Identification, manufacture and use of neoantigens

Abstract

Disclosed herein is a system and method for determining alleles, neoantigens, and vaccine compositions based on tumor mutations in an individual. Systems and methods for obtaining high quality sequencing data from a tumor are also disclosed. In addition, systems and methods for identifying somatic cell changes in polymorphic genomic data are also described herein. Finally, unique cancer vaccines are described herein.

Description

[0001] This application is a divisional application of the invention patent application with application date of December 16, 2016, application number 201680080924.4, and invention name “Identification, manufacture and use of new antigens”.

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] This application claims the benefit of and priority to U.S. Provisional Application No. 62 / 268,333 filed December 16, 2015, U.S. Provisional Application No. 62 / 317,823 filed April 4, 2016, U.S. Provisional Application No. 62 / 379,986 filed August 26, 2016, U.S. Provisional Application No. 62 / 394,074 filed September 13, 2016, and U.S. Provisional Application No. 62 / 425,995 filed November 23, 2016, each of which is incorporated by reference in its entirety for all purposes. Background Art

[0004] Therapeutic vaccines based on tumor-specific neoantigens hold great promise as a new 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 probability of generating neoantigens. 4,5 Early evidence suggests that neoantigen-based vaccination can elicit T cell responses 6 And cell therapies targeting neoantigens can in some cases induce tumor regression in selected patients. 7

[0005] One question in neoantigen vaccine design is which of the many coding mutations present within a subject's tumor will produce the "best" therapeutic neoantigen, i.e., one that can elicit antitumor immunity and cause tumor regression.

[0006] The proposed initial approach incorporates mutation-based analysis using next-generation sequencing, RNA gene expression, and MHC binding affinity prediction of candidate neoantigenic peptides 8 However, none of these proposed approaches can mimic the entire epitope generation process, which consists of many steps in addition to gene expression and MHC binding (e.g., TAP transport, proteasome cleavage, and / or TCR recognition). 9 Therefore, existing methods may suffer from 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 can be found on 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 responses fails to improve prediction accuracy for neoantigens with restricted binding. 12

[0008] This low positive predictive value (PPV) of existing presentation prediction methods raises questions about the design of vaccines based on neoantigens. If a prediction method with a low PPV is used to design a vaccine, most patients are unlikely to receive a therapeutic neoantigen, and a few patients may receive more than one neoantigen (even assuming that all presented peptides are immunogenic). Therefore, neoantigen vaccination with current methods is 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 occur in multiple tumor types and result in 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 analysis methods may inadvertently promote sequence artifacts or germline polymorphisms as neoantigens, leading to inefficient use of vaccine capabilities or risk of autoimmunity, respectively. Summary of the invention

[0011] Disclosed herein is an optimization method for identifying and selecting new antigens for personalized cancer vaccines. First, an optimized tumor exome and transcriptome analysis method for identifying new antigen candidates using next generation sequencing (NGS) is proposed. These methods are based on standard NGS tumor analysis methods to ensure that the new antigen candidates with the highest sensitivity and specificity are promoted within all categories of genomic changes. Secondly, a novel method for selecting high PPV new antigens is proposed to overcome specificity problems and ensure that the new antigens intended to be included in the vaccine are more likely to induce anti-tumor immunity. Depending on the implementation scheme, these methods include trained statistical regression or nonlinear deep learning models, which jointly simulate peptide-allele positioning and independent allele motifs (per-allele motif) of peptides of various lengths, with a total statistical strength in peptides of different lengths. Nonlinear deep learning models can be specifically designed and trained to treat different MHC alleles in the same cell as independent, thereby solving the problem that different MHC alleles of linear models interfere with each other. Finally, other issues that need to be considered in the design and manufacture of personalized vaccines based on new antigens are solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] 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:

[0013] Figure 1A Shown are current clinical approaches for identifying neoantigens.

[0014] Figure 1B It was shown that <5% of the predicted binding peptides were presented on tumor cells.

[0015] Figure 1C Showing the impact of the specificity problem of neoantigen prediction.

[0016] Figure 1D showed that binding prediction is insufficient for neoantigen identification.

[0017] Figure 1E The probability of MHC-I presentation is shown as a function of peptide length.

[0018] Figure 1F An exemplary peptide spectrum generated from the Promega dynamic range standard is shown.

[0019] Figure 1G Shows how adding features increases the positive predictive value of the model.

[0020] Figure 2A is a summary of the environment used to identify the potential for peptide presentation in a patient, according to one embodiment.

[0021] Figure 2B and 2C A method of obtaining presentation information according to one embodiment is shown.

[0022] Figure 3 is a high-level block diagram illustrating the computer logic components of a presentation authentication system according to one embodiment.

[0023] Figure 4 A set of example training data is shown according to one embodiment.

[0024] Figure 5 An example network model associated with MHC alleles is shown.

[0025] FIG. 6 shows an example network model of MHC allele sharing.

[0026] Figure 7 The presentation probability of peptides associated with one MHC allele is shown using an example network model.

[0027] Figure 8 The presentation probability of peptides associated with one MHC allele is shown using an example network model.

[0028] Fig. 9 The generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model is shown.

[0029] Fig.10 The generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model is shown.

[0030] Fig.11 The generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model is shown.

[0031] Fig.12 The generation of presentation probabilities for peptides associated with multiple MHC alleles using an example network model is shown.

[0032] FIG13 shows performance results for various example rendering models.

[0033] Fig.14 An example computer for implementing the entities shown in FIGS. 1 and 3 is shown. DETAILED DESCRIPTION

[0034] I. Definitions

[0035] In general, the terms used in the claims and description are intended to be interpreted as having their ordinary meaning 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 meaning and the provided definition, the provided definition shall apply.

[0036] As used herein, the term "antigen" is a substance that induces an immune response.

[0037] 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. The new antigen may include a polypeptide sequence or a nucleotide sequence. The mutation may include a frameshift or non-frameshift insertion or deletion, a missense or nonsense substitution, a splice site change, a genomic rearrangement or gene fusion, or any genomic or expression change that produces a neoORF. The mutation may also include a splice variant. 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 Iligands are proteasome-generated spliced ​​peptides; Science. 2016 Oct 21; 354(6310): 354-358.

[0038] 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.

[0039] As used herein, the term "neoantigen-based vaccine" is a vaccine construct based on one or more neoantigens, such as a plurality of neoantigens.

[0040] As used herein, the term "candidate neoantigen" is a mutation or other abnormality that produces a new sequence that may represent a neoantigen.

[0041] As used herein, the term "coding region" is the portion of a gene that encodes a protein.

[0042] As used herein, the term "coding mutation" is a mutation present in a coding region.

[0043] As used herein, the term "ORF" refers to open reading frame.

[0044] As used herein, the term "NEO-ORF" is a tumor-specific ORF resulting from mutation or other abnormalities such as splicing.

[0045] As used herein, the term "missense mutation" is a mutation that results in the substitution of one amino acid with another amino acid.

[0046] As used herein, the term "nonsense mutation" is a mutation that results in the substitution of an amino acid with a stop codon.

[0047] As used herein, the term "frameshift mutation" is a mutation that results in a change in the protein framework.

[0048] As used herein, the term "indel" is an insertion or deletion of one or more nucleic acids.

[0049] The term "identity" percentage as used herein in the context of two or more nucleic acids or polypeptide sequences refers to when compared and aligned to achieve maximum correspondence, as measured using one of the sequence comparison algorithms described below (e.g., BLASTP and BLASTN, or other algorithms available to the technician) or obtained by visual inspection, that the two or more sequences or subsequences have a specified percentage of nucleotides or amino acid residues that are identical. Depending on application, the "identity" percentage can be present in a certain region of the sequence being compared, for example, in a functional domain, or in the full length of two sequences to be compared.

[0050] 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 the presence or absence of specific amino acids for the translated sequence.

[0051] 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 for 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., see below).

[0052] 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.

[0053] As used herein, the term "no stop or read-through" is a mutation that results in the removal of the natural stop codon.

[0054] As used herein, the term "epitope" is a specific portion of an antigen that is typically bound by an antibody or T-cell receptor.

[0055] As used herein, the term "immunogenicity" is the ability to elicit an immune response, for example, by T cells, B cells, or both.

[0056] As used herein, the terms "HLA binding affinity", "MHC binding affinity" mean the binding affinity between a specific antigen and a specific MHC allele.

[0057] As used herein, the term "bait" is a nucleic acid probe used to enrich a specific DNA or RNA sequence from a sample.

[0058] As used herein, the term "variant" is a difference between a subject's nucleic acid and a reference human genome used as a control.

[0059] As used herein, the term "variant call" is an algorithmic determination of the presence of a variant, typically determined by sequencing.

[0060] As used herein, the term "polymorphism" is a germline variant, ie, a variant found in all DNA-bearing cells of an individual.

