectal cancer, esophageal squamous cell carcinoma, esophageal cancer other than squamous cell carcinoma, gastric cancer, hepatobiliary duct cancer derived from hepatocytes, hepatobiliary duct cancer derived from cells other than hepatocytes, pancreatic cancer, head and neck cancer associated with human papillomavirus, head and neck cancer not associated with human papillomavirus, lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer and lung cancer other than adenocarcinoma or small cell lung cancer, neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma and leukemia. In some embodiments, the cancer type is additionally selected from the following, including: brain cancer, vulvar cancer, vaginal cancer, testicular cancer, pleural mesothelioma, peritoneal mesothelioma and gallbladder cancer.

[0011] In some embodiments, the first disease state includes a first source tissue and the second disease state includes a second source tissue. The first source tissue or the second source tissue is selected from the following, including: breast tissue, thyroid tissue, lung tissue, bladder tissue, cervical tissue, small intestine tissue, colorectal tissue, esophageal tissue, stomach tissue, tonsil tissue, liver tissue, ovarian tissue, fallopian tube tissue, pancreatic tissue, prostate tissue, kidney tissue and uterine tissue. In some embodiments, the first source tissue or the second source tissue is additionally selected from the following, including: brain tissue and cells, endocrine tissue and cells, vascular endothelial tissue and cells, head and neck tissue and cells, exocrine pancreatic tissue and cells, endocrine pancreatic tissue and cells, lymphoid tissue and cells, mesenchymal tissue and cells, myeloid tissue and cells, pleural tissue and cells, muscle tissue and cells, bone marrow tissue and cells, adipose tissue and cells, gallbladder tissue and cells.

[0012] In some embodiments, the first probability model or the second probability model is a constant model, a binomial model, an independent site model, a neural network model, or a Markov model.

[0013] In some embodiments, the methods, systems, or non-transitory computer-readable media of the present disclosure further include determining a methylation rate for each of a plurality of CpG sites in the first plurality of reference sequence reads or the second plurality of reference sequence reads, wherein the first probability model or the second probability model is parameterized by the product of the methylation rates.

[0014] In some embodiments, the methods, systems, or non-transitory computer-readable media of the present disclosure further include determining, for each sequence read in the first plurality of reference sequence reads or the second plurality of sequence reads, whether the sequence read is abnormally methylated; and filtering the first plurality of reference sequence reads or the second plurality of reference sequence reads using p-value filtering by removing sequence reads having a p-value below a threshold from the first plurality of reference sequence reads or the second plurality of reference sequence reads.

[0015] In some embodiments, the methods, systems, or non-transitory computer-readable media of the present disclosure further include, for each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, determining whether the sequence read is hypomethylated or hypermethylated by determining whether at least a threshold number of CpG sites having at least a threshold percentage of the CpG sites are unmethylated or methylated, respectively.

[0016] In some embodiments, the methods, systems, or non-transitory computer-readable media of the present disclosure further include determining, for each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, whether the sequence read is abnormally methylated; and filtering the first plurality of reference sequence reads using p-value filtering by removing sequence reads having a p-value below a threshold from the first plurality of reference sequence reads.

[0017] In some embodiments, the first probability model or the second probability model is parameterized by a sum of a plurality of mixture components, each component being associated with a product of the methylation rates. In some embodiments, each mixture component of the plurality of mixture components is associated with a fractional assignment, and wherein the fractional assignments sum to one.

[0018] In some embodiments, training the first probability model or the second probability model includes determining a set of parameters for the probability model, which maximizes the total log-likelihood of the first multiple reference sequence reads or the second multiple reference sequence reads derived from subjects associated with the first disease state or the second disease state, and the disease state is associated with the probability model.

[0019] In some embodiments, the method, system, or non-transitory computer-readable medium of the present disclosure further includes, for each of a plurality of windows: selecting a plurality of the first plurality of reference sequence reads originating from the window, and using the sequence reads originating from the window to train the first probability model for the window; and selecting a plurality of the second plurality of reference sequence reads originating from the window, and using the sequence reads to train the probability model for each window.

[0020] In some embodiments, the method, system or non-transitory computer readable medium of the present disclosure further comprises, for each of the plurality of windows, selecting a subset of the plurality of training sequence reads derived from the window; and comparing the first probability value and the second probability value for each sequence read of the subset to identify the one or more features. In some embodiments, each of the windows is separated by at least a threshold number of base pairs between CpG sites. In some embodiments, each of the plurality of windows comprises about 200 base pairs (bp) to about 10 kilobase pairs (kbp).

[0021] In some embodiments, the one or more features include the count of outlier sequence reads of the plurality of training sequence reads where the first probability value is greater than the second probability value. In some embodiments, the one or more features include a binary count. In some embodiments, the one or more features include a total count of outlier sequence reads. In some embodiments, the one or more features include a total count of anonymously methylated sequence reads. In some embodiments, the one or more features include a count of fragments containing one or more specific methylation patterns. In some embodiments, the output of a discriminant classifier trained within a single genomic region is used to identify the one or more features. In some embodiments, the discriminant classifier is a multilayer perceptron or a convolutional neural network model. In some embodiments, comparing the first probability value and the second probability value includes determining the ratio of the first probability value and the second probability value, and wherein the one or more features include a sequence read count of sequence reads exceeding a ratio threshold. In some embodiments, the first probability value or the second probability value is a log-likelihood value. In some embodiments, the one or more features include sorting information sequence reads based on the rarity of the sequence reads in the first disease state.

[0022] In some embodiments, identifying the one or more features includes: for each sequence read of the plurality of training sequence reads: determining a log-likelihood ratio of the first probability value to the second probability value; and for one or more thresholds, determining a count of the sequence reads having a log-likelihood ratio exceeding the threshold.

[0023] In some embodiments, the method, system, or non-transitory computer-readable medium of the present disclosure further comprises: for each of the one or more features, determining an indicator of the feature that distinguishes the first disease state from the second disease state.

[0024] In some embodiments, determining the indicator for each of the one or more features includes: determining mutual information between the feature and the probability of the first disease state and the second disease state being present. In some embodiments, the method of the present disclosure further includes: filtering the one or more features used to train the classifier by sorting the features based on the indicator.

[0025] In some embodiments, the method, system or non-transitory computer-readable medium of the present disclosure also includes training a classifier from the one or more features, the classifier being trained to predict one or more disease states for a plurality of sequence reads of a test sample from a test subject, wherein the one or more disease states include the presence or absence of a disease, a disease type and / or a disease source tissue. In some embodiments, the classifier is a logistic regression, a multinomial logistic regression, a generalized linear model (GLM), a support vector machine, a multilayer perceptron, a random forest or a neural network classifier. In some embodiments, the classifier is a multilayer perceptron model. In some embodiments, the classifier is generated using L1 or L2 regularized logistic regression. In some embodiments, the method of the present disclosure also includes determining a probability vector for the test sample; and determining a label for the test sample based on the probability vector.

[0026] In some embodiments, the methods, systems, or non-transitory computer-readable media of the present disclosure further comprise determining the accuracy of the classifier using a confusion matrix comprising information describing the success rate of the classifier in identifying each of the plurality of disease states.

[0027] In some embodiments, the first reference sample or the second reference sample is a cell-free nucleic acid sample or a tissue nucleic acid sample from a subject with a known disease state.

[0028] In some embodiments, the known disease state is the presence or absence of the disease, the disease type, and / or the tissue of origin of the disease.

[0029] In some embodiments, the training sample comprises a cell-free nucleic acid sample or a tissue sample. In some embodiments, the test sample comprises a cell-free nucleic acid sample.

[0030] In some embodiments, the first plurality of reference sequence reads, the second plurality of reference sequence reads, the plurality of training sequence reads, or the plurality of sequence reads from the test sample are generated by methylation sequencing (or methylation-aware sequencing). In some embodiments, the methylation sequencing comprises whole genome bisulfite sequencing. In some embodiments, the methylation sequencing comprises targeted sequencing.

[0031] In other aspects, the present disclosure provides a method for generating a classifier to predict a tissue of origin associated with a disease state, the method comprising: generating a first plurality of reference sequence reads from a reference sample having one of a plurality of disease states, each disease state being associated with a tissue of origin; training a plurality of probability models using the first plurality of reference sequence reads, each probability model being associated with a different disease state of the plurality of disease states; for each probability model of the plurality of probability models: for each of a second plurality of sequence reads, applying the probability model to the sequence reads to determine a value based at least on a first probability of the sequence read being derived from a sample associated with the disease state associated with the probability model; and identifying a feature by determining a count of the second plurality of sequence reads having a value exceeding a threshold; and generating a classifier using the feature, the classifier being trained to predict a disease state and / or a tissue of origin associated with one of the plurality of disease states for an input sequence read of a test sample from a test subject. In some embodiments, the plurality of disease states include at least two, at least three, at least four, at least five, or at least ten different disease states.

[0032] In some embodiments, the method further comprises determining a methylation rate for each of a plurality of CpG sites in the first plurality of reference sequence reads, wherein each of a plurality of probability models is parameterized by a product of the methylation rates.

[0033] In some embodiments, each probability model of the plurality of probability models is parameterized by a sum of a plurality of mixture components, each component being associated with a product of the methylation rates. In some embodiments, each mixture component of the plurality of mixture components is associated with a score assignment, and wherein the score assignments sum to one.

[0034] In some embodiments, training the multiple probability models comprises: determining a set of parameters for a probability model of the multiple probability models, the set of parameters maximizing the total log-likelihood of the first multiple reference sequence reads from subjects associated with the disease state, the disease state being associated with the probability model. In some embodiments, the method further comprises determining a probability vector for the test sample; and determining a label for the test sample based on the probability vector.

[0035] In some embodiments, determining the value includes determining the first probability of the sequence reads originating from the sample associated with the disease state associated with the probability model, wherein the disease state is associated with the presence of cancer or the type of cancer; determining a second probability of the sequence reads originating from a healthy sample; and determining a log-likelihood ratio of the first probability to the second probability.

[0036] In some embodiments identifying the feature comprises determining, for a plurality of thresholds, a count of the second plurality of sequence reads having a log likelihood ratio exceeding the threshold.

[0037] In some embodiments, the method further comprises determining, for each of the features, an indicator of the feature that distinguishes a first disease state and a second disease state of the plurality of disease states.

[0038] In some embodiments, determining the indicator of the feature includes determining mutual information between the feature and the probability of the first disease state and the second disease state existing.

[0039] In some embodiments, the first probability of the first disease state is equal to the second probability of the second disease state. In some embodiments, the method further comprises filtering the features used to train the classifier by ranking based on the index of the features.

[0040] In some embodiments, the method further comprises determining the accuracy of the classifier using a confusion matrix comprising information describing the success rate of the classifier in identifying each of the plurality of disease states.

[0041] In some embodiments, the method further comprises determining a plurality of blocks of a reference genome, each of the blocks being separated by at least a threshold number of base pairs between CpG sites, wherein the first plurality of reference sequence reads are generated using the plurality of blocks. In some embodiments, the count of the second plurality of sequence reads having the value exceeding the threshold is determined for a plurality of CpG sites.

[0042] In some embodiments, the reference sample comprises one or more of: a cell-free nucleic acid sample and a tissue sample.

[0043] In some embodiments, the plurality of disease states comprises one or more of: a cancer type, a disease type, and a healthy state.

[0044] In some embodiments, the classifier is a logistic regression, multinomial logistic regression, generalized linear model (GLM), multilayer perceptron, support vector machine, random forest or neural network model classifier. In some embodiments, the classifier is generated using L1 or L2 regularized logistic regression. In some embodiments, the classifier is a multilayer perceptron model.

[0045] In some embodiments, the method further comprises binarizing the feature to indicate the presence or absence of one of the plurality of disease states, wherein the classifier is generated using the binarized feature. The binarized features each have a value of 0 or 1.

[0046] In some embodiments, the method further comprises determining, for the reference sample, a measure of uncertainty in localization; and based on the measure, labeling at least one prediction of the classifier as an uncertain tissue of origin.

[0047] In other aspects, the present disclosure provides a method comprising: generating multiple sequence reads from one or more biological samples; for each position of multiple chromosome positions: using the multiple sequence reads to determine the count of nucleic acid fragments of the one or more biological samples that are within the position and have at least a threshold similarity to fragments associated with the disease state; using the counts of the multiple positions as features to train a machine learning model; and using the trained machine learning model to determine the probability of a test sample having a disease state.

[0048] In some embodiments, the method further comprises binarizing the feature in each of the plurality of locations to indicate the presence or absence of one of the plurality of disease states, wherein a count of at least one nucleic acid fragment in a location indicates the presence of one of the disease states in the location.

[0049] In some embodiments, the method further comprises filtering the plurality of sequence reads according to p-value scores of the plurality of sequence reads, wherein the p-value score of a sequence read indicates a probability of observing methylation in the nucleic acid fragments of the one or more biological samples corresponding to the sequence read.

[0050] In some embodiments, the machine learning model is a multilayer perceptron model. In some embodiments, the machine learning model uses logistic regression. In some embodiments, each of the multiple positions represents a plurality of consecutive base pairs of the chromosome.

[0051] In some embodiments, the multiple sequence reads are processed for multiple regions of the genome. In some embodiments, the multiple sequence reads represent nucleic acid fragments of a target subset of genomic regions. In some embodiments, the multiple sequence reads represent nucleic acid fragments of the entire genome. In some embodiments, the disease state is associated with at least one type of cancer. In some embodiments, the disease state is associated with a stage of at least one type of cancer. In some embodiments, the method further comprises determining treatment using the probability that the test sample has the disease state.

[0052] In other aspects, the present disclosure provides a method comprising: generating multiple sequence reads from nucleic acid fragments of multiple biological samples; determining a first training data set by processing the multiple sequence reads; training a first classifier using the first training data set, the first classifier being trained to predict the presence or absence of at least one disease state in the first test biological sample for a first input sequence read from the first test biological sample; determining that one or more disease states are present in a subset of the multiple biological samples using the prediction of the first classifier; determining a second training data set using the subset of the multiple sequence reads of the nucleic acid fragments corresponding to the subset of the multiple biological samples; and training a second classifier using the second training data set, the second classifier being trained to predict a source tissue associated with the disease state present in the second test biological sample for a second input sequence read from the second test biological sample.

[0053] In some embodiments, the second classifier is a multilayer perceptron comprising at least one hidden layer. In some embodiments, the first classifier does not comprise a hidden layer. In some embodiments, the multilayer perceptron comprises a 100-unit hidden layer or a 200-unit hidden layer. In some embodiments, the multilayer perceptron is fully connected and uses a rectified linear unit activation function. In some embodiments, the second classifier is a logistic regression or a multinomial logistic regression model. In some embodiments, the first classifier is a multilayer perceptron comprising at least one hidden layer. In some embodiments, the multilayer perceptron (first classifier) ​​comprises a hidden layer of 100 units or more units, and wherein the multilayer perceptron is fully connected and uses a rectified linear unit activation function. In some embodiments, the second classifier is a second multilayer perceptron comprising at least one hidden layer. In some embodiments, the first classifier is a logistic regression or a multinomial logistic regression model.

[0054] In some embodiments, the method further comprises performing a first cross validation on the first classifier; retraining the first classifier using a first hyperparameter selected based on the output of the first cross validation; performing a second cross validation on the second classifier; and retraining the second classifier using a second hyperparameter selected based on the output of the second cross validation. In some embodiments, the first hyperparameter and the second hyperparameter are selected using the summed results from all folds of the first cross validation and the second cross validation, respectively. In some embodiments, the second hyperparameter is selected to optimize the accuracy of the source tissue of the second classifier.

[0055] In some embodiments, the first classifier and the second classifier are trained without early stopping. In some embodiments, the second classifier is trained using one or more of the following machine learning techniques: stochastic gradient descent, weight decay, dropout regularization, Adam optimization, Ho initialization, learning rate scheduling, rectified linear unit activation function, leaky rectified linear unit activation function, sigmoid activation function, and boosting.

[0056] In some embodiments, determining the first training data set by processing the plurality of sequence reads comprises determining a probability of observed methylation in the nucleic acid fragments of the plurality of biological samples. In some embodiments, the probability of observed methylation is determined for each of a plurality of CpG sites in the plurality of sequence reads.

[0057] In some embodiments, determining the first training data set by processing the multiple sequence reads includes: determining whether the multiple sequence reads are hypomethylated or hypermethylated by determining, for each of the multiple sequence reads, whether at least a threshold number of CpG sites having at least a threshold percentage of CpG sites are unmethylated or methylated, respectively.

[0058] In some embodiments, determining the first training data set by processing the plurality of sequence reads comprises determining that one or more of the plurality of sequence reads are hypomethylated by determining that a threshold number or percentage of CpG sites corresponding to the one or more of the plurality of sequence reads are unmethylated. In some embodiments, determining the first training data set by processing the plurality of sequence reads comprises determining that one or more of the plurality of sequence reads are hypomethylated by determining that a threshold number or percentage of CpG sites corresponding to the one or more of the plurality of sequence reads are unmethylated.

[0059] In some embodiments, determining the first training data set by processing the multiple sequence reads includes determining that one or more of the multiple sequence reads are abnormally methylated; and filtering the multiple sequence reads using p-value filtering to generate the first training data set, wherein the p-value filtering includes removing sequence reads having a p-value less than a threshold p-value.

[0060] In some embodiments, the method further comprises determining, by the second classifier, a score indicating a probability that the source tissue associated with the disease state is present in the second test biological agent; and correcting the score. In some embodiments, a k-nearest neighbor operation associated with the score is performed by the second classifier using a feature space output. In some embodiments, the feature space comprises predictive markers indicating at least first and second source tissues associated with first and second disease states, respectively, present in the second test biological sample. In some embodiments, the feature space further comprises an indication that a correct source tissue prediction for the second test biological sample is different from the first and second source tissues.

[0061] In some embodiments, correcting the score comprises normalizing the probability using a different probability of presence of at least one disease state present in the second biological sample, the different probability determined by the first classifier.

[0062] In some embodiments, the method further comprises determining, by the first classifier, a probability of the at least one disease state being present in the first test biological sample; and predicting the presence of the at least one disease state in the first test biological sample in response to determining that the probability is greater than a binary threshold. In some embodiments, the specificity of the binary threshold is between 90% and 99.9%. In some embodiments, the second test biological sample has a probability predicted by the first classifier that is greater than the binary threshold.

[0063] In some embodiments, the first test biological sample is the second test biological sample.

[0064] In some embodiments, the method further comprises determining, by the second classifier, a probability of the source tissue associated with the disease state being present in the second test biological sample; and predicting the presence of the source tissue associated with the disease state in the second test biological sample in response to determining that the probability is greater than a source tissue threshold. In some embodiments, the method further comprises determining, by the second classifier, different probabilities of different source tissues associated with different disease states being present in the second test biological sample; and predicting the presence of the different source tissues associated with the different disease states in the second test biological sample in response to determining that the different probabilities are greater than a second source tissue threshold.

[0065] In some embodiments, the method further comprises determining, for the second classifier, a tissue of origin threshold associated with a given disease state by: determining a sensitivity rate of the second classifier at a given specificity rate for a plurality of different probabilities of a candidate tissue of origin threshold. In some embodiments, the sensitivity rate is determined by the first classifier using a scoring output. In some embodiments, the sensitivity rate is determined by the second classifier using a scoring output to stratify samples.

[0066] In some embodiments, the method further comprises optimizing a tradeoff between a sensitivity rate and a specificity rate of the second classifier for a given disease state. In some embodiments, the subset of the plurality of biological samples is labeled as having a cancer of known tissue of origin based on information from a reference sample.

[0067] In various embodiments, a system includes a computer processor and a memory storing computer program instructions that, when executed by the computer processor, cause the processor to perform any of the methods described herein. In various embodiments, a non-transitory computer readable medium stores one or more programs that include instructions that, when executed by an electronic device including a processor, cause the device to perform any of the methods described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a flow chart of a method for generating a classifier to predict a disease state according to various embodiments.

[0069] Figure 2A A flow chart of an apparatus for sequencing a nucleic acid sample according to one embodiment is shown.

[0070] Figure 2B is a block diagram of a processing system for processing sequence reads according to various embodiments.

[0071] Figure 3 is a flow chart describing a process for sequencing a nucleic acid according to various embodiments.

[0072] Figure 4A Shown according to various embodiments Figure 3 As part of the process, nucleic acids are sequenced to obtain methylation information and methylation state vectors.

[0073] Figure 4B Data structures for generating a control group according to various embodiments are shown.

[0074] Figure 4CA flow chart is shown describing a process for determining aberrantly methylated fragments from a sample according to various embodiments.

[0075] Figure 5 Blocks of a reference genome according to various embodiments are shown.

[0076] Figure 6 A process of determining features to train a classifier is shown according to various embodiments.

[0077] Fig. 7A , Figure 7B and Figure 7C A confusion matrix indicating the accuracy of the classifier according to various embodiments is included.

[0078] Figure 8 is a flow chart of a model-based characterization method according to various embodiments.

[0079] Fig.9A and Fig. 9B The sensitivity of the tissue of origin classifier according to one embodiment is shown.

[0080] Fig. 10A and Fig. 10B The sensitivity of the tissue-of-origin classifier at different cancer stages according to one embodiment is shown.

[0081] Fig.11 A performance grid representing source tissue localization accuracy is shown according to one embodiment.

[0082] Fig.12 The accuracy and sensitivity of the tissue of origin classifier at different cancer stages according to one embodiment are shown.

[0083] Fig.13A and Fig. 13B Shown is a ROC curve for a tissue of origin classifier according to one embodiment.

[0084] Fig.14 A data flow diagram for training a model according to various embodiments is described.

[0085] Fig.15 Precision-recall curves for an uncertain recall threshold are shown according to various embodiments.

[0086] Fig.16 is a flow chart of a method for determining a probability that a sample has a disease state, according to various embodiments.

[0087] Fig.17 The performance gain in sensitivity of a multilayer perceptron model according to one embodiment is shown.

[0088] Fig.18 Experimental results of a multilayer perceptron model in determining tissue of origin according to one embodiment are described.

[0089] Fig.19 Experimental results of a multilayer perceptron model in determining tissue of origin by cancer staging according to one embodiment are shown.

[0090] Fig. 20 Experimental results of a multilayer perceptron model across cancer types according to one embodiment are shown.

[0091] Fig.21 A plot showing the likelihood of cancer type for non-cancer samples with greater than 95% specificity.

[0092] Fig. 22 A graph showing methylation sequencing data for non-cancer samples and hematological subtype cancer samples.

[0093] Fig.23A A flow chart is shown describing a process for determining a binary threshold cutoff value for a binary cancer classification in accordance with one or more embodiments.

[0094] Fig. 23B A flow chart is shown describing a process for determining a tissue of origin marker threshold for a binary threshold cutoff value for binary cancer classification in accordance with one or more embodiments.

[0095] Fig.24A and Fig. 24B Shown is a confusion matrix demonstrating the performance of the trained cancer tissue-of-origin classifier with additional hematological cancer subtypes.

[0096] Fig.25A and Fig.25B Shown are plots demonstrating the cancer prediction accuracy of the cancer classifier with and without adjusting the threshold cutoffs for multiple cancer types at various cancer stages.

[0097] Fig.26A Receiver operating curves (ROC) showing sensitivity and specificity for detecting cancer using methylation data of target genomic regions of assay panel A are depicted.

[0098] Fig.26B is a confusion matrix describing the accuracy of classifying cancer type in subjects with cancer using methylation data of the target genomic region of assay panel A.

[0099] Fig.27A Receiver operating curves (ROC) showing sensitivity and specificity for detecting cancer using methylation data of target genomic regions of assay panel B are depicted.

[0100] Fig.27Bis a confusion matrix describing the accuracy of classifying cancer type for subjects with cancer using methylation data of the target genomic region of assay panel B.

[0101] Fig.28 Shown is the classifier performance of a proprietary cancer assay panel (Assay Panel C) according to one embodiment.

[0102] Fig.29 Shown is a tissue of origin (TOO) confusion matrix representing the accuracy of cancer tissue of origin localization for assay panel C, according to one embodiment.

[0103] Fig.30 Shown is the classifier sensitivity performance among individual categories (by stage) for assay panel C according to one embodiment.

[0104] Fig.31 The source tissue accuracy for multiple iterations of training models according to various embodiments is shown.

[0105] Fig.32 A process of stratifying hematology signals into two layers is shown according to various embodiments. DETAILED DESCRIPTION

[0106] Reference will now be made in detail to several embodiments, examples of which are shown in the accompanying drawings. It is noted that, wherever practicable, similar or like reference numerals may be used in the drawings and may indicate similar or like functions. It should also be noted that the contents of all published materials (patent applications, patents, papers, conference proceedings, etc.) cited herein are incorporated herein by reference in their entirety.

[0107] I. Definitions

[0108] Unless otherwise defined, all technical and scientific terms used herein have the meanings commonly understood by those skilled in the art to which this specification belongs. As used herein, the following terms have the meanings given below.

[0109] The term "subject" refers to a human individual. The term "healthy individual" refers to an individual who is assumed not to have cancer or disease.

[0110] The term "subject" refers to an individual whose DNA is being analyzed. A subject can be a test subject whose DNA is evaluated using whole genome sequencing or a targeted panel as described herein to assess whether the person has a disease state (e.g., cancer, cancer type, or cancer origin tissue). A subject can also be part of a control group known not to have cancer or another disease. A subject can also be part of a cancer or other disease group known to have cancer or another disease. Control groups and cancer / disease groups can be used to assist in the design or validation of targeted panels.

[0111] The term "control sample" refers to a sample obtained from a subject with a known disease state.

[0112] The term "training sample" refers to a sample obtained from a known disease state that can be used to generate sequence reads. The training sample can be applied to a probabilistic model to generate features that can be used to classify the disease state.

[0113] The term "test sample" refers to a sample that may have an unknown disease state.

[0114] The term "sequence read" refers to a nucleotide sequence read from a sample obtained from an individual. A sequence read can be generated from a nucleic acid fragment in a sample. A sequence read can be a folded sequence read generated from multiple sequence reads of multiple amplicons derived from a single original nucleic acid molecule. In some embodiments, a sequence read can be a de-duplicated sequence read. A sequence read can be obtained by various methods known in the art.

[0115] The term "disease state" refers to the presence or absence of a disease, disease type, and / or disease-derived tissue. For example, in one embodiment, the present disclosure provides methods, systems, and non-transitory computer-readable media for detecting cancer (i.e., the presence or absence of cancer), cancer type, or cancer-derived tissue.

[0116] The term "tissue of origin" or "TOO" refers to an organ, group of organs, body region, or cell type that may generate or cause a disease state. For example, identification of the tissue of origin or cancer cell type often allows identification of the appropriate next steps to further diagnose, stage, and determine treatment.

[0117] As used herein, the term "methylation" refers to a chemical process by which a methyl group is added to a DNA molecule. Two of the four bases of DNA, cytosine ("C") and adenine ("A"), can be methylated. For example, the hydrogen atoms on the pyrimidine ring of the cytosine base can be converted into methyl groups to form 5-methylcytosine. Methylation often occurs at the dinucleotides of cytosine and guanine (referred to herein as "CpG sites"). In other cases, methylation can occur at the cytosine of a non-CpG site portion or at another non-cytosine nucleotide; however, these situations are relatively rare. In the present disclosure, for the sake of clarity, methylation is discussed with reference to CpG sites. However, the principles described herein are equally applicable to detecting methylation in a non-CpG background, including non-cytosine methylation. For example, adenine methylation has been observed in bacterial, plant, and mammalian DNA, although it has received much less attention.

[0118] In such embodiments, the wet lab assays used to detect methylation may be different from those described herein that are known in the art. In addition, the methylation status vector may contain elements of the vector that are typically sites where methylation occurs or does not occur (even if those sites are not specific CpG sites). Where substitution is used, the remaining processes described herein are the same, and thus the inventive concepts described herein are applicable to those other forms of methylation.

[0119] The term "CpG site" refers to a region of a DNA molecule where a cytosine nucleotide is followed by a guanine nucleotide in a linear sequence of bases in the 5' to 3' direction. "CpG" is short for 5'-phospho-G-3', i.e., cytosine and guanine are separated by only one phosphate group; a phosphate ester links any two nucleotides in DNA together. Cytosine in a CpG dinucleotide can be methylated to form 5-methylcytosine.

