Detecting loss of heterozygosity in HLA alleles using machine learning models

Abstract

Provided is a method for detecting heterozygosity loss in HLA alleles.The method can include accessing a machine learning model, and the machine learning model is trained with a training data set, the training data set includes at least one training data set, the training data set includes an adjusted B allele frequency corresponding to the ratio between a first B allele frequency of the heterozygous allele in a tumor sample corresponding to a genomic region and a second B allele frequency of the heterozygous allele in the genomic region and associated with one or more control samples.The method can also include using the machine learning model to generate a result corresponding to the probability of whether or not there is heterozygosity loss in the HLA allele identified in a biological sample of a particular subject by processing sequence data with the machine learning model.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims priority to U.S. Provisional Application No. 63 / 178,151, entitled "Detecting Loss Of Heterozygosity In HLA Alleles Using Machine Learning Models," filed April 22, 2021, which is incorporated by reference in its entirety for all purposes. [Background technology]

[0002] In general, immune checkpoints refer to a set of inhibitory pathways built into the immune system that regulate the duration and amplitude of physiological immune responses. When activated, immune checkpoint molecules (such as PD-1) suppress the immune system to prevent it from indiscriminately attacking cells. Although immune checkpoints are generally effective, tumor cells can manipulate these mechanisms to prevent the immune system from eliminating tumor cells.

[0003] Immune checkpoint blockade therapy is a recent treatment to combat the mechanisms of tumor cells. Immune checkpoint blockade therapy uses drugs such as immune checkpoint inhibitors to activate the immune system to recognize and eliminate cancerous cells. Immune checkpoint blockade therapy allows the immune system to properly recognize and eliminate tumor cells that present neoantigens via major histocompatibility complex (MHC). Despite such early success, a high percentage of subjects do not respond to these therapies. It has proven more difficult than initially expected to unravel the causes of such immune checkpoint blockade resistance, which is due to complex intrinsic and extrinsic tumor mechanisms of tumor cells that resist and evade immune checkpoint blockade therapy.

[0004] One of the mechanisms that causes immune checkpoint blockade resistance can be loss of heterozygosity of human leukocyte antigen (HLA) genes. Neo-antigens corresponding to mutated genes of tumor cells can bind to HLA proteins encoded by specific HLA alleles and be presented on the cell surface. When the presented neo-antigens are detected, the immune system can respond by deploying T cells that identify and eliminate tumor cells by detecting the presented neo-antigens. Thus, the effectiveness of the immune system can depend on whether neo-antigens are presented on the tumor cell surface. Conversely, if the presentation of neo-antigens is prevented, T cells may not be able to detect the corresponding tumor cells.

[0005] Various studies suggest that tumor cells often have heterozygosity loss in HLA genes, and the corresponding HLA proteins of the deleted HLA alleles are not available to present neoantigens on the tumor cell surface. For example, each human subject has six different HLA alleles that can present a diverse set of antigens to the immune system. Germline sequence diversity of HLA alleles may affect tumor evolution by mediating neoantigen presentation to the immune system. The effect of this HLA sequence diversity seems to be more pronounced in the presence of immune checkpoint blockade therapy. As tumor cells mutate, somatic loss of heterozygosity occurs within the HLA allele region, which leads to a reduction in HLA sequence diversity. Such heterozygosity loss of HLA alleles is increasingly recognized as a cause of immune checkpoint blockade resistance by tumor cells.

[0006] Thus, detecting loss of heterozygosity of HLA alleles from sequencing data can be beneficial in predicting immune checkpoint blockade resistance for a given subject and developing corresponding therapy. However, conventional techniques may be insufficient in accurately detecting loss of heterozygosity of HLA alleles. For example, conventional techniques for detecting HLA loss of heterozygosity may include performing genome-wide matching to detect copy number. In this technique, copy number reduction around an HLA gene may indicate its loss of heterozygosity. However, this conventional technique may be unreliable in detecting loss of HLA heterozygosity from sequencing data for at least the following reasons. First, the polymorphic nature of the mutated gene leads to poor alignment of the corresponding sequence reads to the reference genome. Second, the complexity of sequence variation may obscure the specific HLA allele that is deleted, which is crucial information for designing neo-antigen therapy.

[0007] Another conventional technique may involve identifying copy number polymorphisms of HLA genes after aligning sequence data to HLA allele-specific reference sequences. However, most allele-specific alignment techniques that conventional copy number variant algorithms rely on do not consider HLA-specific challenges such as differences in exome probe capture between alleles. In addition, copy number variant algorithms are known to be disadvantageous for biological samples with low tumor purity, and subclonal deletions may be difficult to detect, raising concerns about the sensitivity and accuracy of these algorithms. Thus, despite growing interest, conventional techniques ultimately rely on deletion of adjacent regions surrounding HLA allele regions instead of HLA heterozygosity loss, rather than developing algorithms specific for HLA heterozygosity loss. In view of the above, accurate detection of HLA heterozygosity loss is difficult.

[0008] Moreover, the performance verification of HLA heterozygosity loss detection algorithm is a further challenge in the field.For example, conventional techniques include evaluating the agreement between HLA heterozygosity loss determination and standard CNV algorithm copy number determination in the region adjacent to each HLA gene.Another conventional technique includes designing primers that capture the region surrounding HLA genes, which are applied to PCR to reveal the copy number loss of HLA alleles in subjects.However, none of these approaches verify what specific HLA alleles may be lost, nor address the accuracy of determination for samples with low tumor purity or samples with HLA heterozygosity subclonal loss. Summary of the Invention

[0009] In some embodiments, a method for detecting loss of heterozygosity in an HLA allele is provided. The method includes accessing a trained machine learning model, the machine learning model being trained with a training dataset including one or more training feature sets corresponding to an HLA allele identified in a tumor sample corresponding to a subject of a set of subjects. The first training feature set includes, for a genomic region of the HLA allele, (i) an adjusted B allele frequency corresponding to the ratio between a first B allele frequency of the heterozygous allele in the tumor sample corresponding to the genomic region and a second B allele frequency of the heterozygous allele in the genomic region and associated with one or more control samples, and (ii) a ratio between a first allele-specific coverage of the tumor sample corresponding to the genomic region and a second allele-specific coverage of one or more control samples corresponding to the genomic region. The second training feature set includes, for the HLA allele, an indication of whether at least a portion of the adjacent genomic region surrounding the HLA allele has been deleted.

[0010] The method can also include receiving sequence data corresponding to the biological sample of the particular subject. The method can also include using the machine learning model to generate a result corresponding to the probability of whether loss of heterozygosity is present at the identified HLA allele in the biological sample of the particular subject by processing the sequence data with the machine learning model. The method can also include outputting the result.

[0011] Some embodiments of the present disclosure include a system including one or more data processors. In some embodiments, the system includes a non-transitory computer-readable storage medium that includes instructions that, when executed on one or more data processors, cause one or more data processors to perform part or all of one or more methods and / or part or all of one or more processes disclosed herein. Some embodiments of the present disclosure include a computer program product tangibly embodied in a non-transitory machine-readable storage medium, the computer program product including instructions configured to cause one or more data processors to perform part or all of one or more methods and / or part or all of one or more processes disclosed herein.

[0012] The terms and expressions used are used as terms of description and not as terms of limitation, and the use of such terms and expressions is not intended to exclude any equivalents of the features shown and described or portions thereof, but it is recognized that various modifications are possible within the scope of the invention as claimed. Thus, although the invention as claimed has been specifically disclosed by certain embodiments and optional features, modifications and variations of the concepts disclosed herein may be left to those skilled in the art, and such modifications and variations are considered to be within the scope of the invention as defined by the appended claims.

[0013] The present disclosure is described in conjunction with the accompanying figures. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 shows a schematic for preparing sequence alignment data for detecting loss of heterozygosity in HLA alleles, according to some embodiments. [Diagram 2] FIG. 2 shows a diagram corresponding to an exemplary feature set of a training dataset for training a machine learning model for detecting loss of heterozygosity in HLA alleles, according to some embodiments. [Diagram 3] FIG. 3 shows various features identified from a training dataset for training a machine learning model for detecting loss of heterozygosity in HLA alleles, according to some embodiments. [Figure 4] FIG. 4 shows exemplary data generated by a machine learning model trained to detect loss of heterozygosity in HLA alleles, according to some embodiments. [Diagram 5] FIG. 5 shows an example of a cell line-based limit of detection technique for validating a machine learning model trained to detect loss of heterozygosity in HLA alleles, according to some embodiments. [Figure 6] FIG. 6 shows an exemplary set of results that correspond to the performance levels of the DASH model based on in silico cell line-based limit of detection analysis. [Figure 7] FIG. 7 shows an example of an allele-specific genomic technique for validating a machine learning model trained to detect loss of heterozygosity in HLA alleles, according to some embodiments. [Figure 8] FIG. 8 shows an exemplary set of results that correspond to the performance level of the DASH model based on allele-specific genomic validation using digital PCR. [Figure 9]FIG. 9 shows an overview of the use of immunopeptidomic data to validate machine learning models trained to detect loss of heterozygosity in HLA alleles, according to some embodiments. [Figure 10] FIG. 10 shows an example of quantitative immunopeptidomic data corresponding to the relationship between control samples in which HLA heterozygosity loss is not predicted and samples in which HLA heterozygosity loss is predicted, according to some embodiments. [Figure 11] FIG. 11 shows an example of quantitative immunopeptidomic data corresponding to a control sample in which HLA loss of heterozygosity is not predicted, according to some embodiments. [Figure 12] FIG. 12 shows an example of quantitative immunopeptidomic data corresponding to a sample predicted to have HLA loss of heterozygosity, according to some embodiments. [Figure 13] FIG. 13 shows predicted HLA loss of heterozygosity across tumor types, according to some embodiments. [Figure 14] FIG. 14 shows a set of experimental results 1400 illustrating the relationship between HLA loss of heterozygosity and antigen presentation across various tumor types, according to some embodiments. [Figure 15] FIG. 15 shows experimental data identifying neoantigen expansion in response to immune checkpoint inhibitor therapy in subjects with head and neck squamous cell carcinoma, according to some embodiments. [Figure 16] FIG. 16 shows additional experimental data corresponding to HLA loss of heterozygosity in immunotherapy treated tumors, according to some embodiments. [Figure 17] FIG. 17 includes a flow chart illustrating an example of a method for predicting loss of heterozygosity in an HLA allele, according to certain embodiments. [Figure 18] FIG. 18 illustrates an example of a computer system for implementing some embodiments disclosed herein. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] I. Overview As mentioned above, accurate detection of HLA heterozygosity loss can significantly improve the accuracy and efficacy of cancer immunotherapy, including immune checkpoint blockade therapy.Although allele-specific alignment techniques can be an improvement over other genome matching techniques, the variable exome capture across alleles and the relatively short sequence of HLA genes introduce additional difficulties in identifying copy number changes.In addition, traditional techniques cannot accurately and comprehensively determine the detection limit, sensitivity and specificity of HLA heterozygosity loss detection algorithms.