[0061] As used herein, the term "somatic variant" is a variant arising in non-germline cells of an individual.

[0062] As used herein, the term "allele" is one form of a gene, or one form of a gene sequence, or one form of a protein.

[0063] As used herein, the term "HLA type" is the complementary sequence of the HLA gene alleles.

[0064] As used herein, the term "nonsense-mediated decay" or "NMD" is the degradation of mRNA by cells caused by premature stop codons.

[0065] As used herein, the term "trunk mutation" is a mutation that originates early in tumor development and is present in most tumor cells.

[0066] 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.

[0067] As used herein, the term "exome" is a subset of a genome that encodes proteins. An exome may be all exons of a genome.

[0068] As used herein, the term "logistic regression" is a statistically derived regression model for binary data in which the logit of the odds that a dependent variable is equal to 1 is modeled as a linear function of the dependent variable.

[0069] 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 back-propagation.

[0070] 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.

[0071] As used herein, the term "peptide group" is the collection of all peptides presented on the cell surface by MHC-I or MHC-II. The peptide group can refer to the characteristics of a cell or a group of cells (e.g., the tumor peptidome, meaning the union of the peptide groups of all cells constituting a tumor).

[0072] As used herein, the term "ELISPOT" means enzyme-linked immunosorbent spot assay, which is a common method used to monitor immune responses in humans and animals.

[0073] As used herein, the term "dextramer" is a dextran-based peptide-MHC multimer used for antigen-specific T cell staining in flow cytometry.

[0074] As used herein, the term "tolerance or immune tolerance" is a state of immune unresponsiveness to one or more antigens, such as self-antigens.

[0075] 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).

[0076] 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 differentiation of these T cells into Tregs.

[0077] 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.

[0078] 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.

[0079] 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.

[0080] 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 features of the subject, such as but not limited to age and gender.Clinical factors can be scores, values ​​or a set of values ​​that can be obtained by evaluating a sample (or sample group) or subject from a subject under certain 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.

[0081] 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.

[0082] 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.

[0083] Any term not directly defined herein should be understood to have the meaning commonly associated therewith as understood in the field of the present invention. Certain terms discussed herein are intended to provide additional guidance to practitioners describing 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 represented in more than one way. Therefore, alternative wordings and synonyms can be used for any one or more of the terms discussed herein. It does not matter whether the term is elaborated or discussed herein. Some synonyms or replaceable methods, materials, etc. are provided. Unless explicitly 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 term examples, are for illustrative purposes only and are not intended to limit the scope and meaning of various aspects of the present invention herein.

[0084] 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.

[0085] II. Methods for Identifying Neoantigens

[0086] Disclosed herein are methods for identifying new antigens from a subject's tumor that may be presented on the surface of a tumor cell and / or may be immunogenic. For example, one such method may include the following steps: obtaining at least one of exome, transcriptome, or whole genome tumor nucleotide sequencing data from a subject's tumor cell, wherein the tumor nucleotide sequencing data is used to obtain data representing a peptide sequence of each new antigen in a set of new antigens, and wherein the peptide sequence of each new antigen comprises at least one change that makes it different from the corresponding wild-type parent peptide sequence; inputting the peptide sequence of each new antigen into one or more presentation models to generate a set of digital probabilities that each of the new antigens is presented by one or more MHC alleles on the surface of a tumor cell of a subject's tumor cell or on cells present in the tumor, the set of digital probabilities having been identified based at least on received mass spectrometry data; and selecting a subset of the set of new antigens based on the set of digital probabilities to generate a set of selected new antigens.

[0087] The presentation model may include a statistical regression or machine learning (e.g., deep learning) model trained on a reference data set (also referred to as a training data set) comprising a corresponding set of markers, wherein the reference data set is obtained from each of a plurality of different subjects, wherein optionally, some subjects may have tumors, and wherein the reference data set comprises at least one of: data representing exome nucleotide sequences from tumor tissue, data representing exome nucleotide sequences from normal tissue, data representing transcriptome nucleotide sequences from tumor tissue, data representing proteome sequences from tumor tissue, and data representing MHC peptide group sequences from tumor tissue, and data representing MHC peptide group sequences from normal tissue. The reference data may further include monoallelic cell lines engineered to express predetermined MCH alleles and subsequently exposed to synthetic proteins, normal and tumor human cell lines, and mass spectrometry data, sequencing data, RNA sequencing data, and proteomic data of fresh and frozen original samples, and T cell assays (e.g., ELISPOT). In certain aspects, the reference data set includes each form of reference data.

[0088] The presentation model may comprise a feature set derived at least in part from a reference data set, and wherein the feature set comprises at least one of an allele-dependent feature and an allele-independent feature. In certain aspects, each feature is included.

[0089] Dendritic cells presenting natural T cell characteristics may include at least one of the following: the characteristics described above. The dose and type of antigen in the vaccine (e.g., peptide, mRNA, virus, etc.): (1) the route of antigen type uptake by dendritic cells (DC) (e.g., endocytosis, microcytosis); and / or (2) the efficacy of antigen uptake by DC. The dose and type of adjuvant in the vaccine. The length of the vaccine antigen sequence. The number and site of vaccine administration. Baseline patient immune function (e.g., measured by recent infection history, blood counts, etc.). For RNA vaccines: (1) the conversion rate of mRNA protein products in dendritic cells; (2) the translation rate of mRNA after dendritic cell uptake as measured in in vitro or in vivo experiments; and / or (3) the number or number of mRNA translations after dendritic cell uptake as measured in in vivo or in vitro experiments. The presence of protease cleavage motifs in the peptide, optionally providing additional weight to proteases that are commonly expressed in dendritic cells (e.g., measured by RNA-seq or mass spectrometry). The expression level of proteasomes and immunoproteasomes in typical activated dendritic cells (can be measured by RNA-seq, mass spectrometry, immunohistochemistry or other standard techniques). The expression level of a specific MHC allele in a related individual (e.g., measured by RNA-seq or mass spectrometry), optionally specifically measured in activated dendritic cells or other immune cells. The probability of a specific MHC allele presenting a peptide in other individuals expressing a specific MHC allele, optionally specifically measured in activated dendritic cells or other immune cells. The probability of a peptide presented by an MHC allele in other individuals by the same family molecule (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP), optionally specifically measured in activated dendritic cells or other immune cells.

[0090] Immune tolerance escape signatures may include at least one of the following: Direct measurement of the self-peptide group by protein mass spectrometry of one or several cell types. Estimation of the self-peptide group by taking the union of all k-mer (e.g., 5-25) substrings of self-proteins. Estimation of the self-peptide group using a presentation model similar to that applied to all non-mutated self-proteins described above, optionally taking into account germline variants.

[0091] The ranking can be performed based at least in part on the numerical likelihood using a plurality of neoantigens provided by at least one model. After the ranking, a selection can be performed based on a selection criterion to select a subset of the ranked neoantigens. After the selection, a subset of ranked peptides can be provided as an output.

[0092] The number of selected sets of neoantigens can be 20.

[0093] The presentation model can represent the dependency between: the presence of a particular pair of MHC alleles and a particular amino acid at a particular position in a peptide sequence; and the likelihood that such a peptide sequence comprising a particular amino acid at that particular position will be presented on the surface of tumor cells by a particular allele in the pair of MHC alleles.

[0094] The methods disclosed herein may also include applying one or more presentation models to the peptide sequence of the corresponding neoantigen to generate a dependency score for each of the one or more MHC alleles, wherein the dependency score indicates whether the MHC allele will present the corresponding neoantigen based at least on the position of the amino acid in the peptide sequence of the corresponding neoantigen.

[0095] The methods disclosed herein may also include transforming the dependency scores to generate a corresponding independent allele probability for each MHC allele, thereby indicating the likelihood that the corresponding MHC allele will present the corresponding neoantigen; and combining the independent allele probabilities to generate a numerical likelihood.

[0096] This step of transforming the dependency scores can model the presentation of peptide sequences of corresponding neoantigens as mutually exclusive.

[0097] The methods disclosed herein may also include transforming combinations of dependency scores to produce a numerical likelihood.

[0098] This step of transforming the combination of dependency scores can model the presentation of peptide sequences of the corresponding neoantigens as the presence of interference between MHC alleles.

[0099] The set of digital possibilities can be further identified by at least one allele non-interaction feature, and the methods disclosed herein can also include applying alleles that do not interact with one of the one or more presentation models to the allele non-interaction feature to generate a dependency score for the allele non-interaction feature, wherein the dependency score indicates whether a peptide sequence of the corresponding neoantigen will be presented based on the allele non-interaction feature.

[0100] The methods disclosed herein may also include combining the dependency score for each MHC allele of the one or more MHC alleles with the dependency score of the allele non-interaction feature; transforming the combined dependency score for each MHC allele to generate a corresponding independent allele likelihood for the MHC allele, thereby indicating the likelihood that the corresponding MHC allele will present the corresponding neoantigen; and combining the independent allele probabilities to generate a numerical likelihood.