[0120] The term "methylation site" refers to a single site of a DNA molecule to which a methyl group can be added. A "CpG" site is the most common methylation site, but the methylation site is not limited to a CpG site. For example, DNA methylation can occur in the cytosine of CHG and CHH, where H is adenine, cytosine or thymine. The method and procedure of the present disclosure can also be used to assess cytosine methylation in the form of 5-hydroxymethylcytosine (see, e.g., WO 2010 / 037001 and WO 2011 / 127136, which are incorporated herein by reference) and its characteristics. The term "hypomethylation" or "hypermethylation" refers to the methylation state of a DNA molecule containing multiple CpG sites (e.g., more than 3, 4, 5, 6, 7, 8, 9, 10, etc.), wherein a higher percentage of CpG sites (e.g., more than 80%, 85%, 90% or 95%, or any other percentage within the range of 50%-100%) are unmethylated or methylated, respectively.

[0121] The terms "cell-free deoxyribonucleic acid," "cell-free DNA," or "cfDNA" refer to fragments of deoxyribonucleic acid that circulate in body fluids such as blood, sweat, urine, or saliva and originate from one or more healthy cells and / or one or more cancer cells.

[0122] The term "circulating tumor DNA" or "ctDNA" refers to fragments of deoxyribonucleic acid derived from tumor cells or other types of cancer cells that can be released into an individual's body fluids (such as blood, sweat, urine, or saliva) as a result of biological processes (such as apoptosis or necrosis of dying cells) or actively released by living tumor cells.

[0123] II. Method Overview

[0124] Figure 1is a flow chart of a method 100 of identifying a plurality of features for generating a classifier for predicting a disease state (eg, presence or absence of a disease, disease type, and / or tissue of origin of a disease) according to various embodiments. Figure 2B is a block diagram of a processing system 200 for processing sequence reads according to various embodiments. In some embodiments, the processing system 200 performs method 100 to process sequence reads of fragments from a nucleic acid sample. Method 100 includes, but is not limited to, the following steps: generating sequence reads; training probability models associated with each of a plurality of different disease states (e.g., different cancer types); applying the probability models to determine values ​​based on the probability that the sequence reads originate from a sample associated with each of a plurality of disease states associated with each probability model; identifying features by determining counts of sequence reads having values ​​exceeding a threshold; generating a classifier using the features, and optionally applying the classifier to predict a disease state and / or a tissue of origin associated with a disease state. Each step is described in terms of components of the processing system 200, and see Figures 2-6. In Figure 2B In the embodiment shown in , the processing system 200 includes a sequence processor 210, a machine learning engine 220, a probability model 230, and a classifier 240.

[0125] In step 110, the sequence processor 210 generates a first set of sequence reads from a plurality of samples, each sample having a known or suspected disease state, such as the presence or absence of a disease, a disease type, and / or a disease-derived tissue. For example, in some embodiments, the plurality of samples may include any number of cancer samples from individuals known to have cancer and / or non-cancer samples from healthy individuals. In addition, the sample may include any cell-free nucleic acid sample (e.g., cfDNA), a solid tumor sample, and / or other types of samples. Those skilled in the art will appreciate that a new generation sequencing program can generate multiple sequence reads from a single original nucleic acid molecule. Therefore, in some embodiments, the sequence processor 210 may use known methods to perform deduplication and / or fold sequence reads to remove duplicate sequence reads and identify single sequence reads of a single original nucleic acid molecule, from which one or more original sequence reads are generated.

[0126] II.A. Assay Protocol

[0127] Figure 3 is a flow chart describing a nucleic acid sequencing process 300 according to one embodiment. In some embodiments, as Figure 1 As part of step 110 of method 100 , process 300 is performed to generate sequence reads.

[0128] In step 310, a nucleic acid sample (e.g., DNA or RNA) is extracted from the subject. In the present disclosure, unless otherwise specified, DNA and RNA can be used interchangeably. That is, the embodiments described herein are applicable to nucleic acid sequences of DNA and RNA types. However, for the purpose of clarity and explanation, the examples described herein can focus on DNA. The sample can include any subset from the human genome, including nucleic acid molecules of the whole genome. The sample can include blood, plasma, serum, urine, feces, saliva, other types of body fluids, or any combination thereof. In some embodiments, the invasiveness of the method for extracting a blood sample (e.g., syringe or finger puncture) may be less than the invasiveness of the procedure for obtaining a tissue biopsy, or surgery may be required. The extracted sample may include cfDNA and / or ctDNA. If the subject has a disease state, such as cancer, the cell-free nucleic acid (e.g., cfDNA) in the sample extracted from the subject generally includes a detectable level of nucleic acid that can be used to assess the disease state.

[0129] In step 315, the extracted nucleic acid (e.g., comprising cfDNA fragments) is treated to convert unmethylated cytosine to uracil. In some embodiments, method 300 uses bisulfite treatment of the sample, which converts unmethylated cytosine to uracil without converting methylated cytosine. For example, a commercially available kit such as EZ DNA Methylation TM –Gold,EZ DNA Methylation TM – Direct or EZ DNA Methylation TM - Kit (available from Zymo Research Corp (Irvine, CA)) for bisulfite conversion. In another embodiment, an enzymatic reaction is used to convert unmethylated cytosine to uracil. For example, the conversion can be performed using a commercially available kit to convert unmethylated cytosine to uracil, e.g., APOBEC-Seq (NEBiolabs, Ipswich, MA).

[0130] In step 320, a sequencing library is prepared. In some embodiments, the preparation includes at least two steps. In the first step, a ssDNA aptamer is added to the 3'-OH end of the bisulfite-converted ssDNA molecule using a ssDNA ligation reaction. In some embodiments, the ssDNA ligation reaction uses CircLigase II (Epicentre) to connect the ssDNA aptamer to the 3'-OH end of the bisulfite-converted ssDNA molecule, wherein the 5'-end of the aptamer is phosphorylated and the bisulfite-converted ssDNA has been dephosphorylated (i.e., the 3' end has a hydroxyl group). In another embodiment, the ssDNA ligation reaction uses a thermostable 5'App DNA / RNA ligase (available from New England BioLabs (Ipswich, MA)) to connect the ssDNA aptamer to the 3'-OH end of the bisulfite-converted ssDNA molecule. In this example, the first UMI aptamer is adenylated at the 5' end and blocked at the 3' end. In another embodiment, the ssDNA ligation reaction uses T4 RNA ligase (available from New England BioLabs) to ligate the ssDNA aptamer to the 3'-OH end of the bisulfite converted ssDNA molecule.

[0131] In the second step, the second strand DNA is synthesized in an extension reaction. For example, an extension primer (which hybridizes to a primer sequence contained in an ssDNA aptamer) is used in a primer extension reaction to form a bisulfite-converted double-stranded DNA molecule. Optionally, in some embodiments, the extension reaction uses an enzyme that can read uracil residues in the bisulfite-converted template strand.

[0132] Optionally, in a third step, a dsDNA aptamer is added to the double-stranded DNA molecule converted by bisulfite. The double-stranded DNA converted by bisulfite can then be amplified to add the sequencing aptamer. For example, PCR amplification using a forward primer comprising a P5 sequence and a reverse primer comprising a P7 sequence is used to add the P5 and P7 sequences to the bisulfite-converted DNA. Optionally, during library preparation, a unique molecular identifier (UMI) can be added to a nucleic acid molecule (e.g., a DNA molecule) by aptamer ligation. The UMI is a short nucleic acid sequence (e.g., 4-10 base pairs) that is added to the end of a DNA fragment during aptamer ligation. In some embodiments, the UMI is a degenerate base pair that serves as a unique tag that can be used to identify sequence reads derived from a specific DNA fragment. During PCR amplification after aptamer ligation, the UMI is replicated with the attached DNA fragment, which provides a way to identify sequence reads from the same original fragment in downstream analysis.

[0133] In optional step 325, nucleic acids (e.g., fragments) can be hybridized. Hybridization probes (also referred to herein as "probes") can be used to target and pull down nucleic acid fragments that provide information about disease states. For a given workflow, the probes can be designed to anneal (or hybridize) to a target (complementary) DNA or RNA strand. The target strand can be a "plus" strand (e.g., a strand that is transcribed into mRNA and subsequently translated into protein) or a complementary "minus" strand. The probes can range in length from tens, hundreds, or thousands of base pairs. Moreover, the probes can cover overlapping portions of the target region.

[0134] In optional step 330, the hybridized nucleic acid fragments are captured and can be enriched, for example, using PCR amplification. In some embodiments, the targeted DNA sequence can be enriched from the library. This is used, for example, in situations where a targeted panel assay is performed on a sample. For example, the target sequence can be enriched to obtain an enriched sequence that can be subsequently sequenced. In general, any method known in the art can be used to separate and enrich the target nucleic acid hybridized to the probe. For example, as is well known in the art, a biotin moiety can be added to the 5' end of the probe (i.e., biotinylation) to facilitate separation of the target nucleic acid hybridized to the probe using a streptavidin-coated surface (e.g., streptavidin-coated beads).

[0135] In step 335, a sequence is generated from a nucleic acid sample (e.g., an enriched sequence) to read. Sequencing data can be collected from the DNA sequence of the enrichment by methods well known in the art. For example, the method may include a new generation of sequencing (NGS) technology, including synthesis technology (Illumina), pyrophosphate sequencing (454 Life Sciences), ion semiconductor technology (Ion Torrent sequencing), single molecule real-time sequencing (Pacific Biosciences), connection sequencing (SOLiD sequencing), nanopore sequencing (Oxford Nanopore Technologies) or paired end sequencing. In some embodiments, sequencing by synthesis is used to perform large-scale parallel sequencing using a reversible dye terminator.

[0136] The sequence processor 210 may generate methylation information using the sequence reads in step 340. The methylation information determined from the sequence reads may then be used to generate a methylation state vector. Figure 4B A process 360 is shown according to one embodiment, which starts with Figure 3The process 300 of sequencing a cfDNA molecule in 310 begins to obtain a methylation state vector 352. As an example, the analysis system receives a cfDNA molecule 312, in which the molecule contains three CpG sites. As shown, the first and third CpG sites of the cfDNA molecule 312 are methylated 314. During processing step 315, the cfDNA molecule 312 is converted to produce a converted cfDNA molecule 322. During processing 315, the cytosine of the unmethylated second CpG site is converted to uracil. However, the first and third CpG sites are unconverted.

[0137] After conversion, a sequencing library 330 is prepared, and sequencing is performed to generate a sequence read 342. The analysis system compares the sequence read 342 with a reference genome 344 (not shown). The reference genome 344 provides a background about which position in the human genome the fragment cfDNA is derived from. In this simplified example, the analysis system compares the sequence read 342 so that three CpG sites are related to CpG sites 23, 24 and 25 (any reference identifier used for convenience of description). Therefore, the analysis system generates information about the methylation status of all CpG sites on the cfDNA molecule 312 and the position of the CpG sites mapped to the human genome. As shown, the methylated CpG sites on the sequence read 342 can be interpreted as cytosine. In this example, cytosine only appears in the first and third CpG sites of the sequence read 342, which makes it possible to infer that the first and third CpG sites in the original cfDNA molecule are methylated. The second CpG site is read as thymine (U is converted to T during sequencing), so it can be inferred that the second CpG site is not methylated in the original cfDNA molecule. Using these two pieces of information, methylation status and position, the analysis system 200 generates a methylation state vector 352 for the fragment cfDNA 312. In this example, the resulting methylation state vector 352 is <M 23 , U 24 、M 25 >, where M corresponds to methylated CpG sites, U corresponds to unmethylated CpG sites, and the subscript numbers correspond to the position of each CpG site in the reference genome.

[0138] II.B. Identification of aberrant fragments

[0139] In some embodiments, the analysis system uses the methylation state vector of the sample to determine the abnormal fragment of the sample. For example, for each nucleic acid molecule or fragment in the sample, the analysis system uses the methylation state vector corresponding to the nucleic acid molecule to determine whether the nucleic acid molecule or fragment is an abnormal methylated molecule or fragment relative to the expected methylation state vector from a healthy sample (by reading and analyzing the sequence derived therefrom). In one embodiment, the analysis system calculates the p-value score of each methylation state vector, describing the probability of observing the methylation state vector or other methylation state vectors in a healthy control group or even lower (as described in, for example, U.S. Patent Application Publication No. 2019 / 0287652, which is incorporated herein by reference). The process of calculating the p-value score is also discussed in the following section II.BiP-value filtering. The analysis system can determine and optionally filter out the sequence reading of the nucleic acid molecule or fragment with a methylation state vector, which has a p-value score below a threshold value as an abnormal fragment. In another embodiment, the analysis system further marks the fragments of the CpG sites with at least a certain number of methylation or unmethylation percentages exceeding a certain threshold percentage as high methylation and low methylation fragments, respectively. Hypermethylated fragments or hypomethylated fragments may also be referred to as abnormal methylated fragments (UFXM). In other embodiments, the analysis system may execute various other probability models to determine abnormal molecules or fragments. Examples of other probability models include hybrid models, deep probability models, etc. In some embodiments, the analysis system may use any combination of the following processes to identify abnormal fragments. Using the identified abnormal fragments, the analysis system can filter the methylation state vector set of the sample for other processes, for example, for training and deploying a cancer classifier.

[0140] II. BIP-value filtering

[0141] In one embodiment, the analysis system calculates a p-value score for each methylation state vector compared to the methylation state vector of the fragments from the healthy control group. The p-value score describes the probability of observing a nucleic acid molecule with a methylation state that matches the methylation state vector in the healthy control group. In order to determine abnormally methylated DNA fragments, the analysis system uses a healthy control group, in which the majority of fragments are normally methylated. When performing this probabilistic analysis of determining abnormal fragments, the determination is of great significance compared to the group of control subjects that constitute the healthy control group. In order to ensure the robustness of the healthy control group, the analysis system can select a certain threshold number of healthy individuals to obtain a sample containing DNA fragments. The following Figure 4B Methods are described for generating a data structure for a healthy control group that can be used by an analysis system to calculate a p-value score. Figure 4C Methods for calculating p-value scores using the resulting data structure are described.

[0142] Figure 4B 4 is a flowchart of a process 400 for generating a data structure for a healthy control group according to one embodiment. To generate the data structure for a healthy control group, the analysis system receives a plurality of DNA fragments (e.g., cfDNA) from a plurality of healthy individuals. The methylation state vector of each fragment is identified by, for example, process 360.

[0143] Using the methylation state vector for each fragment, the analysis system subdivides 405 the methylation state vector into strings of CpG sites. In one embodiment, the analysis system subdivides 405 the methylation state vector so that the resulting strings are all less than a given length. For example, a methylation state vector of length 11 can be subdivided into strings of length less than or equal to 3, which will produce 9 strings of length 3, 10 strings of length 2, and 11 strings of length 1. In another example, subdividing a methylation state vector of length 7 into strings of length less than or equal to 4 will produce 4 strings of length 4, 5 strings of length 3, 6 strings of length 2, and 7 strings of length 1. If the length of the methylation state vector is less than or equal to the specified string length, the methylation state vector can be converted into a single string containing all CpG sites of the vector.

[0144] The analysis system 200 counts 410 the strings by counting the number of strings present in a control group having a specified CpG site as the first CpG site in the string and having a methylation state possibility for each possible CpG site and methylation state possibility in the vector. For example, at a given CpG site and considering a string length of 3, there are 2^3 or 8 string configurations. At a given CpG site, for each of the 8 possible string configurations, the analysis system counts 410 the number of times each methylation state vector possibility appears in the control group. Continuing with this example, for each initial CpG site x in the reference genome, this may involve counting the following quantities: <M x 、M x+1 、M x+2 >, <M x 、M x+1 , U x+2 >, ..., x , U x+1 , U x+2 >. The analysis system generates 415 a data structure storing statistical counts of each initial CpG site and string likelihood.

[0145] ​Setting the upper limit of string length has several benefits. First, according to the maximum length of the string, the size of the data structure created by the analysis system can be significantly increased. For example, the maximum string length is 4 fingers, and for a string of length 4, each CpG site has a minimum of 2^4 statistics. The maximum string length is increased to 5 fingers, and each CpG site has another 2^4 or 16 statistics, thereby doubling the statistics (and requiring computer memory) compared with the previous string length. Reducing the string size helps to keep the generation and performance of the data structure (for example, for later access, as described below) reasonable in terms of calculation and storage. Secondly, the statistical consideration of limiting the maximum string length is to avoid overfitting the downstream model using string counts. If the long CpG site string does not have a strong biological effect on the outcome (for example, abnormal prediction of the presence of cancer prediction), it may be problematic to calculate the probability based on increasing the CpG site string because it requires a large amount of data that may not be available, so the model is too sparse to be properly executed. For example, calculating the conditional abnormality / cancer probability on the first 100 CpG sites would require counts of strings in a data structure of length 100, and ideally, some strings that perfectly match the previous 100 methylation states. If only sparse counts are available for strings of length 100, then the data is insufficient to determine whether a given string of length 100 is abnormal in a test sample.

[0146] Figure 4C is a flow chart describing a process 420 for identifying aberrantly methylated fragments from an individual, according to one embodiment. In process 420, the analysis system generates methylation state vectors 352 from the subject's cfDNA fragments. The analysis system processes each methylation state vector as follows.

[0147] For a given methylation state vector, the analysis system enumerates 430 all possibilities (i.e., CpG site sets) of methylation state vectors having the same initial CpG site and the same length. Because each methylation state is typically either methylated or unmethylated, there are actually two possible states at each CpG site, and thus the count of different possibilities of the methylation state vector is determined according to a power of 2, so that a methylation state vector of length n involves 2 n In the case where the methylation state vector contains an uncertain state for one or more CpG sites, the analysis system can enumerate 430 possible methylation state vectors, where only CpG sites with observed states are considered.

[0148] By accessing the healthy control group data structure, the analysis system 200 calculates 440 the probability of observing each methylation state vector possibility for the identified initial CpG site and the methylation state vector length. In one embodiment, calculating the probability of observing a given possibility uses Markov chain probability to model the joint probability calculation. In other embodiments, a calculation method other than Markov chain probability is used to determine the probability of each observed methylation state vector possibility.

[0149] The analysis system uses the calculated probabilities for each possibility to calculate 450 a p-value score for the methylation state vector. In one embodiment, this includes identifying the calculated probability corresponding to the possibility of matching the methylation state vector in question. Specifically, this is the probability of having the same set of CpG sites as the methylation state vector or similarly having the same initial CpG sites and length. The analysis system sums the calculated probabilities for any possibility whose probability is less than or equal to the identified probability to generate a p-value score.

[0150] The p-value represents the probability of observing the methylation state vector of the fragment or other methylation state vectors that are extremely unlikely in the healthy control group. Therefore, a low p-value score usually corresponds to a methylation state vector that is rare in healthy individuals and causes the fragment to be marked as abnormally methylated relative to the healthy control group. A high p-value score is usually related to the methylation state vector being expected to exist in healthy individuals in a relative sense. If the healthy control group is a non-cancerous group, a low p-value mark indicates that the fragment is abnormally methylated relative to the non-cancerous group, and thus may indicate the presence of cancer in the test subject.

[0151] As described above, the analysis system calculates a p-value score for each of a plurality of methylation state vectors, each methylation state vector representing a cfDNA fragment in a test sample. To identify which fragments are abnormally methylated, the analysis system may filter 460 the set of methylation state vectors based on the p-value score. In one embodiment, filtering is performed by comparing the p-value score to a threshold value and retaining only those fragments below the threshold value. The threshold p-value score may be 0.1, 0.01, 0.001, 0.0001, or a similar order of magnitude.

[0152] Based on the example results from the process, the analysis system yielded a result that the median number (range) of segments with abnormal methylation patterns in participants without cancer in training was 2,800 (1,500-12,000), and the median number (range) of segments with abnormal methylation patterns in participants with cancer in training was 3,000 (1,200-220,000). These filtered sets of segments with abnormal methylation patterns can be used for downstream analysis as described below.

[0153] In one embodiment, the analysis system uses a 455 sliding window to determine the likelihood of the methylation state vector and calculate a p-value. Rather than enumerating the likelihood and calculating the p-value for the entire methylation state vector, the analysis system enumerates the likelihood and calculates the p-value only for a window of consecutive CpG sites, where the length of the window (CpG site) is at least shorter than the length of some fragments (otherwise, the window would be useless). The window length can be static, user-determined, dynamic, or otherwise selected.

[0154] When calculating the p-value of a methylation state vector that is larger than a window, the window identifies a sequential set of CpG sites from the vector starting from the first CpG site in the vector within the window. The analysis system calculates the p-value score of the window containing the first CpG site. The analysis system then "slides" the window to the second CpG site in the vector and calculates another p-value score for the second window. Therefore, for a window size of l and a methylation vector length of m, each methylation state vector will generate m-l+1 p-value scores. After completing the p-value calculations for each vector portion, the lowest p-value score in all sliding windows can be considered as the overall p-value score of the methylation state vector. In another embodiment, the analysis system summarizes the p-value scores of the methylation state vector to generate an overall p-value score.

[0155] The use of a sliding window helps to reduce the number of listed possibilities for the methylation state vector and the corresponding probability calculations that need to be performed separately. In a realistic example, a fragment may have more than 54 CpG sites. Instead of calculating the probabilities of 2^54 (~1.8×10^16) possibilities to produce a single p score, the analysis system can instead use a window of size 5 (for example), which can calculate 50 p values ​​for each of the 50 methylation state vector windows for the fragment. Each of the 50 calculations lists 2^5 (32) methylation state vector possibilities, which calculate a total of 50×2^5 (1.6×10^3) ​​probabilities. This can greatly reduce the calculations performed and does not significantly affect the accurate identification of abnormal fragments.

[0156] In embodiments with uncertain states, the analysis system can calculate a p-value score for CpG sites with uncertain states in the methylation state vector of the summary fragment. The analysis system identifies all possibilities that are consistent with all methylation states of the methylation state vector (excluding the uncertain state). The analysis system can assign a probability to the methylation state vector as the sum of the probabilities of the identified possibilities. As an example, the analysis system assigns the methylation state vector <M 1 ,I 2 , U 3 The probability of > is calculated as the methylation state vector <M 1 、M 2 , U 3 > and <M1 , U 2 , U 3 > is the sum of the probabilities of the likelihoods of methylation state vectors having one or more uncertain states, because the methylation states of CpG sites 1 and 3 are observed and are consistent with the fragment methylation states of CpG sites 1 and 3. This method of summing the CpG sites with uncertain states uses the calculation of the likelihood probabilities up to 2^i, where i represents the number of uncertain states in the methylation state vector. In other embodiments, a dynamic programming algorithm can be performed to calculate the probabilities of methylation state vectors having one or more uncertain states. Advantageously, the dynamic programming algorithm runs with current computing time.

[0157] In one embodiment, by caching at least some of the calculations, the computational burden of calculating probabilities and / or p-values ​​can be further reduced. For example, the analysis system can cache the probability calculations of the likelihood of a methylation state vector (or its window) in a transient or permanent memory. If other fragments have the same CpG site, caching the likelihood probability allows the p-score value to be effectively calculated without recalculating the potential likelihood probability. Similarly, the analysis system can calculate each possible p-value score of the methylation state vector associated with the CpG site set from the vector (or its window). The analysis system can cache the p-value score for determining the p-value score of other fragments containing the same CpG site. Typically, the p-value score of the likelihood of a methylation state vector with the same CpG site can be used to determine the p-value scores of different possibilities from the same CpG site set.

[0158] II.B.II. Hypermethylated and Hypomethylated Fragments

[0159] In some embodiments, the analysis system determines the abnormal fragment as a fragment having more than a threshold number of CpG sites and having more than a threshold percentage of methylated CpG sites or having more than a threshold percentage of unmethylated CpG sites; the analysis system identifies such fragments as hypermethylated fragments or hypomethylated fragments. Exemplary thresholds for the length of fragments (or CpG sites) include greater than 3, 4, 5, 6, 7, 8, 9, 10, etc. Example percentage thresholds for methylation or unmethylation include greater than 80%, 85%, 90% or 95% or any other percentage in the range of 50%-100%.

[0160] II.C. Exemplary Sequencers and Analysis Systems

[0161] Figure 2A and Figure 2B 2 is a flow chart of a system and apparatus for sequencing a nucleic acid sample according to one embodiment. The illustrative flow chart includes apparatus such as a sequencer 270 and an analysis system 200. The sequencer 270 and the analysis system 200 can work in tandem to perform one or more steps in the process described herein.

[0162] In various embodiments, the sequencer 270 receives the enriched nucleic acid sample 260. Figure 2A As shown in , the sequencer 270 may include a graphical user interface 275 (which enables the user to interact with a specific task (e.g., start sequencing or terminate sequencing)) and one or more loading stations 280 for loading a sequencing cartridge containing an enriched fragment sample and / or for loading a buffer required for performing a sequencing assay. Thus, once the user of the sequencer 270 has provided the required reagents and sequencing cartridge to the loading station 280 of the sequencer 270, the user may start sequencing by interacting with the graphical user interface 275 of the sequencer 270. Once started, the sequencer 270 performs sequencing and outputs sequence reads of the enriched fragments from the nucleic acid sample 260.

[0163] In some embodiments, the sequencer 270 is coupled to the analysis system 200 in a communication manner. The analysis system 200 includes a certain number of computing devices, which are used to process sequence reads for various applications (such as evaluating the methylation status of one or more CpG sites, variant calling, or quality control). The sequencer 270 can provide sequence reads in the form of BAM files to the analysis system 200. The analysis system 200 can be coupled to the sequencer 270 in a communication manner through wireless communication technology, wired communication technology, or a combination of wireless and wired communication technology. Typically, the analysis system 200 is configured to have a processor and a non-transitory computer-readable storage medium, which stores computer instructions that, when executed by a processor, cause the processor to process sequence reads or perform one or more steps of any method or process disclosed herein.

[0164] In some embodiments, the sequence reads can be compared with the reference genome using methods known in the art to determine the alignment position information. The alignment position can generally describe the starting position and end position of the region corresponding to the starting nucleotide base and the end nucleotide base of a given sequence read in the reference genome. Corresponding to methylation sequencing, the alignment position information can be summarized as indicating the first CpG site and the last CpG site contained in the sequence read according to the comparison with the reference genome. The alignment position information can further indicate the methylation state and position of all CpG sites in a given sequence read. The region in the reference genome may be associated with a gene or a fragment of a gene; therefore, the analysis system 200 can read one or more gene marker sequences aligned with the sequence read. In one embodiment, the fragment length (or size) is determined from the starting position and the end position.

[0165] In various embodiments, for example, when using a paired end sequencing method, the sequence read includes a read pair represented as R_1 and R_2. For example, the first read R_1 can be sequenced from the first end of a double-stranded DNA (dsDNA) molecule, and the second read R_2 can be sequenced from the second end of the double-stranded DNA (dsDNA). Therefore, the nucleotide base pairs of the first read R_1 and the second read R_2 can always be aligned with the nucleotide base pairs of the reference genome (for example, in the opposite direction). The alignment position information derived from the read pair R_1 and R_2 may include the starting position corresponding to the end of the first read (for example, R_1) in the reference genome and the end position corresponding to the end of the second read (for example, R_2) in the reference genome. In other words, the starting position and the end position in the reference genome represent the possible positions corresponding to the nucleic acid fragments in the reference genome. In one embodiment, the read pair R_1 and R_2 can be assembled into fragments, and the fragments can be used for subsequent analysis and / or classification. An output file in the form of a SAM (sequence alignment map) or a BAM (binary) form can be generated and output for further analysis.

[0166] Now refer to Figure 2B , Figure 2B 2 is a block diagram of an analysis system 200 for processing a DNA sample according to one embodiment. The analysis system includes one or more computing devices for analyzing a DNA sample. The analysis system 200 includes a sequence processor 210, a sequence database 215, a model database 225, one or more probability models 230 and / or one or more classifiers 240, and a parameter database 235. In some embodiments, the analysis system 200 performs one or more steps in a method or process disclosed herein.

[0167] The sequence processor 210 generates a methylation state vector from the sample fragments. At each CpG site on the fragment, the sequence processor 210 generates a methylation state vector from the sample fragments. Figure 4B The process 360 generates a methylation state vector for each fragment, which specifies the position of the fragment in the reference genome, the number of CpG sites in the fragment, and the methylation state of each CpG site in the fragment (methylated, unmethylated, or uncertain). The sequence processor 210 can store the methylation state vector of the fragment in the sequence database 215. The data in the sequence database 215 can be organized so that the methylation state vectors from the samples are related to each other.

[0168] In addition, a plurality of different models 230 can be stored in the model database 225 or retrieved for use with a test sample. In one example, the model is a trained cancer classifier 240 that is used to determine a cancer prediction for a test sample using a feature vector derived from an abnormal segment. The training and use of a cancer classifier is discussed elsewhere herein. The analysis system 200 can train one or more models 230 and / or one or more classifiers 240 and store various trained parameters in a parameter database 235. The analysis system 200 stores the models 230 and / or classifiers and functions in the model database 225.

