Neoantigen identification using generic allele model

Abstract

A method for identifying neoantigens that may be presented by an MHC allele on the surface of a tumor cell of a subject. A peptide sequence of a tumor neoantigen and a peptide sequence of an MHC allele are obtained by sequencing tumor cells of the subject. The peptide sequence of the tumor neoantigen and the peptide sequence of the MHC alleles are input into a machine learning presentation model to produce presentation possibilities of the tumor neoantigen, each presentation likelihood representing a likelihood that a neoantigen is presented by at least one of the MHC alleles on a tumor cell surface of the subject. A subset of the neoantigens is selected based on the presentation likelihood.

Description

[0001] This application is a divisional application of the patent application with application number 201980019430.9 and invention name “New Antigen Identification Using a Pan-Allele Model”.

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] This application claims the benefit of and priority to U.S. Provisional Application No. 62 / 636,061, filed February 27, 2018. The contents of the above-referenced application are incorporated by reference in their entirety. Background Art

[0004] Therapeutic vaccines and T cell therapies based on tumor-specific neoantigens hold great promise as the next generation of personalized cancer immunotherapy. 1–3 Cancers with high mutational burdens, such as non-small cell lung cancer (NSCLC) and melanoma, are particularly attractive targets for such therapies given their relatively high likelihood of generating neoantigens. 4,5 Early evidence suggests that neoantigen-based vaccination can elicit T cell responses 6 And T cell therapy targeting neoantigens can, in some cases, induce tumor regression in selected patients. 7 Both MHC class I and class II influence T cell responses 70-71 .

[0005] However, identification of neoantigens and T cells that recognize them have become a key component in assessing tumor response. 77,110 , check tumor progression 111 and designing the next generation of personalized therapies 112 Current techniques for identifying neoantigens are time-consuming and labor-intensive. 84,96 , or not precise enough 87,91–93 Although it has recently been shown that T cells that recognize neoantigens are the main component of TIL 84,96,113,114 and circulates in the peripheral blood of cancer patients 107 However, current methods for identifying neoantigen-reactive T cells have a combination of the following three limitations: (1) they rely on clinical specimens that are difficult to obtain, such as TILs; 97,98 or leukaphereses 107 (2) It requires screening of impractically large peptide libraries 95 Or (3) it relies on MHC multimers, which are practically available only for a small number of MHC alleles.

[0006] In addition, the proposed preliminary approach incorporates mutation-based analysis using next-generation sequencing, RNA gene expression, and MHC binding affinity prediction of candidate neoantigen peptides. 8However, none of these proposed methods can mimic the entire epitope generation process, which, in addition to gene expression and MHC binding, comprises many steps (e.g., TAP transport, proteasomal cleavage, MHC binding, transport of the peptide-MHC complex to the cell surface and / or TCR recognition of MHC-I; endocytosis or autophagy, cleavage by extracellular or lysosomal proteases (e.g., cathepsins), competition with CLIP peptides for HLA-DM-catalyzed HLA binding, transport of the peptide-MHC complex to the cell surface and / or TCR recognition of MHC-II). 9 Therefore, existing methods may have the problem of low positive predictive value (PPV). Figure 1A )

[0007] In fact, analyses of peptides presented by tumor cells by multiple research groups have shown that less than 5% of the peptides predicted to be presented using gene expression and MHC binding affinity are actually found on the tumor surface MHC. 10,11 ( Figure 1B This low correlation between binding prediction and MHC presentation is further supported by the recent observation that targeting the number of mutations alone for checkpoint inhibitor response fails to improve prediction accuracy for binding-restricted neoantigens. 12

[0008] This low positive predictive value (PPV) of existing presentation prediction methods raises questions about the design of neoantigen-based vaccines and neoantigen-based T cell therapies. If prediction methods with low PPV are used to design vaccines, most patients are unlikely to receive therapeutic neoantigens, and a few patients may receive more than one neoantigen (even assuming that all presented peptides are immunogenic). Similarly, if therapeutic T cells are designed based on predictions with low PPV, most patients are unlikely to receive T cells reactive to tumor neoantigens, and the time and physical resource costs of using downstream laboratory techniques to identify predictive neoantigens after prediction may be prohibitive. Therefore, neoantigen vaccination and T cell therapy with current methods are unlikely to be successful in a large number of subjects with tumors. ( Figure 1C )

[0009] Furthermore, previous approaches have used only cis-acting mutations to generate candidate neoantigens, largely neglecting to consider other sources of neo-ORFs, including splicing factor mutations that arise in multiple tumor types and lead to aberrant splicing of many genes. 13 , and mutations that create or remove protease cleavage sites.

[0010] Finally, standard approaches to tumor genome and transcriptome analysis may miss somatic mutations that generate candidate neoantigens due to suboptimal conditions for library construction, exome and transcriptome capture, sequencing, or data analysis. Similarly, standard tumor profiling methods may inadvertently promote sequence artifacts or germline polymorphisms as neoantigens, leading to inefficient use of vaccine capabilities or the risk of autoimmunity, respectively. Summary of the Invention

[0011] Disclosed herein is an optimization method for identifying and selecting neoantigens for personalized cancer vaccines, for T cell therapy, or both. First, a method for analyzing tumor exomes and transcriptomes using next-generation sequencing (NGS) to identify the optimization of neoantigen candidates is proposed. These methods are based on standard NGS tumor analysis methods to ensure that the highest sensitivity and specificity of neoantigen candidates are promoted within all categories of genomic changes. Secondly, a novel method for selecting high PPV neoantigens is proposed to overcome specificity problems and ensure that neoantigens intended to be included in vaccines and / or as targets for T cell therapy are more likely to trigger anti-tumor immunity. Depending on the embodiment, these methods include trained statistical regression or nonlinear deep learning models that are configured to predict the presentation of peptides of various lengths, thereby sharing the statistical strength of peptides across different lengths on the basis of pan-alleles. The model is able to predict the likelihood that a peptide will be presented by any MHC allele (including unknown MHC alleles that the model has not previously encountered during training). Nonlinear deep learning models can be specifically designed and trained to treat different MHC alleles in the same cell as independent, thus resolving the problem of linear models where different MHC alleles interfere with each other. Finally, this addresses other issues that need to be considered in the design and manufacture of personalized vaccines based on neoantigens, as well as the production of personalized neoantigen-specific T cells for T cell therapy.

[0012] The models disclosed herein outperform state-of-the-art predictors trained on binding affinity and earlier predictors based on MS peptide data by up to an order of magnitude. By more reliably predicting peptide presentation, the model can identify neoantigen-specific or tumor antigen-specific T cells for personalized therapy in a more time- and cost-effective manner using a clinical practice approach that uses a limited amount of patient peripheral blood, screens a small number of peptides per patient, and does not necessarily rely on MHC multimers. However, in another embodiment, the models disclosed herein can use MHC multimers to identify tumor antigen-specific T cells in a more time- and cost-effective manner by reducing the number of peptides bound to MHC multimers that need to be screened in order to identify neoantigen- or tumor antigen-specific T cells.

[0013] The predictive performance of the model disclosed in this paper on a TIL neoepitope dataset and a prospective neoantigen-reactive T cell identification task demonstrates that it is now possible to obtain therapeutically useful neoepitope predictions by modeling HLA processing and presentation. In summary, this work provides practical in silico antigen identification for antigen-targeted immunotherapy, thereby accelerating the process of curing patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] These and other features, aspects and advantages of the present invention will be better understood with reference to the following description and accompanying drawings, in which:

[0015] Figure 1A Shown are current clinical methods used to identify neoantigens.

[0016] Figure 1B It was shown that <5% of the predicted binding peptides were presented on tumor cells.

[0017] Figure 1C Showing the impact of the specificity problem in neoantigen prediction.

[0018] Figure 1D showed that binding prediction is insufficient for neoantigen identification.

[0019] Figure 1E The probability of MHC-I presentation is shown as a function of peptide length.

[0020] Figure 1F Shown are exemplary peptide spectra generated from the Promega dynamic range standard.

[0021] Figure 1G Shows how adding features increases the positive predictive value of the model.

[0022] Figure 2A is a summary of the environment used to identify the potential for peptide presentation in a patient, according to one embodiment.

[0023] Figure 2B and 2C A method of obtaining presentation information according to one embodiment is shown.

[0024] Figure 3 is a high-level block diagram illustrating the computer logic components of a presentation authentication system according to one embodiment.

[0025] Figure 4 A set of example training data is shown according to one embodiment.

[0026] Figure 5 An example network model associated with MHC alleles is shown.

[0027] Figure 6 An example network model NN shared by MHC alleles is shown according to one embodiment. H (·).

[0028] Figure 7 Shown is the generation of presentation probabilities for peptides associated with one MHC allele using an example network model.

[0029] Figure 8 Shown is the generation of presentation probabilities for peptides associated with one MHC allele using an example network model.

[0030] Figure 9 Shown is the generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model.

[0031] Figure 10 Shown is the generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model.

[0032] Figure 11 Shown is the generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model.

[0033] Figure 12 Shown is the generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model.

[0034] Figure 13 An example network model NN shared by MHC alleles is shown according to one embodiment. H (·).

[0035] Figure 14 An example network model is shown that is not associated with MHC alleles.

[0036] Figure 15 Shown is the use of an example network model shared by MHC alleles to generate presentation probabilities for peptides associated with MHC alleles.

[0037] Figure 16 Shown are the precision / recall curves for the first test sample output by a pan-allele model comprising a neural network and trained on samples comprising the tested HLA allele, and the precision / recall curves for the first test sample output by a pan-allele model comprising a neural network but not trained on samples comprising the tested HLA allele.

[0038] Figure 17Shown are the precision / recall curves for the second test sample output by a pan-allele model comprising a neural network and trained on samples comprising the tested HLA allele, and the precision / recall curves for the second test sample output by a pan-allele model comprising a neural network but not trained on samples comprising the tested HLA allele.

[0039] Figure 18 Shown are the precision / recall curve for the third test sample output by a pan-allele model comprising a neural network and trained on samples comprising the tested HLA allele, and the precision / recall curve for the third test sample output by a pan-allele model comprising a neural network but not trained on samples comprising the tested HLA allele.

[0040] Figure 19 Shown are precision / recall curves output by a pan-allelic model comprising a neural network, a random forest model, a quadratic discriminant model, and an MHC Flurry model trained on samples comprising the tested HLA alleles.

[0041] Figure 20 Shown are precision / recall curves for the first test sample output by a pan-allele model comprising a neural network, a random forest model, a quadratic discriminant model, and an MHC Flurry model trained on samples comprising the tested HLA allele.

[0042] Figure 21 Shown are precision / recall curves for the second test sample output by a pan-allele model comprising a neural network, a random forest model, a quadratic discriminant model, and an MHC Flurry model trained on samples comprising the tested HLA allele.

[0043] Figure 22 Shown are precision / recall curves for the third test sample output by a pan-allele model comprising a neural network, a random forest model, a quadratic discriminant model, and an MHC Flurry model trained on samples comprising the tested HLA allele.

[0044] Figure 23A Shown are the sample frequency distributions of mutation burden in NSCLC patients.

[0045] Figure 23B Shown are the numbers of neoantigens presented in simulated vaccines for patients selected based on whether they met inclusion criteria for a minimum mutation load, according to one embodiment.

[0046] Figure 23C According to one embodiment the number of neoantigens presented in the simulated vaccines is compared between selected patients associated with a vaccine comprising a treatment subset identified based on the presentation model and selected patients associated with a vaccine comprising a treatment subset identified by the state of the art model.

[0047] Figure 23D The number of neoantigens presented in the simulated vaccines was compared between selected patients associated with a vaccine comprising a treatment subset identified based on a single independent allele presentation model for HLA-A*02:01 and selected patients associated with a vaccine comprising a treatment subset identified based on a dual independent allele presentation model for HLA-A*02:01 and HLA-B*07:02. According to one embodiment, the vaccine capacity was set to v = 20 epitopes.

[0048] Figure 23E According to one embodiment the number of neoantigens presented in a simulated vaccine was compared between patients selected based on mutational load and patients selected by expected utility score.

[0049] Figure 24 The positive predictive value (PPV) of the pan-allelic presentation model at 40% recall was compared when the pan-allelic presentation model using the presentation hotspot parameter was tested on five hold-out test samples with the pan-allelic presentation model not using the presentation hotspot parameter.

[0050] Figure 25A The proportion of somatic mutations recognized by T cells (e.g., pre-existing T cell responses) among the top 5, top 10, and top 20 somatic mutations identified using standard HLA binding affinity predictions at a >2 TPM threshold for gene expression as determined by RNA-seq, allele-specific neural network models, and pan-allelic neural network models was compared for a test set of 12 different test samples, each taken from a patient with at least one pre-existing T cell response.

[0051] Figure 25BThe proportion of minimal neoepitopes recognized by T cells (e.g., pre-existing T cell responses) among the top 5, top 10, and top 20 minimal neoepitopes identified using standard HLA binding affinity predictions at a threshold of >2 TPM for gene expression as determined by RNA-seq, allele-specific neural network models, and pan-allelic neural network models was compared for a test set of 12 different test samples, each taken from a patient with at least one pre-existing T cell response.

[0052] Figure 26A Depicted are the detection of T cell responses to patient-specific neoantigen peptide libraries for 9 patients.

[0053] Figure 26B Depicted are the detection of T cell responses to individual patient-specific neoantigen peptides for four patients.

[0054] Figure 26C Example images of ELISpot wells from patient CU04 are depicted.

[0055] Figure 27A Depicted are results from control experiments with neoantigens in HLA-matched healthy donors.

[0056] Figure 27B Depicted are results from control experiments with neoantigens in HLA-matched healthy donors.

[0057] Figure 28 Describes the Figure 26A For each donor and each in vitro expansion depicted, T cell responses were detected against a PHA positive control.

[0058] Figure 29A Depicted is the detection of T cell responses for patient CU04 to each individual patient-specific neoantigen peptide in pool #2.

[0059] Figure 29B Depicted are the detection of T cell responses to individual patient-specific neoantigen peptides for each of three visits for patient CU04 and each of two visits for patient 1-024-002, each visit occurring at a different time point.

[0060] Figure 29C Depicted are the detection of T cell responses to individual patient-specific neoantigen peptides and patient-specific neoantigen peptide pools for each of two visits for patient CU04 and for each of two visits for patient 1-024-002, with each visit occurring at a different time point.

[0061] Figure 30 Describes the Figure 26A Detection of T cell responses to two patient-specific neoantigen peptide pools and a DMSO negative control.

[0062] Figure 31A Plotted are the precision-recall curves for the pan-allelic model and the allele-specific model for each test sample 0 containing an MHC class II allele.

[0063] Figure 31B Plotted are the precision-recall curves for the pan-allele model and the allele-specific model for each test sample 1 containing an MHC class II allele.

[0064] Figure 31C Plotted are the precision-recall curves for the pan-allelic model and the allele-specific model for each test sample 2 containing an MHC class II allele.

[0065] Figure 31D Plotted are the precision-recall curves for the pan-allele model and the allele-specific model for each test sample 4 containing an MHC class II allele.

[0066] Figure 32 Described is a method for sequencing the TCRs of neoantigen-specific memory T cells from the peripheral blood of NSCLC patients.

[0067] Figure 33 Depicted are exemplary embodiments of TCR constructs for introducing a TCR into a recipient cell.

[0068] Figure 34 Depicted are exemplary P526 construct backbone nucleotide sequences for cloning TCRs into expression systems for therapeutic development.

[0069] Figure 35 Depicted are exemplary construct sequences for cloning a patient neoantigen-specific TCR, clonotype 1 TCR, into an expression system for therapy development.

[0070] Figure 36 Depicted are exemplary construct sequences used to clone a patient neoantigen-specific TCR, clonotype 3, into an expression system for therapy development.

[0071] Figure 37 is a flow chart of a method for providing customized neoantigen-specific therapy to a patient, according to an embodiment.

[0072] Figure 38 An example computer for implementing the entities shown in Figures 1 and 3 is shown. DETAILED DESCRIPTION

[0073] I. Definition

[0074] Generally, the terms used in the claims and description are intended to be interpreted as having their ordinary meanings as understood by one of ordinary skill in the art. For clarity, certain terms are defined below. If there is a conflict between the ordinary meanings and the provided definitions, the provided definitions shall apply.

[0075] As used herein, the term "antigen" is a substance that induces an immune response.

[0076] As used herein, the term "neoantigen" is an antigen having at least one change that makes it different from the corresponding wild-type parent antigen, for example, the change is a tumor cell mutation or a tumor cell-specific post-translational modification. Neoantigens may include a polypeptide sequence or a nucleotide sequence. Mutations may include frameshift or non-frameshift insertions and deletions, missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or any genomic or expression changes that produce neoORFs. Mutations may also include splice variants. Tumor cell-specific post-translational modifications may include abnormal phosphorylation. Tumor cell-specific post-translational modifications may also include spliced ​​antigens produced by the proteasome. See Liepe et al., A large fraction of HLA class I ligands are proteasome-generated spliced ​​peptides; Science. 2016 Oct 21; 354(6310): 354-358.

[0077] As used herein, the term "tumor neoantigen" is a neoantigen that is present in tumor cells or tissues of a subject but not in corresponding normal cells or tissues of the subject.

[0078] As used herein, the term "neoantigen-based vaccine" is a vaccine construct based on one or more neoantigens, such as a plurality of neoantigens.

[0079] As used herein, the term "candidate neoantigen" is a mutation or other abnormality that produces a new sequence that may represent a neoantigen.

[0080] As used herein, the term "coding region" is the portion of a gene that encodes a protein.

[0081] As used herein, the term "coding mutation" is a mutation present in a coding region.

[0082] As used herein, the term "ORF" refers to open reading frame.

[0083] As used herein, the term "NEO-ORF" is a tumor-specific ORF that results from a mutation or other aberration such as splicing.

[0084] As used herein, the term "missense mutation" is a mutation that results in the substitution of one amino acid with another amino acid.

[0085] As used herein, the term "nonsense mutation" is a mutation that results in the substitution of an amino acid by a stop codon.

[0086] As used herein, the term "frameshift mutation" is a mutation that results in a change in the protein framework.

[0087] As used herein, the term "indel" is an insertion or deletion of one or more nucleic acids.

[0088] The term "identity" as used herein in the context of two or more nucleic acids or peptide sequences refers to when comparing and aiming to reach maximum correspondence, as measured using one of the sequence comparison algorithms described below (such as BLASTP and BLASTN, or other algorithms available to the technician) or two or more sequences or subsequences obtained by visual inspection that have a specified percentage of nucleotides or amino acid residues are identical. Depending on application, the "identity" percentage can be present in a certain region of the sequence compared, for example, in a functional domain, or in the total length of two sequences to be compared.

[0089] For sequence comparison, usually, a sequence serves as a reference sequence to compare with a test sequence. When using a sequence comparison algorithm, the test sequence and the reference sequence are input into a computer, subsequence coordinates are specified if necessary, and sequence algorithm program parameters are specified. Then, the sequence comparison algorithm calculates the sequence identity percentage of the test sequence relative to the reference sequence based on the specified program parameters. Alternatively, sequence similarity or dissimilarity can be determined by combining specific nucleotides at selected sequence positions (e.g., sequence motifs), or by the presence or absence of specific amino acids for the translated sequence.

[0090] Optimal alignment of sequences for comparison can be performed, for example, by the local homology algorithm of Smith and Waterman, Adv. Appl. Math. 2:482 (1981); the homology alignment algorithm of Needleman and Wunsch, J. Mol. Biol. 48:443 (1970); the search similarity method of Pearson and Lipman, Proc. Nat'l. Acad. Sci. USA 85:2444 (1988); computerized implementations of these algorithms (GAP, BESTFIT, FASTA, and TFASTA in the Wisconsin Genetics Software Package; Genetics Computer Group, 575 Science Dr., Madison, Wis.); or by visual inspection (see generally, Ausubel et al., infra).

[0091] One example of an algorithm suitable for determining percent sequence identity and sequence similarity is the BLAST algorithm described in Altschul et al., J. Mol. Biol. 215:403-410 (1990). Software for performing BLAST analyses is publicly available through the National Center for Biotechnology Information.

[0092] As used herein, the term "null stop or read-through" is a mutation that results in the removal of the natural stop codon.

[0093] As used herein, the term "epitope" is a specific portion of an antigen that is typically bound by an antibody or T-cell receptor.

[0094] As used herein, the term "immunogenicity" is the ability to elicit an immune response, for example, by T cells, B cells, or both.

[0095] As used herein, the terms "HLA binding affinity", "MHC binding affinity" mean the binding affinity between a specific antigen and a specific MHC allele.

[0096] As used herein, the term "bait" is a nucleic acid probe used to enrich specific DNA or RNA sequences from a sample.

[0097] As used herein, the term "variant" is a difference between a subject's nucleic acid and a reference human genome used as a control.

[0098] As used herein, the term "variant call" is an algorithmic determination of the presence of a variant, typically determined by sequencing.

[0099] As used herein, the term "polymorphism" is a germline variant, ie, a variant found in all DNA-bearing cells of an individual.

[0100] As used herein, the term "somatic variant" is a variant that arises in non-germline cells of an individual.

[0101] As used herein, the term "allele" is a form of a gene, or a form of a gene sequence, or a form of a protein.

[0102] As used herein, the term "HLA type" is the complementary sequence of the HLA gene alleles.

[0103] As used herein, the term "nonsense-mediated decay" or "NMD" is the degradation of mRNA by cells caused by premature stop codons.

[0104] As used herein, the term "trunk mutation" is a mutation that originates early in tumor development and is present in the majority of tumor cells.

[0105] As used herein, the term "subclonal mutation" is a mutation that originates late in tumor development and is present in only a small fraction of tumor cells.

[0106] As used herein, the term "exome" is a subset of a genome that encodes proteins. An exome can be all exons of a genome.

[0107] As used herein, the term "logistic regression" is a statistically derived regression model for binary data in which the logit of the probability that a dependent variable is equal to 1 is modeled as a linear function of the dependent variable.

[0108] As used herein, the term "neural network" is a machine learning model for classification or regression consisting of multiple layers of linear transformations followed by element-wise nonlinearities typically trained via stochastic gradient descent and backpropagation.

[0109] As used herein, the term "proteome" is the collection of all proteins expressed and / or translated by a cell, a population of cells, or an individual.

[0110] As used herein, the term "peptide group" is the collection of all peptides presented on the surface of a cell by MHC-I or MHC-II. The peptidome can refer to the characteristics of a cell or a group of cells (e.g., a tumor peptidome, meaning the union of the peptidomes of all cells that make up a tumor).

[0111] As used herein, the term "ELISPOT" means enzyme-linked immunosorbent spot assay, a common method used to monitor immune responses in humans and animals.

[0112] As used herein, the term "dextramer" is a dextran-based peptide-MHC multimer used for antigen-specific T cell staining in flow cytometry.

[0113] As used herein, the term "MHC multimer" is a peptide-MHC complex comprising multiple peptide-MHC monomer units.

[0114] As used herein, the term "MHC tetramer" is a peptide-MHC complex comprising four peptide-MHC monomer units.

[0115] As used herein, the term "tolerance or immune tolerance" is a state of immune unresponsiveness to one or more antigens, such as self-antigens.

[0116] As used herein, the term "central tolerance" is tolerance experienced in the thymus by deleting autoreactive T cell clones or by promoting the differentiation of autoreactive T cell clones into immunosuppressive regulatory T cells (Tregs).

[0117] As used herein, the term "peripheral tolerance" is tolerance experienced in the periphery by downregulating or anergizing autoreactive T cells that survived central tolerance, or by promoting the differentiation of these T cells into Tregs.

[0118] The term "sample" can include a single cell or multiple cells, or cell fragments, or an aliquot of body fluid obtained from a subject by means including venipuncture, excretion, ejaculation, massage, biopsy, needle aspiration, lavage sample, scraping, surgical incision or intervention, or other means known in the art.

[0119] The term "subject" encompasses cells, tissues, or organisms, human or non-human, whether in vivo, ex vivo, or in vitro, male or female. The term subject includes mammals, including humans.

[0120] The term "mammal" encompasses both humans and non-humans, and includes, but is not limited to, humans, non-human primates, canines, felines, murines, bovines, equines, and porcines.

[0121] The term "clinical factor" refers to a subject's condition, such as a measure of disease activity or severity. "Clinical factor" encompasses all markers of a subject's health status, including non-sample markers, and / or other characteristics of the subject, such as, but not limited to, age and gender. A clinical factor can be a score, a value, or a set of values ​​that can be obtained by evaluating a sample (or sample group) or subject from a subject under defined conditions. Clinical factors can also be predicted by markers and / or other parameters such as gene expression surrogates. Clinical factors can include tumor type, tumor subtype, and smoking history.

[0122] Abbreviations: MHC: major histocompatibility complex; HLA: human leukocyte antigen or human MHC locus; NGS: next-generation sequencing; PPV: positive predictive value; TSNA: tumor-specific neoantigen; FFPE: formalin-fixed paraffin-embedded; NMD: nonsense-mediated decay; NSCLC: non-small cell lung cancer; DC: dendritic cell.

[0123] As used in this specification and the appended claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise.

[0124] Any term not directly defined herein should be understood to have the meaning generally associated therewith as understood in the field of the present invention. Certain terms discussed herein are intended to provide additional guidance to practitioners in describing the compositions, devices, methods, etc. of various aspects of the present invention and their preparation or use. It should be understood that the same thing can be expressed in more than one way. Therefore, alternative wordings and synonyms may be used for any one or more of the terms discussed herein. It does not matter whether a term is elaborated or discussed herein. Some synonyms or alternative methods, materials, etc. are provided. Unless expressly stated, the description of one or more synonyms or equivalents does not exclude the use of other synonyms or equivalents. Examples, including the use of the term example, are for illustrative purposes only and are not intended to limit the scope and meaning of various aspects of the present invention herein.

[0125] All references, issued patents, and patent applications cited within the text of this specification are herein incorporated by reference in their entirety for all purposes.

[0126] II. Methods for Identifying Neoantigens

[0127] Disclosed herein is a method for identifying at least one neoantigen from one or more tumor cells of a subject that may be presented on the surface of the tumor cells by one or more MHC alleles. The method comprises obtaining exome, transcriptome, and / or whole-genome nucleotide sequencing data from tumor cells and normal cells of the subject. The nucleotide sequencing data is used to obtain a peptide sequence for each neoantigen in a set of neoantigens. The neoantigen set is identified by comparing the nucleotide sequencing data from the tumor cells with the nucleotide sequencing data from the normal cells. Specifically, the peptide sequence of each neoantigen in the set of neoantigens comprises at least one variation that differs from a corresponding wild-type peptide sequence identified from a normal cell of the subject. The method further comprises encoding the peptide sequence of each neoantigen in the set of neoantigens into a corresponding digital vector. Each digital vector comprises information describing the amino acids that constitute the peptide sequence and the position of the amino acids in the peptide sequence. The method further comprises obtaining exome, transcriptome, and / or whole-genome nucleotide sequencing data from the tumor cells of the subject. The nucleotide sequencing data is used to obtain a peptide sequence for each of the one or more MHC alleles of the subject. The peptide sequence of each of the one or more MHC alleles of the subject is encoded into a corresponding digital vector. Each digital vector contains information describing the amino acids that make up the peptide sequence of an MHC allele and the positions of these amino acids in the peptide sequence of the MHC allele. The method further includes inputting the digital vector encoding the peptide sequence of each of the neoantigens and the digital vector encoding the peptide sequence of each of the one or more MHC alleles into a machine learning presentation model to generate a presentation probability for each neoantigen in the neoantigen set. Each presentation probability represents the probability that the corresponding neoantigen will be presented on the surface of a tumor cell in a subject by one or more MHC alleles. The machine learning presentation model includes multiple parameters and functions. The multiple parameters are identified based on a training dataset. The training dataset includes: for each sample in a plurality of samples, a signature obtained by mass spectrometry of the presence of a peptide that binds to at least one MHC allele in the set of MHC alleles identified as present in the sample, a training peptide sequence encoded as a digital vector containing information describing the amino acids that make up the peptide and the positions of the amino acids in the peptide, and a training peptide sequence encoded as a digital vector containing information describing the amino acids that make up at least one MHC allele that binds to the sample's peptide and the positions of the amino acids in the MHC allele peptide. The function represents a relationship between the digital vector received as input by the machine learning presentation model and a presentation probability generated as output by the machine learning presentation model based on the digital vector and the plurality of parameters. The method further includes selecting a subset of the set of neoantigens based on the presentation probability to generate a set of selected neoantigens; and recovering the set of selected neoantigens.

[0128] In some embodiments, inputting a digital vector of peptide sequences encoding each of the neoantigens and a digital vector of peptide sequences encoding each of the one or more MHC alleles into a machine learning presentation model comprises: applying the machine learning presentation model to the peptide sequences of the neoantigens and to the peptide sequences of the one or more MHC alleles to generate a dependency score for each of the one or more MHC alleles. The dependency score of the MHC allele indicates whether the MHC allele will present the neoantigen based on a specific amino acid at a specific position of the peptide sequence. In further embodiments, inputting a digital vector of peptide sequences encoding each of the neoantigens and a digital vector of peptide sequences encoding each of the one or more MHC alleles into the machine learning presentation model further comprises: transforming the dependency score to obtain a corresponding independent allele probability for each MHC allele, thereby indicating the probability that the corresponding MHC allele will present the corresponding neoantigen; and combining the independent allele probabilities to generate a presentation probability of the neoantigen. In some embodiments, transforming the dependency score models the presentation of the neoantigen as mutually exclusive between one or more MHC alleles. In an alternative embodiment, inputting the digital vector of peptide sequences encoding each of the neoantigens and the digital vector of peptide sequences encoding each of the one or more MHC alleles into the machine learning presentation model further comprises: transforming the combination of dependency scores to generate a presentation likelihood. In such an embodiment, the transforming the combination of dependency scores models presentation of the neoantigen as interference between the one or more MHC alleles.

[0129] In some embodiments, the set of presentation probabilities is further identified by at least one or more allele non-interaction features. In such embodiments, the method further includes applying a machine learning presentation model to the allele non-interaction feature to generate a dependency score for the allele non-interaction feature. The dependency score indicates whether the peptide sequence of the corresponding neoantigen will be presented based on the allele non-interaction feature. In some embodiments, the method further includes combining the dependency score of each MHC allele in one or more MHC alleles with the dependency score of the allele non-interaction feature, transforming the dependency score of the combination of each MHC allele to generate an independent allele probability for each MHC allele, and combining the independent allele probabilities to generate a presentation probability. The independent allele probability of the MHC allele indicates the probability that the MHC allele will present the corresponding neoantigen. In an alternative embodiment, the method further includes combining the dependency score of each MHC allele with the dependency score of the allele non-interaction feature; and transforming the combined dependency score to generate a presentation probability.

[0130] In some embodiments, the one or more MHC alleles include two or more different MHC alleles.