[0101] The methods disclosed herein may also include transforming the combination of the dependency score for each MHC allele and the dependency score for the non-interacting characteristic of the allele to generate a numerical likelihood.

[0102] The set of digital parameters for the presentation model can be trained based on a training data set comprising at least one set of training peptide sequences identified as being present in a plurality of samples and one or more MHC alleles associated with each training peptide sequence, wherein the training peptide sequences are identified by mass spectrometry of separated peptides eluted from MHC alleles derived from the plurality of samples.

[0103] Samples may also include cell lines engineered to express a single MHC class I or class II allele.

[0104] Samples may also include cell lines engineered to express multiple class I or class II MHC alleles.

[0105] Samples may also include human cell lines obtained or derived from multiple patients.

[0106] Samples may also include fresh or frozen tumor samples obtained from multiple patients.

[0107] Samples may also include fresh or frozen tissue samples obtained from multiple patients.

[0108] The sample may also include peptides identified using a T cell assay.

[0109] The training data set may additionally include data related to: the abundance of peptides of the training peptide set present in the sample; the length of peptides of the training peptide set in the sample.

[0110] The training data set may be generated by comparing, via alignment, a set of training peptide sequences to a database comprising a set of known protein sequences, where the set of training protein sequences is longer than and includes the training peptide sequences.

[0111] The training data set can be generated based on performing or having performed nucleotide sequencing on a cell line to obtain at least one of exome, transcriptome, or whole genome sequencing data from the cell line, wherein the sequencing data includes at least one nucleotide sequence containing a variation.

[0112] The training data set can be generated based on obtaining at least one of exome, transcriptome, or whole genome sequencing data from a normal tissue sample.

[0113] The training data set may additionally include data related to proteomic sequences associated with the sample.

[0114] The training data set may additionally include data associated with MHC peptidomimetic sequences associated with the sample.

[0115] The training data set may additionally include data relating to peptide-MHC binding affinity measurements of at least one of the isolated peptides.

[0116] The training data set may additionally include data relating to peptide-MHC binding stability measurements for at least one of the isolated peptides.

[0117] The training dataset may additionally include data related to the transcriptome associated with the sample.

[0118] The training data set may additionally include data associated with the genome associated with the sample.

[0119] The length of the training peptide sequences can be in the range of k-mers, where k is between 8-15 (inclusive) for MHC class I or between 9-30 (inclusive) for MHC class II.

[0120] The methods disclosed herein may also include encoding the peptide sequence using a one-hot encoding scheme.

[0121] The methods disclosed herein may also include encoding the training peptide sequences using a left-padded one-hot encoding scheme.

[0122] The method of treating a subject having a tumor comprises performing the steps of claim 1, and further comprises obtaining a tumor vaccine comprising a collection of selected neoantigens, and administering the tumor vaccine to the subject.

[0123] The present invention also discloses a method for manufacturing a tumor vaccine, which comprises the following steps: obtaining at least one of exome, transcriptome or whole genome tumor nucleotide sequencing data from a subject's tumor cells, wherein the tumor nucleotide sequencing data is used to obtain data representing the peptide sequence of each neoantigen in a set of neoantigens, and wherein the peptide sequence of each neoantigen comprises at least one mutation that makes it different from the corresponding wild-type parent peptide sequence; inputting the peptide sequence of each neoantigen into one or more presentation models to generate a set of digital possibilities that each of the neoantigens is presented on the tumor cell surface of the subject's tumor cells by one or more MHC alleles, the set of digital possibilities having been identified based at least on the received mass spectrometry data; and selecting a subset of the set of neoantigens based on the set of digital possibilities to generate a set of selected neoantigens; and manufacturing or producing a tumor vaccine comprising the set of selected neoantigens.

[0124] The present invention also discloses a tumor vaccine comprising a set of selected neoantigens, wherein the set of selected neoantigens is selected by performing a method comprising the following steps: obtaining at least one of exome, transcriptome or whole genome tumor nucleotide sequencing data from tumor cells of a subject, wherein the tumor nucleotide sequencing data is used to obtain data representing the peptide sequence of each neoantigen in the set of neoantigens, and wherein the peptide sequence of each neoantigen comprises at least one mutation that makes it different from the corresponding wild-type parent peptide sequence; inputting the peptide sequence of each neoantigen into one or more presentation models to generate a set of digital probabilities that each neoantigen is presented on the tumor cell surface of the subject's tumor cells by one or more MHC alleles, the set of digital probabilities having been identified based at least on received mass spectrometry data; and selecting a subset of the set of neoantigens based on the set of digital probabilities to generate a set of selected neoantigens; and manufacturing or producing a tumor vaccine comprising the set of selected neoantigens.

[0125] Tumor vaccines may comprise one or more of nucleotide sequences, polypeptide sequences, RNA, DNA, cells, plasmids or vectors.

[0126] Tumor vaccines can include one or more neoantigens presented on the surface of tumor cells.

[0127] Tumor vaccines can include one or more neoantigens that are immunogenic in a subject.

[0128] Tumor vaccines do not contain one or more neoantigens that induce an autoimmune response against normal tissues in the subject.

[0129] Tumor vaccines may contain an adjuvant.

[0130] Tumor vaccines may comprise excipients.

[0131] The methods disclosed herein may also include selecting, based on the presentation model, neo-antigens that have an increased likelihood of being presented on tumor cells relative to unselected neo-antigens.

[0132] The methods disclosed herein may also include selecting, based on the presentation model, a neo-antigen having an increased likelihood of inducing a tumor-specific immune response in a subject relative to an unselected neo-antigen.

[0133] The methods disclosed herein may also include selecting a neo-antigen having an increased probability of being presented to naive T cells by a professional antigen presenting cell (APC) relative to an unselected neo-antigen based on the presentation model, optionally wherein the APC is a dendritic cell (DC).

[0134] The methods disclosed herein may also include selecting neoantigens that have a reduced likelihood of undergoing central or peripheral tolerance suppression relative to unselected neoantigens based on the presentation model.

[0135] The methods disclosed herein may also include selecting, based on the presentation model, a neo-antigen that has a reduced likelihood of inducing an autoimmune response against normal tissue in a subject relative to an unselected neo-antigen.

[0136] Exome or transcriptome nucleotide sequence data can be obtained by sequencing tumor tissue.

[0137] The sequencing can be next generation sequencing (NGS) or any massively parallel sequencing method.

[0138] The set of digital possibilities can be further identified by at least MHC allele interaction features, which include at least one of the following: predicted affinity of the MHC allele for binding to the neoantigen encoding peptide; predicted stability of the neoantigen encoding peptide-MHC complex; sequence and length of the neoantigen encoding peptide; the probability of presentation of a neoantigen encoding peptide with a similar sequence in cells from other individuals expressing a specific MHC allele as assessed by mass spectrometry proteomics or other means; the expression level of a specific MHC allele in related subjects (e.g., measured by RNA-seq or mass spectrometry); the probability of presentation by a specific MHC allele in other different subjects expressing a specific MHC allele independent of the overall neoantigen encoding peptide sequence; the probability of presentation by an MHC allele in the same family of molecules (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other different subjects independent of the overall neoantigen encoding peptide sequence.

[0139] The set of digital likelihoods is further identified by at least MHC allele non-interacting features, including at least one of the following: C-terminal and N-terminal sequences flanking the neoantigen encoding peptide within the source protein sequence; the presence of a protease cleavage motif in the neoantigen encoding peptide, optionally weighted according to the expression of the corresponding protease in tumor cells (as measured by RNA-seq or mass spectrometry); the turnover rate of the source protein as measured in an appropriate cell type; as measured by RNA-seq or proteomic mass spectrometry, or according to annotation of germline or somatic splicing mutations detected in DNA or RNA sequence data The length of the predicted source protein, optionally taking into account the specific splice variants ("isoforms") that are most highly expressed in tumor cells; the expression level of the proteasome, immunoproteasome, thymic proteasome or other proteases in tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry or immunohistochemistry); the expression of the source gene encoding the new antigen peptide (e.g., measured by RNA-seq or mass spectrometry); the typical tissue-specific expression of the source gene encoding the new antigen peptide during each stage of the cell cycle; a comprehensive feature catalog of the source protein and / or its domains, as can be found, for example, in uniProt or PDB http: / / www.rcsb.org / pdb / home / home.dofeatures that describe the properties of the source protein domain containing the peptide, such as: secondary or tertiary structure (e.g., alpha helix vs. beta sheet); alternative splicing; the probability of presentation of the peptide from the source protein of the related neoantigen encoding peptide in other different subjects; the probability that the peptide will not be detected or over-represented by mass spectrometry due to technical bias; the expression of various gene modules / pathways that provide information about the status of tumor cells, stroma, or tumor infiltrating lymphocytes (TILs) as measured by RNASeq (without the source protein containing the peptide); the number of copies of the source gene of the neoantigen encoding peptide in tumor cells; the probability of the peptide binding to TAP or the measured or predicted value of the binding affinity of the peptide to TAP; the expression level of TAP in tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry, immunohistochemistry); the presence or absence of tumor mutations, which include but are not limited to: known cancer driver genes such as EGFR, KRAS, AL 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-DPA1, HLA-DPB1, HLA-DQ, HLA-DQA1, HLA-DQA2, HLA-DQB1, HLA-DQB2, HLA-DR, HLA-DRA, HLA-DRB1, 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 undergo loss-of-function mutations in the tumor have a reduced chance of presentation; the presence or absence of functional germline polymorphisms, including but not limited to: 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-DPA1, HLA-DPB1, HLA -DQ, HLA-DQA1, HLA-DQA2, HLA-DQB1, HLA-DQB2, HLA-DR, HLA-DRA, HLA-DRB1, HLA-DRB3, HLA-DRB4, HLA-DRB5, or any gene encoding a component of the proteasome or immunoproteasome; tumor type (e.g., NSCLC, melanoma); clinical tumor subtype (e.g., squamous lung cancer versus non-squamous lung cancer); smoking history; typical expression of the source gene of the peptide in the relevant tumor type or clinical subtype; optionally stratified by driver mutations.