[0016] To address at least the deficiencies of the conventional systems described above, the techniques of the present invention can be used. The techniques of the present invention can detect heterozygosity loss in HLA alleles using a machine learning approach. A machine learning model for identifying allele-specific HLA deletions ("DASH model") can be accessed. In particular, the DASH model can be trained with training data for a subject of a subject set, including the following features: (1) allele-specific features, (2) subject-specific features, and (3) whole exome features. The allele-specific features can include, for a genomic region of an HLA allele, an adjusted B allele frequency that corresponds to the ratio between a first B allele frequency of a heterozygous allele in a tumor sample corresponding to the genomic region and a second B allele frequency of a heterozygous allele in the genomic region and associated with one or more control samples, and a ratio between a first allele-specific coverage of the tumor sample corresponding to the genomic region and a second allele-specific coverage of one or more control samples corresponding to the genomic region. By using the above training features together, the DASH model can be trained to accurately detect heterozygosity loss in HLA alleles. In some cases, the allele-specific characteristic corresponds to a genomic region of an HLA allele that has been identified as having a somatic mutation.

[0017] As referred to herein, the B allele frequency is a normalized measure of the allele intensity ratio of two alleles (A and B), where a B allele frequency of 1 or 0 indicates the complete absence of one of the two alleles (e.g., AA or BB), and a B allele frequency of 0.5 indicates the equal presence of both alleles (e.g., AB). For example, a first B allele frequency can indicate the allele intensity ratio between HLA-B*46:01:01 and HLA-B*13:01:01 corresponding to a normal biological sample for a given genomic location. A second B allele frequency can indicate the allele intensity ratio between HLA-B*46:01:01 and HLA-B*13:01:01 corresponding to a tumor sample for the same genomic location. The adjusted B allele frequency can be a ratio obtained by dividing the first B allele frequency by the second B allele frequency (or vice versa).

[0018] The subject-specific features can include an estimated tumor purity value and an estimated tumor ploidy value corresponding to the subject's tumor sample.As used herein, tumor purity refers to the ratio of tumor cells to the total cells in the sample.As used herein, tumor ploidy refers to the average copy number of the entire tumor genome.The whole exome features can include an indication of whether or not an HLA allele has at least a portion of the adjacent genomic region surrounding the HLA allele has been deleted.

[0019] The DASH model trained with the above training features can then be used to process sequence data and generate results corresponding to the probability of whether there is loss of heterozygosity at the identified HLA allele in a biological sample of a particular subject. Sequence data corresponding to a biological sample of a particular subject can be accessed. As used herein, sequence data refers to data corresponding to a biological sequence corresponding to a nucleic acid (e.g., DNA, RNA) or protein (e.g., alanine, arginine). In some cases, the sequence data includes one or more sequence reads. The sequence data can be generated by using whole genome sequencing or whole exome sequencing on the biological sample to generate multiple sequence reads. After the sequence data is generated, one or more HLA alleles can be identified from the sequence data.

[0020] A reference sequence corresponding to the identified HLA allele can be obtained, and the sequence reads can be aligned to the retrieved reference sequence.After alignment, allele-specific data for each of the identified HLA alleles corresponding to the sequence data can be identified.In some cases, the allele-specific data identifies the number of sequence reads that align to each genomic region corresponding to the identified HLA allele.

[0021] The trained DASH model uses the allele-specific data for each of the identified HLA alleles as input to generate a result. Other types of information corresponding to the identified HLA alleles (e.g., an indication of whether at least a portion of the adjacent genomic region surrounding the HLA allele has been deleted) can be used as additional input to the trained DASH model. The trained DASH model, including one or more gradient boosting algorithms, can process the above features of sequence data to generate a result. In some cases, the result is used to predict the reduced efficacy of immune checkpoint blockade therapy administered to a particular subject.

[0022] Thus, some embodiments of the present disclosure provide technical advantages over conventional systems by accurately detecting heterozygosity loss in HLA alleles. For example, the DASH model can accurately detect heterozygosity loss in HLA alleles by processing whole-exome sequence data. This is different from conventional techniques that rely only on sequence reads corresponding to HLA alleles. Considering that HLA genes are relatively short and most deletions involve larger genomic regions, the DASH model uses a whole-exome platform to incorporate sequence information from within as well as around HLA genes. As a result, the detection accuracy of the DASH model can be verified at a sensitivity level of 100% for samples with tumor purity above 8% and a specificity level of 100% for samples across all tumor purity levels. Thus, accurate detection of HLA heterozygosity loss facilitates the study of tumor cell mechanisms that contribute to immune checkpoint blockade resistance and the development of new cancer immunotherapies. Furthermore, some embodiments of the present disclosure can use allele-specific features to accurately identify which genomic regions of HLA alleles may be deleted, thereby detecting heterozygosity loss with increased granularity.

[0023] The following examples are provided to introduce certain embodiments. In the following description, for the purpose of explanation, specific details are set forth to allow the examples of the present disclosure to be fully understood. However, it will be apparent that various examples can be practiced without these specific details. For example, devices, systems, structures, assemblies, methods, and other components may be shown as components in block diagram form so as not to obscure the examples in unnecessary detail. In other examples, well-known devices, processes, systems, structures, and techniques may be shown without necessary details so as to avoid obscuring the examples. The drawings and descriptions are not intended to be limiting. The terms and expressions used in this disclosure are used as terms of description rather than limitation, and the use of such terms and expressions is not intended to exclude any equivalents of the features shown and described or portions thereof. The term "example" is used herein to mean to serve as an example, instance, or illustration. Any embodiment or design described herein as an "example" is not necessarily to be construed as preferred or advantageous over other embodiments or designs.

[0024] II. HLA allele sequence data and allele-specific coverage FIG. 1 shows a schematic diagram 100 for preparing sequence alignment data for detecting loss of heterozygosity in HLA alleles, according to some embodiments. In step 105, tumor and normal biological samples corresponding to each set of subjects can be collected. For example, each pair of tumor and normal samples can be tissue samples (e.g., adjacent normal samples) or blood / plasma samples of the corresponding subjects. In some cases, sample subsets are collected from different tumor types. For each subject, formalin-fixed paraffin-embedded (PFEE) or fresh frozen samples can be profiled to identify various features corresponding to normal and tumor samples. For example, the features can include comprehensive tumor mutation information, gene expression quantification, neoantigen characterization, HLA alleles (types and mutations), and tumor microenvironment profiling.

[0025] In step 110, whole exome library preparation and sequencing can be performed to generate sequence reads corresponding to each biological sample. In some cases, whole genome sequencing is performed to generate sequence reads. Whole exome capture libraries can be constructed using DNA from tumor and peripheral blood mononuclear cells / adjacent samples, and the libraries are constructed based on two whole exome sequencing (WES) capture kits: Agilent SureSelect Human All Exon v5 plus untranslated regions and Agilent SureSelect Clinical Research Exome. Modifications to the sequencing protocol can be made to obtain an average library insert size of approximately 250 bp (for example). In some cases, various polymerases can be used to generate sequence reads (including (for example) KAPA HiFi DNA polymerase and Herculase II DNA polymerase). Sequencing can be performed at a depth of 20G for normal samples and 35G for tumor samples. Using the exemplary sequencing methods described above, exome-wide coverage of over 300-fold may be available for over 20,000 genes, and coverage of over 1000-fold may be available for boost regions that include over 500 cancer-associated genes (including HLA-A, -B, and -C alleles).

[0026] In step 115, HLA genotyping (alternatively referred to as "HLA typing") can be performed to identify one or more HLA alleles from the sequence reads. HLA types can be calculated up to six digits. In some cases, sequence reads corresponding to tumor samples are processed to identify somatic mutations corresponding to one or more HLA alleles. In some cases, additional types of data can be identified from tumor and normal samples. For example, sequencing data can be analyzed to identify allele-specific copy number changes from specific HLA allele types. Additionally or alternatively, sequencing data can be analyzed to estimate tumor purity (alternatively referred to as tumor cellularity) and tumor ploidy.

[0027] In step 120, sequence reads can be aligned to one or more reference sequences (e.g., hs37d5 reference genome build). Subject-specific homologous alleles can be aligned to determine the positions of differences between alleles. Both single nucleotide variants (SNVs) and indels can be detected in the alignment. In some cases, only the first position of each indel can be considered to ensure that SNVs can be appropriately weighted. HLA alleles that differ at less than five positions can be considered homozygous.

[0028] In some cases, the reference sequence corresponding to the identified HLA allele is obtained by querying an HLA allele database. The HLA allele database can use an imputation approach to obtain the reference sequence corresponding to the identified HLA allele. The HLA allele database can be initialized in a specific format (e.g., multiple sequence alignment (MSA) format from IMGTv312). To implement imputation, the cDNA data can be used to impute exons in HLA alleles and sequence-incomplete HLA alleles with reference alleles that have protein-level identity defined by the same four-digit nomenclature. If no such allele exists, a reference sequence from the same HLA subtype can be obtained from the HLA-allele database based on the same two-digit nomenclature. If multiple options with the same two-digit nomenclature exist, the first allele listed in the MSA can be used. The above approach can be adopted using gDNA files to impute the intronic regions of each allele. Thus, the full-length genomic sequence of each allele can be imputed by assembling the exons from the cDNA imputation set and the introns from the gDNA imputation step. In some cases, duplicate reads are removed. Additionally or alternatively, the Genome Analysis toolkit (GATK) may be used to correct base quality scores and improve sequence alignment of sequence reads.

[0029] In some cases, any sequence reads with soft clipping of more than 20% of their total length are excluded. To improve the quality of the coverage information, any reads containing mismatches can be discarded. However, if somatic mutations are identified in HLA alleles, stringency can be relaxed to allow sequence reads with one mismatch.

[0030] In step 125, allele-specific data for the identified HLA alleles corresponding to each of the normal and tumor samples can be determined. The allele-specific data can identify the number of sequence reads that align to each genomic region corresponding to the identified HLA allele. Copy number changes in sequence reads at specific genomic locations can indicate loss of heterozygosity of the corresponding HLA allele. For example, for the B*13:01:01 HLA allele, a decrease in the number of sequence reads can be identified at genomic locations ranging from 1800 to 2000 in the tumor sample compared to the normal sample. Such copy number changes can indicate loss of heterozygosity of the B*13:01:01 HLA allele.

[0031] III. Machine learning models for predicting loss of heterozygosity in HLA alleles (a) Overview As described above, the DASH model can be used to process sequence data and generate results corresponding to the probability of whether there is loss of heterozygosity at a specified HLA allele in a biological sample of a particular subject.To start the process, sequence data corresponding to a biological sample of a particular subject can be generated by using whole genome sequencing or whole exome sequencing on the biological sample.The biological sample can be a tissue sample that can contain DNA from tumor or healthy cells.In some cases, the biological sample contains cell-free DNA, some of which can originate from healthy cells and some of which can originate from tumor cells.

[0032] In some cases, HLA genotyping is performed on multiple sequence reads to identify one or more HLA alleles corresponding to sequence data.A reference sequence corresponding to the identified HLA allele can be obtained, and the sequence reads can be aligned to the retrieved reference sequence.After alignment, the allele-specific coverage in each genomic region can be determined for the identified HLA allele corresponding to sequence data.In some cases, the allele-specific coverage identifies the number of sequence reads that align to each genomic position of the identified HLA allele.

[0033] The trained DASH model can be used to process the allele-specific data for each of the identified HLA alleles to generate a result corresponding to the probability of whether there is a loss of heterozygosity in each of the identified HLA alleles.Other types of information corresponding to the identified HLA alleles (e.g., an indication of whether at least a portion of the adjacent genomic region surrounding the HLA allele is deleted) can be used as input to the trained DASH model.The trained DASH model, including one or more gradient boosting algorithms, can process the above features of sequence data to generate a result.

[0034] In some cases, the allele-specific data of each of the identified HLA alleles is input to the trained DASH model. Other types of information corresponding to the identified HLA alleles (e.g., indication of whether at least a portion of the adjacent genomic region surrounding the HLA allele is deleted) can be used as input to the trained DASH model. The trained DASH model, including one or more gradient boosting algorithms, can process the above features of sequence data to generate a result. In some cases, this result is used to predict the reduced efficacy of immune checkpoint blockade therapy administered to a particular subject.