[0131] In some embodiments, the peptide sequence comprises a peptide sequence that is not 9 amino acids in length.

[0132] In some embodiments, encoding the peptide sequence comprises encoding the peptide sequence using a one-hot encoding scheme.

[0133] In some embodiments, the plurality of samples includes at least one of: a cell line engineered to express a single MHC allele; a cell line engineered to express multiple MHC alleles; a human cell line obtained or derived from a plurality of patients; fresh or frozen tumor samples obtained from a plurality of patients; and fresh or frozen tissue samples obtained from a plurality of patients.

[0134] In some embodiments, the training dataset further comprises at least one of: data associated with measurements of peptide-MHC binding affinity of at least one of the peptides; and data associated with measurements of peptide-MHC binding stability of at least one of the peptides.

[0135] In some embodiments, the presentation probability set is further identified by the expression level of one or more MHC alleles in the subject as measured by RNA-seq or mass spectrometry.

[0136] In some embodiments, the presentation likelihood set is further identified by features comprising at least one of: predicted affinity between neoantigens in the neoantigen set and one or more MHC alleles; and predicted stability of the neoantigen-encoded peptide-MHC complex.

[0137] In some embodiments, the set of numerical possibilities is further identified by features comprising at least one of: a C-terminal sequence flanking the neo-antigen encoding peptide sequence within its source protein sequence; and an N-terminal sequence flanking the neo-antigen encoding peptide sequence within its source protein sequence.

[0138] In some embodiments, selecting the set of selected neoantigens comprises selecting neoantigens having an increased likelihood of being presented on the surface of the tumor cell relative to unselected neoantigens based on a machine learning presentation model.

[0139] In some embodiments, selecting the set of selected neoantigens comprises selecting neoantigens having an increased likelihood of inducing a tumor-specific immune response in a subject relative to unselected neoantigens based on a machine learning presentation model.

[0140] In some embodiments, selecting the set of selected neoantigens comprises selecting neoantigens that have an increased likelihood of being presented to natural T cells by professional antigen presenting cells (APCs) relative to unselected neoantigens based on a presentation model. In such embodiments, the APCs are optionally dendritic cells (DCs).

[0141] In some embodiments, selecting the set of selected neoantigens comprises selecting neoantigens having a reduced likelihood of undergoing central or peripheral tolerance suppression relative to unselected neoantigens based on a machine learning presentation model.

[0142] In some embodiments, selecting the set of selected neoantigens comprises selecting neoantigens that have a reduced likelihood of inducing an autoimmune response against normal tissue in the subject relative to unselected neoantigens based on a machine learning presentation model.

[0143] In some embodiments, the one or more tumor cells are selected from the group consisting of lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, stomach cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia and T-cell lymphocytic leukemia, non-small cell lung cancer and small cell lung cancer.

[0144] In some embodiments, the method further comprises generating an output for constructing a personalized cancer vaccine from the set of selected neoantigens. In such embodiments, the output of the personalized cancer vaccine may comprise at least one peptide sequence or at least one nucleotide sequence encoding the set of selected neoantigens.

[0145] In some embodiments, the machine learning presentation model is a neural network model. In such embodiments, the neural network model can be a single neural network model comprising a series of nodes arranged in one or more layers. The single neural network model can be configured to receive a digital vector encoding a peptide sequence of a plurality of different MHC alleles. In such embodiments, the neural network model can be trained by updating the parameters of the neural network model. In some embodiments, the machine learning presentation model can be a deep learning model comprising one or more layers of nodes.

[0146] In some embodiments, the training peptide sequences encoded as digital vectors comprising a set of information about a plurality of amino acids that constitute at least one MHC allele that binds to peptides of the sample and the positions of the amino acids in the at least one MHC allele do not include: peptide sequences of the subject's MHC alleles that are input into the machine learning presentation model to generate a set of presentation probabilities for the set of neo-antigens.

[0147] In certain aspects disclosed herein, at least one MHC allele that binds to a peptide from each of the plurality of samples in the training dataset belongs to a gene family to which one or more MHC alleles of the subject belong.

[0148] In some embodiments, the at least one MHC allele that binds to a peptide from each of the plurality of samples in the training dataset comprises one MHC allele. In alternative embodiments, the at least one MHC allele that binds to a peptide from each of the plurality of samples in the training dataset comprises more than one MHC allele.

[0149] In some embodiments, the one or more MHC alleles are class I MHC alleles.

[0150] Also disclosed herein is a computer system comprising a computer processor and a memory storing computer program instructions that, when executed by the computer processor, cause the computer processor to perform an embodiment of the method described above.

[0151] III. Identification of Tumor-Specific Mutations in Neoantigens

[0152] Also disclosed herein are methods for identifying certain mutations (e.g., variants or alleles present in cancer cells). Specifically, these mutations can be present in the genome, transcriptome, proteome, or exome of cancer cells in a subject with cancer, but not in the subject's normal tissue.

[0153] If the gene mutations in the tumor result only in changes in the amino acid sequence of the protein in the tumor, then these mutations are considered to be useful for immune targeting of the tumor. Useful mutations include: (1) non-synonymous mutations that result in different amino acids in the protein; (2) read-through mutations, in which the stop codon is modified or deleted, resulting in translation of a longer protein with a new tumor-specific sequence at the C-terminus; (3) splice site mutations that result in the inclusion of introns in the mature mRNA and thus produce a unique tumor-specific protein sequence; (4) chromosomal rearrangements (i.e., gene fusions) that produce chimeric proteins with tumor-specific sequences at the junction of two proteins; (5) frameshift mutations or deletions that produce new open reading frames with new tumor-specific protein sequences. Mutations can also include one or more of non-frameshift insertions and deletions, missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or any genomic or expression changes that produce neoORFs.

[0154] Peptides with mutations in tumor cells or mutant polypeptides resulting from, for example, splice site mutations, frameshift mutations, read-through mutations, or gene fusion mutations can be identified by sequencing DNA, RNA, or protein in tumor and normal cells.

[0155] Mutations can also include previously identified tumor-specific mutations. Known tumor mutations can be found in the Catalogue of Somatic Mutations in Cancer (COSMIC) database.

[0156] A variety of methods can be used to detect the presence of specific mutations or alleles in the DNA or RNA of an individual. Improvements in this area are to provide accurate, easy and cheap large-scale SNP genotyping. For example, several technologies have been described, including dynamic allele-specific hybridization (DASH), microplate array diagonal gel electrophoresis (MADGE), pyrophosphate sequencing, oligonucleotide specific ligation, TaqMan system and various DNA "chip" technologies, such as Affymetrix SNP chips. These methods are generally amplified by PCR target gene region. Some other methods are based on producing small signal molecules by invasive cleavage, followed by mass spectrometry or immobilized padlock probes and rolling circle amplification. Several methods known in the art for detecting specific mutations are summarized below.

[0157] Detection methods based on PCR can include multiple amplification of multiple markers simultaneously. For example, it is well known in the art that PCR primers are selected to produce PCR products that do not overlap in size and can be analyzed simultaneously. Alternatively, primers that are labeled in different ways and can thus be detected in different ways can be used to amplify different markers. Of course, detection methods based on hybridization can detect multiple PCR products in a sample in different ways. Other techniques that can multiplex multiple markers are known in the art.

[0158] Several methods have been developed to promote the analysis of single nucleotide polymorphisms in genomic DNA or cellular RNA. For example, single base polymorphisms can be detected by using dedicated exonuclease-resistant nucleotides, as disclosed in, for example, Mundy, CR (U.S. Patent No. 4,656,127). According to this method, a primer complementary to the allele sequence of the 3' end of the polymorphic site can hybridize with the target molecule obtained from a specific animal or human. If the polymorphic site on the target molecule contains a nucleotide complementary to the specific exonuclease-resistant nucleotide derivative present, the derivative will be incorporated into the end of the hybridization primer. This type of merging makes the primer resistant to exonucleases and thus allows its detection. Since the identity of the exonuclease-resistant derivative of the sample is known, the discovery that the primer produces resistance to exonucleases discloses that the nucleotides present in the target molecule polymorphic site are complementary to the nucleotide derivatives used in the reaction. The advantage of this method is that it does not need to measure a large amount of unrelated sequence data.

[0159] Solution-based methods can be used to determine the identity of the nucleotide at a polymorphic site. Cohen, D. et al. (French Patent 2,650,840; PCT Application No. WO91 / 02087). As described in the Mundy method of U.S. Patent No. 4,656,127, primers are used that are complementary to the allele sequence immediately 3' to the polymorphic site. This method uses labeled dideoxynucleotide derivatives to determine the identity of the nucleotide at that site; if the nucleotide is complementary to the nucleotide at the polymorphic site, the nucleotide is incorporated into the primer end. Goelet, P. et al. (PCT Application No. 92 / 15712) describe an alternative method, known as Genetic Bit Analysis or GBA. Goelet, P. et al.'s method uses a mixture of labeled terminators and primers that are complementary to the sequence at the 3' end of the polymorphic site. The incorporated labeled terminator is thus determined by the nucleotides present in the polymorphic site of the target molecule being evaluated and is complementary to the nucleotides present in the polymorphic site of the target molecule being evaluated. In contrast to the method of Cohen et al. (French Patent 2,650,840; PCT Application No. WO 91 / 02087), the method of Goelet, P. et al. can be a heterogeneous assay in which either the primer or the target molecule is immobilized on a solid phase.

[0160] Several primer-directed nucleotide incorporation procedures for determining polymorphic sites in DNA have been described (Komher, JS et al., Nucl. Acids. Res. 17:7779-7784 (1989); Sokolov, BP, Nucl. Acids Res. 18:3671 (1990); Syvanen, A.-C. et al., Genomics 8:684-692 (1990); Kuppuswamy, MN et al., Proc. Natl. Acad. Sci. (USA) 88:1143-1147 (1991); Prezant, TR et al., Hum. Mutat. 1:159-164 (1992); Ugozzoli, L. et al., GATA 9:107-112 (1992); Nyren, P. et al., Anal. Biochem. 208:171-175 (1993). These methods differ from GBA in that they utilize the incorporation of labeled deoxynucleotides to distinguish bases at polymorphic sites. In this format, since the signal is proportional to the number of deoxynucleotides incorporated, polymorphisms occurring in a run of the same nucleotide sequence can generate a signal proportional to the length of the run (Syvanen, A.-C. et al., Amer. J. Hum. Genet. 52:46-59 (1993)).

[0161] Many approaches directly obtain sequence information from millions of independent DNA or RNA molecules in parallel. Real-time single-molecule sequencing-by-synthesis technology relies on the detection of fluorescent nucleotides because these nucleotides are incorporated into the nascent DNA chain complementary to the sequencing template. In one method, oligonucleotides of 30-50 bases in length are covalently anchored to a glass coverslip at their 5' ends. These anchoring strands perform two functions. First, if the template is configured with a capture tail complementary to the surface-bound oligonucleotide, it acts as a capture site for the target template strand. These anchoring strands also serve as primers for template-guided primer extension, forming the basis for sequence reading. The capture primers serve as fixed sites for sequencing using multiple cycles of synthesis, detection, and dye-linker chemical cleavage to remove the dye. Each cycle consists of adding a polymerase / labeled nucleotide mixture, washing, imaging, and dye cleavage. In an alternative method, the polymerase is modified to have a fluorescent donor molecule and fixed to a glass slide, while each nucleotide is color-coded with an acceptor fluorescent moiety attached to a γ-phosphate. The system detects the interaction between a fluorescently labeled polymerase and the fluorescently modified nucleotides as they are incorporated into a de novo synthesized chain.Other sequencing-by-synthesis technologies also exist.

[0162] Any suitable sequencing while synthesis platform can be used to identify mutations. As described above, there are currently four main sequencing while synthesis platforms: the genome sequencer from Roche / 454 Life Sciences, the 1G analyzer from Illumina / Solexa, the SOLiD system from Applied BioSystems, and the Heliscope system from HelicosBiosciences. Pacific BioSciences and VisiGen Biotechnologies have also described sequencing while synthesis platforms. In some embodiments, multiple nucleic acid molecules sequenced are bound to a support (such as a solid support). In order to fix nucleic acid on a support, capture sequence / universal priming site can be added at the 3' and / or 5' end of the template. Nucleic acid can be bound to a support by hybridizing the capture sequence with a complementary sequence covalently attached to the support. Capture sequence (also known as universal capture sequence) is a nucleic acid sequence complementary to the sequence attached to the support, and the sequence can also serve as a universal primer.

[0163] As an alternative to capture sequences, one member of a coupling pair (such as an antibody / antigen, receptor / ligand, or anti-biotin-biotin pair, as described, for example, in U.S. Patent Application No. 2006 / 0252077) can be attached to each fragment to capture it on a surface coated with the corresponding second member of the coupling pair.

[0164] After capture, the sequence can be analyzed by, for example, single molecule detection / sequencing, including template-dependent sequencing-by-synthesis, as described in the Examples and U.S. Patent No. 7,283,337. In sequencing-by-synthesis, a surface-bound molecule is exposed to a plurality of labeled nucleotide triphosphates in the presence of a polymerase. The template sequence is determined by the order in which the labeled nucleotides are incorporated into the 3' end of the growing chain. This can be performed in real time or in a step-and-repeat format. For real-time analysis, different optical labels can be incorporated into each nucleotide and the incorporated nucleotides can be stimulated using a variety of lasers.

[0165] Sequencing can also include other massively parallel sequencing or next generation sequencing (NGS) technologies and platforms. Other examples of massively parallel sequencing technologies and platforms include Illumina HiSeq or MiSeq, Thermo PGM or Proton, Pac BioRS II or Sequel, Qiagen's Gene Reader and Oxford Nanopore MinION. Other similar massively parallel sequencing technologies currently available, as well as improved versions of these technologies, can be used.

[0166] Any cell type or tissue can be used to obtain nucleic acid samples for use in the methods described herein. For example, DNA or RNA samples can be obtained from a tumor or a body fluid, such as blood obtained using known techniques (e.g., venipuncture), or saliva. Alternatively, nucleic acid testing can be performed on a dry sample (e.g., hair or skin). In addition, one sequencing sample can be obtained from a tumor and another sequencing sample can be obtained from normal tissue, where the normal tissue and the tumor are of the same tissue type. One sequencing sample can be obtained from a tumor and another sequencing sample can be obtained from normal tissue, where the normal tissue and the tumor are of different tissue types.

[0167] The tumor may include one or more of the following: lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, stomach cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia and T-cell lymphocytic leukemia, non-small cell lung cancer and small cell lung cancer.

[0168] Alternatively, protein mass spectrometry can be used to identify or confirm the presence of mutant peptides that bind to MHC proteins on tumor cells. The peptides can be eluted with acid from tumor cells or from HLA molecules immunoprecipitated from tumors and then identified using mass spectrometry.

[0169] IV. Neoantigens

[0170] Neoantigens can include nucleotides or polypeptides. For example, a neoantigen can be an RNA sequence that encodes a polypeptide sequence. Therefore, neoantigens that can be used in vaccines include nucleotide sequences or polypeptide sequences.

[0171] Disclosed herein are isolated peptides comprising tumor-specific mutations identified by the methods disclosed herein, peptides comprising known tumor-specific mutations, and mutant polypeptides or fragments thereof identified by the methods disclosed herein. Neoantigenic peptides can be described in the context of their coding sequence, wherein the neoantigen comprises a nucleotide sequence (e.g., DNA or RNA) encoding the relevant polypeptide sequence.

[0172] The one or more polypeptides encoded by the neoantigen nucleotide sequence may comprise at least one of the following: binding affinity to MHC with an IC50 value of less than 1000 nM; for class I MHC peptides of 8-15, i.e., 8, 9, 10, 11, 12, 13, 14, or 15 amino acids in length, a sequence motif within or near the peptide that promotes proteasomal cleavage; and a sequence motif that promotes TAP transport. For class II MHC peptides of 6-30, i.e., 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 amino acids in length, a sequence motif within or near the peptide that promotes cleavage by extracellular or lysosomal proteases (cathepsins) or HLA-DM-catalyzed HLA binding.

[0173] One or more neoantigens may be presented on the surface of a tumor.

[0174] The one or more neoantigens can be immunogenic in a subject having a tumor, for example, capable of eliciting a T cell response or a B cell response in the subject.

[0175] In the context of generating a vaccine for a subject with a tumor, it may be considered to exclude one or more neoantigens that induce an autoimmune response in the subject.

[0176] The size of at least one neoantigenic peptide molecule may include, but is not limited to, about 5, about 6, about 7, about 8, about 9, about 10, about 11, about 12, about 13, about 14, about 15, about 16, about 17, about 18, about 19, about 20, about 21, about 22, about 23, about 24, about 25, about 26, about 27, about 28, about 29, about 30, about 31, about 32, or about 33. In some embodiments, the neoantigenic peptide molecule comprises about 30, about 33, about 34, about 35, about 36, about 37, about 38, about 39, about 40, about 41, about 42, about 43, about 44, about 45, about 46, about 47, about 48, about 49, about 50, about 60, about 70, about 80, about 90, about 100, about 110, about 120 or more amino acid residues, and any range derivable therefrom. In certain embodiments, the neoantigenic peptide molecule is equal to or less than 50 amino acids.

[0177] Neoantigenic peptides and polypeptides may be: 15 residues or less in length and typically consist of between about 8 and about 11 residues, particularly 9 or 10 residues, for MHC class I; 6-30 residues (inclusive) for MHC class II.

[0178] If necessary, longer peptides can be designed in several ways. In one case, when the likelihood of presentation of a peptide on an HLA allele is predicted or known, the longer peptide can be composed of either: (1) an individually presented peptide extended by 2-5 amino acids toward the N-terminus and C-terminus of each corresponding gene product; or (2) a concatenation of some or all of the presented peptides with the respective extensions. In another case, when sequencing reveals the presence of a longer (>10 residues) neoepitope sequence in the tumor (e.g., caused by frameshifting, readthrough, or inclusion of introns that generate novel peptide sequences), the longer peptide will be composed of: (3) an entire stretch consisting of novel tumor-specific amino acids, thereby bypassing the need for computational or in vitro selection of shorter peptides for the strongest HLA presentation. In both cases, the use of longer chains enables endogenous processing by the patient's cells and can result in more efficient antigen presentation and induction of T cell responses.

[0179] Neoantigenic peptides and polypeptides can be presented on HLA proteins. In some aspects, neoantigenic peptides and polypeptides are presented on HLA proteins with a higher affinity than wild-type peptides. In some aspects, the IC50 value of the neoantigenic peptide or polypeptide can be at least less than 5000 nM, at least less than 1000 nM, at least less than 500 nM, at least less than 250 nM, at least less than 200 nM, at least less than 150 nM, at least less than 100 nM, at least less than 50 nM, or less.

[0180] In some aspects, the neoantigenic peptides and polypeptides do not induce an autoimmune response and / or stimulate immune tolerance when administered to a subject.

[0181] Also provided are compositions comprising at least two or more neoantigenic peptides. In some embodiments, the composition contains at least two different peptides. At least two different peptides can be derived from the same polypeptide. Different polypeptides mean that the peptides are different in length, amino acid sequence, or both. These peptides are derived from any polypeptide that is known or found to contain tumor-specific mutations. Suitable polypeptides that can serve as a source of neoantigenic peptides can be found in, for example, the COSMIC database. COSMIC curates comprehensive information on somatic mutations in human cancers. The peptides contain tumor-specific mutations. In some aspects, tumor-specific mutations are driver mutations for a particular cancer type.

[0182] Neoantigenic peptides and polypeptides with desired activities or properties can be modified to provide certain desired attributes, such as improved pharmacological characteristics, while increasing or at least maintaining substantially all of the biological activities of the unmodified peptide to bind to the desired MHC molecule and activate appropriate T cells. For example, neoantigenic peptides and polypeptides can undergo various changes, such as conservative or non-conservative substitutions, where such changes may provide certain advantages in their use, such as improved MHC binding, stability, and presentation. Conservative substitutions mean that an amino acid residue is replaced by another amino acid residue that is biologically and / or chemically similar, for example, one hydrophobic residue is replaced by another, or one polar residue is replaced by another. Substitutions include combinations such as Gly, Ala; Val, Ile, Leu, Met; Asp, Glu; Asn, Gln; Ser, Thr; Lys, Arg; and Phe, Tyr, etc. The effects of single amino acid substitutions can also be explored using D-amino acids. Such modifications can be made using well-known peptide synthesis procedures, as described, for example, in Merrifield, Science 232:341-347 (1986), Barany & Merrifield, The Peptides, Gross & Meienhofer, eds. (NY, Academic Press), pp. 1-284 (1979); and Stewart and Young, Solid Phase Peptide Synthesis, (Rockford, 111., Pierce), 2nd ed. (1984).

[0183] Modification of peptides and polypeptides with various amino acid mimetics or non-natural amino acids is particularly suitable for increasing the in vivo stability of the peptides and polypeptides. Stability can be determined in a variety of ways. For example, stability is tested using peptidases and various biological media such as human plasma and serum. See, for example, Verhoef et al., Eur. J. Drug Metab Pharmacokin. 11: 291-302 (1986). The half-life of the peptide can be determined using 25% human serum (v / v) and measured in a conventional manner. The protocol is as follows. Before use, pooled human serum (AB type, not heat-inactivated) is defatted by centrifugation. Then, the serum is diluted to 25% with RPMI tissue culture medium and used to test peptide stability. At predetermined time intervals, a small amount of the reaction solution is removed and added to a 6% trichloroacetic acid aqueous solution or ethanol. The turbid reaction sample is cooled (4°C), kept for 15 minutes, and then centrifuged to allow the precipitated serum proteins to aggregate. Then, the presence of the peptide is determined by reverse phase HPLC using stability-specific chromatographic conditions.

[0184] These peptides and polypeptides can be modified to provide the desired properties except for the improved serum half-life. For example, the ability to induce CTL activity can be enhanced by connecting these peptides to a sequence containing at least one epitope that can induce T helper cell responses. Immunogenic peptide / T helper conjugates can be connected by means of spacer molecules. Spacers are generally included in relatively small neutral molecules that are generally uncharged under physiological conditions, such as amino acids or amino acid mimetics. These spacers are generally selected from, for example, Ala, Gly or other neutral spacers composed of non-polar amino acids or neutral polar amino acids. It should be understood that the optional spacer does not need to include the same residues and can therefore be a hetero-oligomer or homo-oligomer. When present, the spacer is generally at least one or two residues, more generally three to six residues. Alternatively, the peptide can be connected to the T helper peptide without a spacer.

[0185] The neoantigenic peptide can be linked to the T helper peptide directly or via a spacer at the amino or carboxyl terminus of the peptide. The amino terminus of the neoantigenic peptide or T helper peptide can be acylated. Exemplary T helper peptides include tetanus toxoid 830-843, influenza 307-319, and malarial circumsporozoite 382-398 and 378-389.

[0186] Proteins or peptides can be prepared by any technology known to those skilled in the art, including expressing proteins, polypeptides or peptides by standard molecular biology techniques, isolating proteins or peptides from natural sources, or chemically synthesizing proteins or peptides. Previously, nucleotides and protein, polypeptide and peptide sequences corresponding to various genes have been disclosed, and computerized databases known to those of ordinary skill in the art can be found. One such database is the Genbank and GenPept databases of the National Center for Biotechnology Information at the National Institutes of Health website. The coding regions of known genes can be amplified and / or expressed using technology disclosed herein or known to those of ordinary skill in the art. Alternatively, various commercially available preparations of proteins, polypeptides and peptides known to those skilled in the art are available.

[0187] On the other hand, the new antigen includes a nucleic acid (e.g., a polynucleotide) encoding a new antigenic peptide or a portion thereof. The polynucleotide can be, for example, single-stranded and / or double-stranded DNA, cDNA, PNA, CAN, RNA (e.g., mRNA), or a natural or stabilized form of a polynucleotide, such as, for example, a polynucleotide having a thiophosphate backbone, or a combination thereof, and the polynucleotide may or may not contain introns. On the other hand, an expression vector capable of expressing a polypeptide or a portion thereof is provided. Expression vectors for different cell types are well known in the art and can be selected without excessive experimentation. In general, DNA is inserted into an expression vector, such as a plasmid, in the appropriate orientation and correct expression reading frame. If necessary, the DNA can be connected to an appropriate transcriptional and translational regulatory control nucleotide sequence that can be recognized by the desired host, but such controls are generally available in expression vectors. The vector is then inserted into the host by standard techniques. Relevant guidance can be found, for example, in Sambrook et al. (1989) Molecular Cloning, A Laboratory Manual, Cold Spring Harbor Laboratory, Cold Spring Harbor, NY

[0188] IV. Vaccine Composition

[0189] Also disclosed herein is an immunogenic composition, such as a vaccine composition, capable of eliciting a specific immune response, such as a tumor-specific immune response. A vaccine composition typically comprises a plurality of neoantigens, such as those selected using the methods described herein. A vaccine composition may also be referred to as a vaccine.

[0190] The vaccine may contain between 1 and 30 peptides, i.e., 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29 or 30 different peptides; 6, 7, 8, 9, 10 11, 12, 13 or 14 different peptides; or 12, 13 or 14 different peptides. The peptides may include post-translational modifications. The vaccine may contain between 1 and 100 or more nucleotide sequences, i.e. 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51 , 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100 or more different nucleotide sequences; 6, 7, 8, 9, 10 11, 12, 13, or 14 different nucleotide sequences; or 12, 13, or 14 different nucleotide sequences. The vaccine may contain between 1 and 30 neoantigen sequences, i.e., 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75 or 12, 13, or 14 different neoantigen sequences.

[0191] In one embodiment, the different peptides and / or polypeptides, or nucleotide sequences encoding the same, are selected so that the peptides and / or polypeptides are capable of associating with different MHC molecules, such as different MHC class I molecules and / or different MHC class II molecules. In some aspects, a vaccine composition comprises coding sequences for peptides and / or polypeptides capable of associating with the most commonly occurring MHC class I molecules and / or MHC class II molecules. Thus, a vaccine composition may comprise different fragments capable of associating with at least two preferred, at least three preferred, or at least four preferred MHC class I molecules and / or MHC class II molecules.

[0192] The vaccine composition can induce a specific cytotoxic T cell response and / or a specific helper T cell response.

[0193] The vaccine composition may also include an adjuvant and / or a carrier. Examples of useful adjuvants and carriers are provided below. The composition may be associated with a carrier, such as, for example, a protein or an antigen presenting cell, such as a dendritic cell (DC) capable of presenting the peptide to a T cell.

[0194] An adjuvant is any substance that is added to a vaccine composition to increase or otherwise alter the immune response to a neoantigen. A carrier can be a scaffold structure, such as a polypeptide or polysaccharide, that can associate with the neoantigen. Optionally, the adjuvant is covalently or non-covalently conjugated.

[0195] The ability of an adjuvant to increase the immune response to an antigen is typically manifested by a significant or substantial increase in an immune-mediated reaction, or a decrease in disease symptoms. For example, an increase in humoral immunity is typically manifested as a significant increase in the titer of antibodies produced against the antigen, and an increase in T cell activity is typically manifested as an increase in cell proliferation, or cytotoxicity, or cytokine secretion. An adjuvant can also alter the immune response, for example, by shifting a predominantly humoral or Th response to a predominantly cellular or Th response.

[0196] Suitable adjuvants include, but are not limited to, 1018ISS, alum, aluminum salts, Amplivax, AS15, BCG, CP-870,893, CpG7909, CyaA, dSLIM, GM-CSF, IC30, IC31, Imiquimod, ImuFactIMP321, ISPatch, ISS, ISCOMATRIX, JuvImmune, LipoVac, MF59, monophosphoryl lipid A, Montanide IMS1312, Montanide ISA 206, Montanide ISA 50V, Montanide ISA-51, OK-432, OM-174, OM-197-MP-EC, ONTAK, PepTel vector system, PLG microparticles, resiquimod, SRL172, virosomes and other virus-like particles, YF-17D, VEGF traps, R848, β-glucan, Pam3Cys, Aquila's saponin-derived QS21 stimulator (Aquila Biotech, Worcester, Mass., USA), mycobacterial extracts and synthetic bacterial cell wall mimics, and other specialized adjuvants such as Ribi's Detox.Quil or Superfos. Adjuvants such as incomplete Freund's adjuvant or GM-CSF are useful. Several immunoadjuvants specifically for dendritic cells (e.g., MF59) and methods for their preparation have been previously described (Dupuis M et al., Cell Immunol. 1998; 186(1):18-27; Allison AC; Dev Biol Stand. 1998; 92:3-11). Cytokines can also be used. Stem cell factors are directly related to: affecting the migration of dendritic cells to lymphoid tissues (e.g., TNF-α); accelerating the maturation of dendritic cells into effective antigen-presenting cells for T lymphocytes (e.g., GM-CSF, IL-1, and IL-4) (U.S. Patent No. 5,849,589, specifically incorporated herein by reference in its entirety) and acting as immunoadjuvants (e.g., IL-12) (Gabrilovich DI et al., J Immunother Emphasis Tumor Immunol. 1996(6):414-418).

[0197] CpG immunostimulatory oligonucleotides have also been reported to enhance the effects of adjuvants in a vaccine setting.Other TLR binding molecules, such as RNA binding TLR 7, TLR 8 and / or TLR 9, may also be used.

[0198] Other examples of useful adjuvants include, but are not limited to, chemically modified CpG (e.g., CpR, Idera), poly (I: C) (e.g., poly i: C12U), non-CpG bacterial DNA or RNA, and immunologically active small molecules and antibodies, such as cyclophosphamide, sunitinib, bevacizumab, celebrex, NCX-4016, sildenafil, tadalafil, vardenafil, sorafinib, XL-999, CP-547632, pazopanib, ZD2171, AZD2171, ipilimumab, tremelimumab, and SC58175, which can act as therapeutic agents and / or as adjuvants. The amounts and concentrations of adjuvants and additives can be readily determined by a skilled artisan without undue experimentation. Other adjuvants include colony stimulating factors, such as granulocyte macrophage colony stimulating factor (GM-CSF, sargramostim).

[0199] Vaccine compositions may contain more than one different adjuvant. In addition, therapeutic compositions may contain any adjuvant substance, including any of the above or a combination thereof. In addition, it is contemplated that the vaccine and adjuvant may be administered together or separately in any appropriate order.

[0200] The carrier (or excipient) may exist independently of the adjuvant. The function of the carrier may be, for example, to increase the molecular weight of a particular mutant to increase activity or immunogenicity; to confer stability, increase biological activity, or increase serum half-life. In addition, the carrier may help present the peptide to T cells. The carrier may be any suitable carrier known to those skilled in the art, such as a protein or an antigen presenting cell. The carrier protein may be, but is not limited to, keyhole limpet hemocyanin, a serum protein such as transferrin, bovine serum albumin, human serum albumin, thyroglobulin or ovalbumin, an immunoglobulin, or a hormone such as insulin or palmitic acid. For human immunization, the carrier is generally a physiologically acceptable carrier for humans and is safe. However, tetanus toxoid and / or diphtheria toxoid are suitable carriers. Alternatively, the carrier may be a dextran, such as agarose.