[0140] The at least one mutation may comprise a frameshift or non-frameshift indel, a missense or nonsense substitution, a splice site change, a genomic rearrangement or a gene fusion, or any genomic or expression change that produces a neoORF.

[0141] The tumor cells may be 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.

[0142] The methods disclosed herein may also include obtaining a tumor vaccine comprising the selected set of neoantigens or a subset thereof, and optionally further comprising administering the tumor vaccine to a subject.

[0143] When in polypeptide form, at least one neoantigen in the selected collection of neoantigens may include at least one of the following: binding affinity to MHC with an IC50 value of less than 1000 nM; for class 1 MHC polypeptides, a length of 8-15, i.e., 8, 9, 10, 11, 12, 13, 14 or 15 amino acids; the presence of a sequence motif within or near the polypeptide in the parent protein sequence that promotes proteasomal cleavage; and the presence of a sequence motif that promotes TAP transport.

[0144] Also disclosed herein is a method for generating a model for identifying one or more new antigens that may be presented on the surface of a tumor cell of a tumor cell, the method comprising the following steps: receiving mass spectrometry data comprising data associated with a plurality of separated peptides eluted from a major histocompatibility complex (MHC) derived from a plurality of samples; obtaining a training data set by at least identifying a set of training peptide sequences present in the sample and one or more MHCs associated with each training peptide sequence; using the training data set comprising the training peptide sequences to train a set of digital parameters for a presentation model, the presentation model providing a plurality of digital probabilities that a peptide sequence from a tumor cell is presented on the surface of the tumor cell by one or more MHC alleles.

[0145] The presentation model can represent the dependency between the presence of a particular amino acid at a particular position in a peptide sequence and the likelihood that a peptide sequence comprising a particular amino acid at a particular position will be presented by one of the MHC alleles on a tumor cell.

[0146] Samples may also include cell lines engineered to express a single MHC class I or class II allele.

[0147] Samples may also include cell lines engineered to express multiple class I or class II MHC alleles.

[0148] Samples may also include human cell lines obtained or derived from multiple patients.

[0149] Samples may also include fresh or frozen tumor samples obtained from multiple patients.

[0150] The sample may also include peptides identified using a T cell assay.

[0151] The training data set may additionally include data related to: the abundance of peptides of the training peptide set present in the sample; the length of peptides of the training peptide set in the sample.

[0152] The methods disclosed herein may also include obtaining a set of training protein sequences based on the training peptide sequence by comparing the set of training peptide sequences to a database comprising a set of known protein sequences through alignment, wherein the set of training protein sequences is longer than the training peptide sequence and includes the training peptide sequence.

[0153] The methods disclosed herein may also include performing or having performed mass spectrometry on the cell line to obtain at least one of exome, transcriptome, or whole genome nucleotide sequencing data from the cell line, the nucleotide sequencing data comprising at least one protein sequence comprising a mutation.

[0154] The methods disclosed herein may also include encoding the training peptide sequences using a one-hot encoding scheme.

[0155] The method disclosed herein may also include obtaining at least one of exome, transcriptome, and whole genome normal nucleotide sequencing data from a normal tissue sample, and training a parameter set of the presentation model using the normal nucleotide sequencing data.

[0156] The training data set may additionally contain data relating to proteomic sequences associated with the sample.

[0157] The training data set may additionally contain data relating to MHC peptidomimetic sequences associated with the sample.

[0158] The training data set may additionally comprise data relating to peptide-MHC binding affinity measurements of at least one of the isolated peptides.

[0159] The training data set may additionally comprise data relating to a peptide-MHC binding stability measurement for at least one of the isolated peptides.

[0160] The training dataset may additionally contain data related to the transcriptome associated with the sample.

[0161] The training dataset may additionally contain data associated with the genome associated with the sample.

[0162] The methods disclosed herein may also include performing a logistic regression on the set of parameters.

[0163] The length of the training peptide sequences can be in the range of k-mers, where k is between 8-15 (inclusive) for MHC class I or between 9-30 (inclusive) for MHC class II.

[0164] The methods disclosed herein may also include encoding the training peptide sequences using a left-padded one-hot encoding scheme.

[0165] The method disclosed herein may also include determining the value of the parameter set using a deep learning algorithm.

[0166] Disclosed herein is a method for identifying one or more neoantigens that may be presented on the surface of a tumor cell of a tumor cell, the method comprising performing the following steps: receiving mass spectrometry data comprising data associated with a plurality of isolated peptides eluted from a major histocompatibility complex (MHC) derived from a plurality of fresh or frozen tumor samples; obtaining a training data set by at least identifying a set of training peptide sequences that are present in the tumor samples and are presented on one or more MHC alleles associated with each training peptide sequence; obtaining a set of training protein sequences based on the training peptide sequences; and using the training protein sequences and the training peptide sequences to train a set of digital parameters of a presentation model that provides a plurality of digital probabilities that a peptide sequence from a tumor cell is presented on the surface of the tumor cell by the one or more MHC alleles.

[0167] The presentation model can represent the dependency between: the presence of a particular pair of MHC alleles and a particular amino acid at a particular position in a peptide sequence; and the likelihood that such a peptide sequence comprising a particular amino acid at that particular position will be presented on the surface of tumor cells by a particular allele in the pair of MHC alleles.

[0168] The methods disclosed herein may also include selecting a subset of neoantigens, wherein the subset of neoantigens is selected because each of them has a higher probability of being presented on the surface of a tumor cell relative to one or more different tumor neoantigens.

[0169] The methods disclosed herein may also include selecting a subset of neoantigens, wherein the subset of neoantigens is selected because each of them has a higher likelihood of being able to induce a tumor-specific immune response in a subject relative to one or more different tumor neoantigens.

[0170] The methods disclosed herein may also include selecting a subset of neoantigens, wherein the subset of neoantigens is selected because each of them has a higher probability of being presented to naive T cells by a professional antigen presenting cell (APC) relative to one or more different tumor neoantigens, optionally wherein the APC is a dendritic cell (DC).

[0171] The methods disclosed herein may also include selecting a subset of neoantigens, wherein the subset of neoantigens is selected because they are each less likely to undergo central or peripheral tolerance suppression relative to one or more different tumor neoantigens.

[0172] The methods disclosed herein may also include selecting a subset of neoantigens, wherein the subset of neoantigens is selected because each of them has a lower likelihood of inducing an autoimmune response against normal tissue in a subject relative to one or more different tumor neoantigens.

[0173] The methods disclosed herein may also include selecting a subset of neoantigens, wherein the subset of neoantigens is selected because each of them is less likely to be post-translationally modified in tumor cells in a manner different from APCs, optionally wherein the APCs are dendritic cells (DCs).

[0174] Unless otherwise indicated, the practice of the methods herein will adopt conventional methods of protein chemistry, biochemistry, recombinant DNA technology and pharmacology within the technical scope of the art. These techniques are fully explained in the literature. See, for example, TECreighton, Proteins: Structures and Molecular Properties (WH Freeman and Company, 1993); A Lehninger, Biochemistry (Worth Publishers, Inc., current edition); Sambrook et al., Molecular Cloning: A Laboratory Manual (2nd edition, 1989); Methods In Enzymology (S. Colowick and N. Kaplan, ed., Academic Press, Inc.); Remington's Pharmaceutical Sciences, 18th edition (Easton, Pennsylvania: Mack Publishing Company, 1990); Carey and Sundberg Advanced Organic Chemistry 3rd edition (Plenum Press) Volumes A and B (1992).

[0175] III. Identification of Tumor-Specific Mutations in Neoantigens

[0176] 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.