[0169] During inference, the machine learning engine 220 uses one or more models 230 and / or classifiers 240 to return outputs. The machine learning engine accesses the models 230 and / or classifiers 240 in the model database 225 and the training parameters from the parameter database 235. According to each model, the machine learning engine 220 receives the appropriate inputs for the model and calculates the outputs based on the received inputs, parameters, and input and output related functions of each model. In some use cases, the machine learning engine 220 further calculates metrics related to the confidence of the calculated outputs from the model. In other use cases, the machine learning engine 220 calculates other intermediate values ​​used in the model.

[0170] II.B. Blocks of the Reference Genome

[0171] Figure 5Blocks of a reference genome according to one embodiment are shown. The sequence processor 210 may segment the reference genome (or a subset of the reference genome) into one or more segments, for example, in a use case involving targeted methylation assays. For example, the sequence processor 210 divides the reference genome into blocks of CpG sites. Each block is defined when the interval between two adjacent CpG sites exceeds a threshold (e.g., greater than 200 base pairs (bp), 300 bp, 400 bp, 500 bp, 600 bp, 700 bp, 800 bp, 900 bp, or 1,000 bp, among other values). Therefore, the base pair sizes of the blocks may be different. For each block, the sequence processor 210 may subdivide the block into windows of a certain length, such as 500 bp, 600 bp, 700 bp, 800 bp, 900 bp, 1,000 bp, 1,100 bp, 1,200 bp, 1,300 bp, 1,400 bp, or 1,500 bp, among other values. In other embodiments, the length of the window may be 200 bp to 10 kilobase pairs (kbp), 500 bp to 2 kbp, or about 1 kbp. A certain number of base pairs or a certain percentage of length (e.g., 10%, 20%, 30%, 40%, 50%, or 60%, among other values) of windows (e.g., adjacent windows) may overlap. The interval between two adjacent CpG sites in the window may exceed a threshold, for example, greater than 200 base pairs (bp), 300 bp, 400 bp, 500 bp, 600 bp, 700 bp, 800 bp, 900 bp, or 1,000 bp, among other values.

[0172] The sequence processor 210 can use a windowing program to analyze sequence readings derived from DNA fragments. In particular, the sequence processor 210 scans blocks window by window and reads the fragments in each window. The fragments may be derived from tissues and / or high signal cfDNA. High signal cfDNA samples may be determined by a binary classification model, by cancer stage, or by another metric. By segmenting the reference genome (e.g., using blocks and windows), the sequence processor 210 can facilitate computational parallelization. In addition, the sequence processor 210 can reduce the computational resources for processing the reference genome by targeting base pair segments containing CpG sites and skipping other segments that do not contain CpG sites.

[0173] III. Model-based Feature Engineering and Classification

[0174] III.A. Model-Based Feature Engineering

[0175] According to one embodiment, Figure 8As shown in, the present disclosure relates to model-based feature engineering for deriving features useful for disease state classification. As described elsewhere herein, disease state can be disease presence or absence, disease type and / or disease tissue or source. For example, as described herein, disease state can be cancer presence or absence, cancer type and / or cancer source tissue. Cancer type and / or cancer source tissue can be selected from the following, including: breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, renal pelvis ureter urothelial carcinoma, renal cancer other than urothelial carcinoma, prostate cancer, anorectal cancer, colorectal cancer, esophageal cancer, gastric cancer, hepatobiliary carcinoma derived from hepatocytes, hepatobiliary carcinoma derived from cells other than hepatocytes, pancreatic cancer, squamous cell carcinoma of the upper digestive tract, upper digestive tract cancer other than squamous cell carcinoma, head and neck cancer, lung cancer (such as lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer and cancer other than adenocarcinoma or small cell lung cancer), neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma and leukemia and other cancer types.

[0176] In step 810, as described elsewhere herein, a first plurality of sequence reads are generated from a first reference sample having a first disease state, and a second plurality of sequence reads are generated from a second reference sample having a second disease state. The first plurality of sequence reads and the second plurality of sequence reads can be more than 10,000, more than 50,000, more than 100,000, more than 200,000, more than 500,000, more than 1,000,000, more than 2,000,000, more than 5,000,000 or more than 10,000,000 sequence reads. As used herein, a "reference sequence" is a sample obtained from a subject with a known disease state. In some embodiments, one or more reference samples with one or more known disease states can be used to train one or more probability models, which can then be used to derive features for classifying the disease state of an unknown test sample. The sample can be a genomic DNA (gDNA) sample or a cell-free DNA (cfDNA) sample. The reference sample can be a blood, plasma, serum, urine, feces, and saliva sample. Alternatively, the reference sample can be whole blood, blood components, tissue biopsy samples, pleural fluid, pericardial fluid, cerebrospinal fluid and peritoneal fluid. In some embodiments, the first reference sample is from a subject known to have cancer and the second reference sample is from a healthy subject or a non-cancer subject. In some embodiments, the first reference sample is from a subject known to have a first type of cancer (e.g., lung cancer) and the second reference sample is from a subject known to have a second type of cancer (e.g., breast cancer). In other embodiments, the first reference sample is from a subject known to have a first disease source tissue (e.g., lung disease) and the second reference sample is from a subject known to have a second disease state source tissue (e.g., liver disease).

[0177] In step 815, the machine learning engine 220 trains a first probability model 230 and a second probability model 230 from the first plurality of sequence reads and the second plurality of sequence reads (generated in step 110), respectively, each probability model being associated with a different disease state of one or more possible disease states. As previously described, the disease state can be the presence or absence of cancer, the type of cancer, and / or the tissue of origin of the cancer. In various embodiments, the training data is divided into K subsets (folds) for K-fold cross-validation. The folds can be balanced for the following factors: cancer / non-cancer status, tissue of origin, cancer stage, age (e.g., grouped in 10-year periods), gender, ethnicity, and smoking status, as well as other factors. Data from K-1 of the folds can be used as training data for the probability model, and the retained folds can be used as test data.

[0178] For a first disease state and a second disease state, the machine learning engine 220 trains the first and second probability models 230 by fitting each of the probability models 230 to the first and second plurality of sequence reads, respectively. For example, in one embodiment, the first probability model is fit using a first plurality of sequence reads derived from one or more samples from subjects known to have cancer, and the first probability model is fit using a second plurality of sequence reads derived from one or more samples from healthy subjects or non-cancer subjects. In other embodiments, the first probability model may be trained for a first type of cancer or a first source tissue, and the second probability model may be trained for a second type of cancer or a second source tissue. As will be appreciated by those skilled in the art, any number of disease state probability models may be trained using sequence reads derived from one or more samples obtained from subjects having any of a number of possible disease states. For example, in some embodiments, other cancer-specific probability models (i.e., models of other types of cancer and or tissue of origin) can be trained for a third, fourth, fifth, sixth, seventh, eighth, ninth, tenth, etc. (e.g., up to twenty, thirty, or more) particular cancer types and used to determine the probability that a sequence read from a training set or an unknown cancer type is more likely to be derived from one cancer type (or cancer's tissue of origin) than another cancer type (or cancer's tissue of origin), as described elsewhere herein.

[0179] As used herein, a "probabilistic model" is any mathematical model that can assign probabilities to sequence reads based on the methylation state at one or more sites on the sequence reads. During training, the machine learning engine 220 fits sequence reads derived from one or more samples from subjects with a known disease and can be used to utilize methylation information or methylation state vectors (e.g., previously generated for Figure 3-4 described) to determine the probability of a sequence read indicating a disease state. In particular, in one embodiment, the machine learning engine 220 determines the observed methylation rate of each CpG site within the sequence read. The methylation rate represents the score or percentage of methylated base pairs within the CpG site. The trained probability model 230 can be parameterized by the product of the methylation rate. In general, any known probability model can be used to assign probabilities to sequence reads from a sample. For example, the probability model can be a binomial model, in which a methylation probability is assigned to each site (e.g., a CpG site) on a nucleic acid fragment, or an independent site model, in which the methylation of each CpG is specified by different methylation probabilities and it is assumed that the methylation at one site is independent of the methylation at one or more other sites on the nucleic acid fragment.

[0180] In some embodiments, the probability model 230 is a Markov model in which the probability of methylation at each CpG site depends on the methylation status of a certain number of previous CpG sites in the sequence read or nucleic acid molecule derived from the sequence read. See, e.g., U.S. Patent Application No. 16 / 352,602, entitled "Anomalous Fragment Detection and Classification," and filed on March 13, 2019.

[0181] In some embodiments, the probability model 230 is a "hybrid model" that is fitted using a mixture component from a base model. For example, in some embodiments, multiple independent site models can be used to determine the mixture components, where it is assumed that the methylation (e.g., methylation rate) at each CpG site is independent of the methylation at other CpG sites. Using the independent site model, the probability of a nucleic acid molecule assigned to a sequence read or derived therefrom is the product of the methylation probability at each CpG site (where the sequence read is methylated), and 1 minus the methylation probability at each CpG site (where the sequence read is unmethylated). According to this embodiment, the machine learning engine 220 determines the methylation rate of each of the mixture components. The mixture model is parameterized by the sum of the mixture components, each component being related to the product of the methylation rates. The probability model Pr of n mixture components can be expressed as:

[0182]

[0183] For the input fragment, m i ∈{0,1} represents the observed fragment methylation state at position i of the reference genome, where 0 indicates unmethylated and 1 indicates methylated. The score of each mixture component k is assigned to f k , where f k ≥0 and The probability of methylation at position i in the CpG site of the mixture component k is β ki Therefore, the probability of unmethylation is 1-β ki The number of mixture components n can be 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, etc.

[0184] In some embodiments, the machine learning engine 220 fits the probability model 230 using maximum likelihood estimation to identify a set of parameters {β ... ki , f k}. The maximum number of N total fragments can be expressed as:

[0185]

[0186] As will be appreciated by those skilled in the art, other means can be used to fit the probability model to identify the parameters of the log-likelihood of all sequence reads derived from the reference sample for maximization. For example, in one embodiment, Bayesian fitting (for example, using Markov chain Monte Carlo method) is used, wherein each parameter is not assigned a single value, and is related to distribution. In other embodiments, gradient-based optimization is used, wherein the gradient about parameter value likelihood (or log-likelihood) is used to pass through the parameter space towards the best step. In other embodiments, expectation maximization is used, wherein the potential parameter set (for example, the attribute of the mixed component of each fragment derived) is set to its expected value under the previous model parameters, and then the model parameters are assigned to maximize the likelihood under the assumption of these potential variables. Then, the two-step process is repeated until convergence.

[0187] At step 820, a plurality of training sequence reads are generated from a training sample. The plurality of training sequence reads may be more than 10,000, more than 50,000, more than 100,000, more than 200,000, more than 500,000, more than 1,000,000, more than 2,000,000, more than 5,000,000, or more than 10,000,000 sequence reads. As used herein, a "training sample" is a sample obtained from a known disease state that can be used to generate sequence reads, which are then applied to a first and / or second probability model to generate features that can be used for disease state classification. In step 825, the processing system 200 applies the first and second probability models 230 to determine a first probability value and a second probability value for each sequence read in the plurality of training sequence reads. The first and second probability values ​​are determined based on the probability that the sequence reads are derived from samples associated with the first disease state and the second disease state, respectively. Processing system 200 can repeat step 130 (not shown) for any additional probability models 230 (eg, trained from sequence reads from a third, fourth, fifth, etc. reference sample).

[0188] At step 830, one or more features are identified by comparing the first probability value and the second probability value of each of the plurality of training sequence reads. In general, many methods can be used to compare the first probability value and the second probability value and identify features. For example, in one embodiment, one or more features include the count of outlier sequence reads in the plurality of training sequence reads whose first probability value is greater than the second probability value. The count can be a binary count, a total count of outlier sequence reads, or a total count of anonymous methylated sequence reads. In another embodiment, one or more features include counts of sequence reads or fragments containing a specific methylation pattern. For example, one or more features can be counts of sequence reads or fragments that are fully methylated at each CpG site, counts of sequence reads or fragments that are partially methylated (e.g., at least 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90% or 95% methylated). In another embodiment, the output of a discriminant classifier trained in a single genomic region (e.g., the discriminant classifier can be a multilayer perceptron or a convolutional neural network model) is used to identify one or more features. In another embodiment, comparing the first probability value and the second probability value includes determining a ratio of the first probability value to the second probability value, and the one or more features include sequence read counts of sequence reads exceeding a ratio threshold.

[0189] In another embodiment, the first probability value or the second probability value is a log-likelihood value. For example, the processing system 200 can use the fitted probability model associated with the first disease state and the second disease state to calculate the log-likelihood ratio R. Specifically, the probability Pr of the methylation pattern observed on the fragments of the sample associated with the first disease state and the second disease state can be used to calculate the log-likelihood ratio:

[0190]

[0191] The processing system 200 can use multiple levels of thresholds to identify features. For example, the levels include thresholds 1, 2, 3, 4, 5, 6, 7, 8, and 9. In some embodiments, a smoothing function can be applied. For example, in response to determining that R is (e.g., significantly) less than the level value, the processing system 200 assigns a feature value of about 0; in response to determining that R is equal to the level value, the processing system 200 assigns a feature value of 0.5; in response to determining that R is (e.g., significantly) greater than the level value, the processing system 200 assigns a feature value of about 1. Each level indicates a different threshold at which a fragment (by which a sequence read is generated) is more likely to come from a sample associated with a disease state than from a healthy sample. The processing system 200 can use a threshold to determine a count of outlier fragments, which can be used as a feature.

[0192] By using threshold filtering, the processing system 200 may treat certain fragments as outliers because these fragments are unlikely to be present in healthy samples. Therefore, outlier fragments may be considered more likely to be associated with (e.g., originate from) a disease state or a cancer sample. The number of features may vary between different levels, for example, one level may have a different number of features than another level based on a corresponding threshold. In other embodiments, the processing system 200 uses a different number of levels or other thresholds. Other ways to identify features or to sort identified features based on an indicator of features that distinguish between different disease states (e.g., using mutual information to determine an indicator of the information content of a feature when distinguishing between two disease states) are described elsewhere herein.

[0193] In other embodiments, the processing system 200 may use different types of ratios or equations to identify multiple features. The machine learning engine 220 may determine segments indicative of a disease state (e.g., cancer) based on whether at least one log likelihood ratio considered for various disease states is above a threshold.

[0194] Subsequently, as described in further detail elsewhere herein, a plurality of features may be used to train a disease state classifier. For example, in some embodiments, a plurality of features may be used to train a classifier to classify the presence or absence of cancer, cancer type, and / or cancer tissue of origin.

[0195] III.B. Classification of Disease State Origin Tissues

[0196] According to another embodiment, Figure 1 As described in step 120 of , the machine learning engine 220 trains probability models 230, each probability model being associated with a different disease state of a set of multiple different disease states. For clarity, Figure 1 Model-based characterization and training of classifiers for classifying tissue of origin of disease states are described. However, as previously described, in various embodiments, the disease state can be the presence or absence of cancer, the type of cancer, and / or the tissue of origin of the cancer. In addition, the disease state can be associated with another type of disease (not necessarily associated with cancer) or be a healthy state (absence of cancer or disease).

[0197] The machine learning engine 220 trains the probability model 230 using one or more sequence read sets, wherein each of the one or more sequence read sets is generated (according to step 110) from a different disease state of a set of a plurality of different disease states. The disease states may include any number of selected cancer types or cancer source tissues including: breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, renal pelvic ureteral urothelial carcinoma, renal cancer other than urothelial carcinoma, prostate cancer, anorectal cancer, colorectal cancer, esophageal cancer, gastric cancer, hepatobiliary carcinoma derived from hepatocytes, hepatobiliary carcinoma derived from cells other than hepatocytes, pancreatic cancer, squamous cell carcinoma of the upper gastrointestinal tract, upper gastrointestinal cancer other than squamous cell carcinoma, head and neck cancer, lung cancer (such as lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer, and cancer other than adenocarcinoma or small cell lung cancer), neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma and leukemia, and other cancer types.

[0198] For each of the multiple disease states, the machine learning engine 220 trains the probability model 230 by fitting the probability model 230 to the sequence reads derived from each sample corresponding to each disease state. For example, in some embodiments, the probability model can be trained for a specific type of cancer. According to this embodiment, a cancer-specific probability model can be trained for a first, second, third, etc. specific cancer type and used to evaluate the cancer type (e.g., about an unknown test sample). For example, a sequence read set derived from one or more samples related to lung cancer can be used to fit a lung cancer-specific probability model. As another example, a sequence read set derived from one or more samples related to breast cancer is used to fit a breast cancer-specific probability model. In some embodiments, a tissue-specific probability model can be trained for a first, second, third, etc. tissue type and used to evaluate the disease state source tissue. For example, a sequence read set derived from a first tissue type (e.g., from a lung tissue sample, such as a lung tissue biopsy) can be used to fit a first source tissue probability model, and a sequence read set derived from a second tissue type (e.g., from a liver tissue sample, such as a liver tissue biopsy) can be used to fit a second source tissue probability model. Alternatively, in some embodiments, a cancer probability model is fitted using a set of sequence reads derived from one or more samples from subjects known to have cancer, and a non-cancer specific probability model is fitted using a set of sequence reads derived from one or more samples from healthy subjects or non-cancer subjects. As will be appreciated by those skilled in the art, any number of disease state probability models can be trained using sequence reads derived from one or more samples obtained from subjects having any of a number of possible disease states. For example, in some embodiments, multiple sequence reads can be generated from 3, 4, 5, 6, 7, 8, 9, 10 or more reference samples, each of which is obtained from one or more subjects having a different disease state (e.g., a different cancer type), and the sequence reads are used to train 3, 4, 5, 6, 7, 8, 9, 10 or more probability models.

[0199] During training, sequence reads indicative of disease states can be used to exploit methylation information or methylation state vectors (e.g., previously targeted Figure 3-4) to train the machine learning engine 220. In particular, the machine learning engine 220 determines the observed methylation rate of each CpG site within the sequence read. The methylation rate represents the fraction or percentage of methylated base pairs within the CpG site. The trained probability model 230 can be parameterized by the product of the methylation rates. As previously described, any known probability model can be used to assign probabilities to sequence reads from a sample. For example, the probability model can be a binomial model, in which a methylation probability is assigned to each site (e.g., a CpG site) on a nucleic acid fragment; or an independent site model, in which the methylation of each CpG is specified by a different methylation probability, and it is assumed that the methylation at one site is independent of the methylation at one or more other sites on the nucleic acid fragment.

[0200] In some embodiments, a Markov model is used in which the probability of methylation at each CpG site depends on the methylation status of a certain number of previous CpG sites in the sequence read or nucleic acid molecule derived from the sequence read. See, e.g., U.S. Patent Application No. 16 / 352,602, entitled "Anomalous Fragment Detection and Classification," and filed on March 13, 2019.

[0201] In some embodiments, the probability model 230 is a "hybrid model" that is fitted using a mixture component from a base model. For example, in some embodiments, multiple independent site models can be used to determine the mixture components, where it is assumed that the methylation (e.g., methylation rate) at each CpG site is independent of the methylation at other CpG sites. Using the independent site model, the probability of a nucleic acid molecule assigned to a sequence read or derived therefrom is the product of the methylation probability at each CpG site (where the sequence read is methylated), and 1 minus the methylation probability at each CpG site (where the sequence read is unmethylated). According to this embodiment, the machine learning engine 220 determines the methylation rate of each of the mixture components. The mixture model is parameterized by the sum of the mixture components, each component being related to the product of the methylation rates. The probability model Pr of n mixture components can be expressed as:

[0202]

[0203] For the input fragment, m i ∈{0,1} represents the observed fragment methylation state at position i of the reference genome, where 0 indicates unmethylated and 1 indicates methylated. The score of each mixture component k is assigned to f k , where f k ≥0 and The probability of methylation at position i in the CpG site of the mixture component k is β kiTherefore, the probability of unmethylation is 1-β ki The number of mixture components n can be 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, etc.

[0204] In some embodiments, the machine learning engine 220 fits the probability model 230 using maximum likelihood estimation to identify a set of parameters {β ... ki , f k}. The maximum number of N total fragments can be expressed as:

[0205]

[0206] In step 130, the processing system 200 applies the probability model 230 to calculate a value for each sequence read in a second set of sequence reads (e.g., different from the first set of sequence reads generated in step 110). The value is calculated based at least on the probability that the sequence read (and corresponding fragment) originates from a sample associated with a disease state of the probability model 230. The processing system 200 can repeat step 130 for each of the different probability models 230. In some embodiments, the processing system 200 calculates the value using a probability model that is fitted and associated with certain disease states for a likelihood ratio R. Specifically, the log-likelihood ratio can be calculated using the probability Pr of the methylation pattern observed on the fragments of samples associated with the disease state and healthy samples:

[0207]

[0208] In other embodiments, the processing system 200 may use different types of ratios or equations to calculate the value. The machine learning engine 220 may determine segments indicative of a disease state (eg, cancer) based on whether at least one log likelihood ratio considered for various disease states is above a threshold.

[0209] III.C. Feature Selection

[0210] Figure 6 2 shows a process for determining features for training a classifier according to one embodiment. As previously described, the machine learning engine 220 trains a probability model 230 associated with a disease state. Figure 6In the example shown in , the probability model 230 ("tissue model") is associated with non-cancer (healthy), breast cancer, and lung cancer. The processing system 200 processes one or more cfDNA and / or tumor samples to obtain fragments and uses the probability model 230 to assign a value to the fragments associated with non-cancer (healthy), breast cancer, and lung cancer. The processing system 200 can use information from the sequence reads of the cfDNA and / or type samples to identify features of the classifier. In some embodiments, the processing system 200 can obtain and assign fragments from each window of the partitioned reference genome, such as Figure 5 The processing system 200 aggregates the segments from the window into a sequence that is used to determine features for the classifier.

[0211] In step 140, the processing system 200 identifies features by determining the counts of sequence reads having values ​​exceeding a threshold. In embodiments where the value is based on a majority likelihood ratio R, the threshold is a threshold ratio. The processing system 200 can use multiple levels of thresholds to identify features. For example, the levels include thresholds 1, 2, 3, 4, 5, 6, 7, 8, and 9. Each level indicates a different threshold at which a fragment (from which a sequence read is generated) is more likely to be from a sample associated with a disease state than from a healthy sample. The processing system 200 can use a threshold to determine the counts of outlier fragments, which can be used as features.

[0212] By filtering using a threshold, the processing system 200 may treat certain fragments as outliers because the fragments are unlikely to exist in a healthy sample. Therefore, outlier fragments may be considered more likely to be associated with (e.g., derived from) a disease state or a cancer sample. The number of features may vary between different levels. In other embodiments, the processing system 200 uses different numbers of levels or other thresholds. In other embodiments, the processing system 200 may use other methods or scores (such as p-values) to filter fragments. In some embodiments, the processing system 200 calculates the p-value of the methylation state vector, which describes the probability of observing the methylation state vector or other methylation state vectors with a very small probability in the healthy control group. To determine abnormally methylated fragments, the processing system 200 uses a healthy control group in which most fragments are normally methylated (see, e.g., U.S. Patent Application No. 16 / 352,602, titled "Anomalous Fragment Detection and Classification", and submitted on March 13, 2019).

[0213] Processing system 200 may repeat steps 130 to 140 for each probability model trained in step 120. Thus, processing system 200 may identify one or more characteristics of a disease state associated with a probability model. Figure 6In the example shown in FIG. 2 , processing system 200 identifies one or more characteristics of breast cancer and lung cancer.

[0214] In some embodiments, the processing system 200 ranks the identified features based on an indicator of the features that distinguish between different disease states. For example, a feature is informative if it can distinguish certain cancer types from other cancer types or healthy samples. The processing system 200 can use the interaction information to determine an indicator of the information content of the features that distinguish between two disease states. For each pair of different disease states, the processing system 200 can designate one disease state (e.g., cancer type A) as a positive type and the other disease state (e.g., cancer type B) as a negative type.

[0215] The mutual information can be calculated using the estimated scores of samples of positive and negative types (e.g., cancer types A and B) whose features are expected to be non-zero in the final determination. For example, if a feature frequently appears in healthy cfDNA, the processing system 200 determines that the feature is unlikely to frequently appear in cfDNA associated with various types of cancer. Therefore, the feature may be a weak indicator for distinguishing disease states. When calculating the mutual information I, the variable X is a feature (e.g., binary) and the variable Y represents the disease state (e.g., cancer type A or B):

[0216]

[0217] p(1|A)=f A +f H -f H f A

[0218] The joint probability mass function of X and Y is p(x, y), and the boundary probability mass functions are p(x) and p(y). The processing system 200 may assume that feature missingness is uninformative and that any disease state is equally probable a priori, e.g., p(Y=A)=p(Y=B)=0.5. The probability of observing a given binary feature of cancer type A (e.g., in cfDNA) is represented by p(1|A), where f A is the probability of observing the feature in a ctDNA sample (or a high signal cfDNA sample) from a tumor associated with cancer type A, and f H is the probability of observing the feature in a healthy or non-cancer cfDNA sample.

[0219] In some embodiments, f is estimated by the fraction of cancer patients whose cfDNA is expected to contain non-zero eigenvalues. AValue. When the training data for cancer type A consists of cfDNA samples, this fraction can be estimated as the fraction of cfDNA samples where only the feature is observed. When the training data contains species samples, corrections can be made to compensate for tumor-derived fragments in cfDNA that have a fraction lower than the tumor. For N fragments in the species sample determined to have a value greater than the threshold (e.g., from step 140), the processing system 200 calculates the probability r of detecting each of those fragments in the cfDNA from the patient as:

[0220]

[0221] The probability of observing at least one fragment in cfDNA from this patient can then be calculated as p(N cfDNA >0) = 1-(1-r) N . To estimate f A , p(N cfDNA >0), where the probability can be assigned as 1 (for cfDNA samples with the feature), 0 (for cfDNA samples lacking the feature), and 1-(1-r) N (For species samples). In some embodiments, the estimate is based on a predetermined hypothetical value of the species fraction in cfDNA of an early-stage cancer patient (e.g., 0.1%), the cfDNA sequencing depth to be applied to the patient's final analysis (e.g., 1000x), and the tumor sequencing depth (e.g., 25x). H , the processing system 200 uses the fraction of positive samples to determine how many other samples will receive a positive detection classification at a greater sequencing depth.

[0222] III.D. Classification

[0223] In step 150, the processing system 200 generates a classifier using the features. The classifier is trained to predict the source tissue associated with the disease state for the input sequence reads of the test sample from the test subject. The processing system 200 can select a predetermined number (e.g., 1024) of grading features to train the classifier for each pair of disease states (e.g., based on mutual information calculation or another calculation index). The predetermined number can be regarded as a hyperparameter selected based on the performance in the cross validation. The processing system 200 can also select features that are determined to be more informative in distinguishing disease state pairs from the regions of the reference genome. In various embodiments, the processing system 200 maintains the best performance level for each region and each cancer type pair (including non-cancer as a negative type).

[0224] In some embodiments, the processing system 200 trains the classifier by inputting a training sample set with a feature vector into the classifier and adjusting the classification parameters so that the function of the classifier accurately associates the training feature vector with its corresponding label. The processing system 200 can group the training samples into one or more training sample sets for iterative batch training of the classifier. After inputting all sets of training samples (including their training feature vectors) and adjusting the classification parameters, the classifier can be fully trained to label the test sample according to the feature vector within a certain error limit. The processing system 200 can train the classifier according to any of many methods, for example, L1 regularized logistic regression or L2 regularized logistic regression (for example, using a logarithmic loss function), a generalized linear model (GLM), a random forest, a multinomial logistic regression, a multilayer perceptron, a support vector machine, a neural network, or any other suitable machine learning technology.

[0225] In various embodiments, the processing system 200 converts the feature value through binary conversion. In particular, feature values ​​greater than 0 are set to 1, so that the feature value is 0 or 1 (indicating the presence or absence of a disease state). In other embodiments, a smoothing function (e.g., to provide a more refined value) may be implemented instead of binarizing to 0 or 1. Fig.14 As shown in , the processing system 200 may binarize the features in a cross-validation before using the features to train a classifier.