[0201] Cytotoxic T cells (CTLs) recognize antigens in the form of peptides bound to MHC molecules, rather than the entire foreign antigen itself. The MHC molecules themselves are located on the cell surface of antigen-presenting cells. Therefore, if a trimeric complex of peptide antigen, MHC molecules, and APCs is present, CTLs may be activated. Accordingly, if the peptide is not only used to activate CTLs, but also if APCs with corresponding MHC molecules are added, it can enhance the immune response. Therefore, in some embodiments, the vaccine composition additionally contains at least one antigen-presenting cell.

[0202] Neoantigens can also be included in vaccine platforms based on viral vectors, such as vaccinia, fowlpox, self-replicating alphaviruses, marabavirus, adenovirus (see, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or lentiviruses, including but not limited to second, third, and / or mixed second / third generation lentiviruses and any generation recombinant lentivirus designed to target specific cell types or receptors (see, e.g., Hu et al., Immunization Delivered by Lentiviral Vectors for Cancer and Infectious Diseases, Immunol Rev. (2011) 239(1):45-61; Sakuma et al., Lentiviral vectors: basic to translational, Biochem J. (2012) 443(3):603-18; Cooper et al., Rescue of splicing-mediated intron loss maximizes expression in lentiviral vectors containing the human ubiquitin C promoter, Nucl. Acids Res. (2015) 43(1): 682-690; Zufferey et al., Self-Inactivating Lentivirus Vector for Safe and Efficient In Vivo Gene Delivery, J. Virol. (1998) 72(12): 9873-9880). Depending on the packaging capacity of the viral vector-based vaccine platform mentioned above, this method can deliver one or more nucleotide sequences encoding one or more neoantigenic peptides.These sequences can be flanked by non-mutated sequences, can be separated by linkers, or can be preceded by one or more sequences that target subcellular compartments (see, e.g., Gros et al., Prospective identification of neoantigen-specific lymphocytes in the peripheral blood of melanoma patients, Nat Med. (2016) 22(4): 433-8; Stronen et al., Targeting of cancer neoantigens with donor-derived T cell receptor repertoires, Science. (2016) 352(6291): 1337-41; Lu et al., Efficient identification of mutated cancer antigens recognized by T cells associated with durable tumor regressions, Clin Cancer Res. (2014) 20(13): 3401-10). Upon introduction into the host, the infected cells express the neoantigens and thereby elicit a host immune (e.g., CTL) response against the peptide. Vaccinia vectors and methods that can be used for immunization protocols are described, e.g., in U.S. Patent No. 4,722,848. Another vector is Bacille Calmette Guerin (BCG). BCG vectors are described in Stover et al. (Nature 351: 456-460 (1991)). Based on the description herein, it will be apparent to those skilled in the art that various other vaccine vectors that can be used for therapeutic administration or immunization of new antigens, for example, Salmonella typhi vectors, can be used.

[0203] IV.A. Additional Considerations for Vaccine Design and Manufacturing

[0204] IV.A.1. Determination of Peptide Panels Covering All Tumor Subclones

[0205] Truncal peptides, meaning peptides presented by all or most tumor subclones, will be preferentially included in the vaccine. 53Optionally, if there are no Torso peptides predicted to be presented with a higher probability and to be immunogenic, or if the number of Torso peptides predicted to be presented with a higher probability and to be immunogenic is small enough that other non-Torso peptides can be included in the vaccine, other peptides can be prioritized by estimating the number and properties of tumor subclones and selecting peptides to maximize the number of tumor subclones covered by the vaccine. 54

[0206] IV.A.2. Neoantigen Prioritization

[0207] After applying all of the above neoantigen filters, there are still many candidate neoantigens that can be included in the vaccine, compared to the amount that vaccine technology can support. In addition, uncertainty about various aspects of neoantigen analysis may remain, and there may be trade-offs between the different properties of candidate vaccine neoantigens. Therefore, it is possible to consider replacing the predetermined filters at each step of the selection process with an integrative multidimensional model that places candidate neoantigens into a space with at least the following axes and uses an integrative approach to optimize selection.

[0208] 1. Risk of autoimmunity or tolerance (germline risk) (lower autoimmunity risk is generally preferred).

[0209] 2. The probability of sequencing artifacts (lower artifact probabilities are generally preferred).

[0210] 3. Probability of immunogenicity (higher probability of immunogenicity is generally preferred).

[0211] 4. Probability of presentation (higher probability of presentation is generally preferred).

[0212] 5. Gene expression (higher expression is generally preferred).

[0213] 6. HLA gene coverage (increasing the number of HLA molecules involved in presenting the new antigen set can reduce the probability of tumors escaping immune attack through downregulation or mutation of HLA molecules).

[0214] 7. HLA class coverage (covering both HLA-I and HLA-II may increase the probability of treatment response and reduce the probability of tumor escape).

[0215] V. Treatment and Manufacturing Methods

[0216] Also provided is a method of inducing a tumor-specific immune response in a subject, vaccinating against a tumor, treating and or ameliorating symptoms of cancer in a subject by administering to the subject one or more neoantigens, such as a plurality of neoantigens identified using the methods disclosed herein.

[0217] In some aspects, the subject is diagnosed with or is at risk of developing cancer. The subject can be a human, dog, cat, horse, or any animal in need of a tumor-specific immune response. The tumor can be any solid tumor, such as a breast tumor, an ovarian tumor, a prostate tumor, a lung tumor, a kidney tumor, a stomach tumor, a colon tumor, a testicular tumor, a head and neck tumor, a pancreatic tumor, a brain tumor, a melanoma, or other tissue or organ tumor; and a hematologic tumor, such as a lymphoma and leukemia, including acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, and B-cell lymphoma.

[0218] The amount of neoantigen administered should be sufficient to induce a CTL response.

[0219] Neoantigens can be administered alone or in combination with other therapeutic agents. The therapeutic agent is, for example, a chemotherapeutic agent, radiation, or immunotherapy. Any suitable therapeutic treatment for a particular cancer can be administered.

[0220] In addition, anti-immunosuppressive / immunostimulatory agents, such as checkpoint inhibitors, can also be administered to the subject. For example, anti-CTLA antibodies or anti-PD-1 or anti-PD-L1 antibodies can also be administered to the subject. Blocking CTLA-4 or PD-L1 with antibodies can enhance the immune response against cancer cells in the patient. Specifically, it has been shown that blocking CTLA-4 is effective when followed with a vaccination regimen.

[0221] The optimal amount and optimal dosage regimen for each neoantigen included in the vaccine composition can be determined. For example, neoantigens or variants thereof can be prepared for intravenous (iv) injection, subcutaneous (sc) injection, intradermal (id) injection, intraperitoneal (ip) injection, intramuscular (im) injection. Injection methods include subcutaneous (sc), intradermal (id), intraperitoneal (ip), intramuscular (im), and intravenous. DNA or RNA injection methods include intradermal, intramuscular, subcutaneous, intraperitoneal, and intravenous. Other methods of administering vaccine compositions are known to those skilled in the art.

[0222] Vaccines can be designed such that the selection, number and / or amount of neoantigens present in the composition are tissue-, cancer- and / or patient-specific. For example, the exact choice of peptide can be guided by the expression pattern of the parent protein in a given tissue. The selection can depend on the specific type of cancer, the disease state, previous treatment regimens, the patient's immune status and, of course, the patient's HLA haplotype. In addition, vaccines can also contain personalized components based on the individual needs of a particular patient. Examples include changing the selection of neoantigens based on the expression of neoantigens in a particular patient or adjusting subsequent treatments following an initial round of treatment.

[0223] For compositions intended for use as cancer vaccines, neoantigens with similar normal self-peptides that are abundantly expressed in normal tissues should be avoided or present in small amounts in the compositions described herein. On the other hand, if a patient's tumor is known to express a particular neoantigen in high amounts, the corresponding pharmaceutical composition for treating that cancer may be present in high amounts and / or may include more than one neoantigen specific for that particular neoantigen or pathway to that neoantigen.

[0224] Compositions comprising neoantigens can be administered to individuals suffering from cancer. In therapeutic applications, the compositions are administered to the patient in an amount sufficient to elicit an effective CTL response against the tumor antigen and to cure or at least partially arrest symptoms and / or complications. An amount suitable for this purpose is defined as a "therapeutically effective dose." The amount effective for this purpose will depend, for example, on the composition, the mode of administration, the stage and severity of the disease being treated, the patient's weight and general health, and the judgment of the prescribing physician. It will be understood that the compositions can generally be used in serious disease states, that is, life-threatening or potentially life-threatening conditions, particularly when the cancer has metastasized. In such cases, given the need to minimize foreign matter and the relatively non-toxic nature of the neoantigens, it is possible and may be felt necessary for the treating physician to administer a substantial excess of these compositions.

[0225] For therapeutic use, administration can be initiated upon detection or surgical removal of a tumor. This is followed by increasing the dosage until symptoms are at least substantially alleviated and for a period of time thereafter.

[0226] Pharmaceutical compositions (e.g., vaccine compositions) for therapeutic treatment are intended for parenteral, topical, nasal, oral, or local administration. Pharmaceutical compositions can be administered parenterally, such as intravenously, subcutaneously, intradermally, or intramuscularly. These compositions can be administered to the site of surgical resection to induce a local immune response against the tumor. Disclosed herein are compositions for parenteral administration, which comprise a neoantigen solution and the vaccine composition is dissolved or suspended in an acceptable carrier, such as an aqueous carrier. A variety of aqueous carriers can be used, such as water, buffered water, 0.9% saline, 0.3% glycine, hyaluronic acid, and the like. These compositions can be sterilized by well-known conventional sterilization techniques, or can be subjected to sterile filtration. The aqueous solution thus obtained can be packaged for use as is, or lyophilized; the lyophilized preparation is combined with a sterile solution before administration. If necessary, these compositions may contain pharmaceutically acceptable auxiliary substances to approximate physiological conditions, such as pH adjusters and buffers, tonicity adjusters, wetting agents, etc., for example, sodium acetate, sodium lactate, sodium chloride, potassium chloride, calcium chloride, sorbitan monolaurate, triethanolamine oleate, etc.

[0227] Neoantigens can also be administered via liposomes, targeting the liposomes to specific cell tissues, such as lymphoid tissue. Liposomes can also be used to increase half-life. Liposomes include emulsions, foams, micelles, insoluble monolayers, liquid crystals, phospholipid dispersions, lamellar layers, and the like. In these formulations, the neoantigen to be delivered is incorporated as part of the liposomes alone or in combination with molecules that bind to, for example, ubiquitous receptors between lymphocytes, such as monoclonal antibodies that bind to the CD45 antigen, or conjugated with other therapeutic or immunogenic compositions. Thus, liposomes filled with the desired neoantigen can be directed to the lymphocyte site, followed by liposome delivery of the selected therapeutic / immunogenic composition. Liposomes can be formed from standard vesicle-forming lipids, which generally include neutral and negatively charged phospholipids and sterols such as cholesterol. The choice of lipids is generally guided by considerations such as liposome size, acid instability, and the stability of the liposomes in the bloodstream. There are various methods available for preparing liposomes as described, for example, in Szoka et al., Ann. Rev. Biophys. Bioeng. 9; 467 (1980); U.S. Patent Nos. 4,235,871; 4,501,728; 4,501,728; 4,837,028; and 5,019,369.

[0228] To target immune cells, the ligands to be incorporated into the liposomes may include, for example, antibodies or fragments thereof that are specific for cell surface determinants of the desired immune system cells. The liposomal suspension may be administered intravenously, topically, or topically, and the dosage may vary depending on, among other things, the mode of administration, the peptide being delivered, and the stage of the disease being treated.

[0229] For treatment or immunization purposes, nucleic acid encoding peptides and optionally one or more peptides described herein can also be administered to the patient. Nucleic acid is often delivered to the patient using a variety of methods. For example, nucleic acid can be directly delivered, such as "naked DNA". This method is described in, for example, Wolff et al., Science 247:1465-1468 (1990) and U.S. Patent No. 5,580,859 and No. 5,589,466. Nucleic acid can also be administered using, for example, the impact delivery method (ballistic delivery) described in U.S. Patent No. 5,204,253. Particles containing only DNA can be administered. Alternatively, DNA can be attached to particles, such as gold particles. The method for delivering nucleic acid sequences can include viral vectors, mRNA vectors and DNA vectors, with or without electroporation.

[0230] Nucleic acids can also be delivered by forming complexes with cationic compounds, such as cationic lipids. Lipid-mediated gene delivery methods are described, for example, in 9618372WOAWO 96 / 18372; 9324640WOAWO 93 / 24640; Mannino and Gould-Fogerite, BioTechniques 6(7):682-691 (1988); U.S. Patent No. 5,279,833; Rose U.S. Patent No. 5,279,833; 9106309WOAWO 91 / 06309; and Felgner et al., Proc. Natl. Acad. Sci. USA 84: 7413-7414 (1987).

[0231] Neoantigens can also be included in vaccine platforms based on viral vectors, such as vaccinia, fowlpox, self-replicating alphaviruses, Maraba virus, adenovirus (see, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or lentiviruses, including but not limited to second, third, and / or mixed second / third generation lentiviruses and any generation of recombinant lentiviruses designed to target specific cell types or receptors (see, e.g., Hu et al., Immunization Delivered by Lentiviral Vectors for Cancer and Infectious Diseases, Immunol Rev. (2011) 239(1):45-61; Sakuma et al., Lentiviral vectors: basic to translational, Biochem J. (2012) 443(3):603-18; Cooper et al., Rescue of splicing-mediated intronloss maximizes expression in lentiviral vectors containing the humanubiquitin C promoter, Nucl. Acids Res. (2015) 43(1): 682-690; Zufferey et al., Self-Inactivating Lentivirus Vector for Safe and Efficient In Vivo Gene Delivery, J. Virol. (1998) 72(12): 9873-9880). Depending on the packaging capacity of the viral vector-based vaccine platform mentioned above, this method can deliver one or more nucleotide sequences encoding one or more neoantigenic peptides.These sequences can be flanked by non-mutated sequences, can be separated by linkers, or can be preceded by one or more sequences that target subcellular compartments (see, e.g., Gros et al., Prospective identification of neoantigen-specific lymphocytes in the peripheral blood of melanoma patients, Nat Med. (2016) 22(4): 433-8; Stronen et al., Targeting of cancer neoantigens with donor-derived T cell receptor repertoires, Science. (2016) 352(6291): 1337-41; Lu et al., Efficient identification of mutated cancer antigens recognized by T cells associated with durable tumor regressions, Clin Cancer Res. (2014) 20(13): 3401-10). Upon introduction into the host, the infected cells express the neoantigens and thereby elicit a host immune (e.g., CTL) response against the peptide. Vaccinia vectors and methods that can be used for immunization protocols are described, e.g., in U.S. Patent No. 4,722,848. Another vector is Bacillus Calmette-Guérin (BCG). BCG vectors are described in Stover et al. (Nature 351: 456-460 (1991)). Based on the description herein, it will be apparent to those skilled in the art that a variety of other vaccine vectors can be used for therapeutic administration or immunization of new antigens.

[0232] The nucleic acid administration approach utilizes minigene constructs encoding one or more epitopes. To generate DNA sequences encoding selected CTL epitopes (minigenes) for expression in human cells, the amino acid sequences of these epitopes are reverse translated. The human codon usage table guides codon selection for each amino acid. These epitope-encoding DNA sequences are directly adjacent to each other to produce a continuous polypeptide sequence. To optimize expression and / or immunogenicity, additional elements can be incorporated into the minigene design. Examples of amino acid sequences that can be reverse translated and included in the minigene sequence include helper T lymphocytes, epitopes, leader (signal) sequences, and endoplasmic reticulum retention signals. In addition, MHC presentation of CTL epitopes can be improved by including synthetic (e.g., polyalanine) or naturally occurring flanking sequences adjacent to the CTL epitopes. The minigene sequence is converted to DNA by assembling oligonucleotides encoding the plus and minus strands of the minigene. Overlapping oligonucleotides (30-100 bases long) are synthesized, phosphorylated, purified, and annealed under appropriate conditions using well-known techniques. The ends of the oligonucleotides are joined using T4 DNA ligase. This synthetic minigene encoding the CTL epitope polypeptide can then be cloned into a desired expression vector.

[0233] A variety of formulations can be used to prepare purified plasmid DNA for injection. The simplest of these methods is to rehydrate lyophilized DNA in sterile phosphate-buffered saline (PBS). Various methods have been described, and new technologies can also be used. As mentioned above, nucleic acids are preferably formulated with cationic lipids. In addition, sugar esters, fusogenic liposomes, peptides, and compounds, collectively referred to as protective, interactive, non-condensable (PINC), can also be made to form complexes with the purified plasmid DNA to affect various variables, such as stability, intramuscular dispersion, or transportation to specific organs or cell types.

[0234] Also disclosed is a method for producing a tumor vaccine, comprising performing the steps of the method disclosed herein; and producing a tumor vaccine comprising a plurality of neoantigens or a subset of the plurality of neoantigens.

[0235] The neoantigens disclosed herein can be manufactured using methods known in the art. For example, the methods disclosed herein for producing neoantigens or vectors (e.g., vectors comprising at least one sequence encoding one or more neoantigens) may include culturing host cells under conditions suitable for expressing the neoantigen or vector, wherein the host cells comprise at least one polynucleotide encoding the neoantigen or vector; and purifying the neoantigen or vector. Standard purification methods include chromatography, electrophoresis, immunoassays, precipitation, dialysis, filtration, concentration, and isoelectric focusing.

[0236] Host cells can include Chinese hamster ovary (CHO) cells, NSO cells, yeast or HEK293 cells. The host cell can be transformed with one or more polynucleotides comprising at least one nucleic acid sequence encoding a novel antigen or vector disclosed herein, optionally wherein the isolated polynucleotide further comprises a promoter sequence operably linked to the at least one nucleic acid sequence encoding the novel antigen or vector. In certain embodiments, the isolated polynucleotide can be a cDNA.

[0237] VI. Neoantigen Identification

[0238] VI.A. Identification of Neoantigen Candidates.

[0239] Methods for analyzing tumor and normal exomes and transcriptomes using NGS have been described and applied in the area of ​​neoantigen identification. 6,14,15 The following examples consider certain optimization measures for higher sensitivity and specificity of new antigen identification in a clinical setting. These optimization measures can be divided into two areas, namely optimization related to laboratory methods and optimization related to NGS data analysis.

[0240] VI.A.1. Laboratory Method Optimization

[0241] The method presented here improves upon the concepts developed for the reliable assessment of cancer driver genes in target cancer panels. 16 Expanding to the whole exome and whole transcriptome environments required for neoantigen identification solves the challenge of high-accuracy neoantigen discovery from clinical samples with low tumor content and small volumes. Specifically, these improvements include:

[0242] 1. Targeting deep (>500×) unique average coverage of the entire tumor exome to detect mutations that are present at low mutant allele frequency due to low tumor content or subclonal status.

[0243] 2. Targeting uniform coverage across the tumor exome, with <5% of bases covered at <100×, thereby minimizing the likelihood of missing neoantigens by, for example:

[0244] a. Using DNA-based capture probes and individual probe QC 17

[0245] b. Include additional bait for areas with less coverage

[0246] 3. Targeting uniform coverage across the normal exome, with <5% of bases covered at <20×, thereby minimizing the likelihood of neoantigens being unclassified for somatic / germline status (and thus unusable as TSNAs)

[0247] 4. To minimize the total amount of sequencing required, sequence capture probes should be designed to target only gene coding regions, as non-coding RNA does not produce new antigens. Other optimizations include:

[0248] a. Supplementary probes targeting HLA genes, which are GC-rich and difficult to capture by standard exome sequencing 18

[0249] b. Exclude genes predicted to produce few or no candidate neoantigens due to factors such as insufficient expression levels, suboptimal proteasomal digestion, or unusual sequence features.

[0250] 5. Tumor RNA will also typically be sequenced at high depth (>100M reads) to enable variant detection, quantification of gene and splice variant ("isoform") expression levels, and fusion detection. RNA from FFPE samples will use probe-based enrichment methods. 19 , extracted using the same or similar probes as those used to capture the exome in DNA.

[0251] VI.A.2. NGS Data Analysis Optimization

[0252] Improvements in analytical methods address the suboptimal sensitivity and specificity of commonly used research mutation calling methods, with specific considerations for customization relevant to neoantigen identification in the clinical setting. These include:

[0253] 1. Use the HG38 reference human genome or subsequent versions for alignment, as this genome contains multiple MHC region assemblies compared to previous genome versions, better reflecting population polymorphism.

[0254] 2. By combining the results obtained by different programs 5 , overcoming the limitations of single variant callers 20

[0255] a. Detect single nucleotide variants and indels in tumor DNA, tumor RNA, and normal DNA using a suite of tools, including programs based on comparisons of tumor and normal DNA, such as Strelka 21 and Mutect 22 and programs that incorporate tumor DNA, tumor RNA, and normal DNA, such as UNCeqR, which are particularly suitable for low-purity samples 23 .

[0256] b. Indels will be determined using programs that perform local reassembly, such as Strelka and ABRA 24 .

[0257] c. Structural rearrangements will be determined using specialized tools such as Pindel 25 or Breakseq 26 .

[0258] 3. To detect and prevent sample swapping, variant calls in samples from the same patient will be compared at a selected number of polymorphic sites.

[0259] 4. Extensive filtering against pseudo calls would be done, for example, by:

[0260] a. Remove variants found in normal DNA, potentially use less stringent detection parameters at low coverage, and use a permissive proximity criterion in the case of indels.

[0261] b. Remove variants caused by low mapping quality or low base quality 27 .

[0262] c. Remove variants that arise from recurring sequencing artifacts, even if not observed in the corresponding normal context. 27 Examples include variants detected predominantly on one strand.

[0263] d. Remove variants detected in irrelevant control sets 27

[0264] 5. Using seq2HLA 28 ATHLATES 29 or Optitype, accurately calls HLA from normal exomes and also combines exomes with RNA sequencing data 28 Other possible optimizations include using assays specific to HLA typing, such as long-read DNA sequencing. 30 , or adjust the method for joining RNA fragments to maintain continuity 31 .

[0265] 6. Robust detection of neo-ORFs arising from tumor-specific splice variants will be achieved by using CLASS 32 、Bayesembler 33 StringTie 34 or similar programs in their reference-guided mode, assembling transcripts from RNA-seq data (i.e., using known transcript structures rather than attempting to reconstruct the entire transcriptome in each experiment). 35 It is often used for this purpose, but it often produces an unreasonably large number of splice variants, many of which are much shorter than the full-length gene, and simple positive controls cannot be recovered. The possibility of coding sequence and nonsense-mediated decay will be analyzed by tools such as SpliceR36 and MAMBA 37 Gene expression will be determined using tools such as Cufflinks 35 Wild-type and mutant-specific expression counts and / or relative levels will be determined using tools developed for these purposes, such as ASE 38 or HTSeq 39 Determination. Possible filtering steps include:

[0266] a. Remove candidate neo-ORFs that are considered underexpressed.

[0267] b. Remove candidate neo-ORFs predicted to trigger nonsense-mediated decay (NMD).

[0268] 7. Candidate neoantigens observed only in RNA that cannot be directly validated as tumor-specific antigens (e.g., neoORFs) will be classified as likely tumor-specific based on additional parameters, such as by considering the following factors:

[0269] a. The presence of cis-acting frameshift or splice site mutations that only support tumor DNA

[0270] b. The presence of trans-acting mutations in splicing factors that are only confirmed in tumor DNA. For example, in three independent published experiments using the R625 mutant SF3B1, although one experiment examined patients with uveal melanoma, 40 The second experiment examined uveal melanoma cell lines 41 , and the third experiment examined breast cancer patients 42 , but the genes showing the greatest splicing differences are consistent.

[0271] c. For novel splice isoforms, there are confirmed “novel” splice-junction reads in the RNASeq data.

[0272] d. For new rearrangements, there are similar exon reads that are confirmed to be present in tumor DNA but not in normal DNA

[0273] e. Lack of gene expression profiles, such as GTEx 43 (i.e., making germline origin unlikely)

[0274] 8. Complement analyses based on reference genome alignments by directly comparing assembled DNA tumor and normal reads (or k-mers derived from these reads) to avoid errors and artifacts based on alignment and annotation (e.g., somatic variants arising near germline variants or repetitive indels).

[0275] In samples with polyadenylated RNA, the presence of viral and microbial RNA in RNA-seq data will be analyzed using RNA CoMPASS 44 or similar methods to identify other factors that may predict patient response.

[0276] VI.B. Isolation and Detection of HLA Peptides

[0277] Isolation of HLA-peptide molecules was performed using the classical immunoprecipitation (IP) method after lysis of cell and tissue samples. 55-58 Clarified lysates were used for HLA-specific IP.

[0278] Immunoprecipitation is performed using antibodies specific for HLA molecules coupled to beads. For pan-class I HLA immunoprecipitation, a pan-class I CR antibody is used, while for class II HLA–DR, an HLA–DR antibody is used. During an overnight incubation, the antibodies are covalently linked to NHS-agarose beads. Following covalent attachment, the beads are washed and aliquoted for IP. 59,60 Immunoprecipitation can also be performed using antibodies that are not covalently bound to magnetic beads. Typically, this is accomplished by immobilizing the antibodies on a chromatography column using agarose or magnetic beads coated with Protein A and / or Protein G. Listed below are some antibodies that can be used to selectively enrich for MHC / peptide complexes.

[0279] Antibody name Specificity W6 / 32 Class I HLA-A, B, C L243 Class II – HLA-DR Tu36 Class II – HLA-DR LN3 Class II – HLA-DR Tu39 Class II – HLA-DR, DP, DQ

[0280] Clarified tissue lysate was added to antibody beads for immunoprecipitation. After immunoprecipitation, beads were removed from the lysate and stored for further experiments, including further IP. The IP beads were washed to remove nonspecific binding and the HLA / peptide complex was eluted from the beads using standard techniques. Protein components were removed from the peptides using molecular weight spin columns or C18 fractionation. The resulting peptides were dried by SpeedVac evaporation and in some cases stored at -20°C for MS analysis.

[0281] The peptide of drying is rehydrated in the HPLC buffer solution that is suitable for reverse phase chromatography and is loaded on C-18 microcapillary HPLC post to carry out gradient elution in Fusion Lumos mass spectrometer (Thermo).In Orbitrap detector, under high resolution, collect the MS1 spectrum of peptide mass / charge ratio (m / z), subsequently after selected ion experience HCD fragmentation, collect MS2 low-resolution scanning spectrum in ion trap detector.In addition, CID or ETD fragmentation method can be used, or any combination of three kinds of technology obtains MS2 spectrum, to reach the higher amino acid coverage of this peptide.Can also in Orbitrap detector, measure MS2 spectrum with high resolution mass accuracy.

[0282] Using Comet 61,62 , the MS2 spectra obtained from each analysis were searched against the protein database using Percolator 63-65 Peptide identifications were scored. Additional sequencing can be performed using PEAKS studio (Bioinformatics Solutions Inc.), and other search engines or other sequencing methods can be used, including spectral matching and de novo sequencing. 75 .

[0283] VI.B.1. Studies of MS detection limits supporting comprehensive HLA peptide sequencing.

[0284] The detection limit was determined using the peptide YVYVADVAAK using different amounts of peptide loaded onto the LC column. The amounts of peptide tested were 1 pmol, 100 fmol, 10 fmol, 1 fmol, and 100 amol. (Table 1) The results are shown in Figure 1F These results indicate that the lowest limit of detection (LoD) is in the attomol range (10 -18 ), the dynamic range spans five orders of magnitude, and the signal-to-noise ratio appears adequate in the low femtomol range (10 -15 ) were sequenced.

[0285]

[0286]

[0287] VII. Present Delivery Model

[0288] VII.A. Systematic Reviews

[0289] Figure 2A is an overview of an environment 100 for identifying the likelihood of presentation of a peptide in a patient, according to one embodiment. The environment 100 provides a background for introducing a presentation identification system 160, which itself includes a presentation information store 165.

[0290] The presentation authentication system 160 is one or more of the following Figure 38The computer model embodied in the computing system discussed herein receives peptide sequences associated with a set of MHC alleles and determines the likelihood that these peptide sequences will be presented by one or more MHC alleles from the set of associated MHC alleles. The presentation identification system 160 can be applied to both class I and class II MHC alleles. This is applicable in a variety of situations. One specific use case for the presentation identification system 160 is that it can receive nucleotide sequences of candidate neoantigens associated with a set of MHC alleles from a tumor cell from a patient 110 and determine the likelihood that these candidate neoantigens will be presented by one or more of the tumor's associated MHC alleles and / or induce an immunogenic response in the patient's 110 immune system. Candidate neoantigens determined by the system 160 to have a high likelihood of presentation can be selected for inclusion in a vaccine 118, and such an anti-tumor immune response can be elicited by the immune system of the patient 110 who provided the tumor cells. Additionally, T cells with TCRs that respond to candidate neoantigens with a high likelihood of presentation can be generated for use in T cell therapy, thereby also eliciting an anti-tumor immune response from the patient's 110 immune system.

[0291] The presentation identification system 160 determines the likelihood of presentation using one or more presentation models. Specifically, the presentation model generates a probability of whether a given peptide sequence will be presented by a set of relevant MHC alleles, and this is generated based on the presentation information stored in the memory 165. For example, the presentation model can generate a probability of whether the peptide sequence "YVYVADVAAK" will be presented on the cell surface of a sample by a set of alleles HLA-A*02:01, HLA-A*03:01, HLA-B*07:02, HLA-B*08:03, HLA-C*01:04. As another example, the presentation model can also generate a probability of whether the peptide sequence "YVYVADVAAK" will be presented by HLA alleles having the HLA allele sequences "AYANGPW," "UIIKNFDL," and "WRTSAOGH." Presentation information 165 contains information about whether the peptides bind to different types of MHC alleles so that they are presented by the MHC alleles, which is determined in the model based on the position of the amino acids in the peptide sequence. The presentation model can predict whether the presentation of an unrecognized peptide sequence will be associated with a set of related MHC alleles based on the presentation information 165. As previously mentioned, the presentation model can be applied to both class I and class II MHC alleles.

[0292] The term "HLA coverage" is used throughout this specification. As used throughout this specification, "HLA coverage" can apply to individuals and / or populations of individuals. When applied to an individual, "HLA coverage" refers to the proportion of HLA alleles found within the individual's genome that have a presentation pattern. For example, for a homozygous individual with HLA types A*02:01, A*02:01, B*07:02, B*07:02, C*07:02, C*07:02, C*07:02, if a presentation pattern exists for alleles A*02:01 and B*07:02, but not for C*07:02, the individual's HLA coverage is 4 / 6.