[0177] If the genetic mutations in the tumor only result 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.

[0178] 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.

[0179] 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.

[0180] 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 techniques have been described, including dynamic allele-specific hybridization (DASH), microplate array diagonal gel electrophoresis (MADGE), pyrophosphate sequencing, oligonucleotide specific connection, TaqMan system and various DNA "chip" technologies, such as Affymetrix SNP chips. These methods are usually amplified by PCR target gene area. Some other methods are based on the production of small signal molecules by invasive lysis, followed by mass spectrometry or immobilized padlock probes (padlock probe) and rolling circle amplification. Several methods known in the art for detecting specific mutations are summarized below.

[0181] Detection means 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 means based on hybridization can detect multiple PCR products in samples in different ways. Other techniques that are capable of multiple analysis of multiple markers are known in the art.

[0182] Several methods have been developed to promote the analysis of single nucleotide polymorphisms in genomic DNA or cell 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 the method, a primer complementary to the allele sequence at the 3' end of the polymorphic site can hybridize with a 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 the method is that it does not need to measure a large amount of irrelevant sequence data.

[0183] Solution-based methods can be used to determine the identity of the nucleotide at the 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, a primer complementary to the allele sequence immediately 3' to the polymorphic site is used. This method uses a labeled dideoxynucleotide derivative to determine the identity of the nucleotide at the site, and if it is complementary to the nucleotide at the polymorphic site, the nucleotide will be incorporated into the primer end.

[0184] Goelet, P. et al. (PCT Application No. 92 / 15712) describe an alternative method called Genetic Bit Analysis or GBA. The method of Goelet, P. et al. uses a mixture of labeled terminators and primers complementary to the sequence at the 3' end of the polymorphic site. The combined labeled terminator is thereby determined by the nucleotide present in the polymorphic site of the target molecule being evaluated and is complementary to the nucleotide present in the polymorphic site of the target molecule being evaluated. Compared to the method of Cohen et al. (French Patent 2,650,840; PCT Application No. WO91 / 02087), the method of Goelet, P. et al. can be a heterogeneous assay in which the primers or the target molecule are fixed to a solid phase.

[0185] 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 the same nucleotide run can produce a signal proportional to the length of the run (Syvanen, A.-C. et al., Amer. J. Hum. Genet. 52:46-59 (1993)).

[0186] Many schemes 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 cover slide with the 5' end. These anchor chains perform two functions. First, if the template is configured to have a capture tail complementary to the oligonucleotide bound to the surface, it serves as a capture site for the target template chain. These anchor chains also serve as primers for template-guided primer extension, forming the basis for sequence reading. The capture primer is used as a fixed site to use multiple cycles of synthesis, detection, and dye-linker chemical cleavage to remove the dye for sequence determination. 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 is fixed on a glass slide, and each nucleotide is color-coded with an acceptor fluorescent part attached to 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.

[0187] 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 / 454Life Sciences, the 1G analyzer from Illumina / Solexa, the SOLiD system from Applied BioSystems, and the Heliscope system from Helicos Biosciences.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 (e.g., a solid support).In order to fix nucleic acid on a support, a 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.

[0188] As an alternative to a capture sequence, 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.

[0189] 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. During sequencing by synthesis, the 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 of the labeled nucleotides incorporated into the 3' end of the growing chain. This can be done in real time or in a step-by-step repeat mode. 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.

[0190] 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 current massively parallel sequencing technologies, as well as improved forms of these technologies, can be used.

[0191] All cell types or tissues 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 a 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 a normal tissue, where the normal tissue and the tumor are of different tissue types.

[0192] 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.

[0193] Alternatively, protein mass spectrometry can be used to identify or verify 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.

[0194] IV. Neoantigens

[0195] The new antigen can include a nucleotide or a polypeptide. For example, the new antigen can be an RNA sequence encoding a polypeptide sequence. Therefore, the new antigen that can be used in a vaccine includes a nucleotide sequence or a polypeptide sequence.

[0196] 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 a related polypeptide sequence.

[0197] 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, the presence of a sequence motif within or near the peptide that promotes proteasomal cleavage; and the presence of a sequence motif that promotes TAP transport.

[0198] One or more neoantigens may be presented on the surface of a tumor.

[0199] 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.

[0200] In the case of generating a vaccine for a subject suffering from a tumor, it may be considered to exclude one or more neoantigens that induce an autoimmune response in the subject.

[0201] The size of at least one new antigenic 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, 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 51, about 52, about 53, about 54, about 55, about 56, about 57, about 58, about 59, about 60, about 61, about 62, about 63, about 64, about 65, about 66, about 67, about 68, about 69, about 70, about 71, about 72, about 73, about 74, about 75 , 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 molecule residues, and any ranges derivable therefrom. In a specific embodiment, the new antigenic peptide molecule is equal to or less than 50 amino acids.

[0202] 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; 15-24 residues for MHC class II.

[0203] If necessary, longer peptides can be designed in several ways. In one case, when the probability of presentation of a peptide on an HLA allele is predicted or known, the longer peptide can be composed of any of the following: (1) an individually presented peptide extending 2-5 amino acids toward the N-terminus and C-terminus of each corresponding gene product; (2) a concatenation of some or all of the presented peptides with the respective extension sequence. In another case, when sequencing reveals the presence of longer (>10 residues) neoepitope sequences in the tumor (e.g., caused by frameshifting, read-through, or inclusion of introns that generate novel peptide sequences), the longer peptide will be composed of the following: (3) an entire stretch consisting of novel tumor-specific amino acids, thereby bypassing the need to select shorter peptides with the strongest HLA presentation based on calculations or in vitro testing. In both cases, the use of longer chains enables endogenous processing by patient cells and can result in more effective antigen presentation and induction of T cell responses.

[0204] 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.

[0205] In some aspects, the neoantigenic peptides and polypeptides do not induce an autoimmune response and / or stimulate immune tolerance when administered to a subject.

[0206] Also provided is a composition comprising at least two or more new antigenic 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 length, amino acid sequence or both of the peptides are different. These peptides are derived from any polypeptide known or found to contain tumor-specific mutations. Suitable polypeptides that can be used as a source of new antigenic 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 specific cancer types.

[0207] 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 biological activities of unmodified peptides to bind to desired MHC molecules 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 substitution means that an amino acid residue is replaced by another amino acid residue that is biologically and / or chemically similar, such as 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 probed 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, Ill., Pierce), 2nd edition (1984).

[0208] 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 determined in a conventional manner. The scheme is roughly as follows. Before use, the 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 taken out and added to a 6% trichloroacetic acid aqueous solution or ethanol. Cool the turbid reaction sample (4°C), keep it for 15 minutes, and then centrifuge it to aggregate the precipitated serum proteins. Then, by reverse phase HPLC, the presence of the peptide is determined using stability-specific chromatographic conditions.

[0209] These peptides and polypeptides can be modified to provide the desired properties except 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 capable of inducing T helper cell response. 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 a 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.

[0210] The new antigenic peptide can be directly or through a spacer at the amino or carboxyl terminus of the peptide to connect to the T helper cell. The amino terminus of the new antigenic peptide or T helper peptide can be acylated. Exemplary T helper peptides include tetanus toxoid 830-843, influenza 307-319, malaria circumsporozoite 382-398 and 378-389.

[0211] Proteins or peptides can be prepared by any technique 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 can be found in computerized databases known to those of ordinary skill in the art. One such database is the Genbank and GenPept databases of the National Center for Biotechnology Information located at the website of the National Institutes of Health. The coding region of a known gene can be amplified and / or expressed using techniques 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 known.

[0212] On the other hand, the new antigen includes a nucleic acid (e.g., a polynucleotide) encoding a new antigen 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 experiments. In general, DNA is inserted into an expression vector, such as a plasmid, with the appropriate orientation and correct expression reading frame. If necessary, 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. Then, the vector is 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

[0213] IV. Vaccine Composition

[0214] 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. The vaccine composition generally comprises a plurality of new antigens, such as those selected using the methods described herein. The vaccine composition may also be referred to as a vaccine.

[0215] 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.

[0216] In one embodiment, the selection of different peptides and / or polypeptides or their encoding nucleotide sequences enables these peptides and / or polypeptides to associate with different MHC molecules, such as different MHC class I molecules. In some aspects, a vaccine composition comprises the coding sequence of peptides and / or polypeptides that are able to associate with the most frequently occurring class I MHC molecules. Thus, the vaccine composition may comprise different fragments that are able to associate with at least 2 preferred, at least 3 preferred, or at least 4 preferred class I MHC molecules.

[0217] The vaccine composition can induce a specific cytotoxic T cell response and / or a specific helper T cell response.

[0218] 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.

[0219] An adjuvant is any substance that is mixed into a vaccine composition to increase or otherwise alter the immune response to a neoantigen. A carrier may be a scaffold structure, such as a polypeptide or polysaccharide, that is capable of associating with a neoantigen. Optionally, the adjuvant is covalently or non-covalently conjugated.