[0035] The DASH model can be trained to accurately detect allele-specific loss of heterozygosity in HLA genes using aligned sequence data and allele-specific coverage.Specifically, the training dataset of the DASH model can include features derived from a pair of tumor and normal samples (either adjacent tissue or peripheral blood mononuclear cells) from a subject.As described above, to identify training features in the training dataset, whole exome sequencing can be applied to the pair of tumor and normal samples corresponding to the subject to identify aligned sequence data and allele-specific coverage corresponding to HLA alleles.

[0036] In some cases, in addition to whole exome sequencing, capture probes that cover specific HLA alleles are applied. Sequence reads corresponding to HLA alleles in each subject can be mapped to each subject-specific HLA reference. From the alignment data, allele-specific coverage can be determined for each genomic region, and training features corresponding to somatic variants of HLA alleles can be identified. The training features can include modified B allele frequencies that take into account the difference in probe capture and the consistency of allele-specific coverage across various HLA alleles. Since the majority of HLA loss of heterozygosity events are large deletions, in some cases, the training features further include information corresponding to genomic regions surrounding the HLA alleles.

[0037] (b) Model selection The DASH model can include one or more gradient boosting algorithms that can be trained to detect loss of heterozygosity in HLA alleles. Gradient boosting refers to a machine learning technique for regression and classification problems that generates a predictive model in the form of an ensemble of weak predictive models. This technique can build models in a stepwise manner and generalizes the model by allowing the optimization of any differentiable loss function. Gradient boosting combines weak learners into a single strong learner in an iterative manner. As each weak learner is added, a new model is fitted to provide a more accurate estimate of the response variable. The new weak learner can be maximally correlated with the negative gradient of the loss function relative to the entire ensemble. Examples of gradient boosting machines include XGBoost and LightGBM. Additionally or alternatively, other types of machine learning techniques (including bagging procedures, boosting procedures, and / or random forest algorithms) can be used to build the combined model.

[0038] (c) Example training data set The exemplary training data included 720 heterozygous HLA gene sets collected from 279 subjects across multiple tumor types. All the above features were generated for each heterozygous gene, and each case of HLA heterozygosity loss was manually labeled. To train the DASH model, 500 heterozygous genes were separated for training, and 220 heterozygous genes were kept separately for testing. With regard to model selection, the DASH model can include a gradient boosting algorithm (e.g., XGBoost) to learn how to detect HLA heterozygosity loss in each pair of HLA alleles from the above features. If HLA heterozygosity loss is detected by the DASH model, the allele with lower coverage is labeled as deleted. Although rare, a small number of cases with biallelic deletions were detected. If the DASH model detects HLA heterozygosity loss and the allele-specific coverage ratio of the allele with higher coverage is below 0.5 in at least 25% of the bins, both HLA alleles are labeled as deleted.

[0039] (d) Characteristics of the training dataset FIG. 2 shows a diagram 200 corresponding to an exemplary training feature set of a training dataset for training a machine learning model for detecting loss of heterozygosity in HLA alleles, according to some embodiments. Sequence alignment data corresponding to normal and tumor samples can be analyzed to identify somatic variants. Diagram 200 shows various training features 205, 210, and 215 used to train the DASH model. First, the allele-specific features 205 can include adjusted B allele frequency, coverage ratio and coverage consistency between tumor and normal samples. The training features can also include subject-specific features 210, such as tumor purity and tumor ploidy. Finally, the whole exome features 215 can include deletions of flanking regions corresponding to the exome of the identified HLA allele.

[0040] From the allele-specific coverage data, allele-specific features 205 can be determined. The allele-specific features 205 include: 1. Adjusted B allele frequency: At each position of mismatch, the B allele frequency can be calculated separately for the tumor and normal samples. The tumor B allele frequency can then be divided by the normal B allele frequency. Because there may be variability in the probe capture of each allele, the normal samples can be used to adjust the B allele frequency. To combine the ratios into a single feature, the allele reference can be divided into bins of 150 base pairs in length. The absolute value of the median adjusted B allele frequency can be calculated for each bin. In some cases, the median across all bins is used as the training feature. The adjusted B allele frequency feature is bounded by 0, with higher values ​​indicating a higher probability of heterozygosity loss in the HLA gene. 2. Allele-specific coverage ratio: At each position of mismatch between the homologous alleles of the tumor and normal samples, the ratio between the coverage of the tumor sample and the coverage of the normal sample is calculated for each allele. Each ratio value can be normalized by dividing the number of exome-wide tumor reads by the number of exome-wide normal reads. Thus, regardless of the variability of sequencing depth between each run, the expected allele-specific coverage ratio is 1 in the absence of copy number polymorphism. Then, for each bin, the median of the coverage ratio can be calculated for each allele, and the lower value of the two alleles is taken as that of the bin. In some cases, the median value across all bins is used as the training feature. The lower limit of the allele-specific coverage ratio can be 0, and the expected value in the absence of copy number polymorphism is 1. A relatively low allele-specific coverage ratio suggests a high probability of heterozygosity loss in the HLA genes. 3. Total coverage ratio: When the composite sequencing depth of a heterozygous pair of alleles is high, allelic imbalance may be driven by large amplification of the allele rather than loss of the allele. At each position of mismatch between homologous alleles, the ratio between coverage in the tumor sample and coverage in the normal sample is calculated for each allele. The sum of both alleles can then be taken as a value representing each bin (e.g., 150 bp bins). Finally, the median value across the bins is used as the total coverage ratio feature. The minimum value of the total coverage ratio is 0, and values ​​tend to be high in genes without HLA heterozygosity loss. 4. Consistency of coverage: The more consistently low the coverage of an HLA allele across all mismatch positions, the more likely it is to be the case for HLA loss of heterozygosity. Thus, a value of 0 or 1 can be assigned to each bin if the coverage of each allele is lower or higher than its homologous allele. If the allele coverage for a bin cannot be determined (no mismatch sites), each allele is given a value of 0.5. Then, for each allele, the assigned values ​​can be averaged across all bins and the higher average value can be assigned to the value for that feature. The percentage coverage function ranges from 0.5 to 1, with values ​​closer to 1 representing a higher likelihood of HLA loss of heterozygosity.

[0041] Additionally, the subject specific features 210 can include: 1. Tumor Purity: Tumor purity characteristics can be estimated. Values ​​range from 0.1 to 1, with 0.1 being the least pure tumor and 1 being the most pure tumor. As used herein, tumor purity refers to the ratio of tumor cells to total cells in a sample. 2. Tumor ploidy: Tumor ploidy characteristics can be estimated. Values ​​are integers equal to or greater than 1. As used herein, tumor ploidy refers to the average copy number of the entire tumor genome.

[0042] Finally, the whole exome signature 215 can include: 1. Deletion of flanking regions: In most cases, HLA loss of heterozygosity is due to large deletions, so features corresponding to deletions in the flanking regions of each HLA gene can be determined to take advantage of information from a larger number of variable sites. B allele deletions can be determined, and then flanking region deletions can be determined from B allele deletions if there is a deletion within 10,000 base pairs in either direction of the HLA gene. This deletion feature can be binary, with 0 representing a deletion.

[0043] Chart 200 further shows a bar plot 220, which indicates that the trained DASH model outperforms other conventional techniques even when using the same sequencing data. In this example, biological samples with tumor purity below 20% are removed from the analysis. As shown in bar plot 220, the trained DASH model (shown in green) has a sensitivity of 100% and a specificity of 99.7% level that is higher than the sensitivity of 91.8% and the specificity of 94.3% level of the LOHHLA algorithm (shown in blue) (the LOHHLA algorithm is an existing conventional technique for detecting loss of heterozygosity, published in McGranahan, Nicholas et al. “Allele-Specific HLA Loss and Immune Escape in Lung Cancer Evolution.” Cell vol.171,6(2017):1259-1271.e11.doi:10.1016 / j.cell.2017.10.001). When considering all biological samples (including those with tumor purity below 20%), the DASH model achieved a specificity of 98.7% and a sensitivity of 92.9% (F-1 score = 0.939), while the LOHHLA algorithm achieved only a specificity of 94.3% and a sensitivity of 78.8% (F-1 score = 0.777). The DASH model also outperforms other existing techniques. For example, Sequenza detects HLA loss of heterozygosity with a specificity of 92.9% and a sensitivity of 95.0% (F-1 score = 0.848). Furthermore, none of the above conventional techniques for detecting loss of heterozygosity were able to identify the specific allele that was lost.

[0044] Figure 3 shows additional features identified from a training dataset 300 for training a machine learning model according to some embodiments. The training dataset 300 can include a scatter plot 305 showing the relationship between adjusted B allele frequency and tumor purity. HLA genes with heterozygosity loss are shown in solid color and HLA genes without heterozygosity loss are shown in partial shading. As shown in the scatter plot, a higher adjusted B allele frequency value can indicate HLA heterozygosity loss. In some cases, the adjusted B allele frequency can be determined by dividing the tumor B allele frequency by the adjacent normal B allele frequency in a given genomic region.

[0045] The training dataset 300 can also include a scatter plot 310 showing the relationship between allele-specific coverage fraction and tumor purity. HLA genes with loss of heterozygosity are shown in solid color and HLA genes without loss of heterozygosity are shown in partial shading. In some cases, the allele-specific coverage fraction is calculated by dividing the tumor coverage of an allele by the adjacent normal coverage of the same allele, normalized by the coverage over the remainder of the exome.

[0046] The training data set 300 may also include a box plot 315 showing the difference in distribution of the consistency of coverage in subjects with or without HLA heterozygosity loss. The box plot 315 can be used to make the observation that alleles with consistently lower coverage than alternative alleles across genes may be missing, while sporadically lower coverage may be due to stochastic variation (p=2.2e-14, paired T-test). In addition, the difference in distribution between the total sequencing depth in subjects with or without heterozygosity loss may be shown. As mentioned above, the total coverage of subjects with HLA heterozygosity loss is relatively lower than the total coverage of subjects without HLA heterozygosity loss. Thus, the total coverage ratio captures the combined coverage of the two alleles to distinguish allelic imbalance driven by heterozygosity loss from allelic imbalance driven by large amplification of alternative alleles. Under the null hypothesis that the distribution of total coverage proportions is the same for subjects with and without HLA loss of heterozygosity, subjects with HLA loss of heterozygosity have a significantly lower total coverage proportion (p=0.0004, paired T-test).

[0047] The training dataset 300 may also include histograms 320 and 325, where histogram 320 shows the distribution of tumor purity in subjects with or without loss of heterozygosity, and histogram 325 shows the distribution of tumor ploidy in subjects with or without loss of heterozygosity. HLA genes with loss of heterozygosity are shown as uncolored, and HLA genes without loss of heterozygosity are shown as shaded. Histograms 320 and 325 showed that tumor purity and tumor ploidy are identical across HLA-A, HLA-B, and HLA-C for a particular subject.

[0048] The training dataset 300 may also include a histogram 330 showing the distribution of heterozygosity loss sizes corresponding to HLA genes across all subjects. Because 73% of copy number changes causing HLA heterozygosity loss are deletions of greater than 1 megabase, genomic regions flanking the gene of interest may provide useful information to complement the intragenic data. Therefore, the whole-exome nature of the training dataset 300 was used to generate features corresponding to deletions in flanking regions that may measure deletions of 10 kb regions surrounding each HLA gene.