[0293] When applied to a population of individuals, "HLA coverage" refers to the proportion of individuals in the population for which each possible level of individual HLA coverage exists for the presented model. In the case of human individuals, each human genome contains six HLA alleles. Thus, possible levels of individual HLA coverage include 0 / 6, 1 / 6, 2 / 6, ..., 6 / 6. Thus, for example, in a population of individuals, if half of the individuals in the population have an individual HLA coverage of 2 / 6 and half of the individuals in the population have an individual HLA coverage of 6 / 6, then for individual HLA coverage of 0 / 6, the HLA coverage of the population is 0%, for individual HLA coverage of 1 / 6, the HLA coverage of the population is 0%, for individual HLA coverage of 2 / 6, the HLA coverage of the population is 50%, for individual HLA coverage of 3 / 6, the HLA coverage of the population is 0%, for individual HLA coverage of 4 / 6, the HLA coverage of the population is 0%, for individual HLA coverage of 5 / 6, the HLA coverage of the population is 0%, and for individual HLA coverage of 6 / 6, the HLA coverage of the population is 50%.

[0294] As described in more detail below with respect to Section VIII, the goal of training the presentation model is to have the highest possible HLA coverage for each individual in the population, and therefore for the HLA coverage of the population, such that the proportion of individuals in the population with higher individual HLA coverage is as high as possible.

[0295] VII.B. Submission of Information

[0296] Figure 2A A method for obtaining presentation information according to one embodiment is shown. Presentation information 165 includes two general information categories: allele interaction information and allele non-interaction information. Allele interaction information includes information that affects the presentation of peptide sequences related to the type of MHC allele. Allele non-interaction information includes information that affects the presentation of peptide sequences unrelated to the type of MHC allele.

[0297] VII.B. 1. Allelic interaction information

[0298] Allele interaction information mainly comprises the peptide sequence through identification, and it is known that these peptide sequences have been presented by one or more MHC molecules through identification from people, mice etc.It should be noted that this may include or may not include the data obtained from tumor sample.Can identify the peptide sequence presented from the cell of expressing single MHC allele.In this case, the peptide sequence presented is generally collected from single allele cell line, and these cell lines are engineered to express predetermined MHC allele and are exposed to synthetic protein subsequently.The peptide presented on MHC allele is separated by techniques such as acid elution and identified by mass spectrometry. Figure 2B An example of this is shown, where an exemplary peptide YEMFNDKSQRAPDDKMF presented on a predetermined MHC allele, HLA-DRB1*12:01, was isolated and identified by mass spectrometry. Since, in this case, the peptide was identified by cells engineered to express a single predetermined MHC protein, the direct association between the presented peptide and the MHC protein to which it binds is known.

[0299] It is also possible to collect presented peptide sequences from cells expressing multiple MHC alleles. Typically, in the human body, a cell expresses 6 different types of MHC-I and up to 12 different types of MHC-II molecules. Peptide sequences presented in this manner can be identified from multi-allelic cell lines engineered to express multiple predetermined MHC alleles. Peptide sequences presented in this manner can also be identified from tissue samples, such as normal tissue samples or tumor tissue samples. In particular, in this case, MHC molecules can be immunoprecipitated from normal or tumor tissues. Peptides presented on multiple MHC alleles can similarly be separated by techniques such as acid elution and identified by mass spectrometry. Figure 2C An example of this is shown, where six example peptides YEMFNDKSF, HROEIFSHDFJ, FJIEJFOESS, NEIOREIREI, JFKSIFEMMSJDSSUIFLKSJFIEIFJ, and KNFLENFIESOFI were presented to identified MHC class I alleles HLA-A*01:01, HLA-A*02:01, HLA-B*07:02, HLA-B*08:01, and MHC class II alleles HLA-DRB1*10:01, HLA-DRB1:11:01 and isolated and identified by mass spectrometry. With respect to monoallelic cell lines, a direct association between the presented peptide and the MHC protein to which it binds may be unknown because the bound peptide is isolated from the MHC molecule prior to identification.

[0300] Allelic interaction information can also include mass spectrometry ion current, which depends on the concentration of the peptide-MHC molecule complex and the peptide ionization efficiency. The ionization efficiency varies with the peptide in a sequence-dependent manner. In general, the ionization efficiency varies with the peptide within about two orders of magnitude, while the concentration of the peptide-MHC complex varies over a larger range than that.

[0301] Allelic interaction information can also include measurements or predictions of the binding affinity between a given MHC allele and a given peptide. (72,73,74) One or more affinity models can generate such predictions. For example, looking back at Figure 1D In the example shown in , presentation information 165 may include a predicted binding affinity of 1000 nM between the peptide YEMFNDKSF and the class I allele HLA-A*01:01. Peptides with IC50 > 1000 nM are rarely presented by MHC, and lower IC50 values ​​increase the probability of presentation. Presentation information 165 may include a predicted binding affinity of the peptide KNFLENFIESOFI and the class II allele HLA-DRB1:11:01.

[0302] Allelic interaction information can also include a measure or prediction of the stability of the MHC complex. One or more stability models can generate such predictions. More stable peptide-MHC complexes (i.e., complexes with longer half-lives) are more likely to be presented in high copy numbers on tumor cells and on antigen-presenting cells that encounter vaccine antigens. For example, looking back at Figure 2C In the example shown in , the presented information 165 may include a stability prediction value of 1 hour for the half-life of the class I molecule HLA-A*01:01. The presented information 165 may include a stability prediction value of 1 hour for the half-life of the class II molecule HLA-DRB1:11:01.

[0303] Allelic interaction information can also include measured or predicted reaction rates for peptide-MHC complex formation. Complexes formed at higher rates are more likely to be presented at higher concentrations on the cell surface.

[0304] Allelic interaction information can also include the sequence and length of the peptide. Class I MHC molecules generally prefer to present peptides between 8 and 15 peptides in length. 60-80% of the presented peptides are 9 peptides in length. Class II MHC molecules generally prefer to present peptides between 6 and 30 peptides.

[0305] Allelic interaction information can also include the presence of a kinase sequence motif on the neoantigen encoding peptide, as well as the absence or presence of specific post-translational modifications on the neoantigen encoding peptide. The presence of a kinase motif can influence the probability of a post-translational modification that may enhance or interfere with MHC binding.

[0306] Allelic interaction information can also include proteins involved in post-translational modification processes, such as the expression level or activity level of kinases (as measured or predicted by RNAseq, mass spectrometry or other methods).

[0307] Allelic interaction information can also include the probability of presentation of peptides with similar sequences in cells from other individuals expressing a particular MHC allele, which can be assessed by mass spectrometry proteomics or other means.

[0308] Allelic interaction information can also include the expression level of a particular MHC allele in the individual in question (e.g., as measured by RNA-seq or mass spectrometry). Peptides that bind most strongly to MHC alleles expressed at high levels are more likely to be presented than peptides that bind most strongly to MHC alleles expressed at low levels.

[0309] Allelic interaction information can also include the probability of presentation by a particular MHC allele in other individuals expressing that particular MHC allele independent of the overall neoantigen encoding peptide sequence.

[0310] Allelic interaction information can also include the probability of presentation by MHC alleles from the same family of molecules (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other individuals, independent of the overall peptide sequence. For example, HLA-C molecules are typically expressed at lower levels than HLA-A or HLA-B molecules, and from this, it can be inferred that the probability of a peptide being presented by HLA-C is lower than that by HLA-A or HLA-B. As another example, HLA-DP is typically expressed at lower levels than HLA-DR or HLA-DQ, and from this, it can be inferred that the probability of a peptide being presented by HLA-DP is lower than that by HLA-DR or HLA-DQ.

[0311] Allelic interaction information may also include protein sequences for specific MHC alleles.

[0312] Any of the MHC allele non-interaction information listed in the following sections can also be modeled in the same way as the MHC allele interaction information.

[0313] VII.B.2. Allele-non-interaction information

[0314] Allele-independent information can include C-terminal sequences flanking the neoantigen-encoding peptide within the source protein sequence. For MHC-I, the C-terminal flanking sequence may affect the proteasomal processing of the peptide. However, the C-terminal flanking sequence is cleaved from the peptide by the proteasome before the peptide is transported to the endoplasmic reticulum and encounters the MHC allele on the cell surface. Therefore, the MHC molecule does not receive information about the C-terminal flanking sequence, and thus, the influence of the C-terminal flanking sequence does not vary with the MHC allele type. For example, see again Figure 2C In the example shown in , the presentation information 165 may include the C-terminal flanking sequence FOEIFNDKSLDKFJI of the presenting peptide FJIEJFOESS identified from the source protein of the peptide.

[0315] Allele non-interaction information can also include mRNA quantitative measurement. For example, mRNA quantitative data of the sample identical with providing mass spectrometry training data can be obtained. As described later, RNA expression level is identified as a strong predictor of peptide presentation. In one embodiment, mRNA quantitative measurement value is identified by software tool RSEM. Detailed embodiments of the RSEM software tool can be found in Bo Li and Colin N.Dewey.RSEM:accurate transcript quantification from RNA-Seq data with or without a reference genome.BMC Bioinformatics, 12:323, August 2011. In one embodiment, mRNA quantitative is to measure the fragment number (FPKM) of each kilobase transcript in each million positioning reads.

[0316] Allele-non-interaction information may also include the N-terminal sequence flanking the peptide within the source protein sequence.

[0317] Allele non-interaction information can also include the source gene of the peptide sequence. The source gene can be defined as the Ensembl protein family of the peptide sequence. In other examples, the source gene can be defined as the source DNA or source RNA of the peptide sequence. The source gene can be represented as a string of nucleotides encoding a protein, or a named set based on a known DNA or RNA sequence known to encode a specific protein will be more directly represented. In another example, allele non-interaction information can also include a source transcript or isoform of the peptide sequence extracted from a database such as Ensembl or RefSeq or a set of potential source transcripts or isoforms.

[0318] Allele-independent interaction information may also include the tissue type, cell type, or tumor type of the cell from which the peptide sequence originated.

[0319] Allelic non-interaction information can also include the presence of a protease cleavage motif in the peptide, optionally weighted according to the expression of the corresponding protease in tumor cells (as measured by RNA-seq or mass spectrometry). Peptides containing protease cleavage motifs are less likely to be presented because these peptides are more easily degraded by proteases and are therefore less stable in cells.

[0320] Allelic non-interaction information can also include the turnover rate of the source protein as measured in the appropriate cell type. A faster turnover rate (i.e., a shorter half-life) increases the probability of presentation; however, if measured in dissimilar cell types, the predictive power of this feature is lower.

[0321] Allelic non-interaction information can also include the length of the source protein as measured by RNA-seq or proteomic mass spectrometry, or as predicted based on annotation of germline or somatic splicing mutations detected in DNA or RNA sequence data, optionally taking into account the specific splicing variants ("isoforms") that are most highly expressed in tumor cells.

[0322] Allele non-interaction information can also include the expression level of proteasome, immunoproteasome, thymic proteasome or other proteases in tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry or immunohistochemical analysis). Different proteasomes have different cleavage site preferences. The cleavage preference of each type of proteasome proportional to the expression level will be given a greater weight.

[0323] Allele non-interaction information can also include the expression level of the source gene of the peptide (e.g., measured by RNA-seq or mass spectrometry). Possible optimization measures include adjusting the expression level measurements to account for the presence of stromal cells and tumor infiltrating lymphocytes in the tumor sample. Peptides from genes with higher expression levels are more likely to be presented. Peptides from genes with undetectable expression levels can be disregarded.

[0324] Allelic non-interaction information may also include the probability that the source mRNA encoding the neoantigen peptide will undergo nonsense-mediated decay as predicted by a nonsense-mediated decay model, such as the model from Rivas et al., Science 2015.

[0325] Allelic non-interaction information can also include typical tumor-specific expression levels of the source gene of the peptide during various cell cycle stages. Genes that are expressed at lower levels overall (as measured by RNA-seq or plain proteomics) but are known to be expressed at high levels during specific cell cycle stages may produce more presented peptides than genes that are stably expressed at very low levels.

[0326] Allelic non-interaction information can also include a comprehensive catalog of source protein features, such as those provided in uniProt or the PDB http: / / www.rcsb.org / pdb / home / home.do. These features can include, among other things, secondary and tertiary structure, subcellular localization, and Gene Ontology (GO) terms for the protein. Specifically, this information can contain annotations that operate at the protein level, such as 5'UTR length, and annotations that operate at the level of specific residues, such as the helical motif between residues 300 and 310. These features can also include turn motifs, folded motifs, and disordered residues.

[0327] Allelic non-interaction information may also include features describing the properties of the domain of the source protein containing the peptide, such as: secondary or tertiary structure (eg, alpha helix versus beta sheet); alternative splicing.

[0328] Allele non-interaction information can also include the association between the peptide sequence of the neoantigen and one or more k-mer units of the multiple k-mer units of the neoantigen source gene (as present in the nucleotide sequencing data of the subject). During the training of the presentation model, these associations between the peptide sequence of the neoantigen and the k-mer units of the neoantigen nucleotide sequencing data are input into the model and are used in part by the model to learn model parameters representing the presence or absence of presentation hotspots of the k-mer units associated with the training peptide sequences. Then, during the use of the model after training, the association between the test peptide sequence and the one or more k-mer units of the source gene of the test peptide sequence is input into the model, and the parameters learned by the model during the training process enable the presentation model to make more accurate predictions of the likelihood of presentation of the test peptide sequence.

[0329] In general, the model parameter representing the presence or absence of a presentation hotspot for a k-mer unit represents the residual tendency of the k-mer unit to produce a presented peptide after controlling for all other variables (such as peptide sequence, RNA expression, amino acids typically found in HLA-binding peptides, etc.). The parameter representing the presence or absence of a presentation hotspot for a k-mer unit can be a binary coefficient (e.g., 0 or 1) or a simulated coefficient along a scale (e.g., between 0 and 1, including 0 and 1). In either case, a larger coefficient (e.g., close to 1 or 1) indicates that the k-mer unit has a greater likelihood of producing a presented peptide controlling for other factors, while a lower coefficient (e.g., close to 0 or 0) indicates that the k-mer unit has a lower likelihood of producing a presented peptide. For example, a k-mer unit with a low hotspot coefficient may be a k-mer unit from a gene with high RNA expression, where amino acids are typically found in HLA-binding peptides, where the source gene produces many other presented peptides, but presented peptides rarely appear in the k-mer unit. Because other sources of peptide presence may already be accounted for by other parameters (e.g., RNA expression on a k-mer or larger basis, which is commonly found in HLA-binding peptides), these hotspot parameters provide new, separate information that does not “double count” information captured by other parameters.

[0330] Allele-non-interaction information may also include the probability of presentation of peptides from the source protein of the related peptide in other individuals (after adjusting for the expression levels of the source protein in these individuals and the effects of different HLA types of these individuals).

[0331] Allelic non-interaction information can also include the probability that the peptide is not detected or is overrepresented by mass spectrometry due to technical bias.

[0332] The expression of various gene modules / pathways measured by gene expression assays such as RNASeq, microarrays, targeted panels such as Nanostring, or single / multi-gene representation of gene modules measured by assays such as RT-PCR (without the need to contain the source protein of the peptide) provides information about the status of tumor cells, stroma, or tumor infiltrating lymphocytes (TILs).

[0333] Allelic non-interaction information can also include the copy number of the source gene of the peptide in the tumor cell.For example, a peptide of a gene that has undergone homozygous deletion in the tumor cell can be assigned a presentation probability of zero.

[0334] Allele-independent interaction information may also include the probability of a peptide binding to TAP or a measured or predicted value of the binding affinity of the peptide to TAP. Peptides that are more likely to bind to TAP, or peptides that bind to TAP with higher affinity, are more likely to be presented by MHC-I.

[0335] Allele-independent interaction information can also include the expression level of TAP in tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry, or immunohistochemistry). For MHC-I, higher TAP expression levels increase the presentation probability of all peptides.

[0336] Allele-independent interaction information may also include the presence or absence of tumor mutations, including but not limited to:

[0337] i. Driver mutations in known cancer driver genes, such as EGFR, KRAS, ALK, RET, ROS1, TP53, CDKN2A, CDKN2B, NTRK1, NTRK2, and NTRK3

[0338] ii. Mutations in genes encoding proteins involved in the antigen presentation machinery (e.g., B2M, HLA-A, HLA-B, HLA-C, TAP-1, TAP-2, TAPBP, CALR, CNX, ERP57, HLA-DM, HLA-DMA, HLA-DMB, HLA-DO, HLA-DOA, HLA-DOBHLA-DP, HLA-DPAl, HLA-DPBl, HLA-DQ, HLA-DQA1, HLA-DQA2, HLA-DQBl, HLA-DQB2, HLA-DR, HLA-DRA, HLA-DRBl, HLA-DRB3, HLA-DRB4, HLA-DRB5, or any gene encoding a component of the proteasome or immunoproteasome). Peptides whose presentation depends on components of the antigen presentation machinery that have undergone loss-of-function mutations in the tumor have a reduced probability of presentation.

[0339] The presence or absence of functional germline polymorphisms, including but not limited to:

[0340] i. Functional germline polymorphisms in genes encoding proteins involved in the antigen presentation machinery (e.g., B2M, HLA-A, HLA-B, HLA-C, TAP-1, TAP-2, TAPBP, CALR, CNX, ERP57, HLA-DM, HLA-DMA, HLA-DMB, HLA-DO, HLA-DOA, HLA-DOB, HLA-DP, HLA-DPAl, HLA-DPBl, HLA-DQ, HLA-DQA1, HLA-DQA2, HLA-DQBl, HLA-DQB2, HLA-DR, HLA-DRA, HLA-DRBl, HLA-DRB3, HLA-DRB4, HLA-DRB5, or any gene encoding a component of the proteasome or immunoproteasome)

[0341] Allele non-interaction information can also include tumor type (eg, NSCLC, melanoma).

[0342] Allele non-interaction information can also include the known function of the HLA allele, as reflected by, for example, the suffix of the HLA allele. For example, the N suffix in the allele name HLA-A*24:09N indicates a null allele that is not expressed and therefore cannot present epitopes; the complete HLA allele suffix nomenclature is described at https: / / www.ebi.ac.uk / ipd / imgt / hla / nomenclature / suffixes.html.

[0343] Allele non-interaction information can also include clinical tumor subtype (eg, squamous lung cancer versus non-squamous lung cancer).

[0344] Allele non-interaction information can also include smoking history.

[0345] Allele-independent interaction information may also include a history of sunburn, sun exposure, or exposure to other mutagens.

[0346] Allelic non-interaction information can also include the typical expression of the peptide's source gene in relevant tumor types or clinical subtypes, optionally stratified by driver gene mutations. Genes that are typically expressed at high levels in relevant tumor types are more likely to be presented.

[0347] Allele-independent interaction information can also include mutation frequencies in all tumors, or in tumors of the same type, or in tumors from individuals with at least one shared MHC allele, or in tumors of the same type within individuals with at least one shared MHC allele.

[0348] For mutated tumor-specific peptides, the feature list for predicting presentation probability can also include mutation annotations (e.g., missense, read-through, frameshift mutations, fusions, etc.) or predictions of whether the mutation will cause nonsense-mediated decay (NMD). For example, peptides from protein segments that are not translated in tumor cells due to homozygous early termination mutations can be assigned a presentation probability of zero. NMD reduces mRNA translation, thereby reducing the presentation probability.

[0349] VII.C. Presentation Identification System

[0350] Figure 3is a high-level block diagram illustrating the computer logic components of a presentation identification system 160 according to one embodiment. In this example embodiment, the presentation identification system 160 includes a data management module 312, an encoding module 314, a training module 316, and a prediction module 320. The presentation identification system 160 also includes a training data storage 170 and a presentation model storage 175. Some embodiments of the model management system 160 have different modules than those described herein. Similarly, the functional distribution of these modules may differ from the modules described herein.

[0351] VII.C .1. Data Management Module

[0352] The data management module 312 generates a set of training data 170 according to the submitted information 165. Each set of training data contains multiple data instances, where each data instance i contains a set of independent variables z i , these independent variables include at least one presented or unpresented peptide sequence p i , one or more i Associated related MHC alleles a i , and / or one or more peptide sequences p i Associated MHC allele sequence d i ; and a dependent variable y i , the dependent variable represents information that the presentation identification system 160 intends to use to predict new values ​​of the independent variable.

[0353] In a specific embodiment described throughout the rest of this specification, the dependent variable y i is a binary label that indicates the peptide p i Is one or more relevant MHC alleles a i and / or by one or more MHC allele sequences d i However, it should be understood that in other embodiments, depending on the independent variable z i , the dependent variable y i can represent any other type of information that the presentation identification system 160 is interested in predicting. For example, in another embodiment, the dependent variable y i It may also be a numerical value indicative of the mass spectrometric ion current of the identified data instance.

[0354] The peptide sequence p of data instance i i is a k i A sequence of amino acids, where k ican vary within a certain range depending on the data instance i. For example, the range can be 8-15 for class I MHC or 6-30 for class II MHC. In one embodiment of system 160, all peptide sequences p in a training data set are i They may have the same length, for example, 9. The number of amino acids in the peptide sequence may vary depending on the type of MHC allele (e.g., MHC alleles in humans, etc.). i Indicates the presence of the corresponding peptide sequence p i Similarly, in some embodiments, the MHC allele sequence d of data instance i i Indicates which MHC allele sequences are associated with the corresponding peptide sequence p i exist in association.

[0355] The data management module 312 may also include additional allele interaction variables, such as those associated with the peptide sequence p contained in the training data 170. i and related MHC alleles a i The binding affinity of i and stability prediction value s i For example, the training data 170 may contain peptide p i With a i The predicted binding affinity between the indicated MHC molecules b i As another example, the training data 170 may contain i Stability prediction value s for each indicated MHC allele i .

[0356] The data management module 312 may also include allele non-interaction variables w i , such as with the peptide sequence p i Related C-terminal flanking sequences and mRNA quantitative measurements.

[0357] The data management module 312 also identifies peptide sequences that are not presented by the MHC alleles to generate training data 170. Generally, this involves identifying a "longer" sequence of the source protein that includes the presented peptide sequence before presentation. When the presentation information contains an engineered cell line, the data management module 312 identifies a series of peptide sequences in the synthetic protein to which these cells are exposed that are not presented on the cells' MHC alleles. When the presentation information contains a tissue sample, the data management module 312 identifies the source protein that is the source of the presented peptide sequence and identifies a series of peptide sequences in the source protein that are not presented on the MHC alleles of the tissue sample cells.

[0358] The data management module 312 can also artificially generate peptides using random amino acid sequences and identify the generated sequences as peptides that are not presented on MHC alleles. This can be achieved by randomly generating peptide sequences, allowing the data management module 312 to easily generate a large amount of synthetic data related to peptides that are not presented on MHC alleles. Since only a small number of peptide sequences are actually presented by MHC alleles, synthetically generated peptide sequences are likely not to be presented by MHC alleles, even if these sequences are included in proteins processed by cells.

[0359] Figure 4 A set of exemplary training data 170A according to one embodiment is shown. Specifically, the first three data instances in training data 170A indicate peptide presentation information obtained from a monoallelic cell line comprising the allele HLA-C*01:03 and the three peptide sequences QCEIOWAREFLKEIGJ, FIEUHFWI, and FEWRHRJTRUJR. Note that in alternative embodiments of training data 170A, the HLA allele type can be replaced by the HLA allele sequence. For example, the allele type HLA-C*1:03 can be replaced by the amino acid sequence of the allele HLA-C*1:03. The fourth data instance in training data 170A indicates peptide presentation information obtained from a multiallelic cell line comprising the alleles HLA-B*07:02, HLA-C*01:03, HLA-A*01:01, and the peptide sequence QIEJOEIJE. The first data example indicates that the peptide sequence QCEIOWARE is not presented by the HLA-DRB3:01:01 allele. As discussed in the previous two paragraphs, negatively labeled peptide sequences can be randomly generated by the data management module 312 or identified from the source protein presenting the peptide. The training data 170A also includes a binding affinity prediction value of 1000 nM for the peptide sequence-allele pair and a stability prediction value of 1 hour half-life. The training data 170A also includes allele-independent interaction variables, such as the C-terminal flanking sequence of the peptide FJELFISBOSJFIE and 10 2 Quantitative mRNA measurements of TPM. The fourth data example indicates that the peptide sequence QIEJOEIJE is presented by one of the alleles HLA-B*07:02, HLA-C*01:03, or HLA-A*01:01. Training data 170A also includes binding affinity predictions and stability predictions for each allele, as well as the C-terminal flanking sequence of the peptide and quantitative mRNA measurements of the peptide. In further embodiments, training data 170A may also include additional allele-independent variables, such as the peptide family of the presented peptide.

[0360] VII.C.2. Encoding Module

[0361] The encoding module 314 encodes the information contained in the training data 170 into a digital representation that can be used to generate one or more presentation models. In one embodiment, the encoding module 314 uniquely encodes sequences (e.g., peptide sequences and / or C-terminal flanking sequences and / or MHC allele sequences) within a predetermined 20-letter amino acid alphabet. Specifically, a sequence with k i The peptide sequence p of amino acids i Expressed as having 20·k i A row vector of elements, where p i 20·(j-1)+1 ,p i 20·(j-1)+2 ,…,p i 20·j The value of the single element corresponding to the amino acid at position j in the peptide sequence in the alphabet is 1. In addition, the values ​​of the remaining elements are 0. For example, for a given alphabet {A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y}, the peptide sequence EAF with 3 amino acids for data instance i can be represented by a row vector p with 60 elements i =[0 0 0 1 0 0 0 00 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 00 0 0 0 0 0 0 0 0 0 0 0 0 0 0]. C-terminal flanking sequence c i , and the protein sequence of the MHC allele d i And other sequence data in the presentation information can be encoded in a similar manner as described above.

[0362] When the training data 170 contains sequences of different amino acid lengths, the encoding module 314 can also encode the peptides into vectors of equal length by adding PAD characters to expand the predetermined alphabet. For example, this can be done by padding the peptide sequence on the left with PAD characters until the length of the peptide sequence reaches the maximum length of the peptide sequence in the training data 170. Therefore, when the peptide sequence with the maximum length has k max When there are 20 amino acids, the encoding module 314 digitally represents each sequence as a sequence with (20+1)k max For example, for the extended alphabet {PAD,A,C,D,E,F,G,H,I,K,L,M,N,P,Q,R,S,T,V,W,Y} and k max =5, the same example peptide sequence EAF with 3 amino acids can be represented by a row vector with 105 elements: i=[1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 00 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]. C-terminal flanking sequence c i , protein sequences of MHC alleles d i or other sequence data may be encoded in a manner similar to that described above. Thus, the peptide sequence p i 、c i or d i Each independent variable or column of represents the presence of a specific amino acid at a specific position in the sequence.

[0363] Although the above method for encoding sequence data is described with reference to sequences having amino acid sequences, the method can be similarly extended to other types of sequence data, such as DNA or RNA sequence data.

[0364] The encoding module 314 also adds one or more MHC alleles a of the data instance i i is encoded into a row vector with m elements, where each element h = 1, 2, ..., m corresponds to a uniquely identified MHC allele. The value of the element corresponding to the MHC allele of the identified data instance i is 1. In addition, the values ​​of the remaining elements are 0. For example, among m = 4 uniquely identified MHC allele types {HLA-A*01:01, HLA-C*01:08, HLA-B*07:02, HLA-DRB1*10:01}, the alleles HLA-B*07:02 and HLA-DRB1*10:01 corresponding to the data instance i of the multi-allelic cell line can be represented by a row vector with 4 elements: i =[0 0 1 1], where a3 i =1 and a4 i = 1. Although this example is described herein using four identified MHC allele types, the number of MHC allele types may actually be in the hundreds or thousands. As previously discussed, each data instance i typically contains up to six different allele types that are associated with the peptide sequence p. i Related MHC allele types.

[0365] The encoding module 314 also adds the label y of each data instance i iEncoded as a binary variable with values ​​from the set {0,1}, where the value 1 indicates that the peptide x i By the associated MHC allele a i , while a value of 0 indicates that peptide x i Not affected by any related MHC allele a i When the dependent variable y i When representing mass spectrometer ion currents, the encoding module 314 may additionally scale the values ​​using various functions, such as a logarithmic function with a range of (-∞,∞) for ion currents between [0,∞).

[0366] The encoding module 314 can i The interaction variable x between a pair of alleles and the associated MHC allele h h i is represented as a row vector in which the numerical representations of the allele interaction variables are concatenated one after another. For example, the encoding module 314 can h i Expressed as equal to [p i ]、[p i b h i ]、[p i s h i ] or [p i b h i s h i ], where b h i It is peptide p i and the predicted binding affinity of the relevant MHC allele h, and similarly s h i Alternatively, one or more combinations of allele interaction variables may be stored individually (eg, in individual vector or matrix form).

[0367] In one example, the encoding module 314 generates the allele interaction variable x by incorporating the measured or predicted value of binding affinity into the allele interaction variable x. h i The binding affinity information is shown in the figure.

[0368] In one example, the encoding module 314 generates a variable x by incorporating a measured or predicted value of binding stability into the allele interaction variable x. h i Indicates binding stability information.

[0369] In one example, the encoding module 314 calculates the binding association rate by incorporating the measured or predicted value into the allele interaction variable xh i The binding and association rate information is represented in .

[0370] In one example, for peptides presented by class I MHC molecules, encoding module 314 represents the peptide length as a vector in is the indicator function, and L k Peptide p k The length of the vector T k can be included in the allele interaction variable x h i In another example, for peptides presented by class II MHC molecules, encoding module 314 represents the peptide length as a vector in is the indicator function, and L k Peptide p k The length of the vector T k can be included in the allele interaction variable x h i middle.

[0371] In one example, the encoding module 314 generates an MHC allele expression level by incorporating the allele interaction variable x based on RNA-seq. h i The RNA expression information of MHC alleles is represented in the middle.

[0372] Similarly, the encoding module 314 can convert the allele non-interaction variable w i is represented as a row vector where the numeric representations of the allele non-interacting variables are concatenated one after the other. For example, w i Can be equal to [c i ] or [c i m i w i ], where w i Is the peptide p i The C-terminal flanking sequence and the quantitative measurement of mRNA associated with the peptide i In addition, a row vector representing any other allele non-interacting variables is also provided. Alternatively, one or more combinations of allele non-interacting variables can be stored individually (e.g., in the form of individual vectors or matrices).

[0373] In one example, the encoding module 314 generates the allele-independent interaction variable w by incorporating the conversion rate or half-life into the allele-independent interaction variable w. i is used to represent the conversion rate of the source protein of the peptide sequence.

[0374] In one example, the encoding module 314 generates protein length by incorporating the protein length into the allele non-interacting variable w i is used to indicate the length of the source protein or isoform.

[0375] In one example, the encoding module 314 includes β1 i β2 i β5 i The average expression level of immunoproteasome-specific proteasome subunits, including the subunits, was incorporated into the allele-independent interaction variable w i The middle represents the activation of immunoproteasome.