[0226] In various embodiments, the processing system 200 uses the folded training data to train a multinomial logistic regression classifier and generates predictions for the retained data. For each of the K folds, the processing system 200 trains a logistic regression for each hyperparameter combination. An example hyperparameter is the L2 penalty, i.e., a regularized form of the weights applied to the logistic regression. Another example hyperparameter is topK, i.e., the number of high-level regions to be retained for each tissue type pair (including non-cancer). For example, in the case of topK=16, the processing system 200 retains the top 16 regions for each tissue type pair, as sorted by the mutual information procedure described herein. By following this procedure, the processing system 200 can generate predictions for each sample in the training set while ensuring that the classifier is not trained using the data that generates the predictions.

[0227] In various embodiments, for each hyperparameter set, the processing system 200 evaluates performance based on the cross-validated predictions of the complete training set, and the processing system 200 selects the hyperparameter set with the best performance to retrain using the complete training set. Performance can be determined based on a logarithmic loss metric. The processing system 200 can calculate the logarithmic loss in the following manner: obtain the negative logarithm of the prediction of the correct critical hits for each sample, and then sum it in the sample. For example, a perfect prediction of 1.0 for the correct label will result in a logarithmic loss of 0 (the lower the more accurate). To generate a prediction for a new sample, the processing system 200 can calculate the feature value using the above method, but is limited to the features (region / positive classification combination) selected under the selected topK value. The processing system 200 can use the generated features and use the trained logistic regression model to generate predictions.

[0228] In optional step 160, the processing system 200 applies a classifier to predict a tissue of origin of the test sample, wherein the tissue of origin is associated with one of the disease states. In some embodiments, the classifier may return a prediction or likelihood of more than one disease state or tissue of origin. For example, the classifier may return the following prediction: the likelihood that the test sample has a breast cancer tissue of origin is 65%, the likelihood that it has a lung cancer tissue of origin is 25%, and the likelihood that it has a healthy tissue of origin is 10%. The processing system 200 may further process the prediction values ​​to produce a single disease state determination.

[0229] III.E. Uncertain Positioning

[0230] In various embodiments, the tumor score can be a covariate in the prediction of a sample by a trained classifier or model. As the tumor score decreases, the score assignment (e.g., based on the aforementioned log-likelihood ratio R) may become less certain until the classification detection limit is reached (i.e., the probability of detection of the cancer / cancer type is 50%). Samples with high cfDNA tumor scores tend to be clearly classified, while samples with low cfDNA tumor scores tend to be more ambiguous. In the case of ambiguous signals, the assignment becomes less reliable and may be correct or incorrect by chance. In the use case of a single positioning, the processing system 200 can identify ambiguous signals and separate those predictions into an "uncertain positioning category".

[0231] For example, in some embodiments, for individuals with a cancer score greater than a specific target threshold, the processing system 200 may determine a post hoc uncertain assignment from a set of source tissue location vectors. The processing system 200 may determine the uncertain assignment under cross-validation. For each sample, the processing system 200 may calculate a metric that captures the location uncertainty of the sample. As an example method, the processing system 200 calculates the metric using the information entropy (bits) of the source tissue location, where a bit value of zero occurs when a prediction is certain. In the most ambiguous case (equal probability for all n categories), the processing system 200 uses log 2 (n) to calculate the place value. As another example method, the processing system 200 uses the difference (delta value) between the top-level score and the second-level score to determine the metric. A delta value of 1 occurs when a prediction is certain. A delta value of 0 occurs in the most ambiguous case. By including uncertain outcomes, the processing system 200 can filter out weak calls that are only correct by chance and increase the accuracy of determining the position call (e.g., the source organization assigns the correct score).

[0232] As an alternative to post-hoc uncertain assignment, the processing system 200 may use expectation maximization during training to determine assignments to uncertain classes. The processing system 200 may also add a second layer to the classifier output to classify cases into uncertain classes.

[0233] Given the metrics and a record of whether each sample was correctly located, the processing system 200 can calculate a precision-recall curve for an uncertain call threshold, such as Fig.18 As shown. It can be (for example) based on the target precision level (such as Fig.18 The cutoff point is selected by the processing system 200 for individual localization markers (e.g., for a certain cancer type) or for all cancer types as a whole. The trade-off is optimized and may depend on the cost of an erroneous localization call and the number of calls that assign an uncertain result (e.g., precision and recall).

[0234] III.F. Preventing Class Imbalance

[0235] In various embodiments, the factor score vector s for a single sample i Contains the posterior probability of signal localization for each predicted category (e.g., disease state). Each feature is scaled by a prior probability proportional to the proportion of training instances for each class:

[0236]

[0237] If the classification is unbalanced, samples with weak signals may be shifted to inappropriate classifications. For example, a training set may contain 99% of samples that detected liver cancer but very few samples of different cancer types. Therefore, a classifier trained using this set may be biased towards liver cancer prediction (or always guess the classification). In addition, if the classification ratio in classifier training is incompatible with the overall frequency of applying the classifier (e.g., where the classification ratio is more balanced), incorrect predictions may be produced.

[0238] To evaluate the ability of the separator to locate cfDNA samples from methylation and / or genomic and / or clinical features, the processing system 200 can target cross-class ratio equivalence. The processing system 200 can calibrate the scores for the incidence of disease states in the screening population, optionally taking into account the detectability of the disease by tumor scores. By revising the priors applied to the classifiers trained using the general training set, the processing system 200 can customize the classifier to improve the prediction for a specific population related to the prior (e.g., indicating the distribution of disease states in the specific population). Different geographic regions or countries may have different priors based on the prevalence of a specific disease state or cancer type in the corresponding individual subpopulation.

[0239] As an example approach, the processing system 200 performs a post hoc recalibration of the model scores. Specifically, the processing system 200 corrects the scores of a classification by dividing the assigned probabilities by the frequency of training set instances of that classification. The correction can optionally be stabilized by adding pseudo counts. The processing system 200 can then normalize each score vector s i , and the sum is one.

[0240] As another approach, the processing system 200 may resample the low frequency training examples to a desired ratio. As yet another approach, the processing system 200 may reweight the loss function in classifier training.

[0241] IV. Multilayer Perceptron Model

[0242] In some embodiments, a multilayer perceptron model ("MLP") can be used instead of logistic regression for classification. For an international regression-based classifier, the MLP classifier can be a single multi-classification classifier that is used to detect cancer and determine both the tissue of origin (TOO) or cancer type of the cancer. For example, a multi-class classifier can be trained to distinguish between two or more, three or more, five or more, ten or more, fifteen or more, or twenty or more different cancer types. In one embodiment, the multi-class cancer MLP model can also include a classification label for non-cancer, and cancer detection can be determined (e.g., in the form of 1-non-cancer). In another embodiment, the multi-layer perceptron model can be a two-stage classifier, the first stage is used for binary classification (e.g., cancer or non-cancer), and the second stage multi-layer perceptron model is used for multi-class classification (e.g., TOO), for example, with one or more hidden layers.

[0243] In one embodiment, the multilayer perceptron includes the following two stage classifiers: a first stage multilayer perceptron (MLP) binary classifier, which has no hidden layers; and a second stage multilayer perceptron (MLP) multiclass classifier, which has a single hidden layer. In one embodiment, samples determined to have cancer using the first stage classifier are then analyzed by the second stage classifier.

[0244] In the first training phase, a binary (two-class) multilayer perceptron model without hidden layers for detecting the presence of cancer can be trained to distinguish cancer samples (regardless of TOO) from non-cancer. For each sample, the binary classifier outputs a prediction score indicating the likelihood of the presence or absence of cancer.

[0245] In the second training stage, a parallel multi-class multilayer perceptron model for determining the cancer type or the tissue of origin of the cancer can be trained. In one embodiment, only cancer samples with scores above a cutoff threshold (e.g., 95% of non-cancer samples in the first stage classifier) ​​can be included in the training of the multi-class MLP classifier. For each cancer sample used in training and testing, the multi-class MLP classifier outputs a prediction value for the classified cancer type, wherein each prediction value is the likelihood that the administered sample has a certain cancer type. For example, a cancer classifier can return a cancer prediction for a test sample, the cancer prediction including a breast cancer prediction score, a lung cancer prediction score, and / or a cancer-free prediction score.

[0246] Fig.16 1 is a flow chart of a method 1600 for determining a probability that a sample has a disease state according to various embodiments. In some embodiments, the processing system 200 performs the method 1600 to process sequence reads of fragments from a nucleic acid sample. The method 1600 includes, but is not limited to, the following steps described for the components of the processing system 200.

[0247] In step 1610, the processing system 200 generates sequence reads from one or more biological samples. In some embodiments, the processing system 200 filters the sequence reads according to the p-value scores of the sequence reads. The p-value scores of the sequence reads indicate the probability of observing methylation in the nucleic acid fragment corresponding to the sequence read in the one or more biological samples.

[0248] In step 1620, the processing system 200 uses the sequence reads to determine, for each position in the set of chromosomal positions, a count of nucleic acid fragments of one or more biological samples that have at least a threshold similarity to disease state associated fragments (e.g., cancer-like fragments) within the position. The disease state can be associated with at least one type of cancer, a cancer stage, or another type of disease or condition.

[0249] Each position can represent the number of consecutive base pairs of a chromosome. The number of base pairs can vary between different positions. Processing system 200 can generate sequence reads of multiple regions of the genome. There can be up to tens of thousands or more regions. Each region can contain hundreds, thousands or more base pairs. Method 1600 can perform whole genome bisulfite sequencing (WGBS) or targeted panel determination.

[0250] In step 1630, the processing system 200 uses the position count as a feature to train the machine learning model. In some embodiments, the processing system 200 binarizes the feature to indicate the presence or absence of a disease state in each position (e.g., a Boolean value). The count of at least one nucleic acid fragment in a certain position indicates that a disease state exists in the position. The count of zero nucleic acid fragments in a certain position indicates that a disease state does not exist in the position. In some embodiments, the machine learning model can be a logistic regression model. In some embodiments, the machine learning model can be a multilayer perceptron model (neural network). As will be readily understood by those skilled in the art, other machine learning models can be used, including, for example, generalized linear models (GLM), multilayer perceptrons, support vector machines, random forests, or neural network classifiers.

[0251] In step 1640, the trained machine learning model determines the probability that the test sample has a disease state. The test sample can be obtained from the patient and can include blood and / or tissue. In optional step 1650, treatment is provided to the patient based on the probability. For example, treatment (e.g., a drug or interventional procedure) may be provided to the patient in response to determining that the probability is greater than a threshold. In another embodiment, in optional step 1650, a test report may be generated to provide the patient with his test results (including the probability that the test sample has a disease).

[0252] The model was trained using samples from the CCGA study (described further below). Figure 17-20 The experimental results shown in .

[0253] Fig.17 The performance gain of sensitivity of a multilayer perceptron model according to one embodiment is shown. The multilayer perceptron model (MLP) shows performance gain of disease detection sensitivity in cancer stages I, II, III, and IV compared to the logistic regression model.

[0254] Fig.18 The experimental results of determining the tissue of origin using a multilayer perceptron model according to one embodiment are shown. Compared with the logistic regression model (LR: 1803 and 1804), the multilayer perceptron model (MLP: 1801 and 1802) has improved accuracy in determining the tissue of origin. The accuracy can be improved when processing sequence reads related to all cancer types in the training set, and when processing sequence reads of the training set (including more than 10 instance sequence reads of each cancer type in the training set).

[0255] Fig.19 Experimental results of a multilayer perceptron model according to one embodiment for determining tissue of origin according to cancer stage are shown. Compared to logistic regression (LR), the multilayer perceptron model (MLP) shows performance gain in tissue of origin (TOO) detection accuracy in cancer stages I, II, III, and IV. Among the cancer stages, the performance gain of the MLP model for stage I is the largest.

[0256] Fig. 20 The experimental results of a multilayer perceptron model according to one embodiment in various types of cancer are shown. Fig. 20 For most cancer types shown in Figure 2, the multilayer perceptron model (MLP) achieved greater accuracy in tissue of origin (TOO) detection than the logistic regression model.

[0257] In some embodiments, the analysis system uses a two-stage model to determine the tissue of origin (TOO) of a cancer or another type of disease state. The analysis system generates sequence reads from nucleic acid fragments of a biological sample. The analysis system processes the sequence reads, for example, using any of the processes described in the section II.A. assay scheme to determine a first training data set. The analysis system can use methylation information to determine the first training data set. For example, the analysis system determines low methylation sequence reads by determining that a threshold number or percentage of CpG sites corresponding to sequence reads are not methylated. In addition, the analysis system determines high methylation sequence reads by determining a threshold number or percentage of CpG sites corresponding to sequence reads that are methylated. The analysis system can also determine that abnormal methylation occurs in sequence reads. In some embodiments, the analysis system filters sequence reads by removing sequence reads whose p-values ​​are less than a threshold p-value.

[0258] The analysis system trains a binary classifier using a first training data set. The binary classifier is trained to predict a binary output, ie, the presence or absence of at least one disease state in the first test biological sample, for input sequence reads from the first test biological sample.

[0259] Using the predictions of the binary classifier, the analysis system can determine that one or more disease states are present in a subset of the biological samples. The binary classifier can be used to train a tissue of origin classifier. In particular, the analysis system uses sequence reads of nucleic acid fragments corresponding to the subset of biological samples to determine a second training data set. The analysis system uses the second training data set to train the tissue of origin classifier. The tissue of origin classifier is trained to predict a tissue of origin associated with a disease state present in the second test biological sample for an input sequence read from the second test biological sample. The first test biological sample and the second test biological sample can be the same sample or different samples.

[0260] In some embodiments, the analysis system uses a source tissue classifier to determine a score indicating the probability that a source tissue associated with a disease state is present in a second test biological sample. The analysis system may calibrate the score to, for example, adjust the output of an overconfidence model. For example, the analysis system uses a feature space output by the source tissue classifier in conjunction with the score to perform a k-nearest neighbor (KNN) operation. In one embodiment, the feature space includes two top-ranked predictive markers (e.g., lung cancer and prostate cancer) from the source tissue classifier and an indication of whether the correct classification is a disease state different from the two most predicted ones. The analysis system may also correct the score by normalizing the probability using the output of a binary classifier, which output indicates different probabilities of the presence of at least one disease state present in the second test biological sample.

[0261] In some embodiments, the source tissue classifier is a multilayer perceptron comprising at least one hidden layer. The source tissue classifier may also include a 100-unit hidden layer or a 200-unit hidden layer, as well as other sizes of hidden layers. The multilayer perceptron may be fully connected and use a rectified linear unit activation function. In some embodiments, the binary classifier is a multilayer perceptron that does not include a hidden layer. In a different embodiment, the binary classifier is a multilayer perceptron that includes at least one hidden layer. In other embodiments, these classifiers may be logistic regression models, multinomial logistic regression models, or other types of machine learning models.

[0262] In addition, the analysis system can use one or more machine learning techniques known to those skilled in the art to train the source tissue classifier and the binary classifier, including, for example, non-early stopping (but selecting a given number of training epochs), stochastic gradient descent, weight decay, dropout regularization, Adam optimization, Ho initialization and learning rate scheduling, rectified linear unit activation function, leaky rectified linear unit activation function, sigmoid activation function, and boosting, among others. Fig.31 As shown in , the tissue-of-origin accuracy of the tissue-of-origin classifier improves with each training iteration. Each iteration can include a different combination of machine learning techniques. In addition, the tissue-of-origin accuracy increases in different cancer stages: I, II, and III.

[0263] In some embodiments, the analysis system cross-validates one or both of the source tissue classifier and the binary classifier. The analysis system can retrain the classifier using hyperparameters selected based on the output of the cross-validation. The analysis system can select hyperparameters by aggregating the results from all folds in the cross-validation. In one embodiment, the analysis system selects hyperparameters to train the source tissue classifier by optimizing the source tissue accuracy (rather than the log-likelihood) because the classifier can be more confident in samples with stronger signals.

[0264] In some embodiments, the analysis system determines the probability that a source tissue associated with a disease state is present in a second test biological sample by a source tissue classifier. The analysis system predicts that the source tissue associated with the disease state is present in the second test biological sample in response to determining that the probability is greater than a source tissue threshold. The analysis system can determine different source tissue thresholds associated with different source tissues. In addition, the analysis system can determine the source tissue threshold associated with a given disease state by iterating through a plurality of different probabilities of candidate source tissue thresholds. For each iteration, the analysis system determines the sensitivity at a given specificity of the source tissue classifier. The analysis system can optimize the trade-off between the sensitivity rate and the specificity rate of the source tissue classifier for a given disease state. The analysis system can determine the sensitivity rate using a score output by a binary classifier or a source tissue classifier. In addition, the analysis system can use the score from the source tissue classifier to stratify the samples.

[0265] In some embodiments, an analysis system trains a binary classifier and a tissue of origin classifier using binarized features, each feature having a value of 0 or 1. 1 is used in the binarization to replace values ​​greater than 1.

[0266] V. Tuning of Binary Classification Thresholds

[0267] The analysis system can tune the trained cancer classifier to prune the samples used to train the cancer classifier. In particular, the analysis system may attempt to remove non-cancer samples with high tissue signals, thereby reducing the sensitivity of the cancer classifier in cancer prediction. High tissue signal refers to cfDNA from the tissue of origin (TOO) with a significant fraction of samples compared to the healthy distribution, for example, determined by a tissue of origin classifier, a multi-class cancer classifier or other means. Non-cancer samples with high tissue signals are outliers in the non-cancer distribution and may be pre-cancer, early cancer or undiagnosed cancer. The analysis system may identify non-cancer samples with high tissue signals in at least one cancer type. In some embodiments, certain cancer types are further divided into cancer subtypes. For example, hematological cancer types can be further divided into various combinations such as: circulating lymphoid subtypes, non-Hodgkin's lymphoma (NHL) indolent subtypes, NHL aggressive subtypes, Hodgkin's lymphoma (HL) subtypes, myeloid subtypes, and plasma cell subtypes.

[0268] See also Fig.21 , Fig.21A graph showing the likelihood of cancer types for non-cancer samples higher than 95% specificity is shown. Calculate the cancer score for each non-cancer sample from multiple non-cancer samples (i.e., samples from healthy individuals who are not currently diagnosed with cancer). The cancer score can be determined by a binary classifier as the likelihood that the sample has cancer under consideration of the methylation sequencing data of the sample. In other embodiments, the cancer score can be calculated according to other methods, the method at least inputs sequencing data (e.g., methylation, single nucleotide polymorphism (SNP), DNA, RNA, etc.) and outputs the likelihood that the sample has cancer based on the input sequencing data. An example of a classifier is a mixed model classifier. The distribution of non-cancer samples can be generated according to the cancer score of the non-cancer sample. A binary threshold cutoff value can be set to ensure a certain degree of binary classification specificity (e.g., true negative rate). Typically, a high specificity cutoff value is used to classify cancer, for example, between 90% and 99.9% or 99.5% specificity or higher. However, many non-cancer samples used to train cancer classifiers and just below the specificity cutoff value may have high tissue signals, thereby deviating from the binary threshold cutoff value.

[0269] To confirm this, non-cancer samples with greater than 95% specificity were selected and then input into a multi-class cancer classifier to determine the probability of each cancer type or tissue of origin (TOO). The cancer types or TOO markers used in this embodiment of the multi-class cancer classifier include circulating lymphoid, myeloid, NHL indolent, colorectal, NHL aggressive, lung, uterine, breast, prostate, pancreas and gallbladder, upper gastrointestinal tract, bladder and urothelial, plasma cell, head and neck, kidney, ovary, sarcoma, liver and bile duct, cervix, other tissues, HL, anorectal, melanoma, thyroid. Fig.21The figure in shows many non-cancer samples with high tissue signals from at least one tissue type. Each point in the tissue type column corresponds to the likelihood of the source tissue of a non-cancer sample above the 95% specificity threshold. It is worth noting that many tissue types have multiple lung cancer sample outliers with significant tissue contributions, which is not common for non-cancer samples. This can occur when such non-cancer samples have cfDNA signals driven by cancer-like methylation, clonal scores and / or growth / turnover rates. It can be inferred that a large number of non-cancer samples used to train cancer classifiers may be pre-cancer, early cancer or undiagnosed cancer. However, these non-cancer samples with significant tissue contributions shift the binary classification cutoff threshold upward, thereby reducing the sensitivity of cancer classification, especially for samples whose tissue signals are significantly lower than the previously set binary classification cutoff threshold. In practice, such signals (e.g., corresponding to circulating lymphoid, myeloid and NHL inertia) can be the main cause of false positive determinations. It should be noted that circulating lymphoid, myeloid, NHL indolent, colorectal, NHL aggressive, lung, uterine, breast, prostate, pancreas and gallbladder, upper gastrointestinal, plasma cell, head and neck, cervix, HL markers have at least one non-cancer sample with a tissue of origin probability higher than 0.1. In particular, circulating lymphoid, myeloid, NHL indolent, and NHL aggressive markers (all hematological subtypes) have two or more non-cancer samples with a tissue of origin probability higher than 0.5.

[0270] See also Fig. 22 , Fig. 22 A diagram showing hematological subtypes based on methylation sequencing data is shown. Fig. 22 The figure shows the ability to model hematological subtypes. This can prove to be beneficial for providing higher precision for multi-class cancer classification (for example, using hematological subtype markers for classification in addition), or as a way to tune cancer classification by pruning non-cancer samples with high hematological subtype signals before training a cancer classifier. As described above, methylation signals can cover multiple CpG sites, thereby generating a high-dimensional vector space. Using hematological subtype samples and non-cancer samples, the analysis system can perform principal component analysis. Principal component analysis identifies orthogonal principal components (or embeddings) of the vector space according to the order of change of methylation signals in the sample. The first principal component (shown as V1 on the horizontal axis on the figure) has the highest variance, and the second principal component (shown as V2 on the vertical axis on the figure) has the second highest variance. Each hematological subtype and non-cancer sample cluster is annotated on the figure 900. The hematological subtypes shown include circulating lymphoid, solid lymphoid, plasma cell and myeloid. The solid lymphoid subtype can be further divided into HLp, NHL indolent and NHL aggressive. The figure shows the possibility of classification according to hematological subtypes - for adding hematological subtypes in a multi-class cancer classification or for modeling each hematological subtype to tune a cancer classifier.

[0271] Removal of VA high signal non-cancer samples

[0272] Fig.23A A flow chart of a process 1000 for determining a binary threshold cutoff for binary cancer classification according to a description of one or more embodiments is shown. The binary classification for predicting cancer and non-cancer evaluates the sample cancer score for the determined binary threshold cutoff, wherein samples with a cancer score below the binary threshold cutoff are determined as non-cancer, and samples with a cancer score of the binary threshold cutoff or higher are determined as cancer. A trained multi-class cancer classifier evaluates the methylation signals (and / or other sequencing data) of the sample to determine the probabilities of a plurality of TPP markers classified by the multi-class cancer classifier. The TPO markers used in the multi-class cancer classifier may be cancer tissue types or cancer tissue subtypes (e.g., the hematological subtypes described above). Process 1000 may be performed or implemented by an analysis system.

[0273] The analysis system receives 1010 sequencing data of more than one biological sample containing cfDNA fragments, wherein the biological sample includes a cancer sample or a non-cancer sample. The sequencing data may be methylation sequencing data, SNP sequencing data, another DNA sequencing data, RNA sequencing data, etc.

[0274] For each non-cancer sample, the analysis system uses a multi-class cancer classifier to classify the non-cancer sample based on the features derived from the sequencing 1020, wherein the multi-class cancer classifier predicts the probability of each of the multiple TOO markers. The analysis system can generate a feature vector for the non-cancer sample, thereby assigning an abnormal score to each CpG site under consideration based on the abnormally methylated cfDNA fragments that overlap at least with the CpG site.

[0275] For each non-cancer sample, the analysis system determines 1030 whether the predicted probability likelihood exceeds a TOO threshold for one or more TOO markers. The TOO threshold is determined as follows: Fig. 23B Further described in .

[0276] The analysis system determines 1040 a binary threshold cutoff for predicting the presence of cancer, the binary threshold cutoff being determined based on the distribution of non-cancer samples (excluding one or more non-cancer samples identified as having a probability likelihood exceeding at least one TOO threshold). Non-cancer samples whose probability likelihood for at least one TOO marker exceeds the TOO threshold corresponding to the TOO marker are excluded. The analysis system then calculates the distribution of non-cancer samples based on the cancer score of each non-cancer sample, and then determines the binary threshold cutoff at the desired specificity (e.g., 99.4-99.9% specificity) from the distribution. It is worth noting that each cancer score can be determined based on sequencing data, for example, a cancer score can be output by a binary cancer classifier to predict the likelihood of cancer based on methylation sequencing data, as described herein. In other embodiments, a cancer score can be calculated according to other methods, the method at least inputs sequencing data (e.g., methylation, single nucleotide polymorphism (SNP), DNA, RNA, etc.), and outputs the likelihood that the sample has cancer based on the input sequencing data.

[0277] Fig. 23B A flow chart is shown describing a process 1005 of thresholding a TOO marker to determine a binary threshold cutoff for a binary cancer classification according to one or more embodiments. The process 1005 can be an embodiment of the process 1000. The binary classification for predicting cancer and non-cancer evaluates the sample cancer score for the determined binary threshold cutoff, wherein samples with a cancer score below the binary threshold cutoff are determined as non-cancer, and samples with a cancer score of the binary threshold cutoff or higher are determined as cancer. A trained multi-class cancer classifier evaluates the methylation signals (and / or other sequencing data) of the sample to determine the probabilities of a plurality of TOO markers classified by the multi-class cancer classifier. The TOO marker can be a cancer tissue type or more particularly a cancer tissue subtype (e.g., the hematological subtype described above). The process 1005 can be performed or implemented by an analysis system.

[0278] The analysis system obtains 1015 a training set including a plurality of samples having cancer or non-cancer markers and a retention set including a plurality of samples having cancer or non-cancer markers (i.e., cancer samples or non-cancer samples, respectively). Each sample in the training set includes methylation sequencing data (e.g., according to Figure 3In other embodiments, each training sample has other sequencing data used in tandem with or in place of the methylation sequencing data. In addition, each sample from the training set and the retention set has a cancer score. As described above, the cancer score can be determined by a binary classifier as the likelihood of the sample having cancer, taking into account the methylation sequencing data of the sample. In other embodiments, the cancer score is calculated according to other methods, which at least input test data (e.g., methylation, single nucleotide polymorphism (SNP), DNA, RNA, etc.), and output the likelihood of the sample having cancer based on the input test data, such as the hybrid model described herein.

[0279] For each non-cancer training sample, the analysis system determines 1025 feature vectors based on the methylation sequencing data. The analysis system can determine the feature vectors for each non-cancer training sample, for example, by determining the abnormality score of each CpG site in the set of CpG sites under consideration. In some embodiments, the analysis system defines the abnormality score of the feature vector using a binary score based on whether there are abnormal fragments covering the CpG site in the set of abnormal fragments. Once all abnormality scores for the sample are determined, the analysis system determines the feature vector as a vector with abnormality scores associated with each CpG site under consideration. The analysis series can additionally normalize the abnormality scores of the feature vectors based on sample coverage.

[0280] The analysis system inputs 1035 the feature vector of each non-cancer training sample into a multi-class cancer classifier to generate a TOO prediction. The multi-class cancer classifier is trained for multiple TOO markers (including cancer types, cancer subtypes, non-cancer or any combination thereof). The multi-class cancer classifier can be trained as described herein. The trained multi-class cancer classifier determines multiple probabilities of the TOO marker as a cancer prediction, wherein the probability of the TOO marker indicates the likelihood of having a cancer corresponding to the TOO marker.

[0281] In some examples, the analysis system scans 1045 or iterates multiple TOO marker probabilities as candidate TOO thresholds, thereby calculating the specificity rate and sensitivity rate within a certain range of TOO marker probabilities. The analysis system can gradually scan the probability range, for example, 0.01, 0.02, 0.03, 0.04, 0.05, etc. Because the analysis system scans a certain range of probabilities, the analysis system filters the non-cancer training samples whose TOO marker probability is the candidate TOO threshold or higher according to the output of the multi-class cancer classifier. As a numerical example, the analysis system considers a candidate TOO threshold of 0.35. Non-cancer training samples with a TOO marker probability of 0.25 or higher are filtered out from the training set. The analysis system determines the adjusted binary threshold cutoff based on the filtered training set. The analysis system uses the adjusted binary threshold cutoff for the retention set to calculate the predicted specificity rate. Specificity refers to the accuracy of identifying non-cancer samples as non-cancer markers. The analysis system also uses the adjusted binary threshold cutoff for the retention set to calculate the predicted sensitivity rate. Sensitivity refers to the accuracy of identifying a cancer sample as a cancer marker. In practice, specificity and / or sensitivity rates may be defined based on true positive rate, false positive rate, true negative rate, false negative rate, another statistical calculation, etc.