[0220] The ability of an adjuvant to increase the immune response to an antigen is usually manifested by a significant or substantial increase in an immune-mediated response, or a reduction in disease symptoms. For example, an increase in humoral immunity is usually manifested by a significant increase in the titer of antibodies produced against the antigen, and an increase in T cell activity is usually manifested by an increase in cell proliferation, or cytotoxicity, or cytokine secretion. An adjuvant can also alter the immune response, for example by changing a predominantly humoral or Th response to a predominantly cellular or Th response.

[0221] 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, ImuFact IMP321, 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 specialty 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 described previously (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. If 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. Pat. 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).

[0222] 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.

[0223] 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: CI2U), 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 a therapeutic agent and / or as an adjuvant. 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).

[0224] 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 vaccines and adjuvants may be administered together or separately in any appropriate order.

[0225] 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, serum proteins such as transferrin, bovine serum albumin, human serum albumin, thyroglobulin or ovalbumin, immunoglobulins, or hormones 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.

[0226] 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 there is a trimeric complex of peptide antigen, MHC molecules and APC, 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 in addition, it can enhance the immune response. Therefore, in some embodiments, the vaccine composition additionally contains at least one antigen presenting cell.

[0227] 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 hybrid 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 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 above-mentioned viral vector-based vaccine platform, this method can deliver one or more nucleotide sequences encoding one or more new antigenic peptides.These sequences can be flanked by non-mutated sequences, can be separated by linkers, or can be preceded by one or more sequences targeting 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). After introduction into the host, the infected cells express the neoantigens, thereby eliciting a host immune (e.g., CTL) response to the peptide. Vaccinia vectors and methods that can be used for immunization protocols are described, e.g., in U.S. Pat. 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 a variety of other vaccine vectors that can be used for therapeutic administration or immunization of new antigens, for example, Salmonella typhi vectors.

[0228] IV.A. Other Considerations Regarding Vaccine Design and Manufacturing

[0229] IV.A.1. Determination of a peptide pool covering all tumor subclones

[0230] 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

[0231] IV.A.2. Neoantigen Prioritization

[0232] 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 can remain, and there may be trade-offs between different properties of candidate vaccine neoantigens. Therefore, it can be considered to replace 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.

[0233] 1. Risk of autoimmunity or tolerance (germline risk) (lower autoimmunity risk is generally preferred).

[0234] 2. The probability of sequencing artifacts (lower probability of artifacts is generally preferred).

[0235] 3. The probability of immunogenicity (a higher probability of immunogenicity is generally preferred).

[0236] 4. Probability of presentation (higher probability of presentation is generally preferred).

[0237] 5. Gene expression (higher expression is usually preferred).

[0238] 6. HLA gene coverage (increasing the number of HLA molecules involved in presenting a new set of antigens can reduce the chance of tumors escaping immune attack through downregulation or mutation of HLA molecules).

[0239] V. Treatment and Manufacturing Methods

[0240] 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.

[0241] In some aspects, the subject is diagnosed with cancer 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 and other tissue and organ tumors; and blood tumors, such as lymphomas and leukemias, including acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia and B-cell lymphoma.

[0242] The amount of neoantigen administered should be sufficient to induce a CTL response.

[0243] The new antigens 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.

[0244] In addition, anti-immunosuppressive / immunostimulatory agents, such as checkpoint inhibitors, may also be administered to the subject. For example, anti-CTLA antibodies or anti-PD-1 or anti-PD-L1 may also be administered to the subject. Blocking CTLA-4 or PD-L1 with antibodies may enhance the immune response against cancer cells in the patient. Specifically, it has been shown that CTLA-4 is effectively blocked when a vaccination regimen is followed.

[0245] The optimal amount and optimal dosage regimen of 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, peritoneal and intravenous. Other methods of administering vaccine compositions are known to those skilled in the art.

[0246] Vaccines can be designed such that the selection, number and / or amount of neoantigens present in the composition is tissue, cancer and / or patient specific. For example, the exact selection of peptides 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 personal 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 a first round of treatment regimens.

[0247] For compositions intended for use as cancer vaccines, neoantigens with similar normal self-peptides that are expressed in large amounts 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 certain neoantigen in large amounts, the corresponding pharmaceutical composition for treating this cancer may be present in large amounts and / or may include more than one neoantigen specific for this particular neoantigen or the pathway to this neoantigen.

[0248] Compositions comprising neoantigens can be administered to individuals suffering from cancer. In therapeutic applications, the compositions are administered to patients 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 achieving this purpose is defined as a "therapeutically effective dose". The amount effective for this purpose will depend, for example, on the composition, mode of administration, stage and severity of the disease being treated, the patient's weight and general health, and the judgment of the prescribing physician. It should 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, it is possible and possible that the treating physician will feel the need to administer substantially excessive amounts of these compositions, given the need to minimize foreign matter and the relatively non-toxic nature of the neoantigens.

[0249] For therapeutic use, administration may be initiated upon detection or surgical removal of a tumor. This is followed by increasing dosages until symptoms are at least substantially alleviated and for a period of time thereafter.

[0250] 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 contain 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, etc. 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.

[0251] Neoantigens can also be administered via liposomes, targeting 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, etc. In these preparations, the neoantigens to be delivered are either alone or with molecules bound to, for example, ubiquitous receptors between lymphocytes, such as monoclonal antibodies bound to CD45 antigens, or conjugated with other therapeutic or immunogenic compositions as part of the liposomes. Therefore, liposomes filled with the desired neoantigens can be directed to the lymphocyte site, followed by liposome delivery of the selected therapeutic / immunogenic composition. Liposomes can be formed by 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 stability of liposomes in the bloodstream. There are a variety of methods that can be used to prepare liposomes, as described, for example, in Szoka et al., Ann. Rev. Biophys. Bioeng. 9; 467 (1980); U.S. Pat. Nos. 4,235,871, 4,501,728, 4,501,728, 4,837,028, and 5,019,369.

[0252] For targeting 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, topically, etc., and the dosage varies depending on the mode of administration, the peptide being delivered, and the stage of the disease being treated.

[0253] For treatment or immunization purposes, nucleic acids encoding peptides and optionally one or more peptides described herein can also be administered to patients. Nucleic acids are often delivered to patients using a variety of methods. For example, nucleic acids can be delivered directly, such as "naked DNA". This method is described in, for example, Wolff et al., Science 247: 1465-1468 (1990) and U.S. Patents No. 5,580,859 and No. 5,589,466. Nucleic acids can also be administered using, for example, the impact delivery method described in U.S. Patent No. 5,204,253 (ballistic delivery). Particles containing only DNA can be administered. Alternatively, DNA can be attached to particles, such as gold particles. Methods for delivering nucleic acid sequences can include viral vectors, mRNA vectors and DNA vectors, with or without electroporation.

[0254] 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. Pat. No. 5,279,833; Rose U.S. Pat. No. 5,279,833; 9106309WOAWO 91 / 06309; and Felgner et al., Proc. Natl. Acad. Sci. USA 84:7413-7414 (1987).

[0255] 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 generation, third generation, and / or hybrid 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 above-mentioned viral vector-based vaccine platform, this method can deliver one or more nucleotide sequences encoding one or more new antigenic peptides.These sequences can be flanked by non-mutated sequences, can be separated by linkers, or can be preceded by one or more sequences targeting 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). After introduction into the host, the infected cells express the neoantigens, thereby causing host immune (e.g., CTL) responses to the peptides. Vaccinia vectors and methods that can be used for immunization protocols are described, e.g., in U.S. Pat. No. 4,722,848. Another vector is Bacillus 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 a variety of other vaccine vectors can be used for therapeutic administration or immunization of new antigens.

[0256] The method of administering nucleic acids uses a minigene construct encoding one or more epitopes. To generate DNA sequences encoding selected CTL epitopes for expression in human cells (minigenes), the amino acid sequences of these epitopes are reverse translated. The human codon usage table is used to guide the codon selection of each amino acid. These epitope encoding DNA sequences are directly adjacent to each other to produce a continuous polypeptide sequence. In order 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 positive and negative 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.

[0257] A variety of preparations can be used to prepare purified plasmid DNA for injection. The simplest method in these methods is to rehydrate the 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 prepared 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 purified plasmid DNA to affect various variables, such as stability, intramuscular dispersion or transportation to specific organs or cell types.

[0258] Also disclosed is a method for producing a tumor vaccine, the method comprising performing the various steps of the method disclosed herein; and producing a tumor vaccine comprising a plurality of new antigens or a subset of the plurality of new antigens.