[0376] In one example, the encoding module 314 calculates the abundance of the source protein by incorporating the allele-independent interaction variable w i The RNA-seq abundance of the source protein of the peptide or the gene or transcript of the peptide is expressed (quantified in FPKM, TPM by techniques such as RSEM).

[0377] In one example, the encoding module 314 incorporates the probability that the source transcript of the peptide will undergo nonsense-mediated decay (NMD) estimated using the model in Rivas et al., Science, 2015 into the allele non-interaction variable w i denoted by .

[0378] In one example, the encoding module 314 quantifies the expression levels of genes in the pathway in units of TPM, for example, by using RSEM of each gene in the pathway, and then calculates summary statistics, such as the mean, of all genes in the pathway to represent the activation state of the gene module or pathway assessed by RNA-seq. The mean can be incorporated into the allele non-interaction variable w i middle.

[0379] In one example, the encoding module 314 generates the allele non-interacting variable w by incorporating the copy number into the allele non-interacting variable w. i The copy number of the source gene is indicated in .

[0380] In one example, the encoding module 314 includes the measured or predicted TAP binding affinity (e.g., in nanomolar units) in the allele non-interacting variable w i The TAP binding affinity is shown in the middle.

[0381] In one example, the encoding module 314 includes TAP expression levels measured using RNA-seq (and quantified in TPM using, for example, RSEM) in the allele-independent interaction variable w i TAP expression levels are shown in the middle.

[0382] In one example, the encoding module 314 generates the allele non-interacting variable w i In the , tumor mutations are represented as a vector of indicator variables (i.e., if the peptide p k From samples with KRAS G12D mutation, d k =1, otherwise 0).

[0383] In one example, the encoding module 314 represents germline polymorphisms in antigen presenting genes as a vector of indicator variables (ie, if peptide p k From samples with species germline polymorphism in TAP, d k =1). These indicator variables can all be included in the allele non-interaction variable w i middle.

[0384] In one example, the encoding module 314 represents the tumor type as a one-hot encoded vector of length one according to an alphabetical list of tumor types (e.g., NSCLC, melanoma, colorectal cancer, etc.). These one-hot encoded variables can all be included in the allele non-interaction variable w i middle.

[0385] In one example, encoding module 314 represents MHC allele suffixes by treating HLA alleles with four digits with different suffixes. For example, for the purposes of this model, HLA-A*24:09N is considered a different allele from HLA-A*24:09. Alternatively, since HLA alleles ending with the N suffix are not expressed, the presentation probability of MHC alleles ending with the N suffix can be set to zero for all peptides.

[0386] In one example, the encoding module 314 represents the tumor subtype as a one-hot encoded vector of length one according to the alphabet of tumor subtypes (e.g., lung adenocarcinoma, lung squamous cell carcinoma, etc.). These one-hot encoded variables can all be included in the allele non-interaction variable w i middle.

[0387] In one example, the encoding module 314 represents smoking history as a binary indicator variable (if the patient has a history of smoking, then d k =1, otherwise 0), this variable can be included in the allele non-interaction variable w iAlternatively, smoking history can be encoded as a variable with a one-hot encoding of length 1, based on an alphabetical list of smoking severity. For example, smoking status can be rated on a 1-5 scale, with 1 indicating a non-smoker and 5 indicating a current heavy smoker. Since smoking history is primarily associated with lung tumors, when training models on multiple tumor types, this variable can also be defined as equal to 1 if the patient has a smoking history and the tumor type is lung tumor, and zero otherwise.

[0388] In one example, the encoding module 314 represents the history of sunburn as a binary indicator variable (if the patient has a history of severe sunburn, then d k =1, otherwise 0), this variable can be included in the allele non-interaction variable w i Since severe sunburn is primarily associated with melanoma, when training a model on multiple tumor types, this variable can also be defined to be equal to 1 if the patient has a history of severe sunburn and the tumor type is melanoma, and zero otherwise.

[0389] In one example, the encoding module 314 represents the expression level distribution of a specific gene or transcript for each gene or transcript in the human genome as a summary statistic (e.g., mean, median) of the expression level distribution using a reference database such as TCGA. Specifically, for a peptide p in a sample whose tumor type is melanoma, k , not only can the peptide p k The gene or transcript expression level measurements of the source genes or transcripts are included in the allele non-interacting variables w i In addition, peptide p-terminal expression in melanoma was measured by TCGA. k The average and / or median gene or transcript expression levels of the source genes or transcripts are determined.

[0390] In one example, the encoding module 314 represents the mutation type as a variable of length one one-hot encoding according to the alphabet of mutation types (e.g., missense mutation, frameshift mutation, NMD-induced mutation, etc.). These one-hot encoded variables can all be included in the allele non-interaction variable w i middle.

[0391] In one example, the encoding module 314 generates the allele non-interacting variable w i In another example, the encoding module 314 represents the protein level feature of the protein as the annotation value of the source protein (such as 5'UTR length). i Include indicator variables to represent p i The source protein annotation at the residue level, that is, if the peptide p i Equal to 1 if it overlaps with the helical motif and 0 otherwise, or if the peptide p iis completely contained within the helical motif is equal to 1. In another example, the peptide p i The proportion of residues included in the helical motif annotation can be characterized by the allele non-interacting variable w i middle.

[0392] In one example, the encoding module 314 represents the type of protein or isoform in the human proteome as an indicator vector o k , the length of this vector is equal to the number of proteins or isoforms in the human proteome, and if the peptide p k From protein i, the corresponding element o k i is 1, otherwise it is 0.

[0393] In one example, the encoding module 314 converts the peptide p i The source gene G=gene(p i ) is represented as a categorical variable with L possible categories, where L represents the upper limit of the number of indexed source genes 1, 2, ...L.

[0394] In one example, the encoding module 314 converts the peptide p i Tissue type, cell type, tumor type or tumor histological type T = tissue (p i ) is represented as a categorical variable with M possible categories, where M represents an upper limit on the number of index types 1, 2, ..., M. Tissue types may include, for example, lung tissue, heart tissue, intestinal tissue, neural tissue, etc. Cell types may include dendritic cells, macrophages, CD4 T cells, etc. Tumor types may include lung adenocarcinoma, lung squamous cell carcinoma, melanoma, non-Hodgkin's lymphoma, etc.

[0395] The encoding module 314 can also convert the peptide p i and the variable z associated with the MHC allele h i The population set of is represented as a row vector, where the allele interaction variables x i and allele non-interaction variable w i For example, the encoding module 314 can concatenate the digital representations of z h i Expressed as equal to [x h i w i ] or [w i x h i ] is a row vector.

[0396] VIII. Training Module

[0397] The training module 316 constructs one or more presentation models that generate the probability of whether a peptide sequence will be presented by the MHC allele associated with the peptide sequence. Specifically, given a peptide sequence p k and peptide sequence p k A group of MHC alleles associated with k and / or MHC allele sequence d k , each presented model generates an estimate u k , indicating the peptide sequence p k will be associated with one or more related MHC alleles a k Possibility of presentation.

[0398] VIII.A. Overview

[0399] The training module 316 constructs one or more presentation models based on the training data set stored in the memory 170 generated from the presentation information stored in 165. In general, regardless of the specific type of presentation model, all presentation models capture the correlation between the independent variables and the dependent variables in the training data 170 to minimize the loss function. Specifically, the loss function The dependent variable y represents one or more data instances S in the training data 170 i∈S and the estimated likelihood u of the data instance S generated by the presented model i∈S In a specific embodiment described throughout the remainder of this specification, the loss function (y i ∈S ,u i∈S ; θ) is the negative log-likelihood function provided by the following equation (1a):

[0400]

[0401] However, in practice, another loss function can be used. For example, when predicting mass spectrometry ion current, the loss function is the mean square loss provided by the following equation 1b:

[0402]

[0403] The presented model can be a parametric model, in which one or more parameters θ mathematically specify the correlation between the independent variable and the dependent variable. i∈S ,u i∈S The various parameters of the parametric presentation model with the smallest θ are determined by a gradient-based numerical optimization algorithm, such as a batch gradient algorithm, a stochastic gradient algorithm, etc. Alternatively, the presentation model can be a non-parametric model, in which the model structure is determined by the training data 170 and is not strictly based on a fixed parameter set.

[0404] VIII.B. Independent Allele Model

[0405] The training module 316 can construct a presentation model on a per-allele basis to predict the presentation likelihood of a peptide. In this case, the training module 316 can train the presentation model based on data instances S in the training data 170 generated by cells expressing a single MHC allele.

[0406] In one embodiment, the training module 316 trains the specific allele h for the peptide p by the following formula: k The estimated probability of presentation u k Modeling:

[0407]

[0408] where x h k Indicates the encoded peptide p k and the allele interaction variable of the corresponding MHC allele h, f(·) is any function and is referred to as the transformation function throughout this article for ease of explanation. In addition, g h (·) is any function, for ease of explanation, referred to herein as the dependency function, and is based on the parameter set θ of the measured MHC allele h h Generate allele interaction variables x h k The parameter set θ for each MHC allele h h The value of θ can be obtained by h It is determined by minimizing the loss function , where i is each instance in the subset S of the training data 170 generated by cells expressing a single MHC allele h.

[0409] Dependency function g h (x h k θ h ) represents the output value of at least the allele interaction feature x h k , and more specifically, based on the peptide p k The dependency score of the amino acid position in the peptide sequence of for MHC allele h indicates that MHC allele h will present the corresponding neoantigen. For example, if MHC allele h is likely to present peptide p k , then the dependency score of the MHC allele h may have a higher value, whereas it may have a lower value if presentation is not possible. The transformation function f(·) will be input, and more precisely in this case will be represented by g h(x h k θ h ) is transformed into appropriate values ​​to indicate the peptide p k The probability that it will be presented by the MHC allele.

[0410] In one particular embodiment referred to throughout the remainder of this specification, f(·) is a function having a range in [0, 1] for a suitable domain range. In one embodiment, f(·) is an expit function provided by:

[0411]

[0412] As another example, when the value of the domain z is equal to or greater than 0, f(·) may also be a hyperbolic tangent function provided by the following formula:

[0413] f(z)=tanh(z) (4)

[0414] Alternatively, when the predicted value of the mass spectrometer ion current exceeds the range [0, 1], f(·) can be any function, such as an identity function, an exponential function, a logarithmic function, etc.

[0415] Therefore, the dependence function g on the MHC allele h can be expressed as h (·) applied to the peptide sequence p k The encoded form of the peptide sequence p is generated to produce the corresponding dependency score k The independent allele probability that will be presented by MHC allele h. The dependency score can be transformed by the transformation function f(·) to produce the peptide sequence p k Independent allelic probabilities to be presented by MHC allele h.

[0416] VIII.B .1 Dependence function on allelic interaction variables

[0417] In one particular embodiment referred to throughout the specification, the dependency function g h (·) is the affine function given by:

[0418]

[0419] This function converts x h k Each allele interaction variable in the set of parameters θ is associated with the measured MHC allele h h The corresponding parameters in are linearly combined.

[0420] In another specific embodiment mentioned throughout this specification, the dependency function g h(·) is the network function provided by:

[0421]

[0422] The network function consists of a network model NN having a series of nodes arranged in one layer or multiple layers. h (·) indicates that a node can be connected to other nodes through connectors, each of which has a parameter set θ h The value at a particular node can be expressed as the sum of the values ​​of the nodes connected to that particular node, weighted by the relevant parameters mapped by the activation function associated with that particular node. Network models are advantageous over affine functions because they can incorporate nonlinear and process data with different amino acid sequence lengths. Specifically, through nonlinear modeling, network models can capture the interactions between amino acids at different positions in the peptide sequence and how these interactions affect peptide presentation.

[0423] Generally speaking, the network model NN h (·) Can be constructed as feedforward networks, such as artificial neural networks (ANNs), convolutional neural networks (CNNs), deep neural networks (DNNs), and / or recurrent networks, such as long short-term memory networks (LSTMs), bidirectional recurrent networks, deep bidirectional recurrent networks, etc.

[0424] In one example mentioned throughout the rest of this specification, each MHC allele in h=1, 2, ..., m is associated with an independent network model, and NN h (·) represents the output from the network model associated with the MHC allele h.

[0425] Figure 5 An example network model NN3(·) associated with an arbitrary MHC allele h=3 is shown. Figure 5 As shown in , the network model NN3(·) for the MHC allele h=3 includes three input nodes at layer l=1, four nodes at layer l=2, two nodes at layer l=3, and one output node at layer l=4. The network model NN3(·) is associated with a set of ten parameters θ3(1), θ3(2), ..., θ3(10). The network model NN3(·) receives three allele interaction variables x3 for the MHC allele h=3. k (1), x3 k (2) and x3 k (3) takes as input the individual data instances of the encoded polypeptide sequence data and any other training data used and outputs the value NN3(x3 k). The network function may also include one or more network models, each network model taking a different allele interaction variable as input.

[0426] In another embodiment, the identified MHC alleles h=1, 2, ..., m are compared with a single network model NN H (·) is associated, and NN h (·) represents one or more outputs of a single network model associated with the MHC allele h. In this instance, the parameter set θ h can correspond to the parameter set of this single network model, and therefore, the parameter set θ h May be common to all MHC alleles.

[0427] Figure 6 An example network model NN showing sharing of MHC alleles h = 1, 2, ..., m H (·).like Figure 6 As shown in the network model NN H (·) includes m output nodes, each corresponding to an MHC allele. The network model NN3(·) receives the allele interaction variable x3 for MHC allele h=3 k And output m values, including the value NN3(x3 corresponding to MHC allele h=3 k ).

[0428] In yet another embodiment, the dependency function g h (·) can be expressed as:

[0429]

[0430] where g' h (x h k ;θ' h ) is a parameter set θ' h Affine functions, network functions, etc., where the deviation parameter θ of the parameter set of allele interaction variables related to MHC alleles is h 0 represents the baseline presentation probability of MHC allele h.

[0431] In another embodiment, the deviation parameter θ h 0 It can be shared by the gene family of MHC allele h. That is, the bias parameter θ of MHC allele h h 0 can be equal to θ 基因(h) 0, where gene (h) is the gene family of the MHC allele h. For example, the class I MHC alleles HLA-A*02:01, HLA-A*02:02, and HLA-A*02:03 can be assigned to the "HLA-A" gene family, and the deviation parameter θ for each of these MHC alleles is h 0 As another example, the class II MHC alleles HLA-DRB1:10:01, HLA-DRB1:11:01, and HLA-DRB3:01:01 can be assigned to the "HLA-DRB" gene family, and the bias parameters θ for each of these MHC alleles are h 0 Can be shared.

[0432] Returning to equation (2), as an example, using the affine dependency function g h (·) Among the m=4 different MHC alleles identified, peptide p k The probability of being presented by MHC allele h=3 can be given by:

[0433]

[0434] where x3 k is the allele interaction variable for the identified MHC allele h=3, and θ3 is the set of parameters for the MHC allele h=3 determined by minimization of the loss function.

[0435] As another embodiment, when using an independent network transformation function g h (·) Among the m=4 different MHC alleles identified, peptide p k The probability that it will be presented by MHC allele h=3 can be obtained from the following formula:

[0436]

[0437] where x3 k is the allele interaction variable identified for the MHC allele h=3, and θ3 is the set of parameters of the network model NN3(·) determined to be associated with the MHC allele h=3.

[0438] Figure 7 The example network model NN3(·) is used to generate the peptide p associated with the MHC allele h=3. k The possibility of submission. Figure 7 As shown in , the network model NN3(·) receives the allele interaction variable x3 for the MHC allele h=3. k And generate output NN3(x3 k). This output is mapped by the function f(·) to produce the estimated presentation probability u k .

[0439] VIII.B.2. Independent Alleles with Allele-Non-Interacting Variables

[0440] In one embodiment, the training module 316 incorporates allele non-interacting variables and uses the following formula to make the peptide p k The estimated probability of presentation u k Modeling:

[0441]

[0442] where w k Peptide p k The non-interacting variable of the coded alleles, g w (·) is the parameter set θ based on the measured allele non-interacting variables w The allele non-interaction variable w k Specifically, the parameter set θ for each MHC allele h h and the set of parameters θ about the non-interacting variables of the alleles w The value of θ can be obtained by h and θ w It is determined by minimizing the loss function , where i is each instance in the subset S of the training data 170 generated by cells expressing a single MHC allele.

[0443] Dependency function g w (w k θ w The output of ) represents the dependence score of the allele non-interacting variable based on the effect of the allele non-interacting variable, which indicates that the peptide p k Whether it will be presented by one or more MHC alleles. For example, if the peptide p k with peptides known to positively affect k The dependence score of the allele-non-interacting variant is likely to have a higher value if the peptide p k with peptides known to adversely affect k The C-terminal flanking sequence of the presented protein may have a lower value.

[0444] According to equation (7), the function g related to MHC allele h can be h (·) applied to the peptide sequence p k The encoded form of the peptide sequence p is used to generate the corresponding dependency score of the allelic interaction variable kThe probability of an independent allele being presented by an MHC allele h. A function of the non-interacting variables for the alleles g w (·) is also applied to the encoded form of the allele-independent variable to produce a dependency score for the allele-independent variable. The two scores are combined and the combined score is transformed by the transformation function f(·) to produce the peptide sequence p k Independent allelic probabilities to be presented by MHC allele h.

[0445] Alternatively, the training module 316 can be implemented by transforming the allele non-interaction variable w k The allele non-interaction variable x added to Equation (2) h k To convert the allele non-interaction variable w k Included in the predicted value. Therefore, the presentation probability can be obtained as follows:

[0446]

[0447] VIII.B.3 Dependence Functions on Non-Interacting Variables of Alleles

[0448] The dependency function g of the interaction variable with the relevant allele h (·) Similarly, the dependence function g on the non-interacting variable of the allele w (·) can be an affine function or a network function, where the independent network model is related to the allele non-interacting variable w k associated.

[0449] Specifically, the dependency function g w (·) is the affine function given by:

[0450] g w (w k θ w )=w k ·θ w .

[0451] This function converts the allele non-interaction variable w k With parameter set θ w The corresponding parameters in are linearly combined.

[0452] Dependency function g w (·) can also be a network function provided by:

[0453] g h (w k θ w )=NN w (w k θ w ).

[0454] The function is given by a parameter set θ w The network model NN with relevant parameters in w (·) indicates that the network function may also include one or more network models, each of which takes a different allele non-interacting variable as input.

[0455] In another embodiment, the dependency function g on the allele non-interacting variable w (·) can be provided by the following formula:

[0456]

[0457] where g' w (w k ;θ' w ) is an affine function with allele non-interacting parameter set θ' w Network functions, etc. k It is peptide p k The quantitative measurement value of mRNA, h(·) is the function that transforms the quantitative measurement value, and θ w m is a parameter in a set of parameters related to allele-independent interactions that is combined with the mRNA quantitative measurement to generate a dependency score related to the mRNA quantitative measurement. In a specific embodiment referred to throughout the remainder of this specification, h(·) is a logarithmic function, although in practice, h(·) can be any of a variety of different functions.

[0458] In yet another example, the dependency function g on the allele non-interacting variable w (·) can be provided by the following formula:

[0459]

[0460] where g' w (w k ;θ' w ) is an affine function with allele non-interaction parameter set θ' w Network functions, etc. k The peptides p in the human proteome described in Section VII.C.2 are k The indicator vectors of proteins and isoforms, and θ w o is a set of parameters in the set of parameters for allele non-interacting variables, combined with the indicator vector. In one variation, when o k The dimension and parameter set θ w oWhen it is obviously higher, the parameter regularization term can be used when determining the parameter value, such as is added to the loss function, where ||·|| represents the L1 norm, L2 norm, combination, etc. The optimal value of the hyperparameter λ can be determined by appropriate methods.

[0461] In yet another example, the dependency function g on the allele non-interacting variable w (·) can be provided by the following formula:

[0462]

[0463] where g' w (w k ;θ' w ) is an affine function with allele non-interaction parameter set θ' w Network functions, etc. is the indicator function, as described above for the allele-independent interaction variable, if the peptide p k comes from the source gene l, then it is equal to 1, and θ w l is a parameter indicating the "antigenicity" of the source gene l. In one variation, when L is significantly higher and thus the parameter θ w l＝1,2,…,L When the number is also significantly higher, the parameter regularization term can be used when determining the parameter value, such as is added to the loss function, where ||·|| represents the L1 norm, L2 norm, combination, etc. The optimal value of the hyperparameter λ can be determined by appropriate methods.

[0464] In yet another example, the dependency function g on the allele non-interacting variable w (·) can be provided by the following formula:

[0465]

[0466] where g' w (w k ;θ' w ) is an affine function with allele non-interaction parameter set θ' w Network functions, etc. Organization (p k )=m) is the indicator function, as described above for the allele non-interacting variables, if the peptide p k From source gene l and if peptide p k from tissue type m, then it is equal to 1, and θ w lmis a parameter indicating the antigenicity of the combination of source gene 1 and tissue type m. Specifically, the antigenicity of gene 1 in tissue type m can represent the residual propensity of cells of tissue m to present peptides from gene 1 after controlling for RNA expression and peptide sequence context.

[0467] In one variation, when L or M is significantly higher and thus the parameter θ w lm＝1,2,…,LM When the number is also significantly higher, the parameter regularization term can be used when determining the parameter value, such as Add to the loss function, where ||·|| represents the L1 norm, L2 norm, or a combination thereof. The optimal value of the hyperparameter λ can be determined by appropriate methods. In another variation, a parameter regularization term can be added to the loss function when determining the parameter value, so that the parameters of the same source gene do not differ significantly between tissue types. For example, the penalty term may be:

[0468]

[0469] The standard deviation of antigenicity between different tissue types can be penalized in the loss function, where is the average antigenicity between tissue types of the source gene l.

[0470] In yet another example, the dependency function g of the allele non-interacting variable w (·) can be given by the following formula:

[0471]

[0472] where g' w (w k ;θ' w ) is an affine function with allele non-interaction parameter θ' w Network functions of a collection of, etc.; is such an indicator function that if the peptide p k from the source gene l as described above with reference to the allele non-interacting variable, then the indicator function is equal to 1; and θ w l is a parameter indicating the “antigenicity” of the source gene 1; and is such an indicator function that if the peptide p k from proteome position m, then the indicator function is equal to 1; and is a parameter indicating the extent to which a proteomic position m is a presentation "hotspot." In one embodiment, a proteomic position may comprise a unit of n adjacent peptides from the same protein, where n is a hyperparameter of the model determined via an appropriate method such as grid search cross validation.

[0473] In practice, additional terms from any of equations (9), (10), (11), (12a), and (12b) can be combined to produce the dependence function g of the allele-independent variable w (·). For example, the term h(·) representing the quantitative measurement of mRNA in equation (9) and the terms representing the antigenicity of the source gene in equations (11), (12a), and (12b) can be added together with any other affine or network functions to generate a dependency function for the allele non-interacting variables.

[0474] Returning to equation (7), as an example, when using the affine transformation function g h (·), g w (·) Among the m=4 different MHC alleles identified, peptide p k The probability of being presented by MHC allele h=3 can be given by:

[0475]

[0476] where w k The identified peptide p k The alleles are non-interacting variables, and θ w is the set of parameters for the measured allele non-interacting variables.

[0477] As another embodiment, when using the network transformation function g h (·), g w (·) Among the m=4 different MHC alleles identified, peptide p k The probability that it will be presented by MHC allele h=3 can be obtained from the following formula:

[0478]

[0479] where w k The identified peptide p k The allele interaction variable, and θ w is the set of parameters for the measured allele non-interacting variables.

[0480] Figure 8 The example network models NN3(·) and NN are shown. w (·) Generate peptide p associated with MHC allele h=3 k The possibility of submission. Figure 8 As shown in , the network model NN3(·) receives the allele interaction variable x3 for the MHC allele h=3. k And generate output NN3(x3 k ). Network model NN w(·) Receive peptide p k The allele non-interaction variable w k And generate output NN w (w k The outputs are combined and mapped by the function f(·) to produce the estimated presentation probability u k .

[0481] VIII.C. Multi-allele model

[0482] The training module 316 can also construct a presentation model to predict the presentation likelihood of a peptide in a multi-allelic environment where two or more MHC alleles are present. In this case, the training module 316 can train the presentation model based on data instances S in the training data 170 generated by cells expressing a single MHC allele, cells expressing multiple MHC alleles, or a combination thereof.

[0483] VIII.C.1. Example 1: Maximum value of the independent allele model

[0484] In one embodiment, the training module 316 trains peptides p associated with a set of multiple MHC alleles H. k The estimated probability of presentation u k The presentation probability u of each MHC allele h in the pool H determined based on cells expressing a single allele k h∈H The variation of is modeled as described above in conjunction with equations (2)-(10). Specifically, the probability u k It can be u k h∈H In one embodiment, as shown in equations (11), (12a) and (12b), the function is a maximum function and presents the probability u k It can be determined as the maximum presentation probability of each MHC allele h in the set H.

[0485]

[0486] VIII.C.2. Example 2.1: Function of and f- Sums model

[0487] In one embodiment, the training module 316 trains the peptide p by the following formula k The estimated probability of presentation u k Modeling:

[0488]

[0489] where element a h k For the peptide sequence p kThe associated multiple MHC alleles H is 1, and x h k Indicates the encoded peptide p k and the allele interaction variables of the corresponding MHC alleles. The parameter set θ for each MHC allele h h The value of θ can be obtained by h The dependency function g is determined by minimizing the loss function where i is each instance in the subset S of the training data 170 generated by cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. h It can be expressed as the dependency function g introduced in Section VIII.B.1 above h Any form of .

[0490] According to equation (13), the dependency function g h (·) applied to the peptide sequence p for each of the relevant MHC alleles H k The encoded form of the peptide sequence p is used to generate the corresponding score of the allelic interaction variable k The probability of presentation by one or more MHC alleles h is then combined and transformed by the transformation function f(·) to produce the peptide sequence p k The presentation probability to be presented by the MHC allele set H.

[0491] The presentation model of Equation (13) differs from the independent allele model of Equation (2) in that each peptide p k The number of related alleles can be greater than 1. In other words, for a peptide sequence p k Multiple MHC alleles associated with H, a h k More than one element in can have the value 1.

[0492] For example, when using the affine transformation function g h (·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0493]

[0494] where x2 k 、x3 k is the allele interaction variable for the identified MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters for the determined MHC alleles h=2, h=3.

[0495] As another example, when using the network transformation function g h (·), g w (·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0496]

[0497] where NN2(·), NN3(·) are the network models for the identified MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters for the measured MHC alleles h=2, h=3.

[0498] Figure 9 The peptide p associated with MHC alleles h=2 and h=3 is generated using the example network models NN2(·) and NN3(·). k The possibility of submission. Figure 9 As shown in , the network model NN2(·) receives the allele interaction variable x2 for the MHC allele h=2. k And generate output NN2(x2 k ), and the network model NN3(·) receives the allele interaction variable x3 for the MHC allele h=3 k And generate output NN3(x3 k The outputs are combined and mapped by the function f(·) to produce the estimated presentation probability u k .

[0499] VIII.C.3. Example 2.2: Functional Model Using the Sum of Allele Non-Interacting Variables

[0500] In one embodiment, the training module 316 incorporates allele non-interacting variables and uses the following formula to make the peptide p k The estimated probability of presentation u k Modeling:

[0501]

[0502] where w k Indicates the encoded peptide p k Specifically, the set of parameters θ for each MHC allele h is h and the set of parameters θ about the non-interacting variables of the alleles w The value of θ can be obtained by h and θ wThe dependency function g is determined by minimizing the loss function where i is each instance in the subset S of the training data 170 generated by cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. w It can be expressed as the dependency function g introduced in Section VIII.B.3 above w Any form of .

[0503] Therefore, according to equation (14), the function g can be h (·) applied to the peptide sequence p for each of the relevant MHC alleles H k The encoded form of the peptide sequence p is used to generate the corresponding dependency score of the allelic interaction variable for each MHC allele h k The probability of presentation by one or more MHC alleles H. Function of the allele-independent variable g w (·) is also applied to the encoded form of the allele-independent interacting variables to produce the dependency score of the allele-independent interacting variables. The scores are combined and transformed by the transformation function f(·) to produce the peptide sequence p k The probability of presentation will be presented by the MHC allele H.

[0504] In the presentation model of Equation (14), each peptide p k The number of related alleles can be greater than 1. In other words, for a peptide sequence p k Multiple MHC alleles associated with H, a h k More than one element in can have the value 1.

[0505] For example, when using the affine transformation function g h (·), g w (·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0506]

[0507] where w k The identified peptide p k The alleles are non-interacting variables, and θ w is the set of parameters for the measured allele non-interacting variables.

[0508] As another embodiment, when using the network transformation function g h (·), g w(·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0509]

[0510] where w k The identified peptide p k The allele interaction variable, and θ w is the set of parameters for the measured allele non-interacting variables.

[0511] Figure 10 The example network models NN2(·), NN3(·) and NN w (·) Generate peptide p associated with MHC alleles h=2 and h=3 k The possibility of submission. Figure 10 As shown in , the network model NN2(·) receives the allele interaction variable x2 for the MHC allele h=2. k And generate output NN2(x2 k The network model NN3(·) receives the allele interaction variable x3 for the MHC allele h=3. k And generate output NN3(x3 k ). Network model NN w (·) Receive peptide p k The allele non-interaction variable w k And generate output NN w (w k The outputs are combined and mapped by the function f(·) to produce the estimated presentation probability u k .

[0512] Alternatively, the training module 316 can be implemented by transforming the allele non-interaction variable w k The allele non-interaction variable x added to Eq. (15) h k To convert the allele non-interaction variable w k Included in the predicted value. Therefore, the presentation probability can be obtained as follows:

[0513]

[0514] VIII.C.4. Example 3 .1 : Model using implicit independent allele likelihood

[0515] In another embodiment, the training module 316 trains the peptide p by the following formula k The estimated probability of presentation uk Modeling:

[0516]

[0517] where element a h k For the peptide sequence p k The associated multiple MHC alleles h∈H are 1, u' k h is the implicit independent allele presentation probability of MHC allele h, and the vector v is a vector in which element v h Corresponding to a h k ·u' k h , s(·) is a function of the mapped elements v, and r(·) is a clipping function that clips the input values ​​to a given range. As described in more detail below, s(·) can be a summation function or a second-order function, but it should be understood that in other embodiments, s(·) can be any function, such as a maximum function. The value of the set of parameters θ related to the implicit independent allele likelihood can be determined by minimizing a loss function with respect to θ, where i is each instance in the subset S of training data 170 generated by cells expressing a single MHC allele and / or cells expressing multiple MHC alleles.

[0518] Let the presentation probability in the presentation model of Equation (16) be varied with the respective peptide p k The implicit independent allele presentation probability u' of the probability that will be presented by an individual MHC allele h k h Implicit independent allele likelihood differs from the independent allele presentation likelihood of Section VIII.B in that parameters for implicit independent allele likelihood can be learned from a multi-allelic environment where, as in a mono-allelic environment, the direct association between the presented peptide and the corresponding MHC allele is unknown. Thus, in a multi-allelic environment, the presentation model can estimate not only the peptide p k will be presented by the set of MHC alleles H as a whole, and can also provide an indication of the most likely presented peptide p k The individual probability u' of the MHC allele h k h ∈ H The advantage is that the presented model can generate implicit probabilities in the absence of training data on cells expressing a single MHC allele.