[0282] The analysis system determines 1055 a TOO threshold for a TOO marker. The analysis system selects a TOO threshold from the candidate TOO thresholds by optimizing the calculated specificity rate and / or sensitivity rate within a range of candidate TOO thresholds. In some instances, the TOO threshold is determined or otherwise applied to certain TOO tissue type categories or subtype categories (e.g., hematology categories). By way of example only, an algorithm for calculating and applying a TOO-specific probability threshold can be used to remove non-cancer samples with blood disease that exceed the signal. The algorithm may include first searching a probability value grid for each predetermined TOO marker, and using a binary detection threshold for each value to evaluate the clinical specificity and clinical sensitivity of the retention set, the binary detection threshold being calculated after removing non-cancer samples with a probability equal to or greater than the specified TOO marker. By iterating through the probability grid, the algorithm will identify a certain combination of TOO thresholds for predetermined TOO markers that optimizes the trade-off between clinical specificity and clinical sensitivity of the retention set. The final optimized TOO probability threshold is used to filter out non-cancer samples that exceed any value assigned to the TOO marker. The cleaned non-cancer sample set will be used to calculate the cancer-non-cancer detection threshold. Additionally, in some examples, the TOO specificity threshold can be manually set at any cutoff point, for example, a desired specificity level (eg, 99.4-99.9% specificity).

[0283] The analysis system tunes 1065 the binary cancer classification by pruning non-cancer training samples that exceed the TOO threshold before determining the binary threshold cutoff. The analysis system filters out non-cancer training samples from the training set according to the determined TOO threshold for the TOO marker. The analysis system sets the binary threshold cutoff according to the filtered training set. For example, the analysis system determines a new binary threshold cutoff based on the filtered score distribution. In other embodiments, the analysis system can determine the TOO threshold for any TOO marker according to steps 1010, 1020, 1030, and 1040 to tune the binary cancer classification.

[0284] VB stratifies sample distribution according to TOO signal

[0285] In one or more embodiments, the analysis system tunes the cancer classifier by stratifying the sample distribution according to the TOO signal to determine a binary threshold cutoff value for each stratum. The analysis system can stratify the sample distribution according to the signal of one or more TOO markers, which is determined according to the TOO prediction output by the multi-class cancer classifier.

[0286] As used herein, "high tissue signal" refers to a sample whose tissue signal (e.g., generally for any tissue type or for a specific cancer type - also referred to as a TOO marker) exceeds a predetermined threshold. The tissue signal can be determined by a multi-class cancer classifier or other means compared to a healthy distribution. Non-cancer samples with high tissue signals are outliers in the non-cancer distribution. Some of these non-cancer samples can be pre-cancer, early cancer, or undiagnosed cancer. The analysis system can identify non-cancer samples with high tissue signals in at least one TOO marker. In one way to determine a high tissue signal, the predicted value of the TOO marker output by the multi-class cancer classifier is compared with a tissue signal threshold. Samples with predicted values ​​higher than the tissue signal threshold can be considered to have a high tissue signal for the TOO marker; and samples with predicted values ​​lower than the tissue signal threshold can be considered to not have a high tissue signal (or a low tissue signal) for the TOO marker. In another way, one or more top-ranked predictions in the TOO predictions are considered. For example, the TOO prediction of the sample has a first prediction for the colorectal TOO marker, a second prediction for the breast TOO marker, and a third prediction for the head and neck TOO marker. If the top ranked predictions are considered, the sample can be considered to have a high tissue signal for the TOO marker in the first prediction, which in the example is the colorectal TOO marker. If the top two ranked predictions are considered, there is a high tissue signal in both the colorectal TOO marker and the breast TOO marker. Other ways to determine tissue signal may include other models trained to determine tissue signal for one or more TOO markers. Such models may include classifiers trained to determine tissue signal for a subset of TOO markers. For example, a hematology-specific classifier can be trained and used to determine tissue signal for one or more hematology subtypes. Other models include deconvolution models that can deconvolute tissue signals from methylation sequencing data (and / or other types of sequencing data).

[0287] See now Fig.32 , Fig.32 A process of stratifying a hematology signal into two layers is shown in accordance with one or more embodiments. Although the following description describes stratification using hematology signals, the principles can be readily applied to other TOO signals.

[0288] The analysis system stratifies 1300A the holdout set of cancer and non-cancer samples into a low signal layer 1310 and a high signal layer 1320 according to the hematological signal. Each sample of the holdout set has a cancer score determined by a binary cancer classifier and a TOO prediction determined by a multi-class cancer classifier. In one embodiment, the hematological signal of the sample is determined according to the TOO prediction output by the multi-class cancer classifier. In one embodiment, when considering one or more top-ranked predictions (e.g., first, second, etc.), if at least one of the top-ranked predictions considered is a hematological subtype (e.g., a lymphoid neoplasm subtype and a myeloid neoplasm subtype), a high hematological signal can be determined. Other hematological subtypes may be included. Therefore, if the sample has a TOO prediction in which at least one of the top-ranked predictions can be considered a lymphoid neoplasm subtype or a myeloid neoplasm subtype, the sample can be determined to have a high hematological signal. Otherwise, the sample is determined not to have a high hematological signal.

[0289] The analysis system determines the binary threshold cutoff value of each layer to predict the presence or absence of cancer in the sample. The analysis system uses the samples in the low signal layer 1310 to determine 1305 the binary threshold cutoff value for predicting the absence or presence of cancer in the samples in the low signal layer 1310. The binary threshold cutoff value 1305 is determined according to the false positive budget set of the low signal layer 1310. Using the cancer score of the sample in the low signal layer 1310, the analysis system scans a series of candidate binary threshold cutoff values, and evaluates the true positive rate (also called sensitivity) and false positive rate at each candidate binary threshold cutoff value. The candidate binary threshold cutoff value with the closest false positive within the false positive budget is determined as the candidate binary threshold cutoff value. The analysis system performs similar operations to determine 1315 the binary threshold cutoff value of the high signal layer 1320. The false positive budget of the low signal layer 1310 and the false positive budget of the high signal layer 1320 can be set according to the statistical true positive rate ratio of the layer. This ratio is intended to suppress the false positive rate in the high signal layer 1320.

[0290] For the test sample, the analysis system places the test sample in a low signal layer 1310 or a high signal layer 1320 according to the hematology signal. If the test sample is placed in the low signal layer 1310, the analysis system applies 1315 the binary threshold cutoff value of the low signal layer 1310 to the cancer score of the test sample. If the cancer score is greater than or equal to the binary threshold cutoff value of the low signal layer 1310, the analysis system returns a prediction of cancer present in the test sample and additionally returns a prediction of no cancer. If the test sample is placed in the high signal layer 1320, the binary threshold cutoff value of the high signal layer 1320 is applied 1325 to the cancer score of the test sample. If the cancer score is greater than or equal to the binary threshold cutoff value of the high signal layer 1320, the analysis system returns a prediction of cancer present in the test sample and additionally returns a prediction of no cancer.

[0291] VI. Circulating Cell-free Genome Mapping

[0292] In various embodiments, each predictive cancer model is trained using a training dataset derived from a training subset of patients from the Circulating Cell-Free Genome Atlas (CCGA) study (see Clinical Trial.gov Identifier: NCT02889978 (https: / / www.clinicaltrials.gov / ct2 / show / NCT02889978)), and is subsequently tested using a test or validation dataset derived from a test or validation subset of patients from the CCGA study.

[0293] A plurality of known cancer types from the Circulating Cell-Free Genome Atlas (CCGA) study are used to train the predictive cancer models described herein. The CCGA sample set includes the following cancer types: breast, lung, prostate, colorectal, kidney, uterus, pancreas, esophagus, lymphoma, head and neck, ovary, hepatobiliary, melanoma, cervix, multiple myeloma, leukemia, thyroid, bladder, stomach, and anorectal. Thus, the model can be a multi-cancer model (or multi-cancer classifier) ​​for detecting one or more, two or more, three or more, four or more, five or more, ten or more, or 20 or more different types of cancer.

[0294] A predictive cancer model can be trained using a refined training dataset derived from a first subset of patients from the CCGA study, and subsequently tested using a refined test dataset derived from a second subset of patients from the CCGA study.

[0295] VII. Cancer Assay Panel

[0296] In various embodiments, the predictive cancer models described herein use samples enriched using a cancer assay panel comprising a plurality of probes or a plurality of probe pairs. Various targeted cancer assay panels are well known in the art, for example, as described in WO 2019 / 195268 filed on April 2, 2019, PCT / US2019 / 053509 filed on September 27, 2019, and PCT / US2020 / 015082 filed on January 24, 2020 (which are incorporated herein by reference). For example, in some embodiments, a cancer assay panel can be designed to include a plurality of probes (or probe pairs) that can capture fragments that together can provide information relevant to cancer diagnosis. In some embodiments, the panel comprises at least 50, 100, 500, 1,000, 2,000, 2,500, 5,000, 6,000, 7,500, 10,000, 15,000, 20,000, 25,000, or 50,000 pairs of probes. In other embodiments, the panel comprises at least 500, 1,000, 2,000, 5,000, 10,000, 12,000, 15,000, 20,000, 30,000, 40,000, 50,000, or 100,000 probes. The plurality of probes together may comprise at least 0.1 million, 0.2 million, 0.4 million, 0.6 million, 0.8 million, 1 million, 2 million, 3 million, 4 million, 5 million, 6 million, 7 million, 8 million, 9 million, or 10 million nucleotides. The probes (or probe pairs) are specifically designed to target one or more regions of the genome that are differentially methylated in cancer and non-cancer samples. The target genomic regions can be selected to maximize classification accuracy according to a size budget (which is determined by the prediction budget and the desired sequencing depth).

[0297] Targeted sequencing can be performed on samples enriched using a cancer assay panel. Samples enriched using a cancer assay panel can generally be used to detect the presence or absence of cancer and / or provide a cancer classification (e.g., cancer type, cancer stage, such as I, II, III, or IV) or provide a source tissue believed to be the source of the cancer. According to the present, the panel can contain probes (or probe pairs) targeting genomic regions with different methylation in general cancerous (pan-cancer) samples and non-cancerous samples or only in cancerous samples of a specific cancer type (e.g., lung cancer specific targets). Specifically, the cancer assay panel is designed based on bisulfite sequencing data generated from cell-free DNA (cfDNA) or genomic DNA (gDNA) from cancer and / or non-cancer individuals.

[0298] In some embodiments, a cancer assay panel designed by the methods provided herein comprises at least 1,000 pairs of probes, each pair comprising two probes configured to overlap with each other a 30-nucleotide fragment comprising overlapping sequences. The 30-nucleotide fragment comprises at least 5 CpG sites, wherein at least 80% of the at least 5 CpG sites are CpG or UpG. The 30-nucleotide fragment is configured to bind to one or more genomic regions in a cancerous sample, wherein the one or more genomic regions have at least 5 methylation sites with an abnormal methylation pattern. Another cancer assay panel comprises at least 2,000 probes, each probe being designed as a hybridization probe complementary to one or more genomic regions. Each genomic region is selected based on the following criteria, which include: (i) at least 30 nucleotides; and (ii) at least 5 methylation sites, wherein at least 5 methylation sites have an abnormal methylation pattern and are hypomethylated or hypermethylated.

[0299] Each probe (or probe pair) is designed to target one or more target genomic regions. The target genomic region is selected based on several criteria, which are designed to increase the selective enrichment of relevant cfDNA fragments and reduce noise and non-specific binding. For example, the panel can include probes that can selectively bind and enrich cfDNA with different methylations in cancerous samples. In this case, sequencing of the enriched fragments can provide information related to cancer diagnosis. In addition, the probe can be designed to target genomic regions with abnormal methylation patterns and / or high or low methylation patterns, thereby further providing selectivity and specificity of detection. For example, when a genomic region has a low p-value methylation pattern according to a Markov model genomic region trained for a non-cancerous sample set, a genomic region that covers at least 5 CpGs and 90% of which is methylated or unmethylated can be selected. In other embodiments, a hybrid model as described herein can be used to select a genomic region.

[0300] Each probe (or probe pair) can target a genomic region comprising at least 25 bp, 30 bp, 35 bp, 40 bp, 45 bp, 50 bp, 60 bp, 70 bp, 80 bp, or 90 bp. Genomic regions containing less than 20, 15, 10, 8, or 6 methylation sites can be selected. Genomic regions in which at least 80%, 85%, 90%, 92%, 95%, or 98% of at least 5 methylation (e.g., CpG) sites in non-cancerous or cancerous samples are methylated or unmethylated can be selected.

[0301] The genomic regions can be further filtered to select only CpG sites that may be informative based on methylation patterns, for example, CpG sites that have different methylation between cancerous and non-cancerous samples (e.g., abnormally methylated or unmethylated in cancer vs. non-cancer). For this selection, a calculation can be performed for each CpG site. In some embodiments, a first count is determined, which is the number of cancer-containing samples that contain fragments overlapping the CpG (cancer count); and a second count is determined, which is the number of total samples that contain fragments overlapping the CpG (total). The genomic region can be selected based on the following criteria: it is positively correlated with the number of cancer-containing samples that contain fragments overlapping the CpG (cancer count) and negatively correlated with the number of total samples that contain fragments overlapping the CpG (total).

[0302] In one embodiment, the number of non-cancerous samples with overlapping fragments of CpG sites (n 非癌症 ) and the number of cancerous samples (n 癌症 ) is counted. Then, the probability that the sample is cancer is estimated as (for example) (n 癌症 +1)(n 癌症 +n 非癌症 +2). CpG sites are ranked according to this metric and greedily added to the panel until the panel size budget is exhausted.

[0303] The samples used for cancer enumeration may vary depending on whether the assay is intended to be a pan-cancer assay or a single cancer assay, or depending on the type of flexibility desired in selecting CpG sites that contribute to the panel. A similar procedure can be used to design a panel for diagnosing a specific cancer type (e.g., TOO). In this embodiment, for each cancer type and each CpG site, the information gain is calculated to determine whether to include a probe for the CpG site. The information gain is calculated in samples with the administered cancer type compared to all other samples. For example, two random variables "AF" and "CT". "AF" is a binary variable (yes or no) indicating whether there are abnormal fragments overlapping with a specific CpG site in a specific sample. "CT" is a binary random variable indicating whether the cancer is of a specific type (e.g., lung cancer or cancer other than lung cancer). The mutual information about "CT" can be calculated given "AF". That is, if it is known whether there are abnormal fragments overlapping with a specific CpG site, the number of bits of information about the cancer type (lung vs. non-lung in this example) is obtained. This can be used to rank CpGs based on specificity for a particular cancer type (e.g., TOO). This process is repeated for multiple cancer types. For example, if a particular region typically has different methylation only in lung cancer (and not in other cancer types or non-cancer), then the CpGs in that region tend to have high information gain for lung cancer. For each cancer type, the CpG sites are ranked according to this information gain metric and then greedily added to the panel until the size budget for that cancer type is exhausted.

[0304] Can further filter to select the target genome region that the off-target genome region is less than the threshold value.For example, only select the genome region that is less than 15,10 or 8 off-target genome regions.In other cases, filter to remove the sequence of the target genome region in the genome greater than 5,10,15,20,25 or 30 times.Can further filter to select the target genome region that is less than 15,10 or 8 times in the genome with the sequence of 90%,95%,98% or 99% homology to the target genome region, or remove the target genome region that is greater than 5,10,15,20,25 or 30 times in the genome with the sequence of 90%,95%,98% or 99% homology to the genome region.This is to exclude the repeated probes that can pull down the off-target fragment, which is undesirable and can affect the determination efficiency.

[0305] In some embodiments, a fragment-probe repeat of at least 45 bp is required to achieve a non-negligible amount of pull-down (but this amount may vary depending on the assay details). In addition, it has been found that a mismatch rate greater than 10% between the probe and the fragment sequence in the overlapping region is sufficient to greatly disrupt binding and thereby reduce efficiency. Therefore, sequences that can be aligned with the probe at a match rate of at least 90% along at least 45 bp are candidates for off-target pull-down. Therefore, in one embodiment, the number of such regions is scored. The best probe has a score of 1, which means that it matches only in one place (the expected target region). Probes with low scores (i.e., less than 5 or 10) are acceptable, but any probes above this score are discarded. Other cutoff values ​​can be used for specific samples.

[0306] In various embodiments, the selected target genomic region can be located in various positions in the genome, including but not limited to exons, introns, intergenic regions and other parts. In some embodiments, probes targeting non-human genomic regions (e.g., those targeting viral genomic regions) can be added.

[0307] VIII. Cancer Applications

[0308] In some embodiments, the method of the present invention, the analysis system and / or the classifier can be used to detect the presence (or absence) of cancer, monitor cancer progression or recurrence, monitor treatment response or effectiveness, determine the presence of minimal residual disease (MRD) or monitor the disease, or any combination thereof. In some embodiments, the analysis system and / or the classifier can be used to identify the source tissue of cancer. For example, the system and / or the classifier can be used to identify cancer as any one of the following cancer types: head and neck cancer, upper digestive tract cancer, pancreatic / gallbladder cancer, colorectal cancer, ovarian cancer, lung cancer, multiple myeloma, lymphoid neoplasms, melanoma, sarcoma, breast cancer and uterine cancer. For example, as described herein, a classifier can be used to generate a likelihood or probability score (e.g., 0 to 100) of a sample feature vector from a subject suffering from cancer. In some embodiments, the probability score is compared with a threshold probability to determine whether the subject suffers from cancer. In other embodiments, the likelihood or probability score can be evaluated at different time points (e.g., before or after treatment) to monitor disease progression or monitor treatment effectiveness (e.g., treatment effectiveness). In yet other embodiments, the likelihood or probability score can be used to make or image clinical decisions (e.g., cancer diagnosis, treatment selection, treatment effectiveness evaluation, etc.). For example, in one embodiment, if the likelihood or probability score exceeds a threshold, the physician can prescribe an appropriate treatment. In some embodiments, a test report can be generated to provide the patient with their test results, including, for example, the probability score that the patient has a disease state (e.g., cancer), a disease type (e.g., a cancer type), and / or a disease source tissue (e.g., a cancer source tissue).

[0309] IX.A. Early Detection of Cancer

[0310] In some embodiments, the method and / or classifier of the present invention is used to monitor the presence or absence of cancer in a subject suspected of having cancer. For example, a classifier (as described herein) can be used to determine the likelihood or probability score that a sample feature vector is from a subject suffering from cancer.

[0311] In one embodiment, a probability score of greater than or equal to 60 may indicate that the subject has cancer. In yet other embodiments, a probability score of greater than or equal to 65, greater than or equal to 70, greater than or equal to 75, greater than or equal to 80, greater than or equal to 85, greater than or equal to 90, or greater than or equal to 95 indicates that the subject has cancer. In other embodiments, the probability score may indicate the severity of the disease. For example, a probability score of 80 may indicate a more severe form or more advanced stage of the cancer compared to a score below 80 (e.g., a score of 70). Similarly, an increase in the probability score over time (e.g., at a second, subsequent time point) may indicate disease progression or a decrease in the probability score over time (e.g., at a second, subsequent odds ratio time point) may indicate successful treatment.

[0312] In another embodiment, the cancer log odds ratio of the test subject can be calculated by obtaining the logarithm of the ratio of the cancer probability relative to the non-cancerous probability (i.e., 1 minus the cancer probability), as described herein. According to this embodiment, a cancer log odds ratio greater than 1 can indicate that the subject suffers from cancer. In other embodiments, a cancer log odds ratio greater than 1.2, greater than 1.3, greater than 1.4, greater than 1.5, greater than 1.7, greater than 2, greater than 2.5, greater than 3, greater than 3.5 or greater than 4 indicates that the subject suffers from cancer. In other embodiments, the cancer log odds ratio can indicate the severity of the disease. For example, compared with a score lower than 2 (e.g., score 1), a cancer log odds ratio greater than 2 can indicate a more severe form or more advanced stage of cancer. Similarly, an increase in the cancer log odds ratio over time (e.g., at a second, subsequent time point) can indicate disease progression or a decrease in the cancer log odds ratio over time (e.g., at a second, subsequent time point) can indicate successful treatment.

[0313] According to aspects of the present disclosure, the methods and systems of the present invention can be trained to monitor or classify a variety of cancer indications. For example, the methods, systems and classifiers of the present invention can be used to detect the presence of one or more, two or more, three or more, five or more, or ten or more different types of cancer.

[0314] In some embodiments, the cancer is one or more of head and neck cancer, liver / bile duct cancer, upper gastrointestinal cancer, pancreatic / gall bladder cancer, colorectal cancer, ovarian cancer, lung cancer, multiple myeloma, lymphoid neoplasms, melanoma, sarcoma, breast cancer, and uterine cancer.

[0315] IX.B. Cancer and Treatment Monitoring

[0316] In certain embodiments, the first time point is before cancer treatment (e.g., before resection surgery or therapeutic intervention), the second time point is after cancer treatment (e.g., after resection surgery or therapeutic intervention), and the method is used to monitor treatment effectiveness. For example, if the second likelihood or probability score is lower than the first likelihood or probability score, the treatment may be considered to have been successful. However, if the second likelihood or probability score is higher than the first likelihood or probability score, the treatment may be considered to have not yet been successful. In other embodiments, both the first time point and the second time point are before cancer treatment (e.g., between resection surgery or therapeutic intervention). In other embodiments, both the first time and the second time point are after cancer treatment (e.g., after resection surgery or therapeutic intervention), and the method is used to monitor treatment effectiveness or loss of treatment effectiveness. In other embodiments, cfDNA samples can be obtained from cancer patients at the first time point and the second time point, and typing is performed to, for example, monitor cancer progression, determine whether the cancer is in remission (e.g., after treatment), monitor or detect residual disease or disease recurrence, or monitor treatment (e.g., therapeutic) efficacy.

[0317] Those skilled in the art will readily appreciate that test samples can be obtained from cancer patients at any desired time point and analyzed according to the methods of the present invention to monitor the cancer status in the patient. In some embodiments, the first time point and the second time point are separated by an amount of time between about 15 minutes to about 30 years, for example, about 30 minutes, such as about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23 or about 24 hours, such as about 1, 2, 3, 4, 5, 10, 15, 20, 25 or about 30 days, or such as about 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 or 12 about, or such as about 1, 1.5, 2, 2.5, 3, 3.5, 4, 5 4, 4.5,5, 5.5,6, 6.5,7, 7.5,8, 8.5,9, 9.5,10, 10.5,11, 11.5,12, 12.5,13, ​​13.5,14, 14.5,15, 15.5,16, 16.5,17, 17.5,18, 18.5,19, 19.5,20, 20.5,21, 21.5,22, 22.5,23, 23.5,24, 24.5,25, 25.5,26, 26.5,27, 27.5,28, 28.5,29, 29.5 or about 30 years. In other embodiments, a test sample may be obtained from a patient at least once every 3 months, at least once every 6 months, at least once a year, at least once every 2 years, at least once every 3 years, at least once every 4 years, or at least once every 5 years.

[0318] IX.C. Treatment

[0319] In yet another embodiment, the information obtained from any of the methods described herein (e.g., likelihood or probability scores) can be used to make or image clinical decisions (e.g., cancer diagnosis, treatment selection, evaluation of treatment effectiveness, etc.). For example, in one embodiment, if the likelihood or probability score exceeds a threshold, the physician can prescribe appropriate treatment (e.g., resection, radiation therapy, chemotherapy, and / or immunotherapy). In some embodiments, the information (e.g., likelihood or probability score) can be provided as a readout to a physician or subject.

[0320] A classifier (as described herein) can be used to determine that the sample feature vector is from a subject with cancer. Likelihood or probability score. In one embodiment, when the likelihood or probability exceeds a threshold, a prescription for appropriate treatment (e.g., resection or therapeutic measures) is prescribed. For example, in one embodiment, if the likelihood or probability score is greater than or equal to 60, one or more prescriptions for appropriate treatments are prescribed. In another embodiment, if the likelihood or probability score is greater than or equal to 65, greater than or equal to 70, greater than or equal to 75, greater than or equal to 80, greater than or equal to 85, greater than or equal to 90 or greater than or equal to 95, one or more prescriptions for appropriate treatments are prescribed. In other embodiments, the cancer log odds ratio may indicate the effectiveness of cancer treatment. For example, the cancer log odds ratio increases over time (e.g., in the second, subsequent treatment) and may indicate that treatment is not effective. Similarly, the cancer log odds ratio decreases over time (e.g., in the second, subsequent treatment) and may indicate successful treatment. In another embodiment, if the cancer log odds ratio is greater than 1, greater than 1.5, greater than 2.5, greater than 3, greater than 3.5, or greater than 4, one or more appropriate treatments are prescribed.

[0321] In some embodiments, treatment is one or more selected from the following cancer therapeutic agents, including chemotherapeutic agents, targeted cancer therapeutic agents, differentiation therapeutic agents, hormone therapeutic agents and immunotherapeutic agents. For example, treatment can be one or more selected from the following chemotherapeutic agents, including alkylating agents, antimetabolites, anthracyclines, antitumor antibiotics, cytoskeleton disruptors (taxanes), topoisomerase inhibitors, mitotic inhibitors, corticosteroids, kinase inhibitors, nucleotide analogs, platinum drugs and any combination thereof. In some embodiments, treatment is one or more selected from the following targeted cancer therapeutic agents, including signal transduction inhibitors (e.g., tyrosine kinase and growth factor receptor inhibitors), histone deacetylase (HDAC) inhibitors, retinoic acid receptor agonists, proteasome inhibitors, angiogenesis inhibitors and monoclonal antibody conjugates. In some embodiments, treatment is one or more differentiation therapeutic agents, including retinoids, such as retinoic acid, arithione and bexarotene. In some embodiments, treatment is one or more selected from the following hormone therapeutic agents, including anti-estrogens, aromatase inhibitors, progestins, estrogen, anti-androgens and GnRH agonists or analogs. In one embodiment, the treatment is one or more immunotherapeutic agents selected from the group consisting of monoclonal antibody therapies such as rituximab (RITUXAN) and alemtuzumab (CAMPATH), nonspecific immunotherapies and adjuvants such as BCG, interleukin-2 (IL-2) and interferon alpha, immunomodulatory drugs such as thalidomide and lenalidomide (REVLIMID). The selection of an appropriate cancer therapeutic agent based on characteristics such as tumor type, cancer stage, previous exposure to cancer treatments or therapeutic agents, and other characteristics of the cancer is within the capabilities of a skilled physician or oncologist.

[0322] X. Examples

[0323] XA Example 1 - Whole Genome Bisulfite Sequencing (WBGS)

[0324] First CCGA sub-study: Obtained from the first CCGA sub-study Fig. 7A -Data shown in C, where training data blood samples (N=1785) were collected from individuals diagnosed with untreated cancer (including 20 tumor types and all cancer stages) and healthy individuals not diagnosed with cancer (controls) for plasma cfDNA extraction. Another set of blood samples (N=1,010) was collected for validation. Unless otherwise indicated, whole genome bisulfite sequencing analysis was performed on cell-free DNA (cfDNA) and genomic DNA (gDNA) extracted from the first CCGA sub-study samples.

[0325] During the classification process, the processing system 200 considers the fragment methylation state as being derived from a mixture of potential methylation patterns. The processing system 200 assigns a relative probability of the observed fragment being derived from a tissue of origin of a particular cancer.

[0326] More specifically, as described herein, a probabilistic model is fitted to sequence reads derived from a plurality of regions (or windows) from each cancer type (as well as for non-cancer or healthy samples). In this case, a mixture model is used in which each mixture component is an independent site model (in which the methylation at each CpG is independent of the methylation at other CpGs). The model is fitted using maximum likelihood estimation to identify a parameter set that maximizes the total log likelihood that all fragments are derived from a cancer type (or non-cancer).

[0327] For each region and each cancer type pair (including non-cancer as the negative type), a multinomial logistic regression classifier was trained using the best performing stratum. For each sample (regardless of labeling), in each region, for each cancer type and for each segment, the log-likelihood ratio was calculated as previously described, and for each set of "stratum" values, the R 癌症类型 >Number of fragments at a level. The quantized reads for each level are binarized and used as features for training the classifier.

[0328] Finally, where indicated, to generate predictions for unknown samples, feature values ​​were determined (as described above), and the generated features were used to generate cancer and / or tissue of origin predictions using a trained multinomial logistic regression classifier.