[0259] The neoantigens disclosed herein can be made using methods known in the art. For example, the method disclosed herein for producing a neoantigen or vector (e.g., a vector comprising at least one sequence encoding one or more neoantigens) may include culturing a host cell under conditions suitable for expressing the neoantigen or vector, wherein the host cell comprises at least one polynucleotide encoding the neoantigen or vector; and purifying the neoantigen or vector. Standard purification methods include chromatography, electrophoresis, immunology, precipitation, dialysis, filtration, concentration, and isoelectric focusing.

[0260] The host cell may include a Chinese hamster ovary (CHO) cell, a NS0 cell, a yeast or a HEK293 cell. The host cell may 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 may be a cDNA.

[0261] VI. Neoantigen Identification

[0262] VI.A Identification of Neoantigen Candidates.

[0263] 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 with higher sensitivity and specificity for 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.

[0264] VI.A.1. Laboratory Method Optimization

[0265] The method presented here improves upon the concepts developed for the reliable assessment of cancer driver genes in a panel of target cancers. 16 Expanding to the whole exome and whole transcriptome environments required for new antigen identification solves the challenge of discovering new antigens with high accuracy from clinical samples with low tumor content and small volume.

[0266] Specifically, these improvements include:

[0267] 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.

[0268] 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:

[0269] a. Using DNA-based capture probes and individual probe QC 17

[0270] b. Include additional baits for areas with less coverage

[0271] 3. Targeting uniform coverage across the normal exome, with <5% of bases covered at <20×, thereby leaving minimal neoantigens unclassified for somatic / germline status (and thus unavailable for use as TSNAs)

[0272] 4. In order to minimize the total amount of sequencing required, sequence capture probes should be designed to target only the gene coding region, because non-coding RNA does not produce new antigens. Other optimizations include:

[0273] a. Supplementary probes targeting HLA genes, which are GC-rich and difficult to capture by standard exome sequencing 18

[0274] b. Exclude genes predicted to produce little or no candidate neoantigens due to factors such as insufficient expression levels, suboptimal proteasomal digestion, or unusual sequence features.

[0275] 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 the DNA.

[0276] VI.A.2. NGS Data Analysis Optimization

[0277] Improvements in analytical methods address the suboptimal sensitivity and specificity of commonly used research mutation calling methods, and specifically allow for customization relevant to neoantigen identification in the clinical setting. These include:

[0278] 1. Use the HG38 reference human genome or subsequent versions for comparison, because compared with previous genome versions, this genome contains multiple MHC region assemblies and better reflects population polymorphism.

[0279] 2. By combining the results obtained by different programs 5 , overcoming the limitations of single variant callers20

[0280] a. Detect single nucleotide variants and indels in tumor DNA, tumor RNA, and normal DNA using a suite of tools that includes: programs based on comparisons of tumor to 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 useful for low-purity samples 23 .

[0281] b. Indels will be determined using programs that perform local reassembly, such as Strelka and ABRA 24 .

[0282] c. Structural rearrangements will be determined using dedicated tools such as Pindel 25 or Breakseq 26 .

[0283] 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.

[0284] 4. Extensive filtering against pseudo calls would be done, for example, by:

[0285] 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.

[0286] b. Remove variants caused by low mapping quality or low base quality 27 .

[0287] c. Remove variants that arise from recurring sequencing artifacts, even if they are not observed in the corresponding normal context 27 Examples include variants detected predominantly on one strand.

[0288] d. Remove variants detected in irrelevant control sets 27

[0289] 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 .

[0290] 6. Robust detection of neo-ORFs generated by 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). 35It 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 it makes simple positive controls unrecoverable. Coding sequence and nonsense-mediated decay potential will be determined by tools such as SpliceR 36 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 Possible filtering steps include:

[0291] a. Remove candidate neo-ORFs that are considered under-expressed.

[0292] b. Remove candidate neo-ORFs predicted to trigger nonsense-mediated decay (NMD).

[0293] 7. Candidate neoantigens (e.g., neoORFs) that are only observed in RNA and cannot be directly validated as tumor-specific antigens will be classified as likely tumor-specific based on additional parameters, such as by considering:

[0294] a. The presence of cis-acting frameshift or splice site mutations that only support tumor DNA

[0295] b. The presence of trans-acting mutations in splicing factors that are only confirmed in tumor DNA. For example, in three independently 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 were consistent.

[0296] c. For novel splice isoforms, there are confirmed “novel” splice-junction reads in the RNASeq data.

[0297] d. For new rearrangements, there are similar exon reads that are confirmed to be present in tumor DNA but not in normal DNA

[0298] e. Lack of gene expression profiles, such as GTEx 43 (i.e., making germline origin unlikely)

[0299] 8. Complement analysis based on reference genome alignments by directly comparing assembled DNA tumor to normal reads (or k-mers from these reads) to avoid errors and artifacts based on alignment and annotation (e.g., for somatic variants that arise near germline variants or repetitive sequence indels)

[0300] 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.

[0301] VI.B. Isolation and Detection of HLA Peptides

[0302] Isolation of HLA-peptide molecules is performed using classical immunoprecipitation (IP) methods after lysis of cell and tissue samples (55-58). HLA-specific IP is performed using clarified lysates.

[0303] Immunoprecipitation is performed using an antibody coupled to beads, where the antibody is specific for the HLA molecule. For pan-class I HLA immunoprecipitation, pan-class I CR antibodies are used, and for class II HLA–DR, HLA–DR antibodies are used. During an overnight incubation, the antibodies are covalently linked to NHS-agarose beads. After covalent attachment, the beads are washed and aliquoted for IP. (59,60)

[0304] The clarified tissue lysate is added to the antibody beads for immunoprecipitation. After immunoprecipitation, the beads are removed from the lysate and stored for additional experiments, including additional IPs. The IP beads are washed to remove non-specific binding and the HLA / peptide complex is eluted from the beads using standard techniques. Protein components are removed from the peptides using molecular weight spin columns or C18 fractionation. The resulting peptides are dried by SpeedVac evaporation and stored at -20°C in some cases for MS analysis.

[0305] The dried peptide is rehydrated in the HPLC buffer suitable for reverse phase chromatography and loaded on a C-18 microcapillary HPLC column to perform gradient elution in a Fusion Lumos mass spectrometer (Thermo). The MS1 spectrum of the peptide mass / charge ratio (m / z) is collected at high resolution in an Orbitrap detector, and then after the selected ion experience HCD fragmentation, the MS2 low-resolution scanning spectrum is collected in an ion trap detector. In addition, CID or ETD fragmentation methods, or any combination of the three technologies, can be used to obtain the MS2 spectrum, to reach the higher amino acid coverage of the peptide. The MS2 spectrum can also be measured with high-resolution mass accuracy in an Orbitrap detector.

[0306] The MS2 spectra resulting from each analysis were searched against a protein database using Comet (61, 62) and peptide identifications were scored using Percolator (63-65).

[0307] VI.B.1. MS detection limit studies supporting comprehensive HLA peptide sequencing.

[0308] The detection limit was determined using 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 ) for sequencing.

[0309]

[0310] VII. Presentation Model

[0311] VII.A. Systematic review

[0312] 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 storage 165.

[0313] The presentation authentication system 160 is one or more of the following Fig.14 A computer model embodied in the computing system discussed 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 in the associated MHC allele set. This is applicable in a variety of situations. One specific use case of the presentation identification system 160 is that it is capable of receiving nucleotide sequences of candidate neoantigens associated with a set of MHC alleles from a tumor cell from a patient 110 and determining the likelihood that these candidate neoantigens will be presented by one or more of the associated MHC alleles of the tumor and / or induce an immunogenic response in the immune system of the patient 110. Candidate neoantigens determined by the system 160 to have a high likelihood can be selected for inclusion in a vaccine 118, such anti-tumor immune response can be elicited by the immune system of the patient 110 providing the tumor cells.

[0314] The presentation identification system 160 determines the presentation probability through one or more presentation models. Specifically, the presentation model generates the 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 the probability of whether the peptide sequence "YVYVADVAAK" will be presented on the cell surface of the sample by a set of alleles HLA-A*02:01, HLA-B*07:02, HLA-B*08:03, HLA-C*01:04, HLA-A*06:03, HLA-B*01:04. The presentation information 165 contains information about whether the peptide binds to different types of MHC alleles so that these peptides 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 relevant MHC alleles based on the presentation information 165.

[0315] VII.B. Submission of Information

[0316] FIG. 2 illustrates a method for obtaining presentation information according to one embodiment. Presentation information 165 includes two general information categories: allele interaction information and allele non-interaction information. Allele interaction information includes information affecting presentation of peptide sequences associated with the type of MHC allele. Allele non-interaction information includes information affecting presentation of peptide sequences unrelated to the type of MHC allele.