[0519] In one specific embodiment referred to throughout the rest of this specification, r(·) is a function with a range [0, 1]. For example, r(·) can be a clipping function:

[0520] r(z)=min(max(z,0),1),

[0521] The minimum value between z and 1 is selected as the presentation probability u k In another embodiment, when the value of the domain z is equal to or greater than 0, r(·) is a hyperbolic tangent function provided by:

[0522] r(z)=tanh(z).

[0523] VIII.C.5. Example 3.2: Sum of Functions f- Functions model

[0524] In a specific embodiment, s(·) is a summation function, and the presentation probability is obtained by summing the implicit independent allele presentation probabilities:

[0525]

[0526] In one embodiment, the implicit independent allele presentation probability of MHC allele h is given by:

[0527]

[0528] The probability of presentation is estimated by the following formula:

[0529]

[0530] According to equation (19), the function g h (·) applied to the peptide sequence p for each of the relevant MHC alleles H k The encoded form of the peptide sequence p is used to generate the corresponding dependency score of the allelic interaction variable k The probability of presentation by one or more MHC alleles H. Each dependency score is first transformed by a function f(·) to produce an implicit independent allele presentation probability u' k h . The independent allele probability u' k h merged, and a clipping function can be applied to the merged likelihood to clip the values ​​to the range [0,1] to produce the peptide sequence p k The probability of presentation to be presented by the MHC allele set H. The dependency function g h It can be expressed as the dependency function g introduced in Section VIII.B.1 above hAny form of .

[0531] For example, when using the affine transformation function g h (·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0532]

[0533] where x2 k 、x3 k is the allele interaction variable for the identified MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters for the determined MHC alleles h=2, h=3.

[0534] As another embodiment, when using the network transformation function g h (·), g w (·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0535]

[0536] where NN2(·), NN3(·) are the network models for the identified MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters for the measured MHC alleles h=2, h=3.

[0537] Figure 11 The peptide p associated with MHC alleles h=2 and h=3 is generated using the example network models NN2(·) and NN3(·). k The possibility of submission. Figure 11 As shown in , the network model NN2(·) receives the allele interaction variable x2 for the MHC allele h=2. k And generate output NN2(x2 k ), and the network model NN3(·) receives the allele interaction variable x3 for the MHC allele h=3 k And generate output NN3(x3 k ). Each output is mapped by a function f(·) to produce an estimated presentation probability u k .

[0538] In another embodiment, when predicting the logarithm of the mass spectrometer ion current, r(·) is a logarithmic function and f(·) is an exponential function.

[0539] VIII.C.6. Example 3.3: Sum model using functions of allele non-interacting variables

[0540] In one embodiment, the implicit independent allele presentation probability of MHC allele h is given by:

[0541]

[0542] This results in the presentation probability:

[0543]

[0544] to incorporate the effects of allele-noninteracting variables on peptide presentation.

[0545] According to equation (21), the function g h (·) applied to the peptide sequence p for each of the relevant MHC alleles H k The encoded form of the peptide sequence p is used to generate the corresponding dependency score of the allelic interaction variable for each MHC allele h k The probability of presentation by one or more MHC alleles H. Function of the allele-independent variable g w (·) is also applied to the encoded form of the allele non-interacting variables to produce dependency scores for the allele non-interacting variables. The scores for the allele non-interacting variables are combined with the respective dependency scores for the allele interacting variables. Each combined score is transformed by a function f(·) to produce an implicit independent allele presentation probability. The implicit probabilities are combined, and a clipping function can be applied to the combined output to clip the values ​​to the range [0,1] to produce the peptide sequence p k The probability of presentation to be presented by the MHC allele set H. The dependency function g w It can be expressed as the dependency function g introduced in Section VIII.B.3 above w Any form of .

[0546] For example, when using the affine transformation function g h (·), g w (·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0547]

[0548] where w k The identified peptide p k The alleles are non-interacting variables, and θ w is the set of parameters for the measured allele non-interacting variables.

[0549] As another embodiment, when using the network transformation function g h (·), g w (·) Among the m=4 different MHC alleles identified, peptide p k The probability of presentation by MHC alleles h=2 and h=3 can be obtained from the following formula:

[0550]

[0551] where w k The identified peptide p k The allele interaction variable, and θ w is the set of parameters for the measured allele non-interacting variables.

[0552] Figure 12 The example network models NN2(·), NN3(·) and NN w (·) Generate peptide p associated with MHC alleles h=2 and h=3 k The possibility of submission. Figure 12 As shown in , the network model NN2(·) receives the allele interaction variable x2 for the MHC allele h=2. k And generate output NN2(x2 k ). Network model NN w (·) Receive peptide p k The allele non-interaction variable w k And generate output NN w (w k The outputs are merged and mapped by the function f(·). The network model NN3(·) receives the allele interaction variable x3 for the MHC allele h=3. k And generate output NN3(x3 k ), and again compare the output with the same network model NN w The output NN of (·) w (w k ) are combined and mapped by the function f(·). The two outputs are combined to produce the estimated presentation probability u k .

[0553] In another embodiment, the implicit independent allele presentation probability of MHC allele h is given by:

[0554]

[0555] The presentation probability is thus obtained by the following formula:

[0556]

[0557] VIII.C.7. Example 4: Second-Order Model

[0558] In one embodiment, s(·) is a second order function and the peptide p k The estimated probability of presentation u k It is obtained from the following formula:

[0559]

[0560] where element u' k h is the implicit independent allele probability for MHC allele h. The value of the parameter set θ for the implicit independent allele probability can be determined by minimizing a loss function with respect to θ, where i is each instance in the subset S of training data 170 generated by cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. The implicit independent allele presentation probability can be in any of the forms shown in equations (18), (20), and (22) described above.

[0561] On the one hand, the model of equation (23) may imply the presence of peptide p k The possibility of simultaneous presentation by two MHC alleles, where presentation by the two HLA alleles is statistically independent.

[0562] According to equation (23), the peptide sequence p k The probability of presentation by one or more MHC alleles H can be calculated by combining the implicit independent allele presentation probabilities and subtracting from the sum the probability that each pair of MHC alleles will simultaneously present the peptide p k The possibility to generate the peptide sequence p k Will be generated by the presentation probability presented by the MHC allele H.

[0563] For example, when using the affine transformation function g h (·) Among the m=4 different HLA alleles identified, peptide p k The probability of presentation by HLA alleles h=2 and h=3 can be obtained from the following formula:

[0564]

[0565] where x2 k 、x3 k is the allele interaction variable for the identified HLA alleles h=2, h=3, and θ2, θ3 are the sets of parameters for the determined HLA alleles h=2, h=3.

[0566] As another embodiment, when using the network transformation function g h (·), g w (·) Among the m=4 different HLA alleles identified, peptide p k The probability of presentation by HLA alleles h=2 and h=3 can be obtained from the following formula:

[0567]

[0568] where NN2(·), NN3(·) are the network models for the identified HLA alleles h=2, h=3, and θ2, θ3 are the sets of parameters for the measured HLA alleles h=2, h=3.

[0569] VIII.D. Pan-allele model

[0570] Compared to independent allele models, pan-allele models are presentation models that can predict the likelihood of presentation of a peptide based on pan-alleles. Specifically, unlike independent allele models (which can predict the probability that a peptide will be presented by one or more known MHC alleles that have been previously used to train the independent allele model), pan-allele models are presentation models that can predict the probability that a peptide will be presented by any MHC allele (including unknown MHC alleles that the model has not previously encountered during training).

[0571] Briefly, the pan-allele model is trained by the training module 316. Similar to the training of the independent allele model, the training module 316 can train the pan-allele presentation model based on data instances S in the training data 170 generated by cells expressing a single MHC allele, cells expressing multiple MHC alleles, or a combination thereof. However, the training module 316 does not use a specific MHC allele or a specific set of MHC alleles. k h Instead of training a pan-allele presentation model, we use all available MHC allele peptide sequences in the training data 170 h Specifically, the training module 316 trains the pan-allele presentation model based on the amino acid positions of the MHC alleles available in the training data 170 .

[0572] After the pan-allele model has been trained, when a peptide sequence and a known or unknown MHC allele peptide sequence are input into the model to determine the probability that the known or unknown MHC allele will present the peptide, the model is able to accurately predict this probability by using the information learned during training with peptide sequences from similar MHC alleles. For example, a pan-allele model trained using training data 170 that does not contain any occurrences of the A*02:07 allele can still accurately predict presentation of the peptide by the A*02:07 allele by utilizing information learned during training with similar alleles (e.g., alleles in the A*02 gene family). In this way, a single presentation pan-allele model can predict the likelihood of presentation of a peptide on any MHC allele.

[0573] VIII.D.2. Advantages of the Pan-Allelic Model

[0574] The major advantage of the pan-allele presentation model is that the pan-allele presentation model has greater versatility than the independent allele presentation model. As mentioned above, the independent allele model can predict the probability that a peptide will be presented by one or more identified MHC alleles that are used to train the independent allele model. In other words, the independent allele model is associated with a limited set of one or more known MHC alleles.

[0575] Thus, given a sample containing a particular set of one or more MHC alleles, to determine the probability of a peptide being presented by that particular set of MHC alleles, an independent allele model trained using that particular set of MHC alleles is selected. In other words, when relying on the independent allele model to predict the probability of a peptide being presented by an MHC allele, predictions can only be made for MHC alleles that appear in the training data 170. Because there are a large number of MHC alleles (especially for minor variations within the same gene family), a very large number of training samples would be required to train the independent allele presentation model to make peptide presentation predictions for all MHC alleles.

[0576] In contrast, the pan-allele model is not limited to predicting the specific set of one or more MHC alleles trained thereon. Instead, in use, the pan-allele model can accurately predict the probability that previously seen and / or previously unseen MHC alleles will present a given peptide by using the information learned during training with similar MHC allele peptide sequences. The result is that the pan-allele model is not associated with the specific set of one or more MHC alleles and can predict the probability that a peptide will be presented by any MHC allele. This versatility of the pan-allele model means that a single model can be used to predict the possibility that any peptide will be presented by any MHC allele. Therefore, using the pan-allele model reduces the amount of the training data required for maximizing the individual HLA coverage and colony HLA coverage as defined in part VII.A above.

[0577] VIII.D.3. Use of Pan-Allele Models

[0578] The following discussion in Sections VIII.D.4.-VIII.D.7. relates to using a pan-allele model to predict the probability that a peptide will be presented by one or more MHC alleles. For simplicity, this discussion assumes that training module 316 has trained the pan-allele model. Training of the pan-allele model is discussed in detail below in Section VIII.D.8.

[0579] Additionally, the discussion below in Sections VIII.D.4.–VIII.D.6. relates to using a pan-allele model to predict the likelihood that a peptide will be presented by a single MHC allele in a given sample and / or by multiple MHC alleles in a given sample. However, as described in detail below in Section VIII.D.7., there are subtle differences between using a pan-allele model to predict the likelihood that a peptide will be presented by a single MHC allele in a sample and using a pan-allele model to predict the likelihood that a peptide will be presented by multiple MHC alleles in a sample.

[0580] Briefly, when using a pan-allele model to predict the likelihood that a peptide will be presented by a single MHC allele, as described below, a set of inputs is provided to the pan-allele model, and the pan-allele model produces a single output.

[0581] On the other hand, when using the pan-allelic model to predict the probability that a peptide will be presented by multiple MHC alleles, the pan-allelic model is iteratively used for each MHC allele in multiple MHC alleles. Specifically, when using the pan-allelic model to predict the probability that a peptide will be presented by multiple MHC alleles, a first group of inputs associated with a first MHC allele in multiple MHC alleles is provided to the pan-allelic model, and the pan-allelic model produces a first output for the first MHC allele. Then, a second group of inputs associated with a second MHC allele in multiple MHC alleles is provided to the pan-allelic model, and the pan-allelic model produces a second output for the second MHC allele. This process is iteratively performed for each MHC allele in multiple MHC alleles. Finally, the output of each MHC allele in multiple MHC alleles that the pan-allelic model produces is combined to produce a single probability that multiple MHC alleles present a given peptide, as described in section VIII.D.7.

[0582] VIII.D.4. Overview of the Pan-Allele Model

[0583] In one embodiment, a pan-allelic model is used to estimate the effect of allele h on peptide p k The possibility of submission u k In some embodiments, the pan-allelic model is represented by the following equation:

[0584]

[0585] where p k represents the peptide sequence, d h represents the peptide sequence of MHC allele h, f(·) is any transformation function, and g H (·) is any dependency function. The pan-allelic model is based on a set of shared parameters θ determined for all MHC alleles H Generate peptide sequence p k and MHC allele peptide sequences d h The group shares the parameter θ H The value of is learned during pan-allele model training and is discussed in detail in Section VIII.D.8. below.

[0586] Dependency function g H ([p k d h ];θ H The output of ) represents the dependency score of MHC allele h, which indicates whether MHC allele h will be based on at least the peptide sequence p k The amino acid positions and MHC allele peptide sequences ofh The amino acid positions of the peptide p k For example, if MHC allele h is given an input MHC allele peptide sequence d h In the case of possible presentation of peptide p k , then the dependency score of the MHC allele h may have a high value, while it may have a low value if it is less likely to be presented. The transformation function f(·) transforms the input, more specifically, in this case g H ([p k d h ];θ H ) is converted to a suitable value to indicate the peptide p k The probability that it will be presented by MHC allele h.

[0587] In a specific embodiment referred to throughout the rest of the specification, f(·) is a function having a range within [0, 1] for a suitable domain range. In one example, f(·) is an expit function. As another example, when the value of the domain z is equal to or greater than 0, f(·) may also be a hyperbolic tangent function. Alternatively, when the mass spectrometer ion current is predicted to have a value outside the range [0, 1], f(·) may be any function, such as an identity function, an exponential function, a logarithmic function, etc.

[0588] Therefore, the dependency function g can be H (·) applied to the peptide sequence p k The coding form and its application to MCH allele peptide sequence d h The encoded form of the peptide sequence p is generated to produce the corresponding dependency score k The probability that the peptide sequence p will be presented by the MHC allele h. The dependency score can be transformed by the transformation function f(·) to produce k The probability that it will be presented by MHC allele h.

[0589] VIII.D.5. Dependence Functions of Allelic Interaction Variables

[0590] In one specific embodiment mentioned throughout the specification, the dependency function g H (·) is the affine function obtained from the following formula:

[0591]

[0592] where α is the intercept; Indicated in peptide p k The residue at position i; d hjrepresents the residue at position j of the MHC allele h; 1[] represents an indicator variable, which has a value of 1 if the condition in the square brackets is true and 0 otherwise; if the peptide p k The amino acid at position i is amino acid k, then True, false otherwise; if the amino acid at position j of MHC allele h is amino acid l, then d hj = l is true, otherwise it is false; n pep represents the length of the peptide being modeled; n MHC represents the number of MHC residues considered in the model; and θ H,ijkl is the coefficient describing the contribution to the likelihood of presentation of having residue k at position i of the peptide and residue l at position j of the MHC allele. This is a linear model in the one-hot-encoded peptide sequence and one-hot-encoded MHC allele sequence with peptide-residue-MHC-residue interactions for all peptide residues and MHC allele residues.

[0593] In one specific embodiment mentioned throughout the specification, the dependency function g H (·) is the network function obtained by the following formula:

[0594] g H ([p k d h ];θ H )=NN H ([p k d h ];θ H ) (26)

[0595] The network function consists of a network model NN having a series of nodes arranged in one or more layers. H (·) indicates that a node can be connected to other nodes through connections, and these connections are each in the parameter set θ H The value at a particular node can be expressed as the sum of the values ​​of the nodes connected to that particular node, which is weighted by the relevant parameters of the activation function map associated with that particular node. Because the presentation model can incorporate nonlinear and process data with different amino acid sequence lengths, network models are advantageous compared to affine functions. Specifically, through nonlinear modeling, network models can capture the interactions between amino acids at different positions in the peptide sequence and the interactions between amino acids at different positions in the MHC allele peptide sequence, and can capture how these interactions affect peptide presentation.

[0596] Generally speaking, the network model NN H(·) Can be constructed as feedforward networks, such as artificial neural networks (ANNs), convolutional neural networks (CNNs), deep neural networks (DNNs), and / or recurrent networks, such as long short-term memory networks (LSTMs), bidirectional recurrent networks, deep bidirectional recurrent networks, etc.

[0597] In one example, a single network model NN H (·) can be the coding peptide sequence p given the MHC allele h k and protein coding sequence d h A network model that can output dependency scores when . In such an instance, the parameter set θ H It can correspond to the parameter set of a single network model, so the parameter set θ H can be shared by all MHC alleles. Therefore, in such an instance, when providing any input [p k d h ], NN H (·) can represent a single network model NN H As discussed above, this network model is advantageous because it can predict the peptide presentation probability of unknown MHC alleles in the training data simply by identifying the protein sequence of the MHC allele.

[0598] Figure 13 Example network model NN showing MHC allele sharing H (·).like Figure 13 As shown, the network model NN H (·) Peptide sequence p that receives MHC allele h k and protein sequence d h takes as input, and outputs the dependency score NN corresponding to the MHC allele h H ([p k d h ]).

[0599] Figure 14 An example network model NN is shown H (·).like Figure 14 As shown in the network model NN H (·) includes four input nodes at layer l=1, five nodes at layer l=2, two nodes at layer l=3, and one output node at layer l=4. In an alternative embodiment, the network model NN H (·) can contain any number of layers, and each layer can contain any number of nodes. Network model NN H (·) with 13 non-zero parameters θ H (1),θ H (2),……,θH (13). These parameters are used to transform the values ​​propagated from one node to another through the network model.

[0600] like Figure 14 As shown, in the network model NN H The four input nodes at layer l=1 of (·) receive input values ​​including encoded polypeptide sequence data and encoded MHC allele peptide sequence data. The encoded polypeptide sequence data contains the amino acid sequence of the peptide, while the encoded MHC allele peptide sequence data contains the amino acid sequence of the MHC allele that may (or may not) present the peptide. In certain embodiments, once the network model NN is input via the input node at layer l=1, H (·), the encoded polypeptide sequence is connected to the network model NN H (·) in front of the MHC allele peptide sequence encoded within a layer. These input values ​​are then passed through the network model NN according to the parameter values. H (·) propagation. In some embodiments, the network model NN H The layer of (·) comprises two fully connected dense network layers. In a further embodiment, the first of the two fully connected dense network layers comprises 64-128 nodes with a calibrated linear unit activation function. In an even further embodiment, the second of the two fully connected dense network layers comprises a single node with a linear output. In such an embodiment, the single node may be the network model NN H (·) output node. Finally, the network model NN H (·) Output value NN H ([p k d h ]). This output represents the dependency score for MHC allele h, indicating whether MHC allele h will present peptide sequence p k The network function may also include one or more network models, each network model taking a different allele interaction variable (eg, peptide sequence) as input.

[0601] In yet another example, the dependency function g H (·) can be expressed as:

[0602]

[0603] where g' H ([p k d h ];θ' H ) is a parameter set θ' H Affine functions, network functions, etc., where the shared parameter set θ of allele interaction variablesH The bias parameter θ in H 0 represents the baseline probability of presentation of any MHC allele.

[0604] In another embodiment, the deviation parameter θ H 0 It can be shared by the gene family of MHC allele h. That is, the bias parameter θ of MHC allele h H 0 can be equal to θ 基因(h) 0 , where gene (h) is the gene family of the MHC allele h. For example, the class I MHC alleles HLA-A*02:01, HLA-A*02:02, and HLA-A*02:03 can be assigned to the "HLA-A" gene family, and the deviation parameter θ for each of these MHC alleles is H 0 As another example, the class II MHC alleles HLA-DRB1:10:01, HLA-DRB1:11:01, and HLA-DRB3:01:01 can be assigned to the "HLA-DRB" gene family, and the deviation parameter θ for each of these MHC alleles is H 0 As discussed above, the gene family may be one of the allele interaction variables associated with the MHC allele h.

[0605] Returning to equation (23), as an example, using the affine dependency function g H (·), peptide p k The probability of being presented by MHC allele h can be given by:

[0606]

[0607] where α is the intercept; Indicated in peptide p k The residue at position i; d hj represents the residue at position j of the MHC allele h; 1[] represents an indicator variable, which has a value of 1 if the condition in the square brackets is true and 0 otherwise; if the peptide p k The amino acid at position i is amino acid k, then True, false otherwise; if the amino acid at position j of MHC allele h is amino acid l, then d hj = l is true, otherwise it is false; n pep represents the length of the peptide being modeled; n MHCrepresents the number of MHC residues considered in the model; and θ H,ijkl is the coefficient describing the contribution to the likelihood of presentation of having residue k at position i of the peptide and residue l at position j of the MHC allele. This is a linear model in the one-hot-encoded peptide sequence and one-hot-encoded MHC allele sequence with peptide-residue-MHC-residue interactions for all peptide residues and MHC allele residues.

[0608] As another example, using the network transformation function g H (·), peptide p k The probability of being presented by MHC allele h can be given by:

[0609]

[0610] where p k represents the peptide sequence, d h represents the peptide sequence of MHC allele h, θ H is the network model NN associated with all MHC alleles H (·) A set of determined parameters.

[0611] Figure 15 shows the use of an example shared network model NN H (·) Produces peptide p associated with MHC allele h k The possibility of submission. Figure 15 As shown, the shared network model NN H (·) Receive peptide sequence p k and MHC allele peptide sequences d h , and produces the output NN H ([p k d h ]). The output is mapped by the function f(·) to generate the estimated presentation probability u k .

[0612] VIII.D.6. Allele Non-Interaction Variables

[0613] As described above, allele-independent interaction variables include information that affects the presentation of peptides that is independent of the type of MHC allele. For example, allele-independent interaction variables may include the protein sequences at the N-terminus and C-terminus of the peptide, the protein family of the presented peptide, the RNA expression level of the source gene of the peptide, and any additional allele-independent interaction variables.

[0614] In one embodiment, training module 316 is with the similar mode described for independent allele model and multiple allele model (multiple allele model), and allele non-interaction variable is incorporated into pan-allele presentation model.For example, in some embodiments, allele non-interaction variable can be typed into the dependency function separated with the dependency function for allele interaction variable as input.In such embodiment, the output of two independent dependency functions can be summed, and the summation of gained can be input in the transformation function to generate and present prediction.Such embodiment that allele non-interaction variable is incorporated into pan-allele model and other models is discussed in part above VIII.B.2., VIII.B.3., VIII.C.3. and VIII.C.6.

[0615] VIII.D.7. Multiple allele samples

[0616] As mentioned above, a test sample can contain multiple MHC alleles rather than a single MHC allele. In fact, most samples obtained from nature contain more than one MHC allele. For example, each human genome contains six MHC class I loci. Therefore, a sample containing a human genome can contain up to six different MHC class I alleles. Therefore, samples containing multiple MHC alleles rather than a single MHC allele are typical of real-life test cases.

[0617] In embodiments where the test sample contains multiple MHC alleles, the pan-allelic model described above in Sections VIII.D.4.–VIII.D.6. can be used to determine the probability that a given peptide from the test sample will be presented by multiple MHC alleles. However, as briefly described above, when using the pan-allelic model to predict the likelihood that a peptide will be presented by multiple MHC alleles, the pan-allelic model is applied iteratively for each MHC allele in the multiple MHC alleles. In other words, for each MHC allele in the multiple MHC alleles, the MHC allele peptide sequence and the peptide sequence are independently input into a dependency function shared by all MHC alleles. Based on these inputs, the dependency function generates an output corresponding to the MHC allele. This process is performed iteratively for each MHC allele in the multiple MHC alleles. Thus, each MHC allele in the multiple MHC alleles is independently associated with an output of the dependency function. The outputs associated with each MHC allele in the multiple MHC alleles are then combined.

[0618] The outputs of the dependency functions associated with each of the plurality of MHC alleles can be combined as described with respect to Sections VIII.C.-VIII.C.7. As described with respect to Sections VIII.C.-VIII.C.7., the manner in which the plurality of outputs of the dependency functions are combined can vary. For example, in some embodiments, the outputs of the dependency function iterations can be summed, and the resulting sum can be input into the transformation function to generate a presentation prediction. An equation capturing such an embodiment can be written as:

[0619]

[0620] Where T is the total number of unique MHC alleles in a sample containing multiple alleles. In an alternative embodiment, each individual output of the dependency function iteration can be input into a transformation function, and the resulting outputs from the transformation function can be summed to generate a presentation prediction. The equation that captures this alternative embodiment can be written as:

[0621]

[0622] Such embodiments, as well as other embodiments in which multiple outputs of dependency functions are combined to predict the likelihood that a peptide will be presented in a polyallelic environment, are discussed further above in Sections VIII.C.-VIII.C7.

[0623] VIII.D.8. Training of Pan-Allele Models

[0624] Training the pan-allele model involves training a set of shared parameters θ associated with the dependency function H Specifically, the optimization parameter θ H , so that the dependency function can output a dependency score that accurately indicates whether one or more given MHC alleles will present a given peptide sequence.

[0625] To optimize the parameter θ HThe training data 170 is used to train the model using the values ​​of . As mentioned above, the training data 170 used to train the model can include training samples containing cells expressing a single MHC allele, training samples containing cells expressing multiple MHC alleles, or training samples containing cells expressing a combination of both a single MHC allele and multiple MHC alleles. Therefore, each data instance i from the training data 170 is input into the pan-allele model, and more specifically, is input into the dependency function of the pan-allele model. For example, in some embodiments, the MHC allele peptide sequence and the peptide sequence can be input into the pan-allele model. The pan-allele model then processes these inputs as if the model were conventionally used as described above with respect to Sections VIII.D.3–VIII.D.7. However, unlike the operation of the pan-allele model described in Sections VIII.D.3–VIII.D.7., during the training of the pan-allele model, the known results of the peptide presentation are also input into the model. In other words, the label y i In embodiments where the training sample input to the pan-allele model contains cells expressing multiple MHC alleles, for each of the multiple MHC alleles in the sample, y is set to i Set to 1.

[0626] After each iteration of the pan-allele model using data instance i, the model determines the probability of the predicted MHC allele presenting the peptide with the known marker y i Then, to minimize this difference, the pan-allelic model modifies the parameter θ H In other words, the pan-allele model minimizes the H The loss function is used to determine the parameter θ H When the pan-allele model achieves a certain level of prediction accuracy, training is complete and the model is ready to be used as described in Sections VIII.D.3–VIII.D.7.

[0627] VIII.D.9. Examples of Pan-Allele Models

[0628] The following example has compared the prediction accuracy (i.e. positive predictive value) of example independent allele presentation model and example pan-allele presentation model.In this example, identical training data set is used to train allele presentation model and pan-allele presentation model.After training, six test samples are used to test independent allele presentation model and pan-allele presentation model.Note that training data set contains sufficient training data of each MHC allele tested in each test sample.Following table 2 shows the prediction accuracy (or positive predictive value) under 40% recall rate when using independent allele model and pan-allele model.Because each MHC allele tested in six samples all has sufficient training data, the independent allele model is slightly better than the pan-allele model by exceeding 0.04 with average accuracy.

[0629]

[0630] Table 2

[0631] However, in the case of Figure 16-22 The ability of the pan-allelic model to predict the presentation probability of MHC alleles that were not included in the training dataset used to train the model was observed in the alternative experiments discussed.

[0632] Figure 16-22 Depicted are the results of an experiment designed to test the ability of a pan-allele model to predict the probability that an untrained MHC allele will present a given peptide. In particular, Figure 16-18 Depicts the results of an experiment designed to test the ability of a pan-allele model comprising a neural network model to predict the probability that an untrained MHC allele will present a given peptide. Figure 19-22 Depicted are the results of an experiment designed to test the ability of a pan-allele model comprising a non-neural network model to predict the probability that an untrained MHC allele will present a given peptide.

[0633] First, let's look at Figure 16-18In an associated experiment, to demonstrate the ability of a pan-allelic model comprising a neural network model to predict the probability that an untrained MHC allele will present a given peptide, the predictions generated by a pan-allelic model comprising a neural network model that had not been trained with the MHC allele being tested were compared with the predictions generated by the same pan-allelic model that had been trained with the MHC allele being tested. In other words, the only difference between the pan-allelic models was the training data set used to train them. The higher the prediction accuracy of a pan-allelic model that had not been trained on a sample comprising the HLA allele being tested relative to the prediction accuracy of a pan-allelic model that had been trained on a sample comprising the HLA allele being tested, the greater the ability of the pan-allelic model to predict the likelihood of presentation of an MHC allele that was not used to train the pan-allelic model.

[0634] As noted above, in Figure 16-18 The pan-allele model used in the related experiments was the same before training with different training datasets. Figure 16-18 Each pan-allele model used in the associated experiment includes a neural network model as its dependency function. The neural network model used in the pan-allele model contains a single hidden layer. The activation function between the hidden layers of the neural network model is a rectified linear unit (ReLU) function, f(x) = max(0,x). The last layer of the neural network model contains a linear activation layer f(x) = x. The number of hidden units of each subnetwork of the neural network model depends on the input to the neural network model. Specifically, for a neural network model configured to receive mRNA abundance, the number of hidden units in the mRNA abundance subnetwork of the neural network model is 16. For a neural network model configured to receive an encoded flanking sequence, the number of hidden units in the flanking sequence subnetwork of the neural network model is 32. For a neural network model configured to receive an encoded polypeptide sequence, the number of hidden units in the polypeptide sequence subnetwork of the neural network model is 256. For a neural network model configured to receive an encoded polypeptide sequence and an encoded MHC allele peptide sequence (as in the case of a pan-allele model), the number of hidden units in the polypeptide and MHC allele peptide sequence sub-networks of the neural network model is 128.

[0635] and Figure 16-18Each of the associated experiments included a unique test sample, each containing a different HLA allele. To demonstrate that the results generated by these experiments were not limited to specific loci, alleles were selected from each of the three loci A, B, and C. Thus, the first test sample contained HLA-A alleles, the second sample contained HLA-B alleles, and the third sample contained HLA-C alleles. Specifically, the first test sample contained HLA alleles A*02:03, the second test sample contained HLA alleles B*54:01, and the third test sample contained HLA alleles C*08:02. Protein sequences for each of these HLA alleles were obtained from the HLA Protein Sequence Database maintained by the Anthony Nolan Research Institute (https: / / www.ebi.ac.uk / ipd / imgt / hla / ).