[0329] Example confusion matrix: Fig. 7A , Figure 7B and Figure 7C Contains a confusion matrix indicating the accuracy of the classifier according to various embodiments. In some embodiments, the processing system 200 uses a confusion matrix to determine the accuracy of the classifier. The confusion matrix includes information describing the success rate of the classifier in identifying each disease state.

[0330] like Fig. 7A As shown in , matrix 710 includes the example performance of the classifier based on the polynomial model trained using the cfDNA sample set (without tissue samples). Matrix 720 includes the example performance of the classifier based on the hybrid model trained using the same cfDNA sample set by processor 200. The scores along the diagonal of the matrix indicate correct predictions, i.e., where the predicted fragment source tissue matches the true source tissue. Compared to the classifier based on the polynomial model (as a baseline), the classifier based on the hybrid model has greater overall accuracy in predicting the presence of the cancer types shown in the matrix.

[0331] The samples of the training set can be filtered based on one or more criteria (e.g., a particular level of specificity). For example, the training set includes samples that are determined to have cancer based on 98% specificity according to the m-score. For clarity, the remaining (e.g., 2%) of non-cancer samples (erroneously) identified as having cancer are excluded and not displayed in the confusion matrix.

[0332] like Figure 7B As shown in , matrix 730 includes example performance of a classifier based on a hybrid model trained using a cross-validation training set of cfDNA samples (without tissue samples). Matrix 740 includes example performance of a classifier based on a hybrid model trained using a cross-validation training set of cfDNA and tissue samples.

[0333] like Figure 7C As shown in , matrix 750 includes the example performance of the classifier based on the hybrid model trained using the cfDNA sample set (without tissue samples) from the clinical study entitled Circulating Cell-free Genome Atlas Study ("CCGA"). Matrix 740 includes the example names of the classifier based on the hybrid model trained using the cfDNA and tissue sample set from CCGA. The CCGA study is described using Clinical Trial.gov Identifier: NCT02889978 (https: / / www.clinicaltrials.gov / ct2 / show / NCT02889978).

[0334] XB Example 2 - Classification of cancer using targeted bisulfite sequencing with early breakthroughs from the second CCGA sub-study

[0335] Second CCGA sub-study: Early breakthroughs from the second CCGA sub-study Fig.9A -B. Fig. 10A -B. Fig.11 and Fig.12The data shown in , where training data blood samples (N=3,132) were collected from individuals diagnosed with untreated cancer (including 20 types and all cancer stages) and healthy individuals not diagnosed with cancer (controls) for plasma cfDNA extraction. Another blood sample set (N=1,354) was collected for validation. In some embodiments, where indicated, the training set also includes training data from tissue samples (i.e., gDNA). To determine the analysis population, the training data blood samples are filtered based on several factors. For example, 105 samples were excluded due to clinical unlocking; 11 samples were excluded based on eligibility criteria; 58 samples were excluded (unevaluable) due to unconfirmed cancer or treatment status; 4 untreated samples and 72 unevaluable analyses (unevaluable) were excluded; and 581 samples were retained for future analysis. Therefore, the analysis group of 2,301 samples includes 1,422 cancer samples and 879 non-cancer samples.

[0336] Individual participant demographics in the sub-studies are shown in Table 1 below.

[0337]

[0338]

[0339] Table 1: Participant demographics and stage distribution. The cancer and noncancer groups were comparable in age, race, sex, and body mass index (not shown). * Includes anorectal cancer, bladder cancer, brain cancer, breast cancer, cervical cancer, colorectal cancer, esophageal cancer, gastric cancer, head and neck cancer, hepatobiliary cancer, lung cancer, lymphoid neoplasms (chronic lymphocytic leukemia, lymphoma), multiple myeloma, myeloid neoplasms (acute myeloid leukemia, chronic myeloid leukemia), ovarian cancer, pancreatic cancer, prostate cancer, kidney cancer, sarcoma, and uterine cancer. Thirty-eight participants with missing smoking status information were excluded. Two participants with missing BMI values ​​were excluded. §Invasive cancer only. No phase information available.

[0340] To identify cancer-defining and tissue-defining methylation signals, bisulfite sequencing analysis was performed on the extracted cfDNA to target the most informative regions of the methylome, as identified from GRAIL's proprietary whole-genome bisulfite sequencing analysis and methylation database.

[0341] A methylation database was used to interrogate whole genome fragment-level methylation patterns in 811 cancer cell methylomes representing 21 tumor types (97% of SEER cancer incidence). To generate a methylation database of cancer-defining methylation signals, whole genome bisulfite sequencing analysis was performed on genomic DNA from formalin-fixed, paraffin-embedded (FFPE) tumor tissue and isolated cells from species. Panel design and training were performed using this methylation database to optimize the performance of the classifier as described herein. Larger methylation sequence databases of cancer and non-cancer were generated to enable targeted selection of a single test that can classify multiple cancers with high specificity and identify the tissue of origin.

[0342] Target selection and panel design: As described in this article, a methylation sequence database from the CCGA study was used to select target genomic regions. Specifically, the cfDNA sequences in the database were filtered based on the p-value using the non-cancer distribution, and only fragments with p<0.001 were retained. The selected cfDNA was further filtered to retain only at least 90% methylated or 90% unmethylated. Next, for each CpG site in the selected fragments, the number of cancer samples or non-cancer samples containing fragments overlapping with the CpG site was counted. Specifically, P(cancer|overlapping fragments) was calculated for each CpG, and genomic sites with high P values ​​were selected as general cancer targets. By design, the selected fragments have extremely low noise (i.e., less overlap with non-cancer fragments).

[0343] To find cancer type specific targets, a similar selection process is performed. CpG sites are ranked based on their information gain, comparing one cancer type to all other samples (i.e., non-cancer plus other cancer types). A cancer assay panel is generated that includes probes targeting selected genomic regions, as described herein. Specifically, the panel is designed to detect the presence of cancer in general (i.e., versus non-cancer) or a specific cancer type (e.g., TOO). The panel includes probe sets targeting each of the selected genomic regions.

[0344] Probes are designed to overlap any CpG site contained within the start / end of any targeted region (eg, aberrant fragment).

[0345] Classification: During classification, the processing system 200 treats the fragment methylation state as a mixture of potential methylation patterns. The processing system 200 assigns a relative probability of originating from cancer to the observed fragments. For tissue of origin classification, the processing system 200 assigns a relative probability of originating from a specific tissue to the observed fragments. The processing system 200 combines the fragment characteristics of cancer and tissue of origin in the targeted region to classify cancer vs. non-cancer and / or identify the tissue of origin. For binary cancer classification, the processing system 200 estimates sensitivity at 99% specificity.

[0346] More specifically, as described in Example VI.a, a probabilistic model is fitted to sequence reads derived from a plurality of regions (or windows) from each cancer type (as well as for non-cancer or healthy samples), the identified features, and a trained multinomial logistic regression classifier. To generate predictions for positional samples, feature values ​​are determined (as described above), and the generated features are used to generate cancer and / or tissue of origin predictions using a trained multinomial logistic regression classifier.

[0347] Fig.9A and Fig. 9B The sensitivity of the tissue of origin classifier generated by the methods described in this disclosure is shown. Sensitivity is reported at 99% specificity, and 95% confidence intervals are indicated. Fig.9A Model predictions for a predetermined list of cancers are shown. Fig. 9B Model predictions for other cancers included in the CCGA study are shown. Demographic information alone (baseline modeling) correctly classified <5% of participants. The overall sensitivity in the predetermined list of cancers (anorectal cancer, breast cancer [HR negative], colorectal cancer, esophageal cancer, gastric cancer, head and neck cancer, hepatobiliary cancer, lung cancer, lymphoid neoplasms [chronic lymphocytic leukemia, lymphoma], multiple myeloma, ovarian cancer, pancreatic cancer) was 76.1% (95% CI: 73.1-78.9%). In this cohort, the sensitivity in early (I-III) cancer was 68.8% (95% CI: 64.8-72.6%). The overall sensitivity in all cancer types and stages was 55.1% (95% CI: 52.5-57.7%). In early (I-III) cancer, the sensitivity was 43.8% (95% CI: 40.7-46.8%).

[0348] Fig. 10A and Fig. 10B The sensitivity of the tissue of origin classifier at different cancer stages is shown. The overall sensitivity based on the individual stages (as indicated in the legend) of the predetermined cancer of interest is reported at 99% specificity. The numbers in the boxes represent the total number of samples included in each stage. The 95% confidence interval is indicated. "Lymphoid neoplasms" include lymphomas (stages I-IV) and chronic lymphocytic leukemia (unstaged, included as "IN").

[0349] Fig.11A performance grid representing the accuracy of tissue-of-origin localization is shown. Using the tissue-of-origin classifier, the true (x-axis) tissue-of-origin of each sample was consistent with the predicted (y-axis) tissue-of-origin, using the methylation database in stage I-IV samples. The gradient legend corresponds to the correct (x-axis) proportion of the predicted tissue-of-origin (y-axis). The analysis showed that the accuracy of tissue-of-origin localization (the correct fraction of all TOO predictions) was higher (p=0.0066) using the methylation database. This is consistent with stage I-III predictions: 89.9% (384 / 427), as further confirmed in Table 2.

[0350]

[0351] Table 2: Improved performance of source tissues when including a methylation database. *P values ​​calculated using the Stuart-Maxwell test. An uncertain call was defined as a sample detected as cancer without a confident tissue of origin assignment. Samples not called by tissue-of-origin analysis were classified as non-cancer.

[0352] An effective multi-cancer test would ideally detect clinically significant cancers of all stages simultaneously with extremely high specificity (and would thus have a single fixed, low false-positive rate) and accurately identify the tissue of origin. Fig.12 The list of predefined cancer types is shown in Figure 2 for simultaneous detection of individual stages (reported sensitivity at 99% specificity) and tissue of origin determination as a whole. Fig.12 The accuracy and sensitivity of the tissue-of-origin classifier at different cancer stages are shown.

[0353] Fig.13A and Fig. 13B Receiver operating characteristic (ROC) curves for the tissue of origin classifier are shown. The receiver operating characteristic (ROC) curves show classifier performance at 99% specificity, with 55% sensitivity for all cancers and 76% sensitivity for multiple cancers.

[0354] These data demonstrate that a classification approach using targeted methylation signatures simultaneously detects multiple early-stage cancer types at a specificity suitable for population screening (99%). Detection of multiple cancers was achieved at a single, fixed, low false-positive rate. This approach also accurately localizes the tissue of origin, which will simplify downstream diagnostic post-processing. Furthermore, incorporation of data from larger methylation databases may improve classifier performance.

[0355] Overall, this demonstrates that the methods described in this disclosure are clinically applicable as early multi-cancer detection tests for a number of clinically significant cancer types.

[0356] XC Example 3 - Classification of cancer using targeted bisulfite sequencing from the complete second CCGA sub-study

[0357] Generation of hybrid model classifiers: To maximize performance, the predictive cancer model described in this example was trained using sequence data obtained from the following samples: multiple samples of known cancer types and non-cancer from two CCGA sub-studies (CCGA1 and CCGA2); multiple tissue samples of known cancer obtained from CCGA1; and multiple non-cancer samples from the STRIVE study (see Clinical Trail.gov Identifier: NCT03085888 ( / / clinicaltrials.gov / ct2 / show / NCT03085888)). The STRIVE study is a prospective, multicenter, observational cohort study to validate an early detection assay for breast cancer and other invasive cancers, from which additional non-cancer training samples were obtained to train the classifier described herein. The known cancer types included in the CCGA sample set include the following types: breast cancer, lung cancer, prostate cancer, colorectal cancer, kidney cancer, uterine cancer, pancreatic cancer, esophageal cancer, lymphoma, head and neck cancer, ovarian cancer, hepatobiliary cancer, melanoma, cervical cancer, multiple myeloma, leukemia, thyroid cancer, bladder cancer, gastric cancer and anorectal cancer. Therefore, the model can be a multi-cancer model (or multi-cancer classifier) ​​for detecting one or more, two or more, three or more, four or more, five or more, ten or more, or 20 or more different types of cancer. 4,841 participants from the CCGA study (2,836 cancers; 2,005 non-cancers) and 2,202 non-cancer participants from the STRIVE study were included in this pre-specified analysis. Of these participants, 3,133 samples from CCGA were assigned for training (1,742 cancer; 1,391 non-cancer), and 1,354 samples were assigned for validation (740 cancer, 614 non-cancer); 1,587 samples from STRIVE were assigned for training, and 615 samples were assigned for validation. The participant distribution is shown. Overall, 3,052 training samples (1,531 cancer; 1,521 non-cancer) and 1,264 validation samples (654 cancer; 610 non-cancer) could be analyzed in the predefined primary analysis population.Additional details about the CCGA2 substudy and the analyses detailed in this example are described in an Annals of Oncology journal article titled “Sensitive and specific multi-cancer detection and localization using methylation signatures in cell-free DNA,” which was published online on March 30, 2020 (https: / / www.annalsofoncology.org / article / S0923-7534(20)36058-0 / fulltext).

[0358] The classifier performance data shown below are reported for locked classifiers trained on cancer and non-cancer samples obtained from CCGA2 (a substudy of CCGA) and non-cancer samples from STRIVE. The individuals in the CCGA2 substudy were different from the individuals in the CCGA1 substudy whose cfDNA was used to select the target genome (as described in WO 2019 / 195268, filed April 2, 2019, PCT / US2019 / 053509, filed September 27, 2019, and PCT / US2020 / 015082, filed January 24, 2020 (which are incorporated herein by reference)). In the CCGA2 study, blood was collected from individuals diagnosed with untreated cancer (including 20 tumor types and all cancer stages) and healthy individuals not diagnosed with cancer (controls). In STRIVE, blood samples were collected from women within 28 days after a screening mammogram. Cell-free DNA (cfDNA) was extracted from each sample and treated with bisulfite to convert unmethylated cytosine to uracil. Informative cfDNA molecules in the bisulfite-treated cfDNA were enriched using hybridization probes designed to enrich for bisulfite-converted nucleic acids derived from each of multiple targeted genomic regions in the following three cancer analysis panels: (1) Pan-Cancer Assay Panel No. 4 as described and disclosed in WO 2019 / 195268 (labeled herein as Assay Panel A); (2) Pan-Cancer Assay Panel No. 5 as described and disclosed in WO 2019 / 195268 (labeled herein as Assay Panel B); and (3) a larger exclusive Pan-Cancer Assay Panel (Assay Panel C, described below). The enriched bisulfite-converted nucleic acid molecules were sequenced using paired-end sequencing on the Illumina platform (San Diego, CA) to obtain a sequence read set for each training sample, and the resulting read pairs were aligned to the reference genome, assembled into fragments, and methylated and unmethylated CpG sites were identified.

[0359] Mixture Model-Based Characterization

[0360] For each cancer type (including non-cancer), a probabilistic mixture model was trained and used to assign a probability to each fragment from each cancer and non-cancer sample based on the likelihood of observing the fragment in a given sample type.

[0361] Fragment level analysis

[0362] Briefly, for each sample type (cancer and non-cancer samples) and each region (where each region was used as is when less than 1 kb, or was additionally subdivided into 1 kb length regions with 50% overlap between adjacent regions (e.g., 500 base pairs overlap)), a probabilistic model was fitted to fragments derived from each type of cancer and non-cancer training samples. The probabilistic model trained for each sample type was a mixture model, where each of the three mixture components was an independent site model where methylation at each CpG was assumed to be independent of methylation at other CpGs. Fragments were excluded from the model if: their p-value (from the non-cancer Markov model) was greater than 0.01; they were marked as duplicate fragments; the bag size of the fragment was greater than 1 (only for targeted methylation samples); they did not cover at least one CpG site; or the fragment length was greater than 1000 bases. If the retained training fragment overlapped with at least one CpG from a region, it was assigned to that region. If a fragment overlapped with CpGs in multiple regions, it was assigned to all regions.

[0363] Local Source Model

[0364] Each probability model was fitted using maximum likelihood estimation to identify the parameter set that maximized the log-likelihood of all fragments derived from each sample type according to the regularization penalty. Specifically, in each classification region, a set of probability models was trained, one model for each training marker (i.e., one model for each cancer type and one model for non-cancer). Each model took the form of a Bernoulli mixture model using three components. Mathematically,

[0365]

[0366] Where n is the number of mixture components and is set to 3; m i ∈{0,1} is the observed fragment methylation at position i; f k is the fractional value of component k (where f k ≥0 and ∑f k =1); and β kiis the methylation score of component k at CpG i. The product about i only includes those positions whose methylation status can be identified from sequencing. The {f for each model is estimated by using the rprop algorithm (e.g., as described in Riedmiller M, Braun H. RPROP-A Fast Adaptive Learning Algorithm. Proceedings of the International Symposium on Computer and Information Science VII, 1992) k , β ki} to maximize the total log-likelihood of a training labeled segment, and to use the β-distribution prior form for β ki Regularization penalty is applied. Mathematically, the maximization amount is

[0367]

[0368] where r is the regularization strength and is set to 1.

[0369] Characterization

[0370] Once the probabilistic model is trained, a set of numerical features is calculated for each sample. Specifically, for each cancer type and non-cancer sample, in each region, features are extracted for each fragment from each training sample. The extracted features are statistical values ​​of outlier fragments (i.e., abnormally methylated fragments), which are defined as those whose log-likelihood under the first cancer model exceeds the log-likelihood under the second cancer model or the non-cancer model by at least a threshold level value. Outlier fragments are counted separately for each genomic region, sample model (i.e., cancer type), and level (for levels 1, 2, 3, 4, 5, 6, 7, 8, and 9), resulting in 9 features for each region of each sample type. In this way, each feature is defined by three properties: genomic region; a "positive" cancer type marker (excluding non-cancer); and a level value, which is selected from the set {1, 2, 3, 4, 5, 6, 7, 8, 9}. The numerical value of each feature is defined as the number of fragments in the region, so that

[0371]

[0372] where the probability is defined by equation (1) using the maximum likelihood estimated parameter values ​​corresponding to a "positive" cancer type (in the numerator of the algorithm) or non-cancer (in the denominator).

[0373] Feature sorting

[0374] For each paired feature set, the features are ranked based on the ability to distinguish the first cancer type (which defines the log-likelihood model of the derived features) from the second cancer type or non-cancer using mutual information. Specifically, two feature ranking lists are compiled for each unique class marker pair: one list uses the first marker assigned as "positive" and the second marker assigned as "negative", and the other list uses the swapped positive / negative assignments (except for the "non-cancer" marker, which is only allowed as a negative marker). For each of these ranked lists, only the features of the positive cancer type markers (as in equation (3)) that match the considered positive markers are included in the ranking. For each such feature, the scores of the training samples with non-zero feature values ​​are calculated separately for the positive markers and the negative markers. The features with the larger scores in the positive markers are ranked according to the mutual information about the class marker pair.

[0375] 256 top features from each pairwise comparison were identified and added to the final feature set for each cancer type and non-cancer type. To avoid redundancy, if more than one feature was selected from the same positive type and genomic region (i.e., for multiple negative types), only the lowest (most informative) rank assigned to the cancer type pair was retained, thereby interrupting the hierarchy by selecting a higher rank value. The features in the final feature set for each sample (cancer type and non-cancer type) were binarized (any feature value greater than 0 was set to 1, so that all features were 0 or 1).

[0376] Classifier training

[0377] The training samples were then divided into different 5-fold cross-validation training sets, and a two-stage classifier was trained for each fold, using in each case 4 / 5 of the training samples for training and the remaining 1 / 5 for validation.

[0378] In the first training phase, a binary (two-class) logistic regression model for detecting the presence of cancer was trained to distinguish between cancer samples (regardless of TOO) and non-cancer. When training this binary classifier, sample weights were assigned to male non-cancer samples to offset the gender imbalance in the training set. For each sample, the binary classifier outputs a prediction score indicating the likelihood of the presence or absence of cancer.

[0379] In the second stage of training, TOO is used as a target marker to train a scoring multi-class logistic regression model to determine the tissue of origin of the cancer. Only cancer samples with scores higher than 95% of non-cancer samples in the first stage classifier are included to train the multi-class classifier. For each cancer sample used to train the multi-class classifier, the multi-class classifier outputs a predicted value of the classified cancer type, wherein each predicted value is the likelihood that a given sample has a certain cancer type. For example, a cancer classifier can return a cancer prediction for a test sample, and the cancer prediction includes a breast cancer prediction score, a lung cancer prediction score, and / or a cancer-free prediction score.

[0380] Both binary and multi-class classifiers were trained using mini-batches by stochastic gradient descent, and in each case, training was stopped early when performance on the validation fold (assessed by cross entropy loss) began to decline. To predict samples outside the training set, in each stage, the scores assigned by the 5 cross-validated classifiers were averaged. Scores assigned to sex-inappropriate cancer types were set to 0, and the remaining values ​​were renormalized to sum to one.

[0381] The scores assigned to the validation folds in the training set are retained for assigning cutoff values ​​(thresholds) to target certain performance metrics. In particular, the probability scores assigned to the non-cancer samples of the training set are used to define the thresholds corresponding to specific specificities. For example, for a desired target specificity of 99.4%, the threshold is set to the cross-validated cancer detection probability score of the non-cancer samples in the 99.4% assigned value training set. Training samples with probability scores exceeding the threshold are called cancer positive.

[0382] Subsequently, for each training sample determined to be cancer positive, a TOO or cancer type assessment is performed from a multi-class classifier. First, a multi-class logistic regression classifier assigns a probability score set to each sample, one probability score for each prospective cancer type. Next, the confidence of these scores is evaluated as the difference between the highest score and the second highest score assigned to each sample by the multi-class classifier. Then, the cross-validated training set score is used to identify the lowest threshold, so that in the cancer samples where the difference between the first two scores in the training set exceeds the threshold, 90% have been assigned the correct TOO marker and as its highest score. In this way, the score assigned to the validation fold during training is further used to determine the second threshold for distinguishing confidence calls from uncertain TOO calls.

[0383] At the prediction time, samples with scores from the binary (first stage) classifier below a predetermined specificity threshold are assigned a "non-cancer" label. For the remaining samples, those with a difference of the top two TOO scores from the second stage classifier below a second predetermined threshold are assigned an "uncertain cancer" label. The remaining samples are assigned the cancer label to which the TOO classifier assigns the highest score.

[0384] Classifier performance on a panel of target genomic regions

[0385] The differential values ​​of the target genomic regions of assay panels AC were evaluated by testing the ability of the cancer classifier to detect cancer and any of 20 different cancer types according to the methylation status of these target genomic regions. For assay panels AB, performance was evaluated in a training set of 1,531 cancer samples and 1,521 non-cancer samples used to train the classifier, as shown in Table 1. For assay panel C, 1,264 validation samples (654 cancers; 610 non-cancers) were used to evaluate performance on a classifier using a training set of the same 3,052 samples (1,531 cancers; 1,521 non-cancers) used to train assay panels AB. For each sample, a decoy set (including all target genomic regions included in assay panels AC) was used to enrich for cfDNA with different methylation. The classifier was then constrained to provide a cancer determination based only on the methylation status of the target genomic regions of the evaluated list. The two-stage classifier embodiment includes a binary (two-class) logistic regression classifier model for detecting the presence of cancer, which is trained to distinguish cancer samples (regardless of TOO) from non-cancer; and a second stage trained multi-class logistic regression classifier model for determining the tissue of origin of the cancer, which is trained using TOO as a target marker, as previously described in this example. Both classifier models were trained and validated using model-based characterization, also as previously described.

[0386] Table 1 - Cancer diagnosis of individuals whose cfDNA was used to train the classifier

[0387]

[0388]

[0389] Assay Panels A and B: Results from the classifier performance analysis of Assay Panels A and B are shown in Fig.26A and Fig.27A In each figure, part A is a receiver operating curve (ROC) showing true positive results and false positive results for determining cancer or non-cancer. The asymmetric shape of these ROC curves shows that the classifier is designed to minimize false positive results. For both assay panels, the area under the curve for assay panels A and B is 0.83.

[0390] The classifier was used to determine the cancer type (ie, TOO) for all samples that tested positive for cancer. Fig.26B and Fig.27BIncluded are confusion matrices indicating the TOO accuracy for each assay panel A and B. The confusion matrix includes information describing the success rate of the classifier in identifying each cancer type and excluding ambiguous cancer calls.

[0391] like Fig.26B and Fig.27B As shown in , the TOO confusion matrix shows the performance of the multi-class logistic regression classifier as described above. The agreement between the actual (x-axis) and predicted (y-axis) tissue of origin for each sample obtained using the targeted methylation classifier is described. The scores along the diagonal of the matrix indicate correct predictions, i.e., where the predicted tissue of origin of the fragment matches the true tissue. Fig.26B As shown in , when excluding uncertain cancer calls, cancer assay panel A has a TOO accuracy of approximately 90.8% (711 / 783). In addition, Fig.27B It was shown that when the uncertain cancer calls were excluded, analysis panel B had a TOO accuracy of approximately 90.3% (705 / 781).

[0392] These classifier results are further summarized in Tables 2-3, which indicate the accuracy of cancer detection and cancer type determination using a specificity of 0.990, indicating a false positive rate of 1%. These results are described according to cancer stage. It shows that samples from individuals with later stage cancer (e.g., stage III) result in improved cancer detection and cancer type determination compared to samples from individuals with earlier stage cancer (e.g., stage II). For all cancer stages (no stage distinction), the accuracy of cancer type determination for both assay panels A and B (including uncertain cancer calls) was approximately 89%.

[0393] Table 2: Classification accuracy of genomic regions using assay panel A. Data for cancer presence and cancer type at a specificity of 0.990 are shown as percent accuracy, 95% confidence interval (in square brackets), and number of correct assignments out of total (in round brackets).

[0394] Staging Cancer exists Cancer Type I 20.4％[16.6-24.5](86 / 422) 71.8％[60.5-81.4](56 / 78) II 44.6％[39.6-49.7](173 / 388) 87.2％[81.1-91.9](143 / 164) III 81.5％[76.7-85.6](255 / 313) 90.5％[86.1-93.9](220 / 243) IV 90.9％[87.5-93.7](330 / 363) 93.3％[90-95.8](294 / 315) All 56.5％[54-59](866 / 1532) 89.1％[86.8-91.2](731 / 820)

[0395] Table 3: Classification accuracy of genomic regions using assay panel B. Data for cancer presence and cancer type at a specificity of 0.990 are shown as percent accuracy, 95% confidence interval (in square brackets), and number of correct assignments out of total (in round brackets).

[0396] Staging Cancer exists Cancer Type I 19.9％[16.2-24](84 / 422) 72.7％[60.4-83](48 / 66) II 45.1％[40.1-50.2](175 / 388) 84.8％[78.2-90](134 / 158) III 81.2％[76.4-85.3](254 / 313) 91.3％[86.9-94.6](211 / 231) IV 90.9％[87.5-93.7](330 / 363) 93.2％[89.8-95.7](287 / 308) All 56.3％[53.7-58.8](862 / 1532) 89.2％[86.9-91.3](697 / 781)

[0397] Assay Panel C: As described above, a third, larger exclusive broad cancer assay panel was also tested. Assay Panel C was designed from WGBS data obtained from the first CCGA sub-study CCGA1 using the feature selection methods disclosed in PCT / US2019 / 053509, filed September 27, 2019, and PCT / US2020 / 015082, filed January 24, 2020, which are incorporated herein by reference. The larger exclusive targeted methylation panel covers 103,456 different regions (17.2 Mb) and covers 1,116,720 CpGs. Assay Panel C contains 363,033 CpGs in 68,059 regions (7.5 Mb) covered by probes targeting hypomethylated fragments; 585,181 CpGs in 28,521 regions (7.4 Mb) covered by probes targeting hypermethylated fragments; and 218,506 CpGs in 6,876 regions (2.3 Mb) targeting both types of fragments. Individual aberrant target regions contained 1 to 590 CpGs, with a median CpG count of 3 for hypomethylated target regions and 6 for hypermethylated target regions. CpGs are present in the following genomic regions: 193,818 (17%) are present in the region 1 to 5 kpb upstream of the transcription start site (TSS); 278,872 (24%) are present in promoters (<1 kbp upstream of TSS); 500,996 (43%) are present in introns; 292,789 (25%) are present in exons; 247,752 (21%) are present in intron-exon boundaries; 134,144 (11%) are present in 5'-untranslated regions; 182,174 (16%) are present between genes; and the remaining 1,817 (<1%) are unannotated. Percentages are relative to the total number of CpGs and do not sum to 100% because each CpG may receive multiple annotations due to overlapping genes and / or transcripts.