[0317] VII.B.1. Allelic interaction information

[0318] Allele interaction information mainly includes identified peptide sequences, and it is known that these peptide sequences have been presented by one or more identified MHC molecules from people, mice, etc. It is noteworthy that this may or may not include data obtained from tumor samples. The presented peptide sequences can be identified from cells expressing a single MHC allele. In this case, the presented peptide sequences are generally collected from single allele cell lines, which are engineered to express predetermined MHC alleles and are subsequently exposed to synthetic proteins. The peptides presented on the MHC allele are separated by techniques such as acid elution and identified by mass spectrometry. Figure 2B An example of this is shown, where an example peptide YEMFNDKS presented on a predetermined MHC allele HLA-A*01:01 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 with certainty.

[0319] The peptide sequences presented can also be collected from cells expressing multiple MHC alleles. Typically, in the human body, one cell expresses 6 different types of MHC molecules. The peptide sequences presented in this way can be identified from a multi-allelic cell line engineered to express multiple predetermined MHC alleles. The peptide sequences presented in this way 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 be similarly separated by techniques such as acid elution and identified by mass spectrometry. Figure 2C An example of such a situation is shown, where six example peptides YEMFNDKSF, HROEIFSHDFJ, FJIEJFOESS, NEIOREIREI, JFKSIFEMMSJDSSU and KNFLENFIESOFI were presented to the identified MHC alleles HLA-A*01:01, HLA-A*02:01, HLA-B*07:02, HLA-B*08:01, HLA-C*01:03 and HLA-C*01:04 and separated and identified by mass spectrometry. With respect to monoallelic cell lines, the direct association between the presented peptide and the MHC protein to which it binds may be unknown because the bound peptide is separated from the MHC molecule prior to identification.

[0320] Allele interaction information can also include mass spectrometry ion currents, which depend 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 within a larger range than it.

[0321] Allelic interaction information can also include a measurement or prediction of the binding affinity between a given MHC allele and a given peptide. One or more affinity models can generate such predictions. Figure 1D In the example shown in , the presentation information 165 may include a predicted value of 1000 nM binding affinity between the peptide YEMFNDKSF and the allele HLA-A*01:01. Peptides with IC50>1000 nM are rarely presented by MHC, and lower IC50 values ​​increase the probability of presentation.

[0322] 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 2CIn the example shown in , the presented information 165 may include a stability prediction value of 1 hour for the half-life of the molecule HLA-A*01:01.

[0323] Allelic interaction information can also include measured or predicted peptide-MHC complex formation reaction rates. Complexes formed at higher rates are more likely to be presented on the cell surface at higher concentrations.

[0324] 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. A histogram of the lengths of presented peptides from several cell lines is shown in Figure 5 middle.

[0325] Allelic interaction information can also include the presence of a kinase sequence motif on the neoantigen encoding peptide, and the absence or presence of a specific post-translational modification on the neoantigen encoding peptide. The presence of a kinase motif can affect the probability of a post-translational modification that may enhance or interfere with MHC binding.

[0326] Allelic interaction information may also include proteins involved in post-translational modification processes, such as expression levels or activity levels of kinases (as measured or predicted by RNA seq, mass spectrometry or other methods).

[0327] Allelic interaction information may 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.

[0328] 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.

[0329] Allelic interaction information may also include the probability of presentation by a particular MHC allele in other individuals expressing the particular MHC allele independent of the overall neoantigen encoding peptide sequence.

[0330] Allelic interaction information can also include the probability of presentation by MHC alleles in 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 generally expressed at lower levels than HLA-A or HLA-B molecules, and it can be inferred that the probability of a peptide being presented by HLA-C is lower than that by HLA-A or HLA-B11.

[0331] Allelic interaction information may also include protein sequences for specific MHC alleles.

[0332] Any of the MHC allele non-interaction information listed in the following sections can also be modeled in the same manner as the MHC allele interaction information.

[0333] VII.B.2. Allele Non-Interaction Information

[0334] Allele-independent information may include C-terminal sequences flanking the neoantigen encoding peptide within the source protein sequence. 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.

[0335] Allele non-interaction information can also include mRNA quantitative measurements. For example, mRNA quantitative data for the same sample as the mass spectrometry training data can be obtained. Fig.13H As described, RNA expression levels are identified as strong predictors of peptide presentation. In one embodiment, mRNA quantitative measurements are identified by the software tool RSEM. Detailed implementations 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 quantification is measured in units of fragments per kilobase of transcript per million mapped reads (FPKM).

[0336] Allele-non-interaction information may also include the N-terminal sequence flanking the peptide within the source protein sequence.

[0337] 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.

[0338] 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 a dissimilar cell type, the predictive power of this feature is lower.

[0339] 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 annotations of germline or somatic splicing mutations detected in DNA or RNA sequence data, optionally taking into account specific splicing variants ("isoforms") that are most highly expressed in tumor cells.

[0340] 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.

[0341] 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.

[0342] Allele 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.

[0343] Allele-independent 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 overall lower levels (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.

[0344] Allele non-interaction information can also include, for example, uniProt or PDB http: / / www.rcsb.org / pdb / home / home.do A comprehensive catalog of source protein features provided in . These features may include, among others: secondary and tertiary structure of the protein, subcellular localization11, Gene Ontology (GO) terms. Specifically, this information may 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 may also include turn motifs, folded motifs, and disordered residues.

[0345] 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.

[0346] Allele-non-interaction information may also include features describing the presence or absence of presentation hotspots at the position of the peptide in the source protein of the peptide.

[0347] Allele-independent interaction information may also include the presentation probability of peptides from the source protein of the relevant peptide in other individuals (after adjusting for the expression level of the source protein in these individuals and the influence of different HLA types of these individuals).

[0348] Allelic non-interaction information may also include the probability that the peptide is not detected or is over-represented by mass spectrometry due to technical bias.

[0349] 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).

[0350] Allele 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 undergoes homozygous deletion in tumor cells can be assigned a presentation probability of zero.

[0351] 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 a 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.

[0352] 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, immunohistochemistry analysis). Higher TAP expression levels will increase the presentation probability of all peptides.

[0353] Allele-independent interaction information may also include the presence or absence of tumor mutations, including but not limited to:

[0354] i. Driver mutations in known cancer driver genes, such as EGFR, KRAS, ALK, RET, ROS1, TP53, CDKN2A, CDKN2B, NTRK1, NTRK2, and NTRK3

[0355] 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 chance of presentation.

[0356] The presence or absence of functional germline polymorphisms, including but not limited to:

[0357] 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-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)

[0358] Allele non-interaction information may also include tumor type (eg, NSCLC, melanoma).

[0359] 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 in https: / / www.ebi.ac.uk / ipd / imgt / hla / nomenclature / suffixes.html .

[0360] Allele non-interaction information can also include clinical tumor subtype (eg, squamous lung cancer versus non-squamous lung cancer).

[0361] Allele non-interaction information may also include smoking history.

[0362] Allele-independent interaction information may also include a history of sunburn, sun exposure, or exposure to other mutagens.

[0363] Allele-independent interaction information can also include 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.

[0364] 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 common MHC allele, or in tumors of the same type within individuals with at least one common MHC allele.

[0365] For mutated tumor-specific peptides, the feature list for predicting the probability of presentation can also include mutation annotations (e.g., missense, read-through, frameshift mutations, fusions, etc.) or predict 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 probability of presentation.

[0366] VII.C. Presentation Identification System

[0367] Figure 3 is a high-level block diagram illustrating 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 described herein. Similarly, the functional distribution of these modules may be different from the modules described herein.

[0368] VII.C.1. Data Management Module

[0369] 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, wherein each data instance i contains a set of independent variables z i , these independent variables include at least one presenting or non-presenting peptide sequence p i , one or more i Associated MHC alleles i ; and a dependent variable y i , the dependent variable represents information that the presentation identification system 160 intends to predict new values ​​of the independent variable.

[0370] In a specific embodiment described throughout the rest of this specification, the dependent variable y i It is a binary label indicating peptide p i Is it affected by one or more relevant MHC alleles a 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 spectrometer ion current of the identified data instance.

[0371] The peptide sequence p of data instance i iis a i A sequence of amino acids, where k i can vary within a certain range depending on the data instance i. For example, the range can be 8-15 for class I MHC, or 9-30 for class II MHC. In one embodiment of system 160, all peptide sequences p in a training data set are i can 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.). MHC allele a of data example i i Indicates the presence of the corresponding peptide sequence p i Related MHC alleles.

[0372] The data management module 312 may also include additional allele interaction variables, such as peptide sequences 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 i In another embodiment, the training data 170 may contain i Stability prediction values ​​for each MHC allele indicated i .

[0373] 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.

[0374] 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 presenting 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 MHC alleles of the cells. When the presentation information contains a tissue sample, the data management module 312 identifies the source protein that is the source of the presenting 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.

[0375] 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 the 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 about peptides that are not presented on the MHC alleles. Since only a small number of peptide sequences are actually presented by the MHC alleles, it is very likely that the synthetically generated peptide sequences will not be presented by the MHC alleles, even if these sequences are included in proteins processed by the cell.

[0376] Figure 4 A set of exemplary training