[0049] (e) Training process FIG. 4 shows exemplary data 400 generated by a machine learning model trained to detect heterozygosity loss in HLA alleles according to some embodiments. The exemplary data 400 shows a bar plot 405 showing the number of heterozygous gene pairs with or without HLA heterozygosity loss in the training and test sets of the DASH model. For the exemplary training process, six features for 720 heterozygous alleles were collected from 279 subjects. Homozygous alleles were excluded because heterozygous loss cannot occur in homozygous alleles. Training labels were added to the 720 alleles, and missing alleles were hand-curated. The exemplary data was divided into a training dataset (n=500, about 70%) and a test dataset (n=220, about 30%) with roughly equivalent class distribution. A DASH model including a gradient boosting regression (XGBoost) algorithm was trained to predict missing HLA alleles.

[0050] Bar plot 410 shows the impact of each feature in the trained DASH model for detecting loss of heterozygosity in HLA alleles. In some cases, the impact of each feature is measured based on a game theory model. Bar plot 410 reveals that all six of our features contribute independently to the DASH model, with flanking region deletion and adjusted B allele frequency having the most significant impact on outcome compared to other features (e.g., tumor ploidy).

[0051] Scatter plot 415 shows the distribution of the probability of HLA heterozygosity loss returned by the trained DASH model for biological samples corresponding to the test data set and manually annotated as having or not having heterozygosity loss at HLA alleles. In the scatter plot 415, the shaded area shows the ambiguous calls by the DASH model. Since the XGBoost algorithm returns a continuous metric, the calls of HLA heterozygosity loss were divided into high and low confidence calls. In the test data set, the trained DASH model reached a specificity of 96% and a sensitivity of 89%, considering high confidence calls (>0.8 heterozygosity loss prediction cutoff). As mentioned above, the DASH model can exhibit higher specificity and sensitivity than other conventional techniques. For comparison, the LOHHLA algorithm reached a specificity of 92% and a sensitivity of 76% using the same test data set.

[0052] Histogram 420 shows the distribution of tumor purities for HLA genes for which the DASH model made ambiguous calls (>0.2 and <0.8). As shown in histogram 420, the majority of borderline and inaccurate calls were at low tumor purity, highlighting the difficulty of accurately predicting HLA loss of heterozygosity at low levels of tumor purity. When samples with tumor purity below 20% are removed from the test dataset, the performance level of the DASH model increases, as shown in accuracy rescoring curve 425.

[0053] The accuracy rescored curve 425 shows the performance level of the DASH model for the retained dataset (heterozygous genes n=220). In the curve 425, the dotted line shows the performance of all samples, and the solid line shows the performance of samples with at least 20% tumor purity. When samples with tumor purity below 20% were removed from the test dataset, the DASH model reached 97% specificity and 95% sensitivity with a loss of heterozygosity prediction cutoff of >0.8. Continuing with the above example, the DASH model performed better than the LOHHLA algorithm, achieving 94% specificity and 82% sensitivity for the same test dataset excluding samples with tumor purity below 20%. To achieve an optimal balance between sensitivity and specificity, a threshold of 0.2 was applied to the remaining analyses. With this threshold, very strong performance of the DASH model was observed in high purity samples (F1 score=0.93), with a decrease in performance when low purity samples of the test dataset were included (F1 score=0.87).

[0054] Finally, bar plot 430 compares the F1 scores of various DASH models trained on individual features to a DASH model trained on all features of the training dataset (see FIG. 2). As shown in bar plot 430, flanking regions and adjusted b allele frequencies produced higher F-1 scores for predicting loss of heterozygosity at HLA alleles. Nevertheless, it was found that no feature alone could achieve the performance level of a DASH model trained with all features of the training dataset.

[0055] IV. Validation Techniques for Predicting Loss of Heterozygosity in HLA Alleles HLA heterozygosity loss has been observed to occur as a resistance mechanism late in tumor progression. Furthermore, the tumor types most responsive to immune checkpoint blockade (lung, skin) also tend to generate samples with low purity. Therefore, limit of detection analysis with gold standard samples was used to accurately validate the performance of the DASH model in both low clonality and low purity settings.

[0056] (a) Cell line-based detection limit analysis For detection limits, the DASH model can be evaluated across a range of tumor purities and clonality using a pair of tumor-normal lymphoblastoid cell line samples (NCI-H2009). In the NCI-H2009 sample, HLA-A is homozygous, while (i) both HLA-B*51:01 and HLA-C*15:02 alleles are deleted, and (ii) HLA-B*07:02 and HLA-C*07:02 are retained. Deep sequencing can be performed at 50x and 30x coverage on the tumor and normal cell lines, respectively. To simulate realistic sequencing depth, the normal data can be downsampled to reflect 25x sequencing coverage. To create low purity tumor data, an increased proportion of normal reads is mixed with a decreased proportion of tumor reads. The combined normal and tumor reads can be summed to an average of 35x sequencing coverage to surrogate for the tumor sample. As used herein, sequencing coverage refers to the average number of reads that align to a known reference base. Sequencing coverage levels can be used during sequencing to determine whether variant discovery can be performed at a particular base position with a certain degree of confidence. For example, the recommended sequencing coverage for whole genome sequencing can range from 30x to 50x depending on the application and statistical model. In another example, the recommended sequencing coverage for whole exome sequencing can be 100x.

[0057] Using the seqkit library, all combinations of normal and tumor sub-samples can be run with 10 replicates. In some cases, to simulate lower subclonality, the percentage of tumor reads in the mixture was used as the product of the desired tumor purity and subclonality. The tumor purity can then be increased to reflect the desired tumor purity. Samples without HLA heterozygosity loss can be simulated by including only normal reads in the tumor samples and increasing the estimate of tumor purity to reflect the desired range. These runs can be used to estimate specificity.

[0058] (b) Allele-specific digital PCR validation To verify allele-specific HLA loss of heterozygosity in samples, subject-specific primers and probes can be designed and allele-specific DNA depletion can be tested by digital PCR. Because each subject has a unique set of up to six HLA class I alleles, subject-specific primers and probes can be designed for each subject. These primers and probes can bind with high specificity to each allele of interest and discriminate between all other alleles and the rest of the genome. Because some homologous alleles are similar, there is not a good primer and probe for every subject. In some cases, primers and probes are designed to maximize discrimination between alleles for 11 homologous allele pairs with HLA loss of heterozygosity predicted by the DASH model from 10 different subjects and one cell line. In addition, a probe targeting RNase P can also be used to serve as an internal positive control. Different fluors can be assigned to the HLA alleles and RNase P probes to allow multiplexing. A negative control sample (e.g., H2O) may be used.

[0059] To evaluate the efficiency of the primers and probes, digital PCR can be performed in triplicate on DNA from normal and tumor samples (except subject C, which may be performed in duplicate). Three samples can be from the training dataset (B, D, K) and the remaining seven samples can be independent. To analyze the data, both losses and retentions are normalized by the control gene to account for the variation in sample input. Since HLA alleles can be expected to be haploid and RNase P is expected to be diploid, primers and probes can be considered successful if the ratio of HLA allele copies to multiplexed RNase P copies can be 0.5 in the normal sample. For primer designs that meet this requirement, the allele:RNase P ratio in tumor DNA is then compared to the allele:RNase P ratio in normal DNA by a one-tailed T-test to determine if there is a significant reduction in the tumor. This test is performed on both the predicted retained alleles and the predicted lost alleles. Allelic imbalance is determined by measuring the significant difference between the predicted lost alleles and the predicted retained alleles in normal and tumor DNA. Of note, this study focused on specific portions of each gene and is therefore not designed to find small focal deletions in small parts of genes.

[0060] (c) Quantitative immunopeptidemics validation To evaluate the functional impact of HLA heterozygosity loss on peptide presentation by MHC molecules, quantitative immunopeptidomics can be performed on two colon and four lung tumor-normal fresh frozen sample pairs. Samples can be homogenized, normalized for protein content between tumor and normal, and the clarified homogenates can be applied to a pan-MHC-I antibody (W6 / 32)-conjugated immunoaffinity resin. In some cases, the success of immunoprecipitation from lysates is assessed by comparing MHC concentrations before and after IP using ELISA. MHC-associated peptides can be eluted and collected. Eluted peptides from tumor and normal samples can be labeled and analyzed in high-resolution HCD mode in one run for each pair.

[0061] The resulting raw files of all six samples can be processed together. Peptide identification can be performed using de novo identification and then database searching. For example, the parameters of the database search can be: precursor mass tolerance: 10 ppm, fragment mass tolerance: 0.03 Da, protein database: uniprot sequences downloaded in April 2019, enzymatic digestion: none, fixed modifications: carbamidomethylation of cysteines (+57.02 Da) and TMT10plex (+2291.6) on all N-terminal amino acids and lysines, variable modifications: protein N-terminal acetylation (+42.0106) and oxidation of methionines (+15.9949). Peptides can be filtered at 1% FDR and reporter ions can be quantified. The list of quantified peptides can be further filtered to improve the quality of the call by removing peptides without expected TMT n-terminal or lysine modifications, peptides with low intensity (precursor ion intensity less than 10E4), and suspicious peptides with polyamino acids. The intensities can then be log2 transformed and the data median normalized. Finally, the fold change can be calculated from the log2 transformation, with values ​​less than 0 representing depletion of the peptide in the tumor sample and values ​​greater than 0 representing enrichment of the peptide in the tumor sample.

[0062] To assess the overall change in presentation between normal and tumor samples, absolute values ​​corresponding to the logarithm of the fold change were compared between samples. Peptide changes at specific alleles were then estimated. For each subject, each peptide in the peptide set was assigned to an MHC allele. If a peptide could be predicted to bind to multiple alleles, it was considered ambiguous and excluded from the analysis. If the only peptide predicted to bind to one allele could be predicted to bind to multiple alleles, it was included but marked (e.g., with an asterisk). The logarithm of the fold change can be visualized to assess the enrichment or depletion of peptides from specific alleles in tumor samples. Log2-transformed intensity values ​​can be compared with a Wilcoxon rank sum test to assess the statistical significance of any enrichment or depletion. All comparisons with tumor purity use the estimated tumor purity.

[0063] (d) Prediction of HLA-associated neoantigens The MHC class I presentation prediction model can be implemented using a machine learning model trained with a large immunopeptidome dataset and benchmarked against existing binding prediction models (e.g., NetMHCpan 4.0). The output of the trained model can be normalized by allele in a manner similar to NetMHCpan 4.0 to create a rank metric. In some cases, the percentile rank threshold used is 0.1%. All peptide-allele combinations ranked below this threshold can be considered as peptides that bind and are presented on the cell surface.

[0064] (e) Verification results (1) Validation results of cell line-based detection limit analysis FIG. 5 shows an example 500 of a cell line-based detection limit technique for validating a machine learning model trained to detect loss of heterozygosity at HLA alleles, according to some embodiments. In some cases, in silico cell line mixtures and allele-specific genomic validation are used to validate the DASH model. Schematic diagram 505 shows an example of mixing tumor and normal cell lines to simulate low purity sample pairs. Heat map 510 shows the specificity level, and heat map 515 shows the sensitivity level of the DASH model to capture HLA loss of heterozygosity in simulated samples of different purity and clonality. In heat maps 510 and 515, the boxes are shown with different shading to indicate various sensitivity levels, with gray boxes indicating no data available. Thus, heat maps 510 and 515 show a range of read ratios to simulate the potential spectrum of tumor purity and subclonality. As expected, the sensitivity level decreased with lower purity and clonality. However, the DASH model still maintained 100% sensitivity down to tumors with less than 20% lead mixture (observed for several tumor purity and HLA loss of heterozygosity clonality combinations: 20% purity and 100% clonality, 50% purity and 30% clonality, 100% purity and 20% clonality). In contrast, the DASH model maintained 100% specificity across all tumor purity levels. For comparison, heat maps 520 and 525 show the specificity and sensitivity levels of the LOHHLA algorithm across various purity and clonality levels, respectively. As above, boxes are shown with different shading to indicate various sensitivity levels, with grey boxes indicating no data available. As shown, the heat map 515 of the DASH model shows higher sensitivity levels across purity and clonality levels than the heat map 525 of the LOHHLA algorithm, especially at clonality levels below 80%.