[0398] In this evaluation, samples were divided into a training set (n = 4,720) and an independent validation set (n = 1,969). A total of 4,316 participants were analyzable (training: 3,052 [1,531 cancer: stage I: 28%; stage II: 25%; stage III: 20%; stage IV: 24%; missing / unexpected: 3%; 1,521 non-cancer]; validation: 1,264 [654 cancer: stage I: 28%; stage II: 25%; stage III: 21%; stage IV: 23%; missing / unexpected: 3%; 610 non-cancer]) and included in the primary analysis population.

[0399] The results of classifier performance analysis from the training and validation sets are shown in Figure 28-30 middle. Fig.28Panel A shows the specific results for both the training and validation sets, and Panel B shows the sensitivity for predetermined cancers (a subset of 12 high-signal cancers based on the first substudy results and mortality data (anal cancer, bladder cancer, colon / rectum cancer, esophageal cancer, head and neck cancer, liver / bile duct cancer, lung cancer, lymphoma, ovarian cancer, pancreatic cancer, plasma cell neoplasm, gastric cancer)) and all cancer types (>20) in stages I to IV. Fig.28 Panel C of shows the tissue of origin (TOO) accuracy results for both the training and validation sets, and Panel B shows the sensitivity in stages I to IV for the predetermined cancers and all cancer types. Fig.29 The TOO confusion matrix for both the training and validation sets is shown, and Fig.30 Sensitivity results for both the training and validation sets for the predefined transformation types are shown.

[0400] exist Fig.28 In the figure, sensitivity (y-axis) is reported by clinical stage (x-axis) in the predetermined cancer type (left figure) and all cancer types (right figure) for training (orange) and validation (cyan). Tissue of origin accuracy (y-axis) is reported by clinical stage (x-axis) in the predetermined cancer type (left figure) and all cancer types (right figure) for training (orange) and validation (cyan). Numbers indicate samples in the training|validation set.

[0401] like Fig.28As shown in , the classifier consistently achieved high specificity between the cross-validation training set and the independent validation set (99.8% [95% CI: 99.4-99.9%] vs. 99.3% [98.3-99.8%], respectively; P = 0.095); this reflects a single, consistent false positive rate (FPR) of less than 1% across all 20 cancer types. Specificity was similar for CCGA and STRIVE non-cancer samples in the validation set (99.3% [97.4-99.9%] vs. 99.4% [97.9-99.9%], respectively), confirming that performance did not deviate with site or sample selected. Sensitivity was consistent in the training and validation sets. Among all cancers, stage I-III sensitivity was 44.2% (95% CI: 41.3-47.2%) vs. 43.9% (39.4-48.5%), respectively (P = 1.000). For the predefined set of 12 high signal cancers, stage I-III sensitivity was 69.8% (65.6-73.7%) vs. 67.3% (60.7-73.3%), respectively (P=0.988). Similarly, stage I-IV sensitivity in all cancer types was 55.2% (52.7-57.7%) vs. 54.9% (51.0-58.8%), respectively (P=0.897), and in the predefined cancers was 77.9% (75.0-80.7%) vs. 76.4% (71.6-80.7%), respectively (P=0.573).

[0402] Likewise, if Fig.28 As shown in the results, sensitivity increases with increasing disease stage. In validation, the sensitivity in the predefined cancer types was 39% (27–52%) in stage I (n=62), 69% (56–80%) in stage II (n=62), 83% (75–90%) in stage III (n=102), and 92% (86–96%) in stage IV (n=130). Across all cancer types, the sensitivity was 18% (13–25%) in stage I (n=185), 43% (35–51%) in stage II (n=166), 81% (73–87%) in stage III (n=134), and 93% (87–96%) in stage IV (n=148).

[0403] exist Fig.30 Performance in individual tumor types is described in . Sensitivity at 95% confidence intervals and 99.8% specificity (training, orange) or 99.3% specificity (validation, cyan) is reported for individual cancer types using at least 50 samples. Clinical stage and number of samples in training and validation are shown below the graphs.

[0404] like Fig.28As shown in , analysis of the pre-defined TOO accuracy (the fraction of all TOO correctly predicted) found that TOO was predicted in 96% (344 / 359) of samples with cancer-like signatures in the validation set; in these samples, the accuracy was 93% (321 / 344). Accuracy was consistent between the training and validation sets and across stages. The classifier could distinguish between >20 cancer types included in the study, and performance was consistent within individual cancer types.

[0405] Fig.29 Confusion matrices representing the accuracy of tissue-of-origin localization in (A) training and (B) validation sets are shown. The agreement between the actual (x-axis) and predicted (y-axis) tissue-of-origin for each sample obtained using the targeted methylation classifier is shown. The colors correspond to the proportion of predicted tissue-of-origin calls. The included participants (training: n=844, validation: n=359) were predicted to have cancer at 99.8% specificity (training) or 99.3% specificity (validation). Tissue-of-origin calls were assigned in 95% (806 / 844) of the training cases and 96% (344 / 359) of the validation cases; the calls in 92% (744 / 806) of the training cases and 93% (321 / 344) of the validation cases were correct.

[0406] XD Example 4 - Tuning of Binary Classifier Threshold

[0407] According to a generalized embodiment of binary cancer classification, the analysis system determines the cancer score of the test sample based on the test data of the test sample (e.g., methylation sequencing data, SNP sequencing data, other DNA sequencing data, RNA sequencing data, etc.). The analysis system compares the cancer score of the test sample with a binary threshold cutoff value to predict whether the test sample may have cancer. The binary threshold cutoff value can be tuned based on one or more TOO subtypes using TOO criticalization. The analysis system can further generate a feature vector of the test sample for use in a variety of cancer classifiers to determine a cancer prediction indicating one or more possible cancer types.

[0408] Fig.24A A confusion matrix showing the performance of a trained cancer separator according to an example embodiment is shown. The cancer classifier was trained according to the above principles. TOO markers include: lymphoid conjugates, lung cancer, kidney cancer, non-cancer, head and neck cancer, prostate cancer, breast cancer, upper gastrointestinal cancer, liver and bile duct cancer, colorectal cancer, cervical cancer, pancreatic and gallbladder cancer, uterine cancer, sarcoma, bladder and urothelial cancer, ovarian cancer, anorectal cancer, unknown type, melanoma, multiple myeloma, medullary neoplasm, and thyroid cancer. It should be noted that the classification accuracy in the 1,151 samples considered in this maintenance set was 89.1%.

[0409] Fig. 24B A confusion matrix showing the performance of the trained cancer classifier using other hematological cancer subtypes is shown. The cancer classifier was trained according to the above principles. Fig.24A Compared to the TOO markers, adjusted for hematological subtypes. Fig.24A In the context of hematologic subtypes, hematologic subtypes include lymphoid neoplasms, multiple myeloma, and myeloid neoplasms. Fig. 24B In the present study, hematological subtypes included Hodgkin's lymphoma (HL), NHL aggressive, NHL indolent, myeloid, circulating lymphoma (or lymphoid), and plasma cell. It should be noted that the classification accuracy in 1,076 samples was 87.5%.

[0410] Fig.25A and Fig.25B Figures showing the cancer prediction accuracy of multiple cancer types at different cancer stages are shown. In this example, a cancer classifier is trained after pruning non-cancer samples according to the above process 100. The analysis system determines multiple TOO thresholds for hematological subtypes. The analysis system excludes non-cancer samples with at least one TOO probability of or higher corresponding TOO thresholds for hematological subtypes. The displayed figure shows the classification sensitivity of the cancer of the following cancer types at different stages: anorectal cancer, bladder and urothelial cancer, breast cancer, cervical cancer, colorectal cancer, head and neck cancer, liver and bile duct cancer, lung cancer, melanoma, ovarian cancer, pancreatic and gallbladder cancer, prostate cancer, kidney cancer, sarcoma, thyroid cancer, upper digestive tract cancer and uterine cancer. The figure of each cancer type shows the prediction sensitivity of the first cancer classifier (labeled as "locked_v1_orgi") not using TOO thresholding and the second cancer classifier (labeled as "v2_custom") using TOO thresholding at each cancer type stage. It is worth noting that for many cancer types, the second cancer classifier has a higher prediction accuracy and maintains a tight confidence interval at the same time, because more samples can be used for validation. It is particularly noteworthy that the prediction accuracy of many cancer types at the stage I and II level is higher, indicating that the prediction possibility of using TOO thresholding in early cancer has improved.

[0411] XI. Other Considerations

[0412] The above description of the embodiments of the present invention is for illustrative purposes; it is not intended to be exhaustive or to limit the invention to the precise form disclosed. It will be appreciated by those skilled in the relevant art that many modifications and variations are possible based on the above disclosure.

[0413] Certain parts of this specification describe the embodiments of the present disclosure in terms of algorithms and symbolic representations of operations on information. These algorithmic descriptions and representations are often used by technicians in the field of data processing to effectively convey the essence of their work to other technicians in the field. Although these operations are described functionally, computationally, or logically, it should be understood that they are implemented by computer programs or equivalent circuits, microcodes, etc. In addition, without loss of generality, it is sometimes proven to be convenient to refer to these operational arrangements as modules. The described operations and their associated modules can be embodied in software, firmware, hardware, or any combination thereof.

[0414] Any steps, operations or processes described herein may be performed or implemented using one or more hardware or software modules alone or in combination with other devices. In some embodiments, the software module is implemented using a computer program product that includes a computer-readable non-transitory medium containing computer program code that can be executed by a computer processor to perform any or all of the steps, operations or processes described.

[0415] Embodiments may also relate to products produced by the computing processes described herein. Such products may include information generated from the computing process, wherein the information is stored on a non-transitory, tangible computer-readable storage medium and may include any embodiment of a computer program product or other data combination described herein.

[0416] Finally, the language used in the specification is selected primarily for readability and instructional purposes, and may not be selected to describe or limit the subject matter of the present invention. Accordingly, it is intended that the scope of the present invention be limited not by this detailed description, but by any claims that issue based on an application hereof. Accordingly, the embodiments disclosed herein are intended to illustrate, but not to limit, the scope of the present invention, which is set forth in the following claims.

[0417] The present invention seeks protection for:

[0418] 1. A method of analyzing sequence reads to generate features, comprising:

[0419] generating a first plurality of reference sequence reads from a first reference sample from a subject having a first disease state;

[0420] generating a second plurality of reference sequence reads from a second reference sample from a subject having a second disease state,

[0421] training a first probability model using the first plurality of reference sequence reads, the first probability model being associated with the first disease state;

[0422] training a second probability model using the second plurality of reference sequence reads, the second probability model being associated with the second disease state;

[0423] A plurality of training sequence reads are generated from a training sample, and for each sequence read of the plurality of training sequence reads:

[0424] applying the sequence reads to the first probability model to determine a first probability value, the first probability value being a probability of the sequence reads being derived from a sample associated with the first disease state, and

[0425] applying the sequence reads to the second probability model to determine a second probability value, the second probability value being a probability of the sequence reads being derived from a sample associated with the second disease state; and

[0426] One or more features are identified by comparing the first probability value and the second probability value for each sequence read.

[0427] 2. The method of claim 1, wherein the first disease state is cancer and the second disease state is non-cancer.

[0428] 3. The method of claim 1, wherein the first disease state is a first type of cancer and the second disease state is a second type of cancer, and wherein the first type of cancer and the second type of cancer are different.

[0429] 4. The method of item 1, wherein the method further comprises:

[0430] generating a plurality of reference sequence reads from a third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference sample, each of said third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference samples having a different disease state, and wherein each of said different disease state is a different type of cancer; and

[0431] The third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models are trained using the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth plurality of reference sequence reads, wherein each of the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models is respectively associated with a different type of cancer.

[0432] 5. The method of any one of items 2-4, wherein the cancer or the type of cancer is selected from the following, including: breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, renal pelvic ureteral urothelial carcinoma, renal cancer other than urothelial carcinoma, prostate cancer, anorectal cancer, colorectal cancer, esophageal squamous cell carcinoma, esophageal cancer other than squamous cell carcinoma, gastric cancer, hepatobiliary carcinoma originating from hepatocytes, hepatobiliary carcinoma originating from cells other than hepatocytes, pancreatic cancer, head and neck cancer associated with human papillomavirus, head and neck cancer not associated with human papillomavirus, lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer and lung cancer other than adenocarcinoma or small cell lung cancer, neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma and leukemia.

[0433] 6. The method of item 5, wherein the cancer type is additionally selected from the group consisting of brain cancer, vulvar cancer, vaginal cancer, testicular cancer, pleural mesothelioma, peritoneal mesothelioma, and gallbladder cancer.

[0434] 7. The method described in item 1, wherein the first disease state includes a first source tissue and the second disease state includes a second source tissue.

[0435] 8. The method described in item 7, wherein the first source tissue or the second source tissue is selected from the following, including: breast tissue, thyroid tissue, lung tissue, bladder tissue, cervical tissue, small intestine tissue, colorectal tissue, esophageal tissue, gastric tissue, tonsil tissue, liver tissue, ovarian tissue, fallopian tube tissue, pancreatic tissue, prostate tissue, kidney tissue and uterine tissue.

[0436] 9. The method described in item 8, wherein the first source tissue or the second source tissue is additionally selected from the following, including: brain tissue and cells, endocrine tissue and cells, vascular endothelial tissue and cells, head and neck tissue and cells, exocrine pancreatic tissue and cells, endocrine pancreatic tissue and cells, lymphoid tissue and cells, mesenchymal tissue and cells, myeloid tissue and cells, pleural tissue and cells, muscle tissue and cells, bone marrow tissue and cells, adipose tissue and cells, gallbladder tissue and cells.

[0437] 10. The method of any of the preceding items, wherein the first probability model or the second probability model is a constant model, a binomial model, an independent site model, a neural network model or a Markov model.

[0438] 11. The method of any of the preceding items, further comprising:

[0439] A methylation rate for each of a plurality of CpG sites in the first plurality of reference sequence reads or the second plurality of reference sequence reads is determined, wherein the first probability model or the second probability model is parameterized by a product of the methylation rates.

[0440] 12. The method of any of the preceding items, further comprising:

[0441] For each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, determining whether the sequence read is hypomethylated or hypermethylated is performed by determining whether at least a threshold number of CpG sites having at least a threshold percentage of the CpG sites are unmethylated or methylated, respectively.

[0442] 13. The method of any of the preceding items, further comprising:

[0443] determining, for each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, whether the sequence read is aberrantly methylated; and

[0444] The first plurality of reference sequence reads is filtered using p-value filtering by removing sequence reads from the first plurality of reference sequence reads having a p-value below a threshold.

[0445] 14. The method described in item 10, wherein the first probability model or the second probability model is parameterized by the sum of multiple mixed components, each component being related to the product of the methylation rates.

[0446] 15. The method of claim 14, wherein each of the plurality of mixture components is associated with a fractional assignment, and wherein the fractional assignments sum to one.

[0447] 16. The method of any of the preceding clauses, wherein training the first probability model or the second probability model comprises:

[0448] A set of parameters is determined for the probability model, the set of parameters maximizing the total log-likelihood of the first plurality of reference sequence reads or the second plurality of reference sequence reads derived from subjects associated with the first disease state or the second disease state, the disease state being associated with the probability model.

[0449] 17. The method of any of the preceding items, wherein the method further comprises:

[0450] For each of the multiple windows:

[0451] selecting a plurality of the first plurality of reference sequence reads derived from the window, and training the first probability model for the window using the sequence reads derived from the window; and

[0452] A plurality of the second plurality of reference sequence reads derived from the windows are selected, and the probability model for each window is trained using the sequence reads.

[0453] 18. The method of item 17, wherein the method further comprises, for each of the plurality of windows:

[0454] selecting a subset of a plurality of said training sequence reads derived from said window; and

[0455] Comparing the first probability value and the second probability value for each sequence read of the subset identifies the one or more features.

[0456] 19. The method of claim 17, wherein each of the windows is separated by at least a threshold number of base pairs between CpG sites.

[0457] 20. The method of any one of items 17-19, wherein each of the multiple windows comprises about 200 base pairs (bp) to about 10 kilobase pairs (kbp).

[0458] 21. The method of any of the preceding items, wherein the one or more features include a count of outlier sequence reads of the plurality of training sequence reads wherein the first probability value is greater than the second probability value.

[0459] 22. The method of claim 21, wherein the one or more features comprise binary counts.

[0460] 23. A method as described in any of the preceding items, wherein the one or more features include a total count of outlier sequence reads.

[0461] 24. The method of any of the preceding items, wherein the one or more features include a total count of anonymously methylated sequence reads.

[0462] 25. The method of any of the preceding items, wherein the one or more features include counts of fragments comprising one or more specific methylation patterns.

[0463] 26. A method as described in any of the preceding items, wherein the one or more features are identified using the output of a discriminative classifier trained on a single genomic region.

[0464] 27. The method of claim 26, wherein the discriminative classifier is a multilayer perceptron or a convolutional neural network model.

[0465] 28. A method as described in any of the preceding items, wherein comparing the first probability value and the second probability value comprises determining a ratio of the first probability value and the second probability value, and wherein the one or more features comprise a sequence read count of sequence reads that exceeds a ratio threshold.

[0466] 29. The method of any of the preceding items, wherein the first probability value or the second probability value is a log-likelihood value.

[0467] 30. The method of any of the preceding clauses, wherein identifying the one or more features comprises:

[0468] For each sequence read of the plurality of training sequence reads:

[0469] determining a log-likelihood ratio of the first probability value to the second probability value; and

[0470] For one or more thresholds, a count of reads of the sequence having a log-likelihood ratio exceeding the threshold is determined.

[0471] 31. The method of any of the preceding items, further comprising:

[0472] For each of the one or more features, an indicator is determined that distinguishes the feature from the first disease state and the second disease state.

[0473] 32. The method of claim 31, wherein determining the indicator for each of the one or more features comprises:

[0474] Mutual information between the features and probabilities of the first disease state and the second disease state being present is determined.

[0475] 33. The method of item 32, further comprising:

[0476] The one or more features used to train the classifier are filtered by ranking the features based on the metric.

[0477] 34. The method of any of the preceding items, further comprising training a classifier from the one or more features, the classifier being trained to predict one or more disease states based on multiple sequence reads of a test sample from a test subject, wherein the one or more disease states include the presence or absence of a disease, the type of disease and / or the tissue of origin of the disease.

[0478] 35. The method of claim 34, wherein the classifier is a multilayer perceptron model.

[0479] 36. The method described in item 34, wherein the classifier is a logistic regression, support vector machine, multinomial logistic regression, multilayer perceptron, random forest or neural network model classifier.

[0480] 37. The method of claim 34, wherein the classifier is generated using L1 or L2 regularized logistic regression.

[0481] 38. The method of item 34, further comprising:

[0482] determining a probability vector for the test sample; and

[0483] A label for the test sample is determined based on the probability vector.

[0484] 39. The method of item 34, further comprising:

[0485] The accuracy of the classifier is determined using a confusion matrix that includes information describing the success rate of the classifier in identifying each of the plurality of disease states.

[0486] 40. The method of any of the preceding items, wherein the first reference sample or the second reference sample is a cell-free nucleic acid sample or a tissue nucleic acid sample from a subject with a known disease state.

[0487] 41. The method of claim 40, wherein the known disease state is the presence or absence of the disease, the disease type, or the disease-originating tissue.

[0488] 42. The method of any of the preceding items, wherein the training sample comprises a cell-free nucleic acid sample or a tissue sample.

[0489] 43. The method described in item 34, wherein the test sample comprises a cell-free nucleic acid sample.

[0490] 44. The method described in item 34, wherein the first plurality of reference sequence reads, the second plurality of reference sequence reads, the plurality of training sequence reads, or the plurality of sequence reads from the test sample are generated by methylation sequencing.

[0491] 45. The method described in item 44, wherein the methylation sequencing comprises whole genome bisulfite sequencing.

[0492] 46. ​​The method described in item 44, wherein the methylation sequencing includes targeted sequencing.

[0493] 47. A system comprising a computer processor and a memory storing computer program instructions which, when executed by the computer processor, cause the processor to perform steps comprising:

[0494] accessing a first plurality of reference sequence reads from a first reference sample from a subject having a first disease state;

[0495] accessing a second plurality of reference sequence reads from a second reference sample from a subject having a second disease state,

[0496] training a first probability model using the first plurality of reference sequence reads, the first probability model being associated with the first disease state;

[0497] training a second probability model using the second plurality of reference sequence reads, the second probability model being associated with the second disease state;

[0498] Accessing a plurality of training sequence reads from a training sample, and for each sequence read of the plurality of training sequence reads:

[0499] applying the sequence reads to the first probability model to determine a first probability value, the first probability value being a probability of the sequence reads being derived from a sample associated with the first disease state, and

[0500] applying the sequence reads to the second probability model to determine a second probability value, the second probability value being a probability of the sequence reads being derived from a sample associated with the second disease state; and

[0501] One or more features are identified by comparing the first probability value and the second probability value for each sequence read.

[0502] 48. The system of claim 47, wherein the first disease state is cancer and the second disease state is non-cancer.

[0503] 49. The system of claim 47, wherein the first disease state is a first type of cancer and the second disease state is a second type of cancer, and wherein the first type of cancer and the second type of cancer are different.

[0504] 50. The system of item 47, wherein the memory stores other computer instructions, which when executed by the computer processor cause the processor to perform the following steps:

[0505] accessing a plurality of reference sequence reads from a third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference sample, each of said third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference samples having a different disease state, and wherein each of said different disease state is a different type of cancer; and

[0506] The third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models are trained using the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth plurality of reference sequence reads, wherein each of the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models is respectively associated with a different type of cancer.

[0507] 51. The system of any one of items 48-50, wherein the cancer or type of cancer is selected from the following: breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, renal pelvis and ureteral urothelial carcinoma, renal cancer other than urothelial carcinoma, prostate cancer, anorectal cancer, colorectal cancer, esophageal squamous cell carcinoma, esophageal cancer other than squamous cell carcinoma, gastric cancer, hepatobiliary carcinoma originating from hepatocytes, hepatobiliary carcinoma originating from cells other than hepatocytes, pancreatic cancer, head and neck cancer associated with human papillomavirus, head and neck cancer not associated with human papillomavirus, lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer and lung cancer other than adenocarcinoma or small cell lung cancer, neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma and leukemia.

[0508] 52. The system of claim 51, wherein the cancer type is additionally selected from the group consisting of brain cancer, vulvar cancer, vaginal cancer, testicular cancer, pleural mesothelioma, peritoneal mesothelioma, and gallbladder cancer.

[0509] 53. The system of claim 47, wherein the first disease state comprises a first source tissue and the second disease state comprises a second source tissue.

[0510] 54. The system of item 53, wherein the first source tissue or the second source tissue is selected from the group consisting of breast tissue, thyroid tissue, lung tissue, bladder tissue, cervical tissue, small intestine tissue, colorectal tissue, esophageal tissue, gastric tissue, tonsil tissue, liver tissue, ovarian tissue, fallopian tube tissue, pancreatic tissue, prostate tissue, kidney tissue and uterine tissue.

[0511] 55. The system of item 54, wherein the first source tissue or the second source tissue is additionally selected from the group consisting of brain tissue and cells, endocrine tissue and cells, vascular endothelial tissue and cells, head and neck tissue and cells, exocrine pancreatic tissue and cells, endocrine pancreatic tissue and cells, lymphoid tissue and cells, mesenchymal tissue and cells, myeloid tissue and cells, pleural tissue and cells, muscle tissue and cells, bone marrow tissue and cells, adipose tissue and cells, and gallbladder tissue and cells.

[0512] 56. A system as described in any one of items 47-55, wherein the first probability model or the second probability model is a constant model, a binomial model, an independent site model, a neural network model or a Markov model.

[0513] 57. The system of any one of items 47-56, wherein the memory stores other computer instructions that, when executed by the computer processor, cause the processor to perform steps including:

[0514] A methylation rate for each of a plurality of CpG sites in the first plurality of reference sequence reads or the second plurality of reference sequence reads is determined, wherein the first probability model or the second probability model is parameterized by a product of the methylation rates.

[0515] 58. The system of any one of items 47-56, wherein the memory stores other computer instructions that, when executed by the computer processor, cause the processor to perform steps including:

[0516] For each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, determining whether the sequence read is hypomethylated or hypermethylated is performed by determining whether at least a threshold number of CpG sites having at least a threshold percentage of the CpG sites are unmethylated or methylated, respectively.

[0517] 59. The system of any one of items 47-56, wherein the memory stores other computer instructions that, when executed by the computer processor, cause the processor to perform steps including:

[0518] determining, for each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, whether the sequence read is aberrantly methylated; and

[0519] The first plurality of reference sequence reads is filtered using p-value filtering by removing sequence reads from the first plurality of reference sequence reads having a p-value below a threshold.

[0520] 60. The system of claim 56, wherein the first probability model or the second probability model is parameterized by the sum of a plurality of mixed components, each component being related to the product of the methylation rates.

[0521] 61. The system of claim 60, wherein each of the plurality of mixture components is associated with a fractional assignment, and wherein the fractional assignments sum to one.

[0522] 62. The system of any one of items 47-61, wherein training the first probability model or the second probability model comprises:

[0523] A set of parameters is determined for the probability model, the set of parameters maximizing the total log-likelihood of the first plurality of reference sequence reads or the second plurality of reference sequence reads derived from subjects associated with the first disease state or the second disease state, the disease state being associated with the probability model.

[0524] 63. The system of any one of items 47-62, wherein the memory stores other computer instructions that, when executed by the computer processor, cause the processor to perform steps including:

[0525] For each of the multiple windows:

[0526] selecting a plurality of the first plurality of reference sequence reads derived from the window, and training the first probability model for the window using the sequence reads derived from the window; and

[0527] A plurality of the second plurality of reference sequence reads derived from the windows are selected, and the probability model for each window is trained using the sequence reads.

[0528] 64. The system of item 63, wherein the memory stores other computer instructions that, when executed by the computer processor, cause the processor to perform the following steps, for each of the plurality of windows:

[0529] selecting a subset of a plurality of said training sequence reads derived from said window; and

[0530] Comparing the first probability value and the second probability value for each sequence read of the subset identifies the one or more features.

[0531] 65. The system of claim 63, wherein each of the windows is separated by at least a threshold number of base pairs between CpG sites.

[0532] 66. The system of any one of items 63-65, wherein each of the multiple windows comprises about 200 base pairs (bp) to about 10 kilobase pairs (kbp).

[0533] 67. The system of any of items 47-66, wherein the one or more features of the plurality of training sequence reads include a count of outlier sequence reads wherein the first probability value is greater than the second probability value.

[0534] 68. The system of claim 67, wherein the one or more features comprise binary counts.

[0535] 69. The system of any of items 47-68, wherein the one or more features include a total count of outlier sequence reads.

[0536] 70. The system of any of items 47-69, wherein the one or more features include a total count of anonymously methylated sequence reads.

[0537] 71. A system as described in any of items 47-70, wherein the one or more features include counts of fragments comprising one or more specific methylation patterns.

[0538] 72. A system as described in any of items 47-71, wherein the one or more features are identified using the output of a discriminative classifier trained on a single genomic region.

[0539] 73. The system of claim 72, wherein the discriminative classifier is a multilayer perceptron or a convolutional neural network model.

[0540] 74. A system as described in any of items 47-73, wherein comparing the first probability value and the second probability value includes determining a ratio of the first probability value and the second probability value, and wherein the one or more features include a sequence read count of sequence reads that exceeds a ratio threshold.

[0541] 75. A system as described in any of items 47-74, wherein the first probability value or the second probability value is a log-likelihood value.

[0542] 76. The system of any one of items 47-75, wherein identifying the one or more characteristics comprises:

[0543] For each sequence read of the plurality of training sequence reads:

[0544] determining a log-likelihood ratio of the first probability value to the second probability value; and

[0545] For one or more thresholds, a count of reads of the sequence having a log-likelihood ratio exceeding the threshold is determined.

[0546] 77. The system of any one of items 47-76, wherein the memory stores other computer instructions that, when executed by the computer processor, cause the processor to perform steps including:

[0547] For each of the one or more features, an indicator is determined that distinguishes the feature from the first disease state and the second disease state.

[0548] 78. The system of claim 77, wherein determining the indicator for each of the one or more characteristics comprises:

[0549] Mutual information between the feature and the probability of the first disease state and the second disease state being present is determined.

[0550] 79. The system of item 78, wherein the memory stores other computer instructions, which when executed by the computer processor cause the processor to perform the steps comprising:

[0551] The one or more features used to train the classifier are filtered by ranking the features based on the metric.