[0636] For each of the three samples, the protein sequence of the specific HLA allele and the protein sequence of the peptide in question were input into a first pan-allelic model that had not been trained using that HLA allele, and into a second identical pan-allelic model that had been trained using that HLA allele. The pan-allelic model outputs the predicted probability that the HLA allele will present the peptide. These predicted probabilities are compared with the known outcomes of peptide presentation (i.e., the marker y). i ) are compared to produce Figure 16-18 The precision / recall curve shown. Specifically, Figure 16 Corresponds to the data output for the first test sample output by the pan-allelic model. Figure 17 corresponds to the data output for the second test sample output by the pan-allelic model, and Figure 18 Corresponding to the data output for the third test sample output by the pan-allelic model. In each figure, the blue line represents the precision / recall curve of the pan-allelic model trained on samples containing the tested HLA alleles, and the orange line represents the precision / recall curve of the pan-allelic model that has not been trained on any samples containing the tested HLA alleles. In addition, each figure indicates the average prediction accuracy (i.e., positive predictive value) of the trained and untrained pan-allelic models. For example, as shown in Figure 18 As can be seen in Figure 2, the average prediction accuracy of the pan-allele model that has been trained on samples containing the tested HLA allele is 0.256, while the average prediction accuracy of the pan-allele model that has not been trained on samples containing the tested HLA allele is 0.231.

[0637] like Figure 16-18As shown, even though the pan-allele models represented by the orange line had never seen the HLA allele being tested, these pan-allele models were able to achieve comparable performance to the pan-allele models represented by the blue line that had seen the HLA allele being tested during training. Therefore, these results demonstrate that the pan-allele models containing the neural network model can accurately predict the presentation probability of HLA alleles that were not used to train the pan-allele models.

[0638] Next, look at Figure 19-22 In related experiments, to demonstrate the ability of the pan-allele model containing non-neural network models to predict the probability that an untrained MHC allele will present a given peptide, the performance of four models was compared within each experiment. The four models included: Figure 16-18 The pan-allelic representation model described includes a neural network model, an existing random forest model consisting of 1,000 trees, an existing quadratic discriminant analysis (QDA) model fitted with multivariate Gaussians, and the current state-of-the-art MHC class 1 binding affinity model, MHCFlurry, which fits a unique feed-forward fully connected neural network to each allele. Both the random forest model and the quadratic discriminant model are based on a pan-allelic model architecture that includes non-neural network models.

[0639] and Figure 19-22 Each of the associated experiments included test samples, and each test sample contained HLA alleles. To demonstrate that the results produced by these experiments were not limited to specific loci, alleles were selected from each of the three loci A, B, and C. Thus, the first and second test samples contained HLA-A alleles, the third sample contained HLA-B alleles, and the fourth sample contained HLA-C alleles. Specifically, the first and second test samples contained HLA alleles A*02:01, the third test sample contained HLA alleles B*44:02, and the fourth test sample contained HLA alleles C*08:02. The protein sequence of each of these HLA alleles was obtained from the HLA protein sequence database maintained by the Anthony Nolan Research Institute ( https: / / www.ebi.ac.uk / ipd / imgt / hla / ) is obtained.

[0640] During the training of the four models for predicting presentation likelihood for each of the four test samples, the pan-allelic presentation model, random forest model, and quadratic discriminant model were each trained on monoallelic data consisting of 9-mers from 31 different alleles and including HLA-A, HLA-B, and HLA-C. On the other hand, the MHC Furry model was trained by its authors using a subset of the IEDB and BD2013 binding affinity datasets that included alleles from HLA-A, HLA-B, and HLA-C. Each allele was modeled separately using an ensemble of eight neural networks, with the allele name passed directly to the model to select which allele sub-model to use to generate presentation predictions.

[76]

[0641] The specific alleles used to train the four models for each of the four test samples depend on the HLA alleles contained in the given test sample. Specifically, for the first test sample containing the HLA allele A*02:01, the training data used to train the four models to predict the presentation probability of the HLA allele A*02:01 includes the HLA allele A*02:01. For the second test sample containing the HLA allele A*02:01, the training data used to train the four models to predict the presentation probability of the HLA allele A*02:01 does not include the HLA allele A*02:01. For the third test sample containing the HLA allele B*44:02, the training data used to train the four models to predict the presentation probability of the HLA allele B*44:02 does not include the HLA allele B*44:02. For the fourth test sample containing the HLA allele C*08:02, the training data used to train the four models to predict the presentation likelihood of the HLA allele C*08:02 does not include the HLA allele C*08:02.

[0642] During testing of each of the four samples, each model was tested on a set-aside monoallelic dataset that included the HLA alleles in the given sample and consisted of approximately 250,000 peptides (counting both presented and non-presented peptides). Specifically, during testing of each of the four samples, the pan-allelic presentation model, the random forest model, and the quadratic discriminant model each received the same input. In particular, for each of the four samples, the pan-allelic presentation model, the random forest model, and the quadratic discriminant model each received the 34-mer one-hot encoding HLA allele protein sequence of the HLA allele in the sample and the one-hot encoding (i.e., binarized) protein sequence of the 9-mer of the peptide in question. On the other hand, for each of the four samples, the MHC Flurry model received the name of the HLA allele in the sample and the one-hot encoding (i.e., binarized) protein sequence of the 9-mer of the peptide in question. As mentioned above, the difference in input between the models is a result of the fact that the MHCFlurry model is configured to use the name of the allele to select which allele sub-model to use to generate presentation predictions.

[0643] After these inputs are entered into the four models, each of the four models then outputs a predicted probability that the HLA allele will present the peptide. These predicted probabilities are compared with the known outcomes of peptide presentation (i.e., marker y i ) are compared to produce Figure 19-22 The precision / recall curve shown. Specifically, Figure 19 corresponding to the data for the first test sample output by each of the four models, Figure 20 corresponding to the data for the second test sample output by each of the four models, Figure 21 corresponds to the data for the third test sample output by each of the four models, and Figure 22 Corresponding to the data for the fourth test sample output by each of the four models. In each figure, the blue line shows the precision / recall curve of the pan-allele model, the orange line shows the precision / recall curve of the MHC Flurry model, the green line shows the precision / recall curve of the random forest model, and the red line shows the precision / recall curve of the quadratic discriminant model. In addition, each figure indicates the average prediction accuracy (i.e., positive predictive value) of each model. For example, as shown in Figure 19 As can be seen in Figure 2, the average prediction accuracy of the pan-allele model is 0.32.

[0644] like Figure 19-22As shown in , both the random forest model and the quadratic discriminant model used a pan-allele model architecture that included a non-neural network model, and their performance was approximately twice that of the MHC Flurry model. In addition, the performance of the pan-allele presentation model that included a neural network model was approximately twice that of the random forest model and the quadratic discriminant model that used a pan-allele model architecture that included a non-neural network model. In other words, the pan-allele presentation model that included a neural network model achieved the highest accuracy relative to the other models. However, the performance of the random forest model and the quadratic discriminant model that used a pan-allele model architecture that included a non-neural network model was still better than that of the customized independent allele binding affinity model MHC Flurry. Therefore, these results show that the pan-allele model architecture can be well generalized to other non-neural network machine learning models (which are as diverse as decision tree-based random forests and Bayesian methods (such as quadratic discriminant analysis)) while still providing a high level of predictive accuracy.

[0645] In addition, if Figure 20-22 As further shown in , even though the pan-allelic presentation model, random forest model, and quadratic discriminant model had never seen the HLA allele being tested, the models, including the random forest model and quadratic discriminant model that used the pan-allelic model architecture that included the non-neural network model, were able to obtain the same predictions as those corresponding to Figure 19 The results show that the pan-allele model architecture, which includes non-neural networks, can accurately predict the likelihood of presentation for HLA alleles that were not used to train the model.

[0646] IX. Example 5: Prediction Module

[0647] The prediction module 320 receives the sequence data and uses the presentation model to select candidate neoantigens from the sequence data. Specifically, the sequence data can be DNA sequences, RNA sequences, and / or protein sequences extracted from tumor tissue cells of the patient. The prediction module 320 processes the sequence data into a plurality of peptide sequences p having 8-15 amino acids for MHC-I or 6-30 amino acids for MHC-II. k For example, the prediction module 320 may process a given sequence "IEFROEIFJEF" into three peptide sequences of 9 amino acids: "IEFROEIFJ," "EFROEIFJE," and "FROEIFJEF." In one embodiment, the prediction module 320 may identify a portion containing one or more mutations by comparing sequence data extracted from normal tissue cells of a patient with sequence data extracted from tumor tissue cells of the patient, thereby identifying candidate neoantigens as mutant peptide sequences.

[0648] The prediction module 320 applies one or more presentation models to the processed peptide sequences to estimate the presentation likelihood of these peptide sequences. Specifically, the prediction module 320 can select one or more candidate neoantigen peptide sequences that are likely to be presented on tumor HLA molecules by applying the presentation models to the candidate neoantigens. In one embodiment, the prediction module 320 selects candidate neoantigen sequences whose estimated presentation likelihood exceeds a predetermined threshold. In another embodiment, the presentation module selects candidate neoantigen sequences that are likely to be presented on tumor HLA molecules. candidate neoantigen sequences with the highest estimated presentation probability (where Typically, this is the maximum number of epitopes that can be delivered in a vaccine.) A vaccine comprising the candidate neoantigens selected for a given patient can be injected into the patient to induce an immune response.

[0649] X. Example 6: Patient Selection Module

[0650] The patient selection module 324 selects a subset of patients for vaccine treatment and / or T cell therapy based on whether the patient meets the inclusion criteria. In one embodiment, the inclusion criteria are determined based on the presentation probability of the patient's neoantigen candidate generated by the presentation module. By adjusting the inclusion criteria, the patient selection module 324 can adjust the number of patients who will receive the vaccine and / or T cell therapy based on the presentation probability of the patient's neoantigen candidate. Specifically, strict inclusion criteria result in a smaller number of patients to be treated with the vaccine and / or T cell therapy, but can result in a higher proportion of patients treated with the vaccine and / or T cell therapy receiving effective treatment (e.g., receiving one or more tumor-specific neoantigens (TSNAs) and / or one or more neoantigen-responsive T cells). On the other hand, loose inclusion criteria result in a larger number of patients to be treated with the vaccine and / or T cell therapy, but can result in a lower proportion of patients treated with the vaccine and / or T cell therapy receiving effective treatment. The patient selection module 324 modifies the inclusion criteria based on the desired balance between the target proportion of patients to be treated and the proportion of patients receiving effective treatment.

[0651] In some embodiments, the inclusion criteria for selecting patients for vaccine treatment are the same as the inclusion criteria for selecting patients for T cell therapy. However, in alternative embodiments, the inclusion criteria for selecting patients for vaccine treatment may be different from the inclusion criteria for selecting patients for T cell therapy. Sections XA and XB below discuss the inclusion criteria for selecting patients for vaccine treatment and the inclusion criteria for selecting patients for T cell therapy, respectively.

[0652] XA Patient Selection for Vaccine Therapy

[0653] In one embodiment, the patient is associated with a corresponding treatment subset of v neoantigen candidates, which can potentially be included in a vaccine customized for the patient with a vaccine capacity v. In one embodiment, the patient's treatment subset is the neoantigen candidate with the highest presentation probability determined by a presentation model. For example, if a vaccine can contain v=20 epitopes, the vaccine can contain the treatment subset with the highest presentation probability determined by the presentation model for each patient. However, it should be understood that in other embodiments, the patient's treatment subset can be determined based on other methods. For example, the patient's treatment subset can be randomly selected from the patient's set of neoantigen candidates, or can be based in part on a current state-of-the-art model for modeling the binding affinity and stability of peptide sequences, or a combination of factors including presentation probability from a presentation model and affinity or stability information about these peptide sequences.

[0654] In one embodiment, the patient selection module 324 determines that the patient meets the inclusion criteria if the patient's tumor mutation load is equal to or higher than the minimum mutation load. The patient's tumor mutation load (TMB) indicates the total number of non-synonymous mutations in the tumor exome. In one embodiment, if the absolute number of the patient's TMB is equal to or higher than a predetermined threshold, the patient selection module 324 can select the patient for vaccine treatment. In another embodiment, if the patient's TMB is within a threshold percentile in the TMB determined for the patient set, the patient selection module 324 can select the patient for vaccine treatment.

[0655] In another embodiment, if the patient's utility score based on the patient's treatment subset is equal to or greater than a minimum utility score, the patient selection module 324 determines that the patient meets the inclusion criteria. In one embodiment, the utility score is a measure of the estimated number of neoantigens presented in the treatment subset.

[0656] The estimated number of neoantigens presented can be predicted by modeling neoantigen presentation as a random variable with one or more probability distributions. In one embodiment, the utility score for patient i is the expected number of neoantigen candidates presented in the treatment subset, or some function thereof. As an example, the presentation of each neoantigen can be modeled as a Bernoulli random variable, where the probability of presentation (success) is given by the likelihood of presentation of the neoantigen candidate. Specifically, for v neoantigen candidates p i1 、p i2 ,…,p iv Treatment subset S i , each neoantigen candidate has the highest presentation probability u i1 、u i2 、…、u iv , then the new antigen candidate p ijThe presentation is given by the random variable A ij Given, where:

[0657] P(A ij =1)=u ij , P(A ij =0) = 1-u ij (29)

[0658] The expected number of neoantigens presented is given by the sum of the presentation probabilities of each neoantigen candidate. In other words, the utility score for patient i can be expressed as:

[0659]

[0660] The patient selection module 324 selects a subset of patients with utility scores equal to or above a minimum utility for vaccine treatment.

[0661] In another embodiment, the utility score for patient i is the probability that at least a threshold number k of neoantigens will be presented. In one example, the therapeutic subset S of neoantigen candidates i The number of presented neoantigens in is modeled as a Poisson binomial random variable, where the probability of presentation (success) is given by the presentation probability of each epitope. Specifically, the number of presented neoantigens for patient i can be expressed by the random variable N i Given, where:

[0662]

[0663] where PBD(·) denotes the Poisson binomial distribution. The probability that at least a threshold number k of neoantigens will be presented is given by the number of neoantigens presented, N i The sum of the probabilities that are equal to or greater than k gives us . In other words, the utility score for patient i can be expressed as:

[0664]

[0665] The patient selection module 324 selects a subset of patients with utility scores equal to or above a minimum utility for vaccine treatment.

[0666] In another embodiment, the utility score for patient i is the therapeutic subset S of neoantigen candidates i The utility score can be calculated by counting the number of neoantigens that have a binding affinity or predicted binding affinity below a fixed threshold (e.g., 500 nM) for one or more patient HLA alleles. In one example, the fixed threshold is in the range of 1000 nM to 10 nM. Optionally, the utility score can count only those neoantigens detected by RNA-seq.

[0667] In another embodiment, the utility score for patient i is the therapeutic subset S of neoantigen candidates i The utility score is the number of neoantigens whose binding affinity for one or more patient HLA alleles is equal to or lower than a threshold percentile of the binding affinity of a random peptide for that HLA allele. In one example, the threshold percentile is a range from the 10th percentile to the 0.1st percentile. Optionally, the utility score can count only those neoantigens detected by RNA-seq.

[0668] It should be understood that the examples of generating utility scores shown with respect to equations (25) and (27) are merely exemplary, and that the patient selection module 324 may use other statistical or probability distributions to generate utility scores.

[0669] XB for Patient Selection of T Cell Therapy

[0670] In another embodiment, instead of receiving vaccine treatment or in addition to receiving vaccine treatment, the patient can receive T cell therapy. Like vaccine treatment, in the embodiment in which the patient receives T cell therapy, the patient can be associated with the corresponding treatment subset of v new antigen candidates as described above. This treatment subset of v new antigen candidates can be used to identify in vitro T cells from the patient that respond to one or more of the v new antigen candidates. The identified T cells can then be expanded and infused back into the patient for customized T cell therapy.

[0671] Patients can be selected to receive T cell therapy at two different time points. The first point is after the model has been used to predict the treatment subset of v neoantigen candidates for the patient, but before in vitro screening for T cells specific to the predicted treatment subset of v neoantigen candidates. The second point is after in vitro screening for T cells specific to the predicted treatment subset of v neoantigen candidates.

[0672] First, the patient can be selected to receive T cell therapy after a treatment subset of v new antigen candidates has been predicted for the patient, but before T cells specific to the predicted subset of v new antigen candidates from the patient are identified in vitro. Specifically, since in vitro screening of new antigen-specific T cells from patients may be expensive, it may be desirable to select patients to screen new antigen-specific T cells only when the patient is likely to have new antigen-specific T cells. In order to select patients before the in vitro T cell screening step, the same criteria as used to select patients for vaccine treatment can be used. Specifically, in some embodiments, if the patient's tumor mutation load is equal to or higher than the minimum mutation load as described above, the patient selection module 324 can select the patient to receive T cell therapy. In another embodiment, if the patient's utility score for the treatment subset based on the patient's v new antigen candidate as described above is equal to or higher than the minimum utility score, the patient selection module 324 can select the patient to receive T cell therapy.

[0673] Second, in addition to or instead of selecting patients to receive T cell therapy before in vitro identification of T cells with specificity for a predicted subset of v new antigen candidates from the patient, patients can also be selected to receive T cell therapy after in vitro identification of T cells with specificity for a predicted treatment subset of v new antigen candidates. Specifically, if at least a threshold amount of new antigen-specific TCRs are identified for the patient during the in vitro screening process for new antigen recognition of the patient's T cells, the patient can be selected to receive T cell therapy. For example, only when at least two new antigen-specific TCRs are identified for the patient or only when new antigen-specific TCRs are identified for two different new antigens, the patient can be selected to receive T cell therapy.

[0674] In another embodiment, the patient can be selected to receive T cell therapy only when at least a threshold amount of the neoantigens in the treatment subset of the patient's v neoantigen candidates are recognized by the patient's TCR. For example, the patient can be selected to receive T cell therapy only when at least one neoantigen in the treatment subset of the patient's v neoantigen candidates is recognized by the patient's TCR. In other embodiments, the patient can be selected to receive T cell therapy only when at least a threshold amount of the patient's TCR is identified as having neoantigen specificity for neoantigen peptides of a particular HLA restriction class. For example, the patient can be selected to receive T cell therapy only when at least one of the patient's TCR is identified as having neoantigen specificity for neoantigen peptides restricted to class I HLA.

[0675] In even other embodiments, patients can be selected to receive T cell therapy only when at least a threshold amount of neoantigen peptides of a particular HLA-restricted class are recognized by the patient's TCR. For example, patients can be selected to receive T cell therapy only when at least one class I HLA-restricted neoantigen peptide is recognized by the patient's TCR. For another example, patients can be selected to receive T cell therapy only when at least two class II HLA-restricted neoantigen peptides are recognized by the patient's TCR. After in vitro identification of T cells specific for a predicted therapeutic subset of the patient's v neoantigen candidates, any combination of the above criteria can also be used to select patients to receive T cell therapy.

[0676] XI. Example 7: Experimental Results Showing Example Patient Selection Performance

[0677] The effectiveness of the patient selection method described in Section X was tested by performing patient selection on a set of simulated patients, each of which was associated with a test set of simulated neoantigen candidates, where a subset of the simulated neoantigens was known to be present in the mass spectrometry data. Specifically, each simulated neoantigen candidate in the test set was associated with a marker indicating whether the neoantigen was present in the multi-allelic JY cell line HLA-A*02:01 and LA-B*07:02 mass spectrometry dataset from the Bassani-Sternberg dataset (dataset "D1") (data available at www.ebi.ac.uk / pride / archive / projects / PXD0000394). As shown in the following combination Figure 23A As described in more detail, based on the known frequency distribution of mutational load in non-small cell lung cancer (NSCLC) patients, a number of patient-simulated neoantigen candidates were sampled from the human proteome.

[0678] Independent allele presentation models for the same HLA alleles were trained using a training set that was a subset of the monoallelic HLA-A*02:01 and HLA-B*07:02 mass spectrometry data from the IEDB (dataset "D2") (data can be found at http: / / www.iedb.org / doc / mhc_ligand_full.zip). Specifically, the presentation model for each allele was the independent allele model shown in Equation (8), which incorporates the N-terminal and C-terminal flanking sequences as allele-independent variables, and the network dependency function g h (·) and g w(·), and the expit function f(·). The presentation model for the HLA-A*02:01 allele generates the probability that a given peptide will be presented on the HLA-A*02:01 allele, given the peptide sequence as the allele-interacting variable and the N-terminal and C-terminal flanking sequences as the allele-independent interacting variables. The presentation model for the HLA-B*07:02 allele generates the probability that a given peptide will be presented on the HLA-B*07:02 allele, given the peptide sequence as the allele-interacting variable and the N-terminal and C-terminal flanking sequences as the allele-independent interacting variables.

[0679] As in the following examples and with reference to Figures 23A-23E As described, multiple models for peptide binding prediction, such as trained presentation models and current state-of-the-art models, are applied to a test set of new antigen candidates for each simulated patient to identify different treatment subsets of the patient based on the predictions. Patients who meet the inclusion criteria are selected for vaccine treatment and are associated with a customized vaccine containing epitopes in the patient's treatment subset. The size of the treatment subset varies depending on the different vaccine capacities. No overlap is introduced between the training set used to train the presentation model and the test set of simulated new antigen candidates.

[0680] In the following examples, the proportion of selected patients with at least a certain number of presented neoantigens among the epitopes contained in the vaccine was analyzed. This statistic indicates the effectiveness of the simulated vaccine in delivering potential neoantigens that will elicit an immune response in the patient. Specifically, if the neoantigen is present in the mass spectrometry dataset D2, the simulated neoantigens in the test set are presented. A high proportion of patients with presented neoantigens indicates the potential for successful treatment by a neoantigen vaccine by inducing an immune response.

[0681] XI.A. Example 7A: Frequency Distribution of Mutational Load in NSCLC Cancer Patients

[0682] Figure 23A The sample frequency distribution of mutation load in NSCLC patients is shown. The mutation load and mutations of different tumor types including NSCLC can be found in, for example, The Cancer Genome Atlas (TCGA) ( https: / / cancergenome.nih.gov The x-axis represents the number of non-synonymous mutations in each patient, and the y-axis represents the proportion of sample patients with a given number of non-synonymous mutations. Figure 23A The sample frequency distribution showed a range of 3-1786 mutations, with 30% of patients having fewer than 100 mutations. Figure 23A As shown in , studies have shown that smokers have a higher mutation load than non-smokers and that mutation load may be a strong indicator of neoantigen load in patients.

[0683] As introduced at the beginning of Section XI above, each of the many simulated patients is associated with a test set of neoantigen candidates. Figure 23A The frequency distribution shown in i Sampling is used to generate a test set for each patient. For each mutation, a 21-mer peptide sequence from the human proteome is randomly selected to represent the simulated mutation sequence. By identifying each (8, 9, 10, 11)-mer peptide sequence that spans the mutation in the 21-mer, a test set of new antigen candidate sequences is generated for each patient i. Each new antigen candidate is associated with a tag that indicates whether the new antigen candidate sequence exists in the mass spectrometry D1 dataset. For example, a new antigen candidate sequence present in dataset D1 can be associated with a tag "1", while a sequence that does not exist in dataset D1 can be associated with a tag "0". As described in more detail below, Figures 23B to 23E Experimental results showing patient selection based on presented neoantigens among patients in the test set.

[0684] XI.B. Example 7B: Proportion of Selected Patients with Neoantigen Presentation Based on Tumor Mutation Burden Inclusion Criteria

[0685] Figure 23B The number of neoantigens presented in the simulated vaccine is shown for patients selected based on whether they met the inclusion criteria of a minimum mutation load. The proportion of selected patients with at least a certain number of neoantigens presented in the corresponding test was determined.

[0686] exist Figure 23B , the x-axis represents the proportion of patients excluded from vaccine treatment based on a minimum mutation load (as indicated by the label "minimum number of mutations"). For example, the data point at 200 "minimum number of mutations" indicates that the patient selection module 324 only selects a subset of simulated patients with a mutation load of at least 200 mutations. For another example, the data point at 300 "minimum number of mutations" indicates that the patient selection module 324 selected a lower proportion of simulated patients with at least 300 mutations. The y-axis represents the proportion of selected patients associated with at least a certain number of presented neoantigens in the test set without any vaccine capacity v. Specifically, the top graph shows the proportion of selected patients who present at least one neoantigen, the middle graph shows the proportion of selected patients who present at least two neoantigens, and the bottom graph shows the proportion of selected patients who present at least three neoantigens.

[0687] like Figure 23B As shown, the proportion of selected patients with presented neoantigens increased significantly with higher mutational load, suggesting that mutational load as an inclusion criterion can effectively select patients in whom neoantigen vaccines are more likely to induce a successful immune response.

[0688] XI.C. Example 7c: Presentation of Neoantigens in Vaccines Identified by Presentation Models Compared to Prior Art Models Compare

[0689] Figure 23C The number of neoantigens presented in simulated vaccines was compared between selected patients associated with a vaccine containing a treatment subset identified based on the presentation model and selected patients associated with a vaccine containing a treatment subset identified by the current state-of-the-art model. The left panel assumes a limited vaccine capacity v = 10, while the right panel assumes a limited vaccine capacity v = 20. Patients were selected based on a utility score indicating the expected number of neoantigens presented.

[0690] exist Figure 23C In the figure, the solid lines represent patients associated with vaccines that include treatment subsets identified by presentation models based on the alleles HLA-A*02:01 and HLA-B*07:02. The treatment subset for each patient was identified by applying each presentation model to the sequences in the test set and identifying v neoantigen candidates with the highest presentation likelihood. The dotted lines represent patients associated with vaccines that include treatment subsets identified by the current state-of-the-art model NETMHCpan based on the monoallelic HLA-A*02:01. The implementation details of NETMHCpan are described in http: / / www.cbs.dtu.dk / services / NetMHCpan Detailed description is provided in . The treatment subset for each patient is identified by applying the NETMHCpan model to the sequences in the test set and identifying the v new antigen candidates with the highest estimated binding affinity. The x-axis of the two figures represents the proportion of patients excluded from vaccine treatment based on the expected utility score, which indicates the expected number of new antigens presented in the treatment subset identified based on the presentation model. The expected utility score is determined as described in equation (25) in reference X section. The y-axis represents the proportion of selected patients who presented at least a certain number of new antigens (1, 2 or 3 new antigens) contained in the vaccine.

[0691] like Figure 23C As shown in the figure, patients associated with the vaccine comprising the treatment subset based on the presentation model received a vaccine comprising a presented neoantigen at a significantly higher rate than patients associated with the vaccine comprising the treatment subset based on the prior art model. For example, as shown in the figure on the right, 80% of the selected patients associated with the vaccine based on the presentation model received at least one presented neoantigen in the vaccine, compared to only 40% of the selected patients associated with the vaccine based on the current prior art model. The results indicate that the presentation model as described herein is effective for selecting neoantigen candidates for vaccines that are likely to elicit an immune response for treating tumors.

[0692] XI.D. Example 7D: Effect of HLA Coverage on Neoantigen Presentation in Vaccines Identified by Presentation Models

[0693] Figure 23DThe number of neoantigens presented in the simulated vaccine was compared between selected patients associated with a vaccine containing a treatment subset identified by a single independent allele presentation model for HLA-A*02:01 and selected patients associated with a vaccine containing a treatment subset identified by a dual independent allele presentation model for HLA-A*02:01 and HLA-B*07:02. The vaccine capacity was set to v = 20 epitopes. For each experiment, patients were selected based on the expected utility score determined for the different treatment subsets.

[0694] exist Figure 23D In the figure, the solid line represents patients associated with the vaccine for a treatment subset that includes a dual presentation model based on the HLA alleles HLA-A*02:01 and HLA-B*07:02. The treatment subset for each patient was identified by applying each presentation model to the sequences in the test set and identifying v neoantigen candidates with the highest presentation probability. The dashed line represents patients associated with the vaccine for a treatment subset that includes a single presentation model based on the HLA allele HLA-A*02:01. The treatment subset for each patient was identified by applying the presentation model for only a single HLA allele to the sequences in the test set and identifying v neoantigen candidates with the highest presentation probability. For the solid line graph, the x-axis represents the proportion of patients excluded from vaccine treatment based on the expected utility score of the treatment subset identified by the dual presentation model. For the dashed line graph, the x-axis represents the proportion of patients excluded from vaccine treatment based on the expected utility score of the treatment subset identified by the single presentation model. The y-axis represents the proportion of selected patients presenting at least a certain number of neoantigens (1, 2, or 3 neoantigens).

[0695] like Figure 23D As shown in , patients who received a vaccine containing a treatment subset identified by a dual HLA allele presentation model presented neoantigens at a significantly higher rate than patients who received a vaccine containing a treatment subset identified by a single presentation model. The results demonstrate the importance of establishing a presentation model with high HLA allele coverage.

[0696] XI.E. Example 7E: Patients selected by tumor mutational burden versus the expected number of antigens presented Comparison of neoantigen presentation in

[0697] Figure 23E The number of neoantigens presented in the simulated vaccines was compared between patients selected based on tumor mutational burden and those selected by expected utility scores. Expected utility scores were determined based on a treatment subset identified by a presentation model with v = 20 epitopes.

[0698] exist Figure 23EIn the figure, the solid line represents patients selected based on expected utility scores associated with a vaccine containing a treatment subset identified by the presentation model. The treatment subset for each patient is identified by applying the presentation model to the sequences in the test set and identifying v = 20 neoantigen candidates with the highest presentation likelihood. The expected utility score is determined based on the presentation likelihood of the identified treatment subset according to equation (25) in section X. The dotted line represents patients selected based on mutation load associated with a vaccine that also contains a treatment subset identified by the presentation model. The x-axis of the solid line graph represents the proportion of patients excluded from vaccine treatment based on expected utility scores, and the x-axis of the dotted line graph represents the proportion of patients excluded from vaccine treatment based on mutation load. The y-axis represents the proportion of selected patients who received a vaccine containing at least a certain number of presented antigens (1, 2 or 3 neoantigens).

[0699] like Figure 23E As shown in , patients selected based on expected utility scores received vaccines containing presented neoantigens at higher rates compared to patients selected based on mutational load. However, patients selected based on mutational load received vaccines containing presented neoantigens at higher rates than unselected patients. Therefore, mutational load is a valid patient selection criterion for successful neoantigen vaccine therapy, although expected utility scores are more effective.

[0700] XII. Example 8: Evaluation of mass spectrometry trained models based on hold-out mass spectrometry data

[0701] Since HLA peptide presentation by tumor cells is a key requirement for anti-tumor immunity 91,96,97 , generated a large (N = 74 patients) comprehensive dataset of human tumor and normal tissue samples with paired class I HLA peptide sequences, HLA typing, and transcriptome RNA-seq (Methods), with the goal of using these and publicly available data 92,98,99 Training new deep learning models 100 , to predict antigen presentation in human cancers. Samples were selected from several tumor types of interest for immunotherapy development and based on tissue availability. Mass spectrometry identified an average of 3,704 peptides per sample at a peptide level FDR < 0.1 (range 344-11,301). The peptides followed a characteristic class I HLA length distribution: 8-15 aa in length, with a modal length of 9 (56% of peptides). Consistent with previous reports, MHCflurry 90 Most peptides (median 79%) were predicted to bind to at least one patient HLA allele at a standard 500 nM affinity threshold, but there was considerable variability between samples (e.g., 33% of peptides in one sample had predicted affinity > 500 nM). 101"Strong binders" captured a median of only 42% of presented peptides. Transcriptome sequencing generated an average of 131M unique reads per sample, and 68% of genes were expressed at a level of at least 1 transcript per million (TPM) in at least one sample, highlighting the value of large and diverse sample sets for observing the expression of the maximum number of genes. Peptide presentation by HLA was closely correlated with mRNA expression. Significant and reproducible gene-to-gene differences in peptide presentation rates were observed, beyond what could be explained by RNA expression or sequence differences alone. The observed HLA types were consistent with expectations for samples from a patient group primarily of European ancestry.