[0065] FIG. 6 shows an example of results corresponding to the performance levels of the DASH model based on in silico cell line-based detection limit analysis. The exemplary result set 600 is based on validation using in silico cell line mixtures and digital PCR. Line plots 605 and 610 show the sensitivity and specificity levels of the DASH model. Line plots 615 and 620 show the sensitivity and specificity levels of the LOHHLA algorithm. Line plots 605-620 show the respective sensitivity and specificity levels at various purity levels with fully clonal tumors. The shaded areas of each line plot 605-620 indicate the 95% confidence level. As shown in line plots 605-620, the DASH model has a sensitivity level of 100% over tumor purity levels above 20% (e.g., line plot 605), whereas the sensitivity level corresponding to the conventional loss of heterozygosity algorithm is lower at certain tumor purity levels (e.g., line plot 615). Thus, the accuracy and low detection limit of the DASH model can be demonstrated using in silico cell line mixtures and allele-specific genomic data.

[0066] (2) Results of allele-specific genome verification using digital PCR FIG. 7 shows an example 700 of an allele-specific genomic technique for validating a machine learning model trained to detect loss of heterozygosity in HLA alleles, according to some embodiments. After establishing the ability of the DASH model to detect HLA loss of heterozygosity in various tumor and normal samples, digital PCR was used to orthogonally validate the predictive accuracy of the DASH model. Due to the highly polymorphic nature of HLA alleles, PCR primers and probes were designed independently for each pair of alleles in each subject. Probes were designed to avoid targeting other subject-specific alleles in addition to having to target alleles unique to each subject. As described herein, primers for retention-prediction and loss-prediction HLA-C alleles were designed for the cell lines examined using in silico mixtures (NCI-H2009 / NCI-BL2009). For both primer targets, tumor and blood (normal) cell line DNA were mixed to emulate reduced tumor purity levels in triplicate, and digital PCR was performed using both allele-specific designs. The allele primers were multiplexed with RNase P (RPP25) primers as a diploid control. As an illustrative example, the validation technique 700 includes a schematic 705 showing an example of allele-specific genomic validation using paired tumor and adjacent normal fresh frozen samples.

[0067] The bar plot 710 shows the allele-specific copy number of the predicted lost allele for RNase P based on the allele-specific genomic validation using digital PCR. The allele-specific number was measured by digital PCR for a mixture of cell lines of various tumor purity. To confirm the specificity of the primers, both the predicted lost allele copies and the predicted retained allele copies normalized by half the diploid RNase P copies were 1 copy in the normal sample.

[0068] The bar plot 710 includes a dashed line, which indicates the expected value if the copy number is unchanged. The copy number of each allele in the tumor sample was then compared to the normal sample (e.g., using a one-tailed Student's T-test). The asterisk indicates the result indicating whether a statistically significant difference was observed based on the comparison with the copy number of the normal sample. As the tumor purity increases above zero, one or more copies of the lost allele were found in the tumor sample, confirming the allele-specific LOH event. The sensitivity of the digital PCR was 100% at a tumor purity of 10% or more, thereby confirming the sensitivity and reproducibility of the allele-specific digital PCR as an orthogonal method.

[0069] Bar plots 715 and 720 show the ratio of HLA allele digital PCR copies to multiplexed RNase P digital PCR copies, respectively. In bar plots 715 and 720, cell line data is shown on the left and subject data is shown on the right. In addition, gray bars show the ratio in normal DNA and green bars show the ratio in tumor DNA. Alleles predicted to be retained by the DASH model are shown in bar plot 715, while alleles predicted to be lost are shown in bar plot 720. The dashed gray line shows the expected ratio of 0.5 in the absence of copy number changes. Asterisks indicate samples with p-values ​​less than 0.05 by one-tailed Student's T-test. To confirm primer specificity, both copies of the predicted lost allele and copies of the predicted retained allele normalized by RNase P copies were confirmed to have a ratio of 0.5 in the normal samples.

[0070] The ratios of 20 out of 22 primers were found to be highly specific, very close to 0.5. However, the predicted retained allele from subject C and the predicted lost allele from subject K were excluded due to their low specificity. The ratios of each allele were then compared between tumor and normal samples.

[0071] Additionally, significant depletion was found in only 1 of 9 retained predicted alleles and 8 of 9 lost predicted alleles in the tumor samples, as indicated by the asterisks above each bar in bar plots 715 and 720, respectively. The subject with no significant digital PCR depletion in the lost predicted alleles (subject J) appears to have a large digital PCR amplification of the retained allele. This amplification can be confirmed by standard copy number polymorphism testing in the region surrounding the HLA gene.

[0072] Scatter plot 725 shows the distribution of the probability of HLA loss of heterozygosity returned by the DASH model along with their tumor purity. The red areas indicate ambiguous calls by the DASH model. The grey vertical line indicates 20% purity. In scatter plot 725, the retained predicted alleles of subject D appear to be slightly reduced in the tumor. The significant allelic imbalance in the tumor suggests that this may be due to subclonal biallelic deletion or simply amplification of the RNase P control. Excluding calls from subject D, 95% of the predicted retained and lost alleles were correctly identified above the 0.8 threshold. Furthermore, some samples had low tumor content, confirming the accuracy of the DASH model across variable tumor purity. Thus, the subject-specific digital PCR presented herein represents the first allele-specific genomic HLA loss of heterozygosity validation assay.

[0073] FIG. 8 shows an exemplary result set 800 that represents the performance level of the DASH model based on allele-specific genomic validation by digital PCR. Bar plot 805 shows digital PCR copies of alleles predicted to be retained, and bar plot 810 shows digital PCR copy loss of alleles predicted by the DASH model. In both bar plots 805 and 810, multiplexed RNase P was used as a control gene. Cell line data is shown in the left part, and subject data is shown in the right part. For bar plots 805 and 810, data from a set of 11 subjects with matched tumor and normal sample pairs, namely 4 confirmatory subjects from the training data and 7 independent subjects, were profiled. Primers for the retained and lost predicted alleles were designed and digital PCR was performed on DNA from normal samples (adjacent cells or peripheral blood mononuclear cells) and tumor samples.

[0074] Bar plots 815 show the specificity of each primer design as measured by the ratio between allele digital PCR copies and multiplexed RNase P digital PCR copies in normal samples. The grey dashed line in each of the bar plots 815 shows the expected copy number of 1. Bar plots 820 show the ratio between the loss predicted allele (normalized by RNase P copies) and the retention predicted allele (normalized by RNase P copies) to show the allelic imbalance predicted by the DASH model. The grey dashed line in bar plot 630 shows the expected ratio of 1 in normal samples. Deviation below the grey dashed line suggests allelic imbalance. Referring to bar plot 820, the significant allelic imbalance obtained may have resulted in a less confident deletion prediction.

[0075] V. Altered peptide presentation in tumors predicting loss of HLA heterozygosity It has been hypothesized that HLA heterozygosity loss reduces neoantigen load by eliminating surface presentation of neoantigens that bind to specific HLA alleles. Such a hypothesis has been demonstrated using organoids but not in tumor samples from complex subjects. Therefore, to provide functional evidence of reduced peptide presentation in alleles where the DASH model predicts loss, we measured quantitative changes in peptide presentation between adjacent normal samples without HLA heterozygosity loss and tumor samples with HLA heterozygosity loss.

[0076] (a) Overview FIG. 9 shows an overview of the use of immunopeptidomics data to validate a machine learning model trained to detect loss of heterozygosity in HLA alleles, according to some embodiments. For example, schematic 905 shows a process for applying mass spectrometry to determine which HLA alleles have undergone loss of heterozygosity. Specifically, functional immunopeptidomics validation can be performed by (1) obtaining pairs of tumor and adjacent normal fresh frozen samples, (2) purifying each sample to obtain HLA beta 2 microglobulin complexes, and (3) gently eluting peptides from the HLA beta 2 microglobulin complexes. The eluted peptides from each sample can be labeled with one or more chemical labels (e.g., tandem mass tags) and measured using quantitative mass spectrometry.

[0077] The waterfall plot 910 shows the log2 fold change from normal to tumor samples in peptides binding to each of the alleles in the subject. In the waterfall plot 910, dark colors indicate peptides with low frequency in the tumor, and shaded areas indicate peptides with high frequency in the tumor. The grey dashed line represents the midpoint of the plot, and the triangles indicate the intersection points for each allele. Each waterfall plot in the waterfall plot 910 shows whether the subject's HLA allele was deleted or retained. The peptides at each allele are visualized as motifs. Statistical significance is assessed using Wilcoxon paired rank sum tests. In the waterfall plot 910, the tumor samples had significantly fewer predicted binding peptides for three of the four deleted alleles (biallelic deletion predictions) than the adjacent normal samples.

[0078] The box and whisker plot 915 shows the log2 fold change in peptide intensity between loss, retention, and homozygous alleles across HLA-A, HLA-B, and HLA-C alleles. Statistical significance was assessed using a two-tailed Student's T-test. The box and whisker plot 915 shows that for peptides predicted to bind to loss alleles, peptide intensity was reduced in tumor samples compared to normal samples in HLA-A and -B alleles.

[0079] (b) Relationship between samples without HLA loss of heterozygosity and samples predicted to have HLA loss of heterozygosity FIG. 10 shows an example of quantitative immunopeptidomics data corresponding to the relationship between control samples without predicted HLA heterozygosity loss and samples with predicted HLA heterozygosity loss, according to some embodiments. In this example, the quantitative immunopeptidomics data included two control samples without HLA heterozygosity loss and four samples with predicted HLA heterozygosity loss. Bar plot 1005 reveals the number of unique peptides identified in quantitative immunopeptidomics from each of the above samples. Bar plot 1010 shows the percentage of peptides in each sample that are predicted to bind at least one of the HLA alleles of the corresponding subject. Across six samples, bar plot 805 reveals a strong peptide yield (median of about 5000 unique peptides) and bar plot 1010 reveals a high percentage of observed peptides that are predicted to bind at least one of the subject-specific alleles.

[0080] The box and whisker plot 1015 shows the distribution of log2 fold change intensity in samples without HLA loss of heterozygosity (control) and samples with HLA loss of heterozygosity. For samples without any predicted HLA loss of heterozygosity (M and P samples), the differences in surface peptides found between tumor and normal samples were minimal. For example, the box and whisker plot 1015 shows that for samples without any predicted HLA loss of heterozygosity, the quartiles of peptide log fold change ranged from -0.010 to 0.013, with the median peptide fold change close to zero for all alleles. In contrast, samples with predicted HLA loss of heterozygosity (L, C, O, and N samples) showed as much as two-fold variability in peptide presentation between tumor and normal samples, with the quartiles of peptide log fold change ranging from -0.026 to 0.023.