[0552] 80. The system of any one of items 47-79, wherein the system further comprises training a classifier from the one or more features, wherein the classifier is trained to predict one or more disease states for multiple sequence reads of a test sample from a test subject, wherein the one or more disease states include the presence or absence of a disease, the type of disease, and / or the tissue of origin of the disease.

[0553] 81. The system of claim 80, wherein the classifier is a multilayer perceptron model.

[0554] 82. The system of claim 80, wherein the classifier is a logistic regression, a support vector machine, a multinomial logistic regression, a multilayer perceptron, a random forest, or a neural network model classifier.

[0555] 83. The system of claim 80, wherein the classifier is generated using L1 or L2 regularized logistic regression.

[0556] 84. The system of item 80, wherein the memory stores other computer instructions that, when executed by the computer processor, cause the processor to perform the steps comprising:

[0557] determining a probability vector for the test sample; and

[0558] A label for the test sample is determined based on the probability vector.

[0559] 85. The system of item 80, wherein the memory stores other computer instructions, which when executed by the computer processor cause the processor to perform the steps comprising:

[0560] The accuracy of the classifier is determined using a confusion matrix that includes information describing the success rate of the classifier in identifying each of the plurality of disease states.

[0561] 86. The system of any one of items 47-85, wherein the first reference sample or the second reference sample is a cell-free nucleic acid sample or a tissue nucleic acid sample from a subject with a known disease state.

[0562] 87. The system of claim 86, wherein the known disease state is the presence or absence of the disease, the disease type, or the tissue of origin of the disease.

[0563] 88. A system as described in any of items 47-87, wherein the training sample comprises a cell-free nucleic acid sample or a tissue sample.

[0564] 89. The system of claim 80, wherein the test sample comprises a cell-free nucleic acid sample.

[0565] 90. The system of item 80, wherein the first plurality of reference sequence reads, the second plurality of reference sequence reads, the plurality of training sequence reads, or the plurality of sequence reads from the test sample are generated by methylation sequencing.

[0566] 91. The system of claim 90, wherein the methylation sequencing comprises whole genome bisulfite sequencing.

[0567] 92. The system of item 91, wherein the methylation sequencing comprises targeted sequencing.

[0568] 93. A non-transitory computer readable medium comprising instructions which, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0569] accessing a first plurality of reference sequence reads from a first reference sample from a subject having a first disease state;

[0570] accessing a second plurality of reference sequence reads from a second reference sample from a subject having a second disease state,

[0571] training a first probability model using the first plurality of reference sequence reads, the first probability model being associated with the first disease state;

[0572] training a second probability model using the second plurality of reference sequence reads, the second probability model being associated with the second disease state;

[0573] Accessing a plurality of training sequence reads from a training sample, and for each sequence read of the plurality of training sequence reads:

[0574] applying the sequence reads to the first probability model to determine a first probability value, the first probability value being a probability of the sequence reads being derived from a sample associated with the first disease state, and

[0575] applying the sequence reads to the second probability model to determine a second probability value, the second probability value being a probability of the sequence reads being derived from a sample associated with the second disease state; and

[0576] One or more features are identified by comparing the first probability value and the second probability value for each sequence read.

[0577] 94. The non-transitory computer-readable medium of claim 93, wherein the first disease state is cancer and the second disease state is non-cancer.

[0578] 95. The non-transitory computer-readable medium of claim 93, wherein the first disease state is a first type of cancer and the second disease state is a second type of cancer, and wherein the first type of cancer and the second type of cancer are different.

[0579] 96. The non-transitory computer-readable medium of item 93, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0580] accessing a plurality of reference sequence reads from a third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference sample, each of said third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference samples having a different disease state, and wherein each of said different disease state is a different type of cancer; and

[0581] The third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models are trained using the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth plurality of reference sequence reads, wherein each of the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models is respectively associated with a different type of cancer.

[0582] 97. A non-transitory computer-readable medium as described in any of items 94-96, wherein the cancer or type of cancer is selected from the following: breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, renal pelvic ureteral urothelial carcinoma, renal cancer other than urothelial carcinoma, prostate cancer, anorectal cancer, colorectal cancer, esophageal squamous cell carcinoma, esophageal cancer other than squamous cell carcinoma, gastric cancer, hepatobiliary carcinoma originating from hepatocytes, hepatobiliary carcinoma originating from cells other than hepatocytes, pancreatic cancer, head and neck cancer associated with human papillomavirus, head and neck cancer not associated with human papillomavirus, lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer and lung cancer other than adenocarcinoma or small cell lung cancer, neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma and leukemia.

[0583] 98. The non-transitory computer-readable medium of item 97, wherein the cancer type is additionally selected from the group consisting of brain cancer, vulvar cancer, vaginal cancer, testicular cancer, pleural mesothelioma, peritoneal mesothelioma, and gallbladder cancer.

[0584] 99. The non-transitory computer-readable medium of claim 93, wherein the first disease state comprises a first source tissue and the second disease state comprises a second source tissue.

[0585] 100. The non-transitory computer-readable medium of item 99, wherein the first source tissue or the second source tissue is selected from the group consisting of breast tissue, thyroid tissue, lung tissue, bladder tissue, cervical tissue, small intestine tissue, colorectal tissue, esophageal tissue, gastric tissue, tonsil tissue, liver tissue, ovarian tissue, fallopian tube tissue, pancreatic tissue, prostate tissue, kidney tissue, and uterine tissue.

[0586] 101. The non-transitory computer-readable medium of item 100, wherein the first source tissue or the second source tissue is additionally selected from the group consisting of brain tissue and cells, endocrine tissue and cells, vascular endothelial tissue and cells, head and neck tissue and cells, exocrine pancreatic tissue and cells, endocrine pancreatic tissue and cells, lymphoid tissue and cells, mesenchymal tissue and cells, myeloid tissue and cells, pleural tissue and cells, muscle tissue and cells, bone marrow tissue and cells, adipose tissue and cells, and gallbladder tissue and cells.

[0587] 102. The non-transitory computer-readable medium of any one of items 93-101, wherein the first probability model or the second probability model is a constant model, a binomial model, an independent site model, a neural network model, or a Markov model.

[0588] 103. The non-transitory computer-readable medium of any one of items 93-102, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0589] A methylation rate for each of a plurality of CpG sites in the first plurality of reference sequence reads or the second plurality of reference sequence reads is determined, wherein the first probability model or the second probability model is parameterized by a product of the methylation rates.

[0590] 104. The non-transitory computer-readable medium of any one of items 93-103, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0591] For each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, determining whether the sequence read is hypomethylated or hypermethylated is performed by determining whether at least a threshold number of CpG sites having at least a threshold percentage of the CpG sites are unmethylated or methylated, respectively.

[0592] 105. The non-transitory computer-readable medium of any one of items 93-104, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0593] determining, for each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, whether the sequence read is aberrantly methylated; and

[0594] The first plurality of reference sequence reads is filtered using p-value filtering by removing sequence reads from the first plurality of reference sequence reads having a p-value below a threshold.

[0595] 106. The non-transitory computer-readable medium of item 102, wherein the first probability model or the second probability model is parameterized by the sum of a plurality of mixed components, each component being related to the product of the methylation rates.

[0596] 107. The non-transitory computer-readable medium of item 106, wherein each of the plurality of mixture components is associated with a fractional assignment, and wherein the fractional assignments sum to one.

[0597] 108. The non-transitory computer-readable medium of any one of items 93-107, wherein training the first probability model or the second probability model comprises:

[0598] A set of parameters is determined for the probability model, the set of parameters maximizing the total log-likelihood of the first plurality of reference sequence reads or the second plurality of reference sequence reads derived from subjects associated with the first disease state or the second disease state, the disease state being associated with the probability model.

[0599] 109. The non-transitory computer-readable medium of any one of items 93-108, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0600] For each of the multiple windows:

[0601] selecting a plurality of the first plurality of reference sequence reads derived from the window, and training the first probability model for the window using the sequence reads derived from the window; and

[0602] A plurality of the second plurality of reference sequence reads derived from the windows are selected, and the probability model for each window is trained using the sequence reads.

[0603] 110. The non-transitory computer-readable medium of item 109, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising, for each of the plurality of windows:

[0604] selecting a subset of a plurality of said training sequence reads derived from said window; and

[0605] Comparing the first probability value and the second probability value for each sequence read of the subset identifies the one or more features.

[0606] 111. The non-transitory computer-readable medium of claim 109, wherein each of the windows is separated by at least a threshold number of base pairs between CpG sites.

[0607] 112. The non-transitory computer-readable medium of any one of items 109-111, wherein each of the plurality of windows comprises approximately 200 base pairs (bp) to approximately 10 kilobase pairs (kbp).

[0608] 113. The non-transitory computer-readable medium of any one of items 93-112, wherein the one or more features include a count of outlier sequence reads of the plurality of training sequence reads wherein the first probability value is greater than the second probability value.

[0609] 114. The non-transitory computer-readable medium of item 113, wherein the one or more features comprise binary counts.

[0610] 115. The non-transitory computer-readable medium of any of items 93-114, wherein the one or more features include a total count of outlier sequence reads.

[0611] 116. The non-transitory computer-readable medium of any one of items 93-115, wherein the one or more features include a total count of anonymously methylated sequence reads.

[0612] 117. The non-transitory computer-readable medium of any of items 93-116, wherein the one or more features include counts of fragments comprising one or more specific methylation patterns.

[0613] 118. The non-transitory computer-readable medium of any of items 93-117, wherein the one or more features are identified using the output of a discriminative classifier trained on a single genomic region.

[0614] 119. The non-transitory computer-readable medium of item 113, wherein the discriminative classifier is a multilayer perceptron or a convolutional neural network model.

[0615] 120. A non-transitory computer-readable medium as described in any of items 93-119, wherein comparing the first probability value and the second probability value comprises determining a ratio of the first probability value and the second probability value, and wherein the one or more features comprise a sequence read count of sequence reads that exceeds a ratio threshold.

[0616] 121. The non-transitory computer-readable medium of any one of items 93-120, wherein the first probability value or the second probability value is a log-likelihood value.

[0617] 122. The non-transitory computer-readable medium of any one of items 93-121, wherein identifying the one or more characteristics comprises:

[0618] For each sequence read of the plurality of training sequence reads:

[0619] determining a log-likelihood ratio of the first probability value to the second probability value; and

[0620] For one or more thresholds, a count of reads of the sequence having a log-likelihood ratio exceeding the threshold is determined.

[0621] 123. The non-transitory computer-readable medium of any one of items 93-122, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0622] For each of the one or more features, an indicator is determined that distinguishes the feature from the first disease state and the second disease state.

[0623] 124. The non-transitory computer-readable medium of item 123, wherein determining the indicator for each of the one or more characteristics comprises:

[0624] Mutual information between the features and probabilities of the first disease state and the second disease state being present is determined.

[0625] 125. The non-transitory computer-readable medium of item 124, comprising other instructions, which when executed by one or more processors cause the one or more processors to perform steps comprising:

[0626] The one or more features used to train the classifier are filtered by ranking the features based on the metric.

[0627] 126. The non-transitory computer-readable medium of any one of items 93-125, wherein the instructions further include training a classifier from the one or more features, the classifier being trained to predict one or more disease states for multiple sequence reads of a test sample from a test subject, wherein the one or more disease states include the presence or absence of a disease, the type of disease, and / or the tissue of origin of the disease.

[0628] 127. The non-transitory computer-readable medium of item 126, wherein the classifier is a multilayer perceptron model.

[0629] 128. The non-transitory computer-readable medium of item 126, wherein the classifier is a logistic regression, a support vector machine, a multinomial logistic regression, a multilayer perceptron, a random forest, or a neural network model classifier.

[0630] 129. The non-transitory computer-readable medium of claim 126, wherein the classifier is generated using L1 or L2 regularized logistic regression.

[0631] 130. The non-transitory computer-readable medium of item 126, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0632] determining a probability vector for the test sample; and

[0633] A label for the test sample is determined based on the probability vector.

[0634] 131. The non-transitory computer-readable medium of item 126, comprising other instructions that, when executed by one or more processors, cause the one or more processors to perform steps comprising:

[0635] The accuracy of the classifier is determined using a confusion matrix that includes information describing the success rate of the classifier in identifying each of the plurality of disease states.

[0636] 132. The non-transitory computer-readable medium of any one of items 93-131, wherein the first reference sample or the second reference sample is a cell-free nucleic acid sample or a tissue nucleic acid sample from a subject with a known disease state.

[0637] 133. The non-transitory computer-readable medium of claim 132, wherein the known disease state is the presence or absence of the disease, the type of disease, or the tissue of origin of the disease.

[0638] 134. The non-transitory computer-readable medium of any one of items 93-133, wherein the training sample comprises a cell-free nucleic acid sample or a tissue sample.

[0639] 135. The non-transitory computer-readable medium of claim 126, wherein the test sample comprises a cell-free nucleic acid sample.

[0640] 136. The non-transitory computer-readable medium of item 126, wherein the first plurality of reference sequence reads, the second plurality of reference sequence reads, the plurality of training sequence reads, or the plurality of sequence reads from the test sample are generated by methylation sequencing.

[0641] 137. The non-transitory computer-readable medium of claim 136, wherein the methylation sequencing comprises whole genome bisulfite sequencing.

[0642] 138. The non-transitory computer-readable medium of item 136, wherein the methylation sequencing comprises targeted sequencing.

[0643] 139. A method comprising:

[0644] generating a first plurality of reference sequence reads from a reference sample having one of a plurality of disease states, each disease state being associated with a tissue of origin;

[0645] training a plurality of probability models using the first plurality of reference sequence reads, each probability model being associated with a different disease state of the plurality of disease states;

[0646] For each probability model of the plurality of probability models:

[0647] for each of a second plurality of sequence reads, applying the probability model to the sequence read to determine a value based at least on a first probability that the sequence read originated from a sample associated with the disease state associated with the probability model; and

[0648] identifying a feature by determining a count of the second plurality of sequence reads having a value exceeding a threshold; and

[0649] The features are used to generate a classifier that is trained to predict a disease state or a tissue of origin associated with one of a plurality of disease states for an input sequence read of a test sample from a test subject.

[0650] 140. The method described in claim 139, wherein the multiple disease states include at least two, at least three, at least four, at least five or at least ten different disease states.

[0651] 141. The method of item 139 or item 140, further comprising:

[0652] A methylation rate is determined for each of a plurality of CpG sites in the first plurality of reference sequence reads, wherein each of a plurality of probability models is parameterized by a product of the methylation rates.

[0653] 142. The method of item 139 or item 140, further comprising:

[0654] determining, for each sequence read of the first plurality of reference sequence reads or the second plurality of sequence reads, whether the sequence read is aberrantly methylated; and

[0655] The first plurality of reference sequence reads or the second plurality of sequence reads are filtered using p-value filtering by removing sequence reads having a p-value below a threshold from the first plurality of reference sequence reads or the second plurality of sequence reads.

[0656] 143. The method described in item 141, wherein each of the multiple probability models is parameterized by the sum of multiple mixed components, each component being related to the product of the methylation rates.

[0657] 144. The method described in item 143, wherein each of the multiple mixing components is associated with a fractional assignment, and wherein the sum of the fractional assignments is one.

[0658] 145. The method of any one of items 139-144, wherein training the plurality of probability models comprises:

[0659] A set of parameters is determined for a probability model of the plurality of probability models, the set of parameters maximizing a total log-likelihood of the first plurality of reference sequence reads derived from subjects associated with the disease state associated with the probability model.

[0660] 146. The method of any one of items 139-145, further comprising:

[0661] determining a probability vector for the test sample; and

[0662] A label for the test sample is determined based on the probability vector.

[0663] 147. The method of any one of items 139-146, wherein determining the value comprises:

[0664] determining the first probability of the sequence reads derived from the sample being associated with the disease state associated with the probability model, wherein the disease state is associated with the presence of cancer or a type of cancer;

[0665] determining a second probability that the sequence read is derived from a healthy sample; and

[0666] A log-likelihood ratio of the first probability to the second probability is determined.

[0667] 148. The method of claim 147, wherein identifying the characteristic comprises:

[0668] For a plurality of thresholds, a count of the second plurality of sequence reads having a log-likelihood ratio exceeding the threshold is determined.

[0669] 149. The method of any one of items 139-148, further comprising:

[0670] For each of the features, an indicator is determined that distinguishes the feature between a first disease state and a second disease state of the plurality of disease states.

[0671] 150. The method of claim 149, wherein the indicator for determining the characteristic comprises:

[0672] Mutual information between the feature and the probability of the first disease state and the second disease state being present is determined.

[0673] 151. The method described in item 149, wherein the first probability of the first disease state is equal to the second probability of the second disease state.

[0674] 152. The method of item 149, further comprising:

[0675] The features used for training the classifier are filtered by ranking based on the indicators of the features.

[0676] 153. The method of any one of items 139-152, further comprising:

[0677] The accuracy of the classifier is determined using a confusion matrix that includes information describing the success rate of the classifier in identifying each of the plurality of disease states.

[0678] 154. The method of any one of items 139-153, further comprising:

[0679] A plurality of blocks of a reference genome are determined, each of the blocks being separated by at least a threshold number of base pairs between CpG sites, wherein the first plurality of reference sequence reads are generated using the plurality of blocks.

[0680] 155. The method of any one of items 139-154, wherein the counts of the second plurality of sequence reads having the values ​​exceeding the threshold are determined for a plurality of CpG sites.

[0681] 156. The method described in any one of items 139-155, wherein the reference sample comprises one or more of: a cell-free nucleic acid sample and a tissue sample.

[0682] 157. The method of any one of items 139-156, wherein the multiple disease states include one or more of: cancer type, disease type, and health state.

[0683] 158. The method of any one of items 139-157, wherein the classifier is a logistic regression, multinomial logistic regression, multilayer perceptron, support vector machine, random forest or neural network model classifier.

[0684] 159. The method of claim 158, wherein the classifier is generated using L1 or L2 regularized logistic regression.

[0685] 160. The method of any one of items 139-159, further comprising:

[0686] The feature is binarized to indicate the presence or absence of one of the plurality of disease states, wherein the classifier is generated using the binarized feature.

[0687] 161. The method of item 160, wherein the binarized features each have a value of 0 or 1.

[0688] 162. The method of any one of items 139-161, further comprising:

[0689] determining, for the reference sample, a measure of uncertainty in positioning; and

[0690] At least one prediction of the classifier is labeled as an uncertain tissue of origin based on the metric.

[0691] 163. The method of any one of items 139-162, wherein the classifier is a multilayer perceptron model.

[0692] 164. A system comprising a computer processor and a memory storing computer program instructions which, when executed by the computer processor, cause the processor to implement the method of any one of items 139-163.

[0693] 165. A non-transitory computer-readable medium storing one or more programs, wherein the one or more programs contain instructions that, when executed by an electronic device containing a memory, cause the device to implement the method described in any one of items 139-163.

[0694] 166. A method comprising:

[0695] generating a plurality of sequence reads from one or more biological samples;

[0696] For each of multiple positions on a chromosome:

[0697] determining, using the plurality of sequence reads, a count of nucleic acid fragments of the one or more biological samples that are within the locations and that have at least a threshold similarity to fragments associated with the disease state;

[0698] training a machine learning model using the counts of the plurality of locations as features; and

[0699] The trained machine learning model is used to determine the probability of a test sample having a disease state.

[0700] 167. The method of item 166, further comprising:

[0701] The feature is binarized in each of the plurality of locations to indicate the presence or absence of one of the plurality of disease states, wherein a count of at least one nucleic ...

Claims

1. A method for identifying a cancer signal in a test sample analysis sequence read to generate a signature, comprising: obtaining a plurality of sequence reads from a test sample; applying a first probability model associated with a first disease state to each sequence read of the test sample to output a first probability that the sequence read is associated with the first disease state, wherein the first probability model is trained using a first plurality of reference sequence reads derived from one or more samples associated with the first disease state; applying a second probability model associated with a second disease state to each sequence read of the test sample to output a second probability that the sequence read is associated with the second disease state, wherein the second probability model is trained using a second plurality of reference sequence reads derived from one or more samples associated with the second disease state; Determining a feature vector of the test sample based on the first probability and the second probability; and The trained cancer classifier is applied to the feature vector of the test sample to predict the likelihood that the test sample has a cancer signature.

2. The method of claim 1, wherein the first disease state is cancer and the second disease state is non-cancer.

3. The method of claim 1, wherein the first disease state is a first type of cancer and the second disease state is a second type of cancer, and wherein the first type of cancer and the second type of cancer are different.

4. The method of claim 1, wherein the method further comprises: obtaining a plurality of reference sequence reads from a third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference sample, each of said third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth reference samples having a different disease state, and wherein each of said different disease state is a different type of cancer; and The third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models are trained using the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth plurality of reference sequence reads, wherein each of the third, fourth, fifth, sixth, seventh, eighth, ninth and / or tenth probability models is respectively associated with a different type of cancer.

5. The method of any one of claims 2-4, wherein the cancer or type of cancer is selected from the group consisting of: Breast cancer, uterine cancer, cervical cancer, ovarian cancer, bladder cancer, renal pelvis and ureteral urothelial carcinoma, renal cancer other than urothelial carcinoma, prostate cancer, anorectal cancer, colorectal cancer, esophageal squamous cell carcinoma, esophageal cancer other than squamous cell carcinoma, gastric cancer, hepatobiliary duct cancer arising from hepatocytes, hepatobiliary duct cancer arising from cells other than hepatocytes, pancreatic cancer, head and neck cancer associated with human papillomavirus, head and neck cancer not associated with human papillomavirus, lung adenocarcinoma, small cell lung cancer, squamous cell lung cancer and lung cancer other than adenocarcinoma or small cell lung cancer, neuroendocrine cancer, melanoma, thyroid cancer, sarcoma, multiple myeloma, lymphoma and leukemia.

6. The method of claim 5, wherein the cancer type is additionally selected from the group consisting of: Brain cancer, vulvar cancer, vaginal cancer, testicular cancer, pleural mesothelioma, peritoneal mesothelioma and gallbladder cancer.

7. The method of claim 1, wherein the first disease state comprises a first source tissue and the second disease state comprises a second source tissue.

8. The method of claim 7, wherein the first source tissue or the second source tissue is selected from the group consisting of: Breast tissue, thyroid tissue, lung tissue, bladder tissue, cervical tissue, small intestine tissue, colorectal tissue, esophageal tissue, stomach tissue, tonsil tissue, liver tissue, ovarian tissue, fallopian tube tissue, pancreatic tissue, prostate tissue, kidney tissue and uterine tissue.

9. The method of claim 8, wherein the first source tissue or the second source tissue is additionally selected from the group consisting of: Brain tissue and cells, endocrine tissue and cells, vascular endothelial tissue and cells, head and neck tissue and cells, exocrine pancreatic tissue and cells, endocrine pancreatic tissue and cells, lymphoid tissue and cells, mesenchymal tissue and cells, myeloid tissue and cells, pleural tissue and cells, muscle tissue and cells, bone marrow tissue and cells, adipose tissue and cells, gallbladder tissue and cells.

10. The method of claim 1, wherein the first probability model or the second probability model is a constant model, a binomial model, an independent site model, a neural network model, or a Markov model.

11. The method of claim 1, further comprising: A methylation rate for each of a plurality of CpG sites in the first plurality of reference sequence reads or the second plurality of reference sequence reads is determined, wherein the first probability model or the second probability model is parameterized by a product of the methylation rates.

12. The method of claim 11, wherein the first probability model or the second probability model is parameterized by a sum of multiple mixture components, each component being associated with a product of the methylation rates.

13. The method of claim 12, wherein each of the plurality of mixture components is associated with a fractional assignment, and wherein the fractional assignments sum to one.

14. The method of claim 1, further comprising: For each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads from the test sample, determining whether the sequence read is hypomethylated or hypermethylated, respectively, by determining whether at least a threshold number of CpG sites having at least a threshold percentage of the CpG sites are unmethylated or methylated.

15. The method of claim 1, further comprising: determining, for each sequence read in the first plurality of reference sequence reads, the second plurality of sequence reads, or the plurality of training sequence reads, whether the sequence read is aberrantly methylated; and The first plurality of reference sequence reads is filtered using p-value filtering by removing sequence reads from the first plurality of reference sequence reads having a p-value below a threshold.

16. The method of claim 1, wherein training the first probability model comprises: A set of parameters is determined for the probability model, the set of parameters maximizing the total log-likelihood of the first plurality of reference sequence reads derived from subjects associated with the first disease state, the disease state being associated with the probability model.

17. The method of claim 1, wherein training the first probability model and the second probability model comprises: For each of the multiple windows: selecting a first set of sequence reads of the first plurality of reference sequence reads within the window, and training the first probability model for the window using the first set of sequence reads; and A second set of sequence reads derived from the second plurality of reference sequence reads for the window is selected, and the second set of sequence reads is used to train the second probability model for the window.

18. The method of claim 17, wherein at least one window is separated by at least a threshold number of base pairs between CpG sites.

19. The method of any one of claims 17-18, wherein at least one window comprises 200 base pairs (bp) to 10 kilobase pairs (kbp).

20. The method of claim 1, wherein determining the feature variable comprises determining a feature comprising a count of outlier sequence reads of the plurality of sequence reads wherein the first probability value is greater than the second probability value.

21. The method of claim 1, wherein determining the feature variable comprises determining a feature comprising a total count of anonymously methylated sequence reads.

22. The method of claim 1, wherein determining the feature variable comprises determining a feature comprising counts of fragments comprising one or more specific methylation patterns.

23. The method of claim 1, wherein determining the feature variable comprises applying a discriminative classifier trained on a single genomic region to output one or more features based on the first probability and the second probability of the plurality of sequence reads of the test sample.

24. The method of claim 23, wherein the discriminative classifier is a multilayer perceptron or a convolutional neural network model.

25. The method of claim 1, wherein determining the feature variable comprises: comparing the first probability and the second probability comprises determining a ratio of the first probability to the second probability, and One or more features are determined based on sequence read counts having a ratio above a threshold ratio.

26. The method of claim 1, wherein the first probability or the second probability is a log-likelihood value.

27. The method of claim 1, further comprising: determining, for each feature, a measure of the feature that distinguishes the first disease state from the second disease state based on the first plurality of reference sequence reads and the second plurality of reference sequence reads; and A set of features is identified for use in determining the feature vector based on an indicator of features that distinguish the first disease state from the second disease state.

28. The method of claim 1, further comprising: The cancer classifier is trained using feature vectors from a plurality of training samples, wherein a first cohort of the training samples are obtained from individuals diagnosed with cancer and a second cohort of the training samples are obtained from individuals not diagnosed with cancer.

29. The method of claim 28, wherein the classifier is a logistic regression, support vector machine, multinomial logistic regression, multilayer perceptron, random forest, or neural network model classifier.

30. The method of claim 1, wherein obtaining the plurality of sequence reads for the test sample comprises performing targeted methylation sequencing to generate methylation sequence reads indicative of a methylation state of each of a plurality of methylation sites.

31. A non-transitory computer-readable storage medium storing instructions for identifying a cancer signature in a test sample, the instructions, when executed by a processor, causing the processor to perform operations comprising: obtaining a plurality of sequence reads from a test sample; applying a first probability model associated with a first disease state to each sequence read of the test sample to output a first probability that the sequence read is associated with the first disease state, wherein the first probability model is trained using a first plurality of reference sequence reads derived from one or more samples associated with the first disease state; applying a second probability model associated with a second disease state to each sequence read of the test sample to output a second probability that the sequence read is associated with the second disease state, wherein the second probability model is trained using a second plurality of reference sequence reads derived from one or more samples associated with the second disease state; Determining a feature vector of the test sample based on the first probability and the second probability; and The trained cancer classifier is applied to the feature vector of the test sample to predict the likelihood that the test sample has a cancer signature.

32. A system comprising: A sequencing device configured to: performing sequencing to generate a plurality of sequence reads corresponding to a plurality of nucleic acid fragments in the test sample; and A computing device configured to: receiving the plurality of sequence reads from the sequencing device; applying a first probability model associated with a first disease state to each sequence read of the test sample to output a first probability that the sequence read is associated with the first disease state, wherein the first probability model is trained using a first plurality of reference sequence reads derived from one or more samples associated with the first disease state; applying a second probability model associated with a second disease state to each sequence read of the test sample to output a second probability that the sequence read is associated with the second disease state, wherein the second probability model is trained using a second plurality of reference sequence reads derived from one or more samples associated with the second disease state; Determining a feature vector of the test sample based on the first probability and the second probability; and The trained cancer classifier is applied to the feature vector of the test sample to predict the likelihood that the test sample has a cancer signature.

Images