[0702] For each patient, the data points of the positive markers are peptides detected by mass spectrometry, and the data points of the negative markers are peptides not detected by mass spectrometry in the sample from the reference proteome (SwissProt). The data are divided into a training set, a validation set, and a test set (method). The training set consists of 142,844 HLA-presented peptides (FDR < about 0.02) from 101 samples (69 samples newly described in this study and 32 samples previously published). The validation set (for early stopping) consists of 18,004 presented peptides from the same 101 samples. Two mass spectrometry datasets were used for testing: (1) a tumor sample test set consisting of 571 presented peptides from 5 additional tumor samples (2 lung samples, 2 colon samples, 1 ovary sample) that were not included in the training data; and (2) a monoallelic cell line test set consisting of 2,128 presented peptides from genomic location windows (cells) that were adjacent to but distinct from the locations of monoallelic peptides included in the training data (see Methods for more details on the training / testing classification).

[0703] Using these and publicly available HLA peptide data 92,98,99 , a neural network (NN) model was trained to predict HLA antigen presentation. Specifically, in Example 9, the pan-allele model discussed in Section VIII.D above was trained using the above data to predict HLA antigen presentation. On the other hand, in Example 11, the allele-specific model described in detail below was trained using the above data to predict HLA antigen presentation. In Example 10, both the pan-allele model discussed in Section VIII.D above and the allele-specific model described in detail below were trained using the above data to predict HLA antigen presentation.

[0704] In particular, in Examples 10 and 11, to learn allele-specific models from tumor mass spectrometry data (where each peptide may have been presented by any of the six HLA alleles), a novel network architecture was developed that can jointly learn allelic peptide profiles and allele-specific presentation motifs (see Section XVII.B below). The training data identified predictive models for 53 HLA alleles. 92,104 Instead, these models capture the dependence of HLA presentation on each sequence position for peptides of various lengths. The model also correctly learns key dependencies on gene RNA expression and gene-specific presentation propensities, where mRNA abundance and learned allele-specific presentation propensities independently combine to produce up to approximately 60-fold differences in presentation rates between the lowest-expressing (minimum presentation propensity) and the highest-expressing (maximum presentation propensity) genes. It was further observed that the measured stability of the model-predicted HLA / peptide complexes was IEDB-positive even after controlling for predicted binding affinity (p < 0.05 for 8 of the 10 tested alleles). 88 (p<1e-10 for 10 alleles.) Together, these features form the basis for improved prediction of immunogenic class I HLA peptides.

[0705] XIII. Example 9: Experimental Results Including Presentation Hotspot Modeling

[0706] In order to specifically evaluate the benefit of using presentation hotspot parameters in HLA presentation modeling, the performance of the pan-allelic neural network presentation model incorporated into the presentation hotspot parameters was compared with the performance of the pan-allelic neural network presentation model not incorporated into the presentation hotspot parameters. The basic neural network architecture of these two pan-allelic models is identical, and is identical to the pan-allelic presentation model described above in VII-VIII parts. In brief, the pan-allelic model includes peptide and side joint amino acid sequence parameters, RNA sequencing transcription data (TPM), protein family data, every sample identification and HLA-A, B, C type. Each pan-allelic model uses an aggregate of 5 networks. The pan-allelic model comprising presentation hotspot parameters uses equation 12b described above in VIII.B.3. part, and the unit size of the independent gene protein group is 10, and the peptide length is 8-12.

[0707] Two pan-allele models were compared using the mass spectrometry dataset described in Section XII above. Specifically, to fairly evaluate the competing models, five samples were set aside from both model training and validation. The remaining samples were randomly divided into 90% for model training and 10% for validation training.

[0708] Figure 24The positive predictive value (PPV) at 40% recall of the pan-allelic presentation model using the presentation hotspot parameters and the pan-allelic presentation model without the presentation hotspot parameters was compared when the pan-allelic model was tested on five hold-out test samples. Figure 24 As shown, the performance of the pan-allelic model incorporating the presentation hotspot parameter is consistently better than that of the pan-allelic model without the presentation hotspot parameter.

[0709] XIV. Example 10: Model Evaluation of Retrospective Neoantigen T Cell Data

[0710] We then evaluated whether the accurate prediction of HLA peptide presentation by the pan-allelic model could be translated into the ability to identify CD8 T cell epitopes (i.e., immunotherapy targets) of human tumors. Defining a suitable test dataset for this evaluation is challenging because it requires peptides that can be recognized by T cells and presented by HLA on the surface of tumor cells. In addition, formal performance evaluation requires not only positively labeled (i.e., T cell recognized) peptides, but also a sufficient number of negatively labeled (i.e., tested but not recognized) peptides. Mass spectrometry datasets address tumor presentation but not T cell recognition; conversely, T cell assays after priming or vaccination address the presence of T cell precursors and T cell recognition but not tumor presentation (e.g., a strongly binding peptide whose source gene is expressed at too low a level in the tumor to support presentation of the peptide may elicit a strong CD8 T cell response after vaccine administration but is not a therapeutically useful target because it cannot be presented by the tumor).

[0711] To obtain a suitable dataset, we collected published CD8 T cell epitopes from four recent studies that met the required criteria: Study A 140 examined TILs from nine patients with gastrointestinal tumors using a tandem minigene (TMG) approach in autologous DC and reported T cell recognition of 12 / 1,053 somatic SNV mutations as tested by IFN-γ ELISPOT. 84 also used TMG and reported T cell recognition of 6 / 574 SNVs by CD8+PD-1+ circulating lymphocytes from 4 melanoma patients. Study C 141 TILs from three melanoma patients were evaluated using pulse peptide stimulation and found responses to 5 / 381 SNV mutations tested. 108TILs from a breast cancer patient were evaluated using a combination of TMG assays and pulsed with minimal epitope peptides and reported recognition of 2 / 62 SNVs. The combined dataset consisted of 2,023 assayed SNVs from 17 patients (including 26 TSNAs with pre-existing T cell responses). Importantly, since this dataset primarily contained recognition of neoantigens by tumor-infiltrating lymphocytes, successful prediction implies the ability to identify not only the SNVs identified in the literature but also the SNVs identified in the literature. 81 ,82 ,141 These tumors can identify neoantigens that can trigger T cells, and (more strictly speaking) can identify neoantigens that tumors present to T cells.

[0712] We ranked mutations in order of probability of presentation using standard HLA binding affinity predictions at a threshold of >2 TPM on gene expression as determined by RNA-seq, the allele-specific neural network model described in Section VIII.B, and the pan-allele neural network model described in Section VIII.D. Because the power of antigen-specific immunotherapies is limited in the number of specificities that can be targeted (e.g., current personalized vaccines encode approximately 10-20 mutations), we used standard HLA binding affinity predictions at a threshold of >2 TPM on gene expression as determined by RNA-seq, the allele-specific neural network model described in Section VIII.B, and the pan-allele neural network model described in Section VIII.D. 6,81,82 ), we therefore compared prediction methods by counting the number of pre-existing T cell responses to the top 5, top 10, or top 20 mutations in each patient. These results are depicted in Figure 25A Specifically, Figure 25A The proportion of somatic mutations recognized by T cells (e.g., pre-existing T cell responses) among the top 5, top 10, and top 20 somatic mutations identified using standard HLA binding affinity predictions at a >2 TPM threshold for gene expression as determined by RNA-seq, allele-specific neural network models, and pan-allelic neural network models was compared for a test set of 12 different test samples, each taken from a patient with at least one pre-existing T cell response.

[0713] As expected, the binding affinity predictions included only a few pre-existing T cell responses among the prioritized mutations, e.g., 9 / 26 (35%) of the pre-existing T cell responses were among the top 20. In contrast, the majority (19 / 26, 73%) of the pre-existing T cell responses were among the top 20 identified by both the allele-specific model and the pan-allelic NN model ( Figure 25A These results demonstrate that the pan-allelic model can identify human tumor CD8 T cell epitopes with comparable accuracy to the allele-specific model (which was not statistically significant).

[0714] We then evaluated mutations at the minimal neo-epitope level (i.e., identifying 8-11mers that overlap with the mutations), which may be useful for identifying T cells / TCRs for cell therapy. In other words, using standard HLA binding affinity predictions, the minimal neo-epitopes were ranked in order of probability of presentation on gene expression as identified by RNA-seq, the allele-specific neural network model described in Section VIII.B, and the pan-allele neural network model described in Section VIII.D, with a threshold of >2 TPM. As mentioned above, because antigen-specific immunotherapy is technically limited in the number of specificities that can be targeted, the prediction methods were compared by counting the number of pre-existing T cell responses in the top 5, top 10, or top 20 minimal neo-epitopes for each patient with at least one pre-existing T cell response. Positively labeled epitopes were those confirmed as minimally immunogenic epitopes by a peptide-based approach (instead of, or in addition to, a TMG-based assay), while negative examples were all epitopes not recognized in the peptide-based assay, as well as all mutation-spanning epitopes contained in unrecognized minigenes. The results are depicted in Figure 25B middle.

[0715] Specifically, Figure 25B The proportion of minimal neoepitopes recognized by T cells (e.g., pre-existing T cell responses) among the top 5, top 10, and top 20 minimal neoepitopes identified using standard HLA binding affinity predictions at a threshold of >2 TPM for gene expression as determined by RNA-seq, allele-specific neural network models, and pan-allelic neural network models was compared for a test set of 12 different test samples, each taken from a patient with at least one pre-existing T cell response.

[0716] like Figure 25B As shown, when mutations were evaluated at the minimal epitope level, the pan-allelic model continued to perform comparably to the allele-specific model.

[0717] XIV.A. Data

[0718] We learned from Gros et al. 84 , Tran et al. 140 , Stronen et al. 141Mutation calling, HLA typing, and T cell recognition data are available in the supplementary information of Zacharakis et al. Patient-specific RNA-seq data were not available. Presumably, tumor RNA expression is correlated between patients with the same tumor type, so RNA-seq data from tumor-type-matched patients in the TCGA were substituted for neural network prediction and RNA expression filtering before binding affinity prediction. The addition of tumor-type-matched RNA-seq data improved prediction performance.

[0719] For mutation level analysis ( Figure 25A ), the positively labeled data points for Gros et al., Tran et al., and Zacharakis et al. are mutations that are recognized by patient T cells in the TMG assay or the minimal epitope peptide pulse assay. The negatively labeled data points are all other mutations tested in the TMG assay. For Stronen et al., the positively labeled mutations are mutations that are spanned by at least one recognition peptide, and the negative data points are all mutations that are tested but not recognized in the tetramer assay. For the Gros, Tran, and Zacharakis data, mutations are ranked by summing the probabilities of presentation of all spanning mutant peptides or using the minimum binding affinity, because the 25-mer TMG assay of the mutations tests T cell recognition of all spanning mutant peptides. For the Stronen data, mutations are ranked by summing the probabilities of presentation of all spanning mutant peptides tested in the tetramer assay or using the minimum binding affinity.

[0720] For epitope level analysis, positively labeled data points are all minimal epitopes recognized by patient T cells in either peptide pulsing or tetramer assays, and negative data points are all minimal epitopes not recognized by T cells in either peptide pulsing or tetramer assays, as well as all transmutation peptides from the tested TMG that were not recognized by patient T cells. In the cases of Gros et al., Tran et al., and Zacharakis et al., transmutation minimal epitope peptides recognized in the TMG assay that were not tested by the peptide pulsing assay were removed from the analysis because the T cell recognition status of these peptides had not been experimentally determined.

[0721] XV. Examples 11 Identifying neoantigen-reactive T cells in cancer patients

[0722] This example demonstrates that improved prediction can identify neoantigens from routine patient samples. To this end, archived FFPE tumor biopsies and 5-30 ml peripheral blood from 9 patients with metastatic NSCLC receiving anti-PD(L)1 therapy were analyzed (Supplementary Table 1: Figure 26A-C Patient demographics and treatment information for N = 9 patients studied. Key fields include tumor stage and subtype, anti-PD1 therapy received, and an overview of NGS results). Tumor whole-exome sequencing, tumor transcriptome sequencing, and matched normal exome sequencing generated an average of 198 somatic mutations (SNVs and short indels) per patient, of which an average of 118 were expressed (Methods, Supplementary Table 1). The full MS model was applied to prioritize 20 neoepitopes per patient to test against pre-existing anti-tumor T cell responses. To focus the analysis on possible CD8 responses, prioritized peptides were synthesized as 8-11mer minimal epitopes (Methods), and peripheral blood mononuclear cells (PBMCs) were then cultured with the synthesized peptides in a short in vitro stimulation (IVS) culture to expand neoantigen-reactive T cells (Supplementary Table 2). Two weeks later, the presence of antigen-specific T cells was assessed using IFN-γ ELISpot against the prioritized neoepitopes. In the 7 patients for whom sufficient PBMC were available, separate experiments were also performed to fully or partially deconvolute the specific antigens identified. Figure 26A -C and shown in 27A-30.

[0723] Figure 26A Depicts the detection of T cell responses to patient-specific neoantigen peptide libraries for 9 patients. For each patient, the predicted neoantigens were combined into two pools of 10 peptides based on model ranking and any sequence homology (homologous peptides were divided into different pools). Then, for each patient, the patient's in vitro expanded PBMCs were stimulated with the two patient-specific neoantigen peptide pools in IFN-γ ELISpot. Figure 26A The data are expressed as the background (corresponding DMSO negative control) minus the per 10 5 Spot forming units (SFU) of plated cells. Background measurement (DMSO negative control) is shown in Figure 30 For patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002 and CU05, the responses of single wells (patients 1-038-001, CU02, CU03 and 1-050-001) or replicates with mean and standard deviation (all other patients) for homologous peptide pools #1 and #2 are shown. For patients CU02 and CU03, only the number of cells tested for specific peptide pool #1 was allowed. Samples with values ​​>2-fold higher than background were considered positive and were marked with an asterisk (responsive donors included patients 1-038-001, CU04, 1-024-001, 1-024-002 and CU02). Non-responsive donors included patients 1-050-001, 1-001-002, CU05, and CU03. Figure 26CDepicted are photographs of ELISpot wells with in vitro expanded PBMCs from patient CU04 stimulated in an IFN-γ ELISpot with a DMSO negative control, a PHA positive control, CU04-specific neoantigen peptide pool #1, CU04-specific peptide 1, CU04-specific peptide 6, and CU04-specific peptide 8.

[0724] Figure 27A -B depicts results from a control experiment using patient neoantigens from HLA-matched healthy donors. The results of these experiments demonstrate that the in vitro culture conditions only expand pre-existing in vivo primed memory T cells and are not capable of de novo in vitro priming.

[0725] Figure 28 Describes the Figure 26A For each donor and each in vitro expansion described in, T cell responses to the PHA positive control were measured. Figure 26A For each donor and each ex vivo expansion, ex vivo expanded patient PBMCs were stimulated with PHA to maximize T cell activation. Figure 28 The data are expressed as the background (corresponding DMSO negative control) minus the per 10 5 The number of spot forming units (SFU) of plated cells was 100 μg / mL. Responses of single wells or biological replicates are shown for patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002, CU05, and CU03. Patient CU02 was not tested for PHA. Cells from patient CU02 were included in the analysis because of a positive response to peptide pool #1 ( Figure 26A ) indicates viable and functional T cells. Figure 26A As shown, donors who responded to the peptide library included patients 1-038-001, CU04, 1-024-001, and 1-024-002. Figure 26A As shown, donors that did not respond to the peptide pool included patients 1-050-001, 1-001-002, CU05, and CU03.

[0726] Figure 29A Depicted is the detection of T cell responses of patient CU04 to each individual patient-specific neoantigen peptide in pool #2. Figure 29A Also depicted is the detection of T cell responses to a PHA positive control in patient CU04. (This positive control data is also shown in Figure 28) For patient CU04, the patient's in vitro expanded PBMCs were stimulated in IFN-γ ELISpot with the patient-specific individual neoantigen peptides from pool #2 of patient CU04. The patient's in vitro expanded PBMCs were also stimulated in IFN-γ ELISpot with PHA as a positive control. Data are presented as per 10 5 The number of spot-forming units (SFU) per plated cell was determined.

[0727] Figure 29B Depicted are the detection of T cell responses to individual patient-specific neoantigen peptides for each of three visits for patient CU04 and for each of two visits for patient 1-024-002, each visit occurring at a different time point. For both patients, the patient's ex vivo expanded PBMCs were stimulated in IFN-γ ELISpot using the patient-specific individual neoantigen peptides. For each patient, the data for each visit are expressed as per 10 5 The data of patient 1-024-002 are shown as the cumulative SFU of subtracting background of 2 visits.For patient 1-024-002, the data of patient 1-024-002 are shown as the cumulative SFU of subtracting background of 1 month (T0+1 month) after the first visit (T0) and the first visit (T0).For patient 1-024-002, the data of patient 1-024-002 are shown as the cumulative SFU of subtracting background of 2 visits.For patient 1-024-002, the data of patient 1-024-002 are shown as the SFU of subtracting background of the first visit (T0) and the first visit (T0).The sample that value increases>2 times than background is considered as positive, and is marked with asterisk.

[0728] Figure 29C Depicted are the detection of T cell responses to individual patient-specific neoantigen peptides and patient-specific neoantigen peptide pools for each of the two visits for patient CU04 and for each of the two visits for patient 1-024-002, each visit occurring at a different time point. For both patients, the patient's ex vivo expanded PBMCs were stimulated in IFN-γ ELISpot using patient-specific individual neoantigen peptides as well as patient-specific neoantigen peptide pools. Specifically, for patient CU04, the ex vivo expanded PBMCs of patient CU04 were...

Claims

1. A method for identifying at least one neoantigen from one or more tumor cells of a subject that may be presented on the surface of the tumor cells by one or more MHC alleles, the method comprising the steps of: obtaining at least one of exome, transcriptome, or whole genome nucleotide sequencing data from the tumor cells and normal cells of the subject, wherein the nucleotide sequencing data is used to obtain data representing a peptide sequence of each neo-antigen in a set of neo-antigens identified by comparing the nucleotide sequencing data from the tumor cells to the nucleotide sequencing data from the normal cells, wherein the peptide sequence of each neo-antigen comprises at least one change that makes it different from a corresponding wild-type peptide sequence identified from a normal cell of the subject; encoding the peptide sequence of each of the neoantigens into a corresponding digital vector, each digital vector comprising information about a set of a plurality of amino acids constituting the peptide sequence and positions of the amino acids in the peptide sequence; obtaining at least one of exome, transcriptome, or whole genome nucleotide sequencing data from the tumor cells of the subject, wherein the nucleotide sequencing data is used to obtain data representing a peptide sequence for each of the one or more MHC alleles of the subject; encoding a peptide sequence for each of the one or more MHC alleles of the subject into a corresponding digital vector, each digital vector comprising information about a set of a plurality of amino acids constituting the peptide sequence and a position of the amino acids in the peptide sequence; Inputting, using a computer processor, a digital vector encoding a peptide sequence of each of the neoantigens and a digital vector encoding a peptide sequence of each of the one or more MHC alleles into a machine learning presentation model to generate a set of presentation probabilities for the set of neoantigens, each presentation probability in the set representing a likelihood that the corresponding neoantigen is presented on a tumor cell surface of the subject by the one or more MHC alleles, the machine learning presentation model comprising: A plurality of parameters identified based at least on a training data set, the training data set comprising: for each sample in the plurality of samples, measuring by mass spectrometry a signature obtained from the presence of a peptide that binds to at least one MHC allele of the set of MHC alleles identified as being present in the sample; for each sample, a training peptide sequence encoded as a digital vector containing information about a set of a plurality of amino acids constituting the peptide and the positions of the amino acids in the peptide; as well as for each sample, a training peptide sequence encoded as a digital vector containing information about a set of a plurality of amino acids constituting at least one MHC allele to which said peptide of said sample binds and the positions of said amino acids in said at least one MHC allele; a function representing a relationship between a digital vector of peptide sequences encoding each of said neoantigens and a digital vector of peptide sequences encoding each of said one or more MHC alleles received as input and a likelihood of presentation generated as output based on said digital vectors and said parameters; selecting a subset of the set of neoantigens based on the set of presentation probabilities to produce a set of selected neoantigens; as well as Recovering the collection of selected neoantigens.

2. The method of claim 1, wherein inputting a digital vector encoding a peptide sequence for each of the neoantigens and a digital vector encoding a peptide sequence for each of the one or more MHC alleles into the machine learning presentation model comprises: The machine learning presentation model is applied to the peptide sequence of the neoantigen and to the peptide sequence of the one or more MHC alleles to generate a dependency score for each of the one or more MHC alleles, the dependency score indicating whether the MHC allele will present the neoantigen based on a specific amino acid at a specific position of the peptide sequence.

3. The method of claim 2, wherein inputting the digital vector encoding the peptide sequence of each of the neoantigens and the digital vector encoding the peptide sequence of each of the one or more MHC alleles into the machine learning presentation model further comprises: transforming the dependency scores to generate a corresponding independent allele likelihood for each MHC allele, thereby indicating the likelihood that the corresponding MHC allele will present the corresponding neoantigen; and The independent allele probabilities are combined to generate the presentation probability of the neoantigen.

4. The method of claim 3, wherein transforming the dependency score models presentation of the neoantigen as mutually exclusive between the one or more MHC alleles.

5. The method of claim 2, wherein inputting the digital vector encoding the peptide sequence of each of the neoantigens and the digital vector encoding the peptide sequence of each of the one or more MHC alleles into the machine learning presentation model further comprises: The combination of the dependency scores is transformed to generate a presentation probability, wherein transforming the combination of the dependency scores models presentation of the neoantigen as if there is interference between the one or more MHC alleles.

6. The method according to any one of claims 2 to 5, wherein the set of presentation possibilities is further identified by at least one or more allele non-interaction features, and further comprising: The machine learning presentation model is applied to the allele non-interaction feature to generate a dependency score for the allele non-interaction feature, the dependency score indicating whether a peptide sequence of the corresponding neoantigen will be presented based on the allele non-interaction feature.

7. The method according to claim 6, further comprising: Combining the dependency score for each MHC allele of the one or more MHC alleles with the dependency score for the non-interacting signature of the alleles ； transforming the combined dependency scores for each MHC allele to generate an independent allele likelihood for each MHC allele, thereby indicating the likelihood that the corresponding MHC allele will present the corresponding neoantigen; and The independent allele probabilities are combined to generate the presentation probability.

8. The method according to claim 6, further comprising: combining the dependency scores for each of the MHC alleles with the dependency scores for the allele non-interaction signatures; and The combined dependency scores are transformed to generate the presentation likelihoods.

9. The method of any one of claims 1-8, wherein the one or more MHC alleles comprise two or more different MHC alleles.

10. The method of any one of claims 1-9, wherein the peptide sequence comprises a peptide sequence that is not 9 amino acids in length.

11. The method of any one of claims 1-10, wherein encoding a peptide sequence comprises encoding the peptide sequence using a one-hot encoding scheme.

12. The method according to any one of claims 1 to 11, wherein the plurality of samples comprises at least one of the following: (a) one or more cell lines engineered to express a single MHC allele; (b) one or more cell lines engineered to express multiple MHC alleles; (c) one or more human cell lines obtained or derived from a plurality of patients; (d) fresh or frozen tumor samples obtained from multiple patients; and (e) Fresh or frozen tissue samples obtained from multiple patients.

13. The method according to any one of claims 1 to 12, wherein the training data set further comprises at least one of the following: (a) data relating to peptide-MHC binding affinity measurements of at least one of the peptides; and (b) data relating to a measure of peptide-MHC binding stability of at least one of said peptides.

14. The method of any one of claims 1-13, wherein the presentation probability set is further identified by expression levels of the one or more MHC alleles in the subject as measured by at least RNA-seq or mass spectrometry.

15. The method of any one of claims 1-14, wherein the set of presentation possibilities is further identified by features, the features comprising at least one of: (a) predicted affinity between the neo-antigens in the neo-antigen set and the one or more MHC alleles; and (b) Predicted stability of the peptide-MHC complex encoded by the neoantigen.

16. The method of any one of claims 1-15, wherein the set of numerical possibilities is further identified by features, the features comprising at least one of: (a) a C-terminal sequence flanking the neoantigen encoding peptide sequence within the source protein sequence; and (b) The N-terminal sequence of the neoantigen encoding peptide sequence is flanked within the source protein sequence.

17. The method of any one of claims 1-16, wherein selecting the set of selected neo-antigens comprises selecting neo-antigens having an increased likelihood of being presented on the surface of the tumor cell relative to unselected neo-antigens based on the machine learning presentation model.

18. The method of any one of claims 1-17, wherein selecting the set of selected neoantigens comprises selecting neoantigens having an increased likelihood of inducing a tumor-specific immune response in the subject relative to unselected neoantigens based on the machine learning presentation model.

19. The method of any one of claims 1-18, wherein selecting the set of selected neo-antigens comprises selecting neo-antigens that have an increased likelihood of being presented to naive T cells by professional antigen presenting cells (APCs) relative to unselected neo-antigens based on the presentation model, optionally wherein the APCs are dendritic cells (DCs).

20. The method of any one of claims 1-19, wherein selecting the set of selected neoantigens comprises selecting neoantigens having a reduced likelihood of undergoing central or peripheral tolerance suppression relative to unselected neoantigens based on the machine learning presentation model.

21. The method of any one of claims 1-20, wherein selecting the set of selected neoantigens comprises selecting neoantigens that have a reduced likelihood of inducing an autoimmune response against normal tissue in the subject relative to unselected neoantigens based on the machine learning presentation model.

22. The method of any one of claims 1-21, wherein the one or more tumor cells are selected from the group consisting of lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, stomach cancer, colon cancer, testicular cancer, head and neck cancer, pancreatic cancer, brain cancer, B-cell lymphoma, acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia and T-cell lymphocytic leukemia, non-small cell lung cancer and small cell lung cancer.

23. The method of any one of claims 1-22, further comprising generating an output for constructing a personalized cancer vaccine from the set of selected neoantigens.

24. The method of claim 23, wherein the output for the personalized cancer vaccine comprises at least one peptide sequence or at least one nucleotide sequence encoding the set of selected neoantigens.

25. The method of any one of claims 1-24, wherein the machine learning rendering model is a neural network model.

26. The method of claim 25, wherein the neural network model comprises a single neural network model comprising a series of nodes arranged in one or more layers, the single neural network model being configured to receive a digital vector encoding a peptide sequence of a plurality of different MHC alleles.

27. The method of claim 26, wherein the neural network model is trained by updating parameters of the neural network model.

28. The method of any one of claims 25-27, wherein the machine learning rendering model is a deep learning model comprising one or more node layers.

29. The method according to any one of claims 1 to 28, wherein the training peptide sequences encoded as digital vectors containing information about a set of a plurality of amino acids constituting at least one MHC allele that binds to the peptide of the sample and the positions of the amino acids in the at least one MHC allele do not include: Peptide sequences of the subject's MHC alleles that are input into the machine learning presentation model to generate a set of presentation probabilities for the set of neo-antigens.

30. The method of any one of claims 1-29, wherein at least one MHC allele that binds to a peptide of each of the plurality of samples of the training dataset belongs to a gene family to which one or more MHC alleles of the subject belong.

31. The method of any one of claims 1-30, wherein the at least one MHC allele that binds to a peptide of each of the plurality of samples of the training data set comprises one MHC allele.

32. The method of any one of claims 1-30, wherein the at least one MHC allele that binds to a peptide of each of the plurality of samples of the training data set comprises more than one MHC allele.

33. The method of any one of claims 1-32, wherein the one or more MHC alleles are class I MHC alleles.

34. A computer system comprising: Computer processors; A memory for storing computer program instructions that, when executed by the computer processor, cause the computer processor to: obtaining at least one of exome, transcriptome, or whole genome nucleotide sequencing data from tumor cells and normal cells of the subject, wherein the nucleotide sequencing data is used to obtain data representing a peptide sequence of each neo-antigen in a set of neo-antigens identified by comparing the nucleotide sequencing data from the tumor cells to the nucleotide sequencing data from the normal cells, wherein the peptide sequence of each neo-antigen comprises at least one change that makes it different from a corresponding wild-type peptide sequence identified from a normal cell of the subject; encoding the peptide sequence of each of the neoantigens into a corresponding digital vector, each digital vector comprising information about a set of a plurality of amino acids constituting the peptide sequence and positions of the amino acids in the peptide sequence; obtaining at least one of exome, transcriptome, or whole genome nucleotide sequencing data from each of the one or more MHC alleles of the subject, wherein the nucleotide sequencing data is used to obtain data representing a peptide sequence for each of the one or more MHC alleles of the subject; encoding a peptide sequence for each of the one or more MHC alleles of the subject as a corresponding digital vector, each digital vector comprising information about a set of a plurality of amino acids constituting the peptide sequence and a position of the amino acids in the peptide sequence; Inputting, using a computer processor, a digital vector encoding a peptide sequence of each of the neoantigens and a digital vector encoding a peptide sequence of each of the one or more MHC alleles into a machine learning presentation model to generate a set of presentation probabilities for the set of neoantigens, each presentation probability in the set representing a likelihood that the corresponding neoantigen is presented on a tumor cell surface of the subject by the one or more MHC alleles, the machine learning presentation model comprising: A plurality of parameters identified based at least on a training data set, the training data set comprising: for each sample in the plurality of samples, measuring by mass spectrometry a signature obtained from the presence of a peptide that binds to at least one MHC allele of the set of MHC alleles identified as being present in the sample; for each sample, a training peptide sequence encoded as a digital vector containing information about a set of a plurality of amino acids constituting the peptide and the positions of the amino acids in the peptide; as well as for each sample, a training peptide sequence encoded as a digital vector containing information about a set of a plurality of amino acids constituting at least one MHC allele to which said peptide of said sample binds and the positions of said amino acids in said at least one MHC allele; a function representing a relationship between a digital vector of peptide sequences encoding each of said neoantigens and a digital vector of peptide sequences encoding each of said one or more MHC alleles received as input and a likelihood of presentation generated as output based on said digital vectors and said parameters; selecting a subset of the set of neoantigens based on the set of presentation probabilities to produce a set of selected neoantigens; as well as Recovering the collection of selected neoantigens.

Images