[0081] Scatter plot 1020 shows the relationship between estimated tumor purity and standard deviation of log2 fold change of peptide intensity. For scatter plot 1020, green dots represent samples without HLA loss of heterozygosity and blue dots represent samples with HLA loss of heterozygosity. The deviation in intensity between tumor and normal samples increased with increasing tumor purity of the samples, with the purest samples showing a mean deviation of 0.062 (L, 58% tumor purity).

[0082] (c) Immunopeptidemic data corresponding to samples where HLA loss of heterozygosity is not predicted. FIG. 11 shows an example of quantitative immunopeptidomic data corresponding to a control sample not predicted to have HLA heterozygosity loss, according to some embodiments. Waterfall plot 1105 shows immunopeptidomic data corresponding to a first subject not predicted to have HLA heterozygosity loss (e.g., subject M in FIG. 10). Waterfall plot 1110 shows immunopeptidomic data corresponding to a second subject not predicted to have HLA heterozygosity loss (e.g., subject P in FIG. 10). Waterfall plots 1105 and 1110 show the log2 fold change from normal to tumor samples in peptides that bind to each of the alleles in a particular subject. Darker colors indicate peptides with lower frequency in the tumor, and shaded areas indicate peptides with higher frequency in the tumor. Each waterfall plot in waterfall plots 1105 and 1110 shows that the subject HLA alleles are retained. Peptides at each allele are visualized as motifs. Statistical significance is assessed using Wilcoxon's paired rank sum test.

[0083] (d) Immunopeptidemic data corresponding to samples predicted to have HLA loss of heterozygosity FIG. 12 illustrates an example of quantitative immunopeptidomics data corresponding to samples predicted to have HLA heterozygosity loss, according to some embodiments. Waterfall plot 1205 illustrates immunopeptidomics data corresponding to a first subject predicted to have HLA heterozygosity loss (e.g., subject N in FIG. 10). Waterfall plot 1210 illustrates immunopeptidomics data corresponding to a second subject predicted to have HLA heterozygosity loss (e.g., subject O in FIG. 10). Waterfall plot 1215 illustrates immunopeptidomics data corresponding to a third subject predicted to have HLA heterozygosity loss (e.g., subject C in FIG. 10). Waterfall plots 1205-1215 illustrate the log2 fold change from normal to tumor samples in peptides that bind to each of the alleles in a particular subject. Darker colors indicate peptides with lower frequency in the tumor, and shaded areas indicate peptides with higher frequency in the tumor. Shaded boxes indicate deleted alleles, and white boxes indicate retained alleles. Peptides in each allele are visualized as motifs. Statistical significance is assessed using the Wilcoxon paired rank sum test.

[0084] Although the peptide log fold changes for three waterfall plots 1205-1215, representing additional low tumor purity samples predicted to have HLA loss of heterozygosity, were greater than the control samples (e.g., waterfall plots 1105-1110 in FIG. 11), they did not show the same allele-specific peptide depletion predictions (23%, 11%, and 11% for C, O, and N, respectively). However, waterfall plot 1205 (C, 23%) could be validated as having HLA loss of heterozygosity by digital PCR (subject C, represented by waterfall plot 1215), suggesting that other factors also affect peptide presentation. Although expected variability in detection by immunoprecipitation and mass spectrometry limits the level of sensitivity in low purity samples, this is the first of its kind functional validation approach using subject tumor samples to demonstrate that HLA loss of heterozygosity is associated with altered peptide presentation.

[0085] VI. Loss of Heterozygosity in HLA Alleles Across Tumor Types The prevalence data of HLA loss of heterozygosity may indicate that a high percentage of subjects in some tumor types are affected by HLA loss of heterozygosity. Non-small cell lung cancer is known to have a high incidence of HLA loss of heterozygosity, but other types of cancers, including cervical cancer (44%) and head and neck squamous cell carcinoma (40%), also showed high rates of HLA loss of heterozygosity. In contrast, in melanoma (which also has a high mutational burden), only 14% of subjects with HLA loss of heterozygosity were observed. Furthermore, cervical cancer is strongly associated with human papillomavirus (HPV), which may be responsible for the high frequency of HLA loss of heterozygosity. In some cases, subjects have lost multiple HLA alleles at once, which may have a stronger effect on tumor evolution.

[0086] To assess the pervasiveness of HLA loss of heterozygosity as a potential immune escape mechanism, the DASH model was applied to 611 tumors across 15 tumor types. A total of 593 subjects from 14 tumor types were considered for the analysis. Each subject had a tumor and a normal sample, which were sequenced and analyzed. A subset of these samples was used to train the DASH model. The DASH model was applied to each sample to predict genes (HLA-A, -B, and -C) affected by HLA loss of heterozygosity. The co-occurrence rate of HLA loss of heterozygosity between multiple genes within a single subject was calculated based on a reduced cohort including only fully heterozygous subjects.

[0087] (a) Predicted HLA loss of heterozygosity across tumor types FIG. 13 shows predicted HLA heterozygosity loss across tumor types, according to some embodiments. Bar plot 1305 reveals the number of subjects and frequency of HLA heterozygosity loss in each tumor type cohort. In this example, only cohorts with at least 10 subjects are shown. As shown in bar plot 1305, the percentage of subjects with at least one HLA heterozygosity loss occurred ranged from 44% of cervical cancer subjects to 11% of liver cancer subjects. For example, HLA heterozygosity loss was observed in approximately 31% of non-small cell lung cancer adenocarcinomas (non-small cell lung cancer A). In another case, a lower incidence was observed in approximately 34% of non-small cell lung cancer squamous cell carcinomas (non-small cell lung cancer-SCC).

[0088] Bar plot 1310 shows the number of subjects with one, two, or three genes affected by HLA loss of heterozygosity. In this example, only subjects who are fully heterozygous across HLA-A, -B, and -C are shown. In the bar plot 1310, subjects with HLA loss of heterozygosity lost all three genes more frequently (70% of subjects) than lost one gene or two genes (20% and 10% of subjects, respectively).

[0089] The box and whisker plot 1315 shows the distribution of the percentage of each genome affected by HLA loss of heterozygosity. Each tumor type is divided into subjects with and without HLA loss of heterozygosity. Only tumor types with at least 10 subjects affected by HLA loss of heterozygosity are shown. Statistical analysis was performed with the Mann-Whitney U test with Bonferroni correction. Although a high frequency of HLA loss of heterozygosity in a particular tumor type is of interest due to impaired antigen presentation pathways, high frequency alone does not require an evolutionary advantage of loss of heterozygosity. As shown in the box and whisker plot 1315, subjects with HLA loss of heterozygosity had a significantly higher estimated rate of loss of heterozygosity across their genomes, suggesting that some loss of heterozygosity within the HLA region may occur by chance (pan-cancer p<2.2e-14).

[0090] To examine whether the frequency of HLA heterozygosity loss occurs by chance across cancer types, the average estimate of the rate of heterozygosity loss across the genome was compared to the frequency of HLA heterozygosity loss in a given tumor type cohort. If heterozygosity loss occurs randomly in HLA regions, the rates and frequencies were expected to be similar. In particular, if a region has zero estimated copies of the B allele, the region can be considered to have HLA heterozygosity loss. The total number of base pairs affected by heterozygosity loss can be summed for each subject and divided by the total number of base pairs in the entire exome (3.2 billion) to obtain the proportion of the genome with heterozygosity loss. Although this proportion may be an underestimate due to limited coverage of genomic regions without genes, this underestimate can be expected to be consistent across subjects.

[0091] Scatter plot 1320 shows the relationship between the average fraction of the genome affected by heterozygosity loss and the frequency of HLA heterozygosity loss in each tumor type. The gray dashed line indicates x=y. As shown in scatter plot 1320, the frequency of HLA heterozygosity loss is higher than genome-wide heterozygosity loss in almost all tumor types. Although the difference is small in some tumor types, significant enrichment of HLA heterozygosity loss was observed in colorectal cancer, renal clear cell carcinoma, non-small cell lung cancer A, pancreatic cancer, and head and neck squamous cell carcinoma. The data shown in scatter plot 1320 suggests that HLA heterozygosity loss may provide a greater evolutionary advantage in these tumor types than in other tumor types. Alternatively, HLA may be more prone to deletion than the rest of the genome.

[0092] Bar plots 1325-1340 show the difference in neoantigen expression between subjects without HLA loss of heterozygosity (green) and subjects with HLA loss of heterozygosity (blue), with the assumption that tumors with a higher capacity to display neoantigens are under higher selective pressure leading to HLA loss. Bar plot 1325 shows the average difference in neoantigen load between the two subject categories across various cancer types. Bar plot 1325 shows the difference between the two subject categories. Additionally, the remaining bar plots 1330-1340 show the difference between the two subject categories. For example, bar plot 1330 shows a statistically significant difference between the two subject categories for CD274 (PD-L1) expression (p=0.02), bar plot 1335 shows a statistically significant difference between the two subject categories in the percentage of microsatellite unstable sites (p=0.01), and bar plot 1340 shows a statistically significant difference between the two subject categories in the percentage of patients with Fusobacterium nucleatum, an oral bacterium known to be associated with colorectal cancer (p=0.005). Thus, box and whisker plots 1325-1340 indicate that tumors with a greater ability to display neoantigens may be capable of causing HLA loss of heterozygosity.

[0093] (b) Antigen presentation and loss of heterozygosity Because HLA heterozygosity loss affects the ability of tumor cells to present antigens on the cell surface for recognition by the immune system, it was hypothesized that tumors with a high capacity to display neoantigens are under selective pressure to result in HLA loss. Given that subjects with higher HLA evolutionary diversity are able to present a larger immunopeptideome and respond better to checkpoint inhibitors, experiments were performed to determine whether such subjects are more susceptible to HLA loss.

[0094] FIG. 14 shows a set of experimental results 1400 showing the relationship between HLA loss of heterozygosity and antigen presentation across various tumor types, according to some embodiments. The box and whisker plot 1405 shows the distribution of HLA-I evolutionary divergence across subjects with or without HLA loss of heterozygosity. Only tumor types with at least 8 subjects affected by HLA loss of heterozygosity are shown. Statistical analysis was performed with the Mann-Whitney U test, with Bonferroni correction. Germline HLA-I evolutionary divergence scores were calculated from the HLA alleles of each individual subject. The HLA-I evolutionary divergence scores are intended to capture the allelic sequence diversity of the subjects' HLA alleles, with lower scores indicating lower diversity and higher scores indicating higher diversity. The box and whisker plot 1405 shows that there appears to be no substantial correlation between HLA evolutionary diversity and HLA loss of heterozygosity.

[0095] Additionally, the box and whisker plot 1410 shows the corresponding distribution of the number of mutations (e.g., single nucleotide variants, indels, and fusions) across subjects with and without HLA loss of heterozygosity. Only tumor types with at least 8 subjects affected by HLA loss of heterozygosity are shown. Statistical analysis was performed by Mann-Whitney U test with Bonferroni correction. Mutational burden (number of mutations) was determined using tumor-specific genomic events with at least 5% allele fraction verified with transcriptome data. All potential neoepitopes (8-, 9-, 10-, and 11-mers) were generated for each mutation and tested for presentation. Any 8-, 9-, 10-, or 11-mers containing mutations are considered as putative neoepitopes if they are predicted to bind any of the subject-specific alleles. The box and whisker plot 1410 shows that high mutation rates and neoantigen burden may present pressure for cells to lose HLA. Box and whisker plot 1415 shows the percentage of patients with HLA LOH at each ventile of pan-cancer mutational burden. Both box and whisker plots 1410 and 1415 show a "Goldilocks effect" in mutational burden, with diseases with the lowest and highest tumor mutational burdens exhibiting the lowest prevalence of HLA loss of heterozygosity, and tumors in between exhibiting the highest prevalence of HLA loss of heterozygosity.

[0096] Box and whisker plot 1420 shows the distribution of predicted neoepitopes across subjects with and without HLA loss of heterozygosity for each of the various cancer types. Statistical analysis was performed with the Mann-Whitney U test with Bonferroni correction. Box and whisker plot 1420 shows evidence of pan-cancer in correlation with neoantigen load (p=0.03). Additionally, box and whisker plot 1425 shows correlation between HLA loss of heterozygosity and CD274 expression (PD-L1) across each of the various cancer types (p=0.02), and box and whisker plot 1430 shows correlation between HLA loss of heterozygosity and microsatellite instability (MSI) status across each of the various cancer types (p=0.01). More neoantigens were predicted to bind to lost HLA alleles than to their homologous counterparts (Wilcoxon rank sum test, p = 0.01), thereby suggesting that HLA heterozygosity loss contributes to selective exposure of antigens to the immune system.

[0097] VII. Tumor cell response to immune checkpoint blockade therapy The changes in allele-specific neoantigen composition detected in a head and neck squamous cell carcinoma cohort suggest that HLA loss of heterozygosity alters tumor evolution in response to immune checkpoint blockade therapy. This observation confirms that HLA sequence variability is a component of an effective immune checkpoint blockade response by tumor cells. Although larger cohorts with detailed response data will be required to confirm the impact of HLA loss of heterozygosity on subject response and survival, the examples presented below suggest that accurate detection of HLA loss of heterozygosity may be a factor in target selection for checkpoint immunotherapy and cancer vaccines.

[0098] HLA loss of heterozygosity appears to apply limited evolutionary pressure during tumor growth, whereas immune checkpoint inhibitors function to increase immune pressure. Therefore, we investigated the impact of HLA loss of heterozygosity on response to immunotherapy. Because HLA loss of heterozygosity severely reduces the immune peptidome by eliminating several HLA alleles, it was reasoned that HLA loss of heterozygosity should impair response to immunotherapy through reduced MHC presentation.

[0099] To identify HLA loss of heterozygosity in response to immunotherapy, we performed an experiment in a cohort of seven subjects with head and neck squamous cell carcinoma who received a single dose of PD-1 inhibitor (nivolumab) to identify HLA loss of heterozygosity. Corresponding pre- and post-treatment tumor biopsies from the subjects were sequenced. We characterized the interaction between germline variability in HLA sequences and pre-treatment somatic changes to antigen presentation machinery. Using the DASH model, we found four subjects with HLA loss of heterozygosity. In addition, we found beta2 microglobulin loss of heterozygosity in one subject and somatic mutations in HLA alleles in the other two subjects. All three subjects with the most germline HLA sequence diversity suffered from HLA loss of heterozygosity, and the subject with the greatest diversity also had beta2 microglobulin loss of heterozygosity.

[0100] (a) Cohort population Matched pre- and post-intervention normal, tumor, and plasma samples were collected from a cohort of seven subjects with head and neck squamous cell carcinoma. After baseline sampling, all subjects received a single dose of nivolumab followed by curative resection of the primary tumor mass approximately one month later, if feasible, or a second biopsy if resection was not practical. Due to the resection protocol, RECIST criteria were not used to assess response in resected subjects. Solid tumor and matched normal samples were profiled.

[0101] (b) Neoantigen expansion for immunotherapy responses For each subject with pre- and post-treatment samples, the DASH model can be used to predict the occurrence of HLA loss of heterozygosity. In addition, an HLA evolutionary diversity score can be determined to detect HLA somatic mutations and beta 2 microglobulin loss of heterozygosity. The potential epitopes of each mutation detected before or after treatment can be predicted to detect binding with all subject-specific alleles as described above. Neoepitopes can be identified from mutations observed after treatment but not before treatment. In some cases, neoepitopes are predicted to bind to multiple HLA alleles. Some neoepitopes that bind to homozygous alleles can be excluded. A Wilcoxon paired rank sum test can be performed to evaluate the statistical significance of the number of new neoepitopes predicted to bind to lost HLA-A / B alleles and their retained homologous alleles.

[0102] FIG. 15 shows experimental data identifying neoantigen expansion in response to immune checkpoint inhibitor therapy for head and neck squamous cell carcinoma subjects, according to some embodiments. Clones with neoantigens presentable by lost alleles were predicted to expand in response to immunotherapy. Furthermore, clones with neoantigens presentable by intact alleles were predicted to shrink in response to immunotherapy. This experimental data corresponds to the analysis of a head and neck squamous cell carcinoma cohort, due to the high prevalence of HLA loss of heterozygosity (44% of subjects), providing evidence of positive selection. Furthermore, one of the treatment regimens applied to this cohort included a PD-1 inhibitor (pembrolizumab), which is a first-line treatment for patients with PD-L1-positive head and neck squamous cell carcinoma. Previous data suggests that only a small proportion of subjects show a strong response to PD-1 inhibitors. Furthermore, PD-1 inhibitor treatment with pembrolizumab alone resulted in adverse events leading to death in 8% of corresponding subjects. Thus, understanding the mechanisms leading to response can be used to increase the survival rate of subjects.

[0103] In FIG. 15, subject data 1505 shows that in the experimental data, seven subjects (n=7) with head and neck squamous cell carcinoma were biopsied before treatment with Nivolumab. Approximately one month later, each subject underwent either a repeat biopsy or resection. All tumor samples were profiled before and after treatment to reveal various immune-related information (including tumor mutation load, HLA typing, etc.).

[0104] The bars in bar plot 1510 show the HLA-I evolutionary branching scores for each subject in the cohort population identified in Section VII(a) herein. For HLA genes, the shaded boxes indicate somatic loss of heterozygosity, and the shaded boxes in the rows "HLA Mutation" and "B2M LOH" indicate HLA gene mutations or loss of heterozygosity in the beta 2 microglobulin gene, respectively. Homozygous alleles and alleles with very little difference are shown as grey boxes.

[0105] The circle plot 1515 shows the proportion of new post-treatment neoantigens predicted to bind to each HLA allele of the subject. Each portion of each circle plot is shaded differently to reveal whether the HLA allele was lost or retained. The outer circle shows the proportion of neoepitopes predicted to bind to all lost and retained alleles, and the inner circle shows the breakdown by specific allele. The value within the circle corresponds to the number of new neoepitopes predicted after treatment (multiple counts if predicted to be presented by multiple alleles). In the circle plot 1515, neoepitopes presented by homozygous alleles were excluded.

[0106] As shown in the pie plot 1515, it was found that in each subject with HLA loss of heterozygosity, more new post-treatment neo-antigens are predicted to bind to the deleted HLA alleles rather than the retained HLA alleles of the corresponding subject.

[0107] Scatter plot 1520 shows the pairwise relationship between counts of novel post-treatment allele-specific predicted neo-epitopes in retained and deleted HLA alleles. Only HLA-A and -B alleles are shown in scatter plot 1520. Statistical significance is assessed using Wilcoxon paired rank test. Because sequence diversity of HLA-A and -B alleles alone can affect response to immunotherapy, the number of novel post-treatment neo-antigens predicted to bind to HLA-A and -B alleles was compared to their homologous genes and found to be statistically significant across cohorts (p=0.027, Wilcoxon signed rank). This consistent shift in neo-antigen composition suggests that HLA heterozygosity loss acts as an evolutionary force in resistance to response during immunotherapy.

[0108] Figure 16 shows additional experimental data corresponding to HLA heterozygosity loss in immunotherapy treated tumors according to some embodiments. Box plot 1605 shows the distribution of HLA-I evolutionary distance scores in subjects with and without HLA heterozygosity loss. Statistical significance is performed using a Mann-Whitney U test. In the box plot 1605, subjects with the least diversity of germline HLA sequences (HLA-I evolutionary branch < 1) showed no somatic changes to the antigen presentation machinery.

[0109] Box and whisker plot 1610 shows the difference in estimates of tumor infiltrating CD8+ T cell quantification before and after treatment in subjects with or without HLA loss of heterozygosity. Statistical significance is performed using the Mann-Whitney U test. In the box and whisker plot 1610, a trend towards increased CD8+ T cells after treatment is observed in subjects without HLA loss of heterozygosity. In contrast, the same trend was not observed in subjects with HLA loss of heterozygosity. The differences between samples shown in the box and whisker plot 1610 suggest that a decrease in neoantigen diversity may reduce immune infiltration.

[0110] VIII. A Process for Predicting Loss of Heterozygosity in HLA Alleles Using Machine Learning Models FIG. 17 includes a flowchart 1700 illustrating an example of a method for predicting loss of heterozygosity at HLA alleles, according to certain embodiments. The operations described in the flowchart 1700 can be performed by a computer system implementing a trained machine learning model, such as the DASH model, for example. Although the flowchart 1700 may describe the operations as a sequential process, in various embodiments, many of the operations may be performed in parallel or simultaneously. Furthermore, the order of the operations may be rearranged. The operations may have additional steps not shown in the figure. Furthermore, some embodiments of the method can be implemented by hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, the program code or code segments for performing the associated tasks can be stored in a computer-readable medium, such as a storage medium.

[0111] In operation 1710, the computer system accesses the machine learning model. The machine learning model can be trained with a training dataset including (1) allele-specific features, (2) subject-specific features, and (3) whole exome features for a subject of the subject set. The allele-specific features can include, for a genomic region of an HLA allele, an adjusted B allele frequency corresponding to a ratio between a first B allele frequency of a heterozygous allele in the tumor sample corresponding to the genomic region and a second B allele frequency of a heterozygous allele in the genomic region and associated with one or more control samples, and a ratio between a first allele-specific coverage of the tumor sample corresponding to the genomic region and a second allele-specific coverage of one or more control samples corresponding to the genomic region. In some cases, the allele-specific features correspond to a genomic region of an HLA allele identified as having a somatic mutation.

[0112] The subject-specific features can include an estimated tumor purity value and an estimated tumor ploidy value corresponding to the subject's tumor sample. As used herein, tumor purity refers to the ratio of tumor cells to the total cells in the sample. As used herein, tumor ploidy refers to the average copy number of the entire tumor genome. The whole exome features can include an indication of whether, for an HLA allele, at least a portion of the adjacent genomic region surrounding the HLA allele has been deleted. An exemplary embodiment for training a machine learning model can be found in Section III of this disclosure.

[0113] The performance of the trained machine learning model can be evaluated by using one or more validation techniques. For example, the machine learning model can be validated using in silico cell line mixtures, target-specific primers and probes generated using digital PCR, and / or immunopeptidemics data corresponding to the training data set. Exemplary embodiments for validating the machine learning model can be found in Section IV of this disclosure.

[0114] In operation 1720, the computer system accesses sequence data corresponding to a biological sample of a particular subject. The biological sample may be a tissue sample of a particular subject that includes DNA derived from cancer cells. In some cases, the sequence data is derived from the biological sample and a reference sample that does not include cancer cells. The biological sample may include cell-free DNA, some of which may originate from healthy cells and some of which may originate from tumor cells. The sequence data may be profiled to identify various features corresponding to the biological sample. For example, the features may include comprehensive tumor mutation information, gene expression quantification, neo-antigen characterization, HLA alleles (types and mutations), and tumor microenvironment profiling.

[0115] In some cases, the sequence data is generated by using whole genome or whole exome sequencing on the biological sample to generate a plurality of sequence reads. In some cases, HLA genotyping is performed on the plurality of sequence reads to identify one or more HLA alleles corresponding to the sequence data. A reference sequence corresponding to the identified HLA allele can be obtained, and the sequence reads can be aligned to the retrieved reference sequence. After alignment, the allele-specific coverage in each genomic region can be determined for the identified HLA allele corresponding to the sequence data. In some cases, the aligned sequence data can be analyzed to identify allele-specific copy number changes from a particular HLA allele type. Additionally or alternatively, the sequencing data can be analyzed to estimate tumor purity (alternatively referred to as tumor cellularity) and tumor ploidy. Exemplary embodiments for generating sequence data can be found at least in Section I of this disclosure.

[0116] In operation 1730, the computer system processes the sequence data with a machine learning model to generate a result corresponding to the probability of whether there is a loss of heterozygosity at the identified HLA allele in the tissue sample of the particular subject. The machine learning model (e.g., a DASH model) uses the allele-specific data for each of the identified HLA alleles as input to generate a result. Other types of information corresponding to the identified HLA allele (e.g., an indication of whether at least a portion of the adjacent genomic region surrounding the HLA allele has been deleted) can be used as additional input to the trained machine learning model. The machine learning model can include one or more gradient boosting algorithms to process the above features of the sequence data to generate a result.

[0117] In some cases, the results are used to predict the reduced efficacy of immune checkpoint blockade therapy administered to a particular subject. Because tumor samples with predicted HLA heterozygosity loss can be identified as having a particular type of cancer based on the corresponding changes in peptide presentation, the results can be used to predict the particular type of cancer associated with the subject.

[0118] In operation 1740, the computer system outputs the results, after which process 1700 ends.

[0119] IX. Computing Environment FIG. 18 illustrates an example of a computer system 1800 for implementing some embodiments disclosed herein. The computer system 1800 can include a distributed architecture, in which some components (e.g., memory and processor) are part of an end-user device and some other similar components (e.g., memory and processor) are part of a computer server. In some cases, the computer system 1800 is a computer system that uses machine learning models to predict loss of heterozygosity in HLA alleles and includes at least a processor 1802, a memory 1804, a storage device 1806, input / output (I / O) peripherals 1808, communication peripherals 1810, and an interface bus 1812. The interface bus 1812 is configured to communicate, transmit, and transfer data, control, and commands between various components of the computer system 1800. The processor 1802 can include one or more processing units, such as a CPU, a GPU, a TPU, a systolic array, or a SIMD processor. The memory 1804 and the storage devices 1806 include computer-readable storage media, such as RAM, ROM, Electrically Erasable Programmable Read Only Memory (EEPROM), hard drives, CD-ROM, optical storage devices, magnetic storage devices, electronic non-volatile computer storage devices, such as Flash memory, and other tangible storage media. Any such computer-readable storage media can be configured to store instructions or program code embodying aspects of the present disclosure. The memory 1804 and the storage devices 1806 also include computer-readable storage media.

[0120] The computer readable storage medium includes a propagated data signal having computer readable program code embodied therein. Such a propagated signal may take any of a variety of forms, including but not limited to electromagnetic, optical, or any combination thereof. The computer readable storage medium includes any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transmit a program for use with computer system 1800.

[0121] Additionally, memory 1804 includes an operating system, programs, and applications. Processor 1802 is configured to execute stored instructions and includes, for example, logic processing units, microprocessors, digital signal processors, and other processors. For example, computing system 1800 can execute instructions (e.g., program code) that configure processor 1802 to perform one or more operations described herein. Program code includes, for example, code that implements training a DASH model, using a DASH model, accessing sequence data, and / or any other suitable application that performs one or more operations described herein. Instructions may include processor-specific instructions generated by a compiler or interpreter from code written in any suitable computer programming language (including, for example, C, C++, C#, Visual Basic, Java, Python, Perl, JavaScript, and ActionScript).

[0122] The program code can be stored in memory 1804 or any suitable computer readable medium and executed by processor 1802 or any other suitable processor. In some embodiments, all modules in a computer system for predicting loss of heterozygosity at HLA alleles are stored in memory 1804. In additional or alternative embodiments, one or more of these modules from the computer system are stored in different memory devices of different computing systems.

[0123] The memory 1804 and / or the processor 1802 may be virtualized and hosted in another computing system (e.g., a computing system in a cloud network or data center). The I / O peripherals 1808 include user interfaces (e.g., keyboards, screens (e.g., touch screens), microphones, speakers, other input / output devices) and computing components (e.g., image processing units, serial ports, parallel ports, universal serial buses, and other input / output peripherals). The I / O peripherals 1808 connect to the processor 1802 through any of the ports coupled to an interface bus 1812. The communications peripherals 1810 are configured to facilitate communications between the computer system 1800 and other computing devices in a communications network and include, for example, network interface controllers, modems, wireless and wired interface cards, antennas, and other communications peripherals. For example, computing system 1800 can use the network interface device of communications peripheral 1810 to communicate over a data network with one or more other computing devices (e.g., computing devices used to train and validate DASH models, computing devices that display output generated by DASH models).

[0124] Although the subject matter of the present invention has been described in detail with respect to certain embodiments thereof, it will be understood that those skilled in the art, upon gaining the foregoing understanding, can easily generate modifications, variations, and equivalents to such embodiments. It is therefore to be understood that the present disclosure is presented for purposes of illustration and not limitation, and does not exclude the inclusion of modifications, variations, and / or additions to the subject matter of the present invention as would be readily apparent to those skilled in the art. Indeed, the methods and systems described herein may be embodied in other various forms, and furthermore, various omissions, substitutions, and changes in the form of the methods and systems described herein may be made without departing from the spirit of the present disclosure. The appended claims and their equivalents are intended to cover such forms or modifications as are within the scope and spirit of the present disclosure.

[0125] Unless otherwise indicated, throughout this specification, descriptions using terms such as "processing," "computing," "calculating," "determining," and "determining" should be understood to refer to operations or processes of a computing device (e.g., one or more computers, or similar electronic computing device(s)) that manipulates or transforms data represented as physical, electronic, or magnetic quantities in the memory, registers, or other information storage, transmission, or display devices of a computing platform.

[0126] The system(s) discussed herein are not limited to any particular hardware architecture or configuration. A computing device may include any suitable arrangement of components that provides results conditioned on one or more inputs. Suitable computing devices include general-purpose microprocessor-based computing systems that access stored software that programs or configures the computing system from a general-purpose computing device to a specialized computing device that implements one or more embodiments of the inventive subject matter. Any suitable programming, scripting, or other type of language or combination of languages ​​may be used to implement the teachings contained herein into the software used to program or configure a computing device.

[0127] Certain embodiments of the methods disclosed herein may be implemented in operation of such a computing device. For example, the order of the blocks presented in the above examples may be changed, e.g., the blocks may be rearranged, combined, and / or divided into sub-blocks. Certain blocks or processes may be implemented in parallel.

[0128] Conditional language used herein, such as "can," "could," "might," "may," "for example," and the like, among others, is intended to generally convey that certain examples include certain features, elements, and / or steps and other examples do not, unless expressly stated otherwise or understood otherwise within the context in which it is used. Thus, such conditional language is not intended to generally imply that features, elements, and / or steps are in any way required by one or more examples, or that one or more examples necessarily include logic for determining whether those features, elements, and / or steps are included in or performed in any particular example, with or without author input or direction.

[0129] "Comprising," "including," "having," and the like are synonymous and are used in an inclusive, open-ended manner and do not exclude additional elements, features, acts, operations, and the like. The term "or" is also used in an inclusive (rather than exclusive) sense, so that, for example, when used to connect a list of elements, the term "or" means one, some, or all of the elements in the list. The use of "adapted to" or "configured to" herein is intended as an open and inclusive statement that does not exclude devices adapted or configured to perform additional tasks or steps. Additionally, the use of "based on" is intended to be open and inclusive in that a process, step, calculation, or other operation "based on" one or more recited conditions or values ​​may in fact be based on additional conditions or values ​​other than those recited. Similarly, the use of "based at least in part on" is intended to be open and inclusive in that a process, step, calculation, or other operation "based on" one or more recited conditions or values ​​may in fact be based on additional conditions or values ​​other than those recited. The headings, lists, and numbering contained herein are for ease of explanation only and are not meant to be limiting.

[0130] The various features and processes described above may be used independently of one another or may be combined in various ways. All possible combinations and subcombinations are intended to fall within the scope of the present disclosure. Furthermore, in some implementations, certain method or process blocks may be omitted. Also, the methods and processes described herein are not limited to any particular order, and the blocks or states thereof may be performed in other orders as appropriate. For example, the blocks or states described may be performed in an order other than that specifically disclosed, or multiple blocks or states may be combined into a single block or state. The example blocks or states may be performed sequentially, in parallel, or in any other manner. Blocks and states may be added to or deleted from the examples of the present disclosure. Similarly, the example systems and components described herein may be configured differently than described. For example, elements may be added, deleted, or rearranged as compared to the examples of the present disclosure.

Claims

1. (a) accessing a machine learning model, the machine learning model being trained with one or more training feature sets corresponding to human leukocyte antigen (HLA) alleles identified in a tumor sample corresponding to a subject of a subject set, and for genomic regions of the HLA alleles, obtaining one or more of the following (i) through (iii): (i) an adjusted B allele frequency that corresponds to the ratio between a first B allele frequency of a heterozygous allele in the tumor sample corresponding to the genomic region and a second B allele frequency of a heterozygous allele within the genomic region and associated with one or more control samples; (ii) a ratio between a first allele-specific coverage of the tumor sample corresponding to the genomic region and a second allele-specific coverage of the one or more control samples corresponding to the genomic region; and (iii) an indication of whether at least a portion of the flanking genomic region surrounding the HLA allele has been deleted; and determining each of said plurality of (b) receiving sequence data corresponding to a particular subject's biological sample; and (c) processing the sequence data with the machine learning model to generate a result corresponding to the probability of whether loss of heterozygosity is present at the identified HLA allele in the biological sample of the particular subject; and (d) outputting the results; method.

2. The method of claim 1 , wherein the sequence data is whole-exome sequencing data.

3. The method of claim 1 , wherein the sequence data is whole genome sequencing data.

4. 2. The method of claim 1, wherein the machine learning model is trained using the training dataset further comprising a tumor purity value and a tumor ploidy value corresponding to the subject.

5. 2. The method of claim 1, wherein the training dataset is generated using reference sequences for the HLA alleles corresponding to the subject.

6. The method of claim 1 , wherein the machine learning models include one or more trained gradient boosting algorithms.

7. 2. The method of claim 1, further comprising predicting a decrease in efficacy of immune checkpoint blockade therapy to be administered to the particular subject based on the result.

8. 10. The method of claim 1, wherein the biological sample of the particular subject comprises one or more cancer cells.

9. 2. The method of claim 1, further comprising predicting one or more neo-antigens corresponding to the HLA alleles identified in the biological sample of the particular subject based on the results.

10. 2. The method of claim 1, wherein processing the sequence data with the machine learning model comprises determining allele-specific data for identified HLA alleles from the sequence data.

11. The method of claim 10, wherein the HLA alleles are identified from the sequence data by applying HLA genotyping to the sequence data.

12. one or more data processors; a non-transitory computer readable storage medium containing instructions that, when executed on said one or more data processors, cause said one or more data processors to perform the method of claim 1.

13. 13. A computer program product tangibly embodied on a non-transitory machine-readable storage medium, the computer program product comprising instructions configured to cause one or more data processors to perform the method of claim 1.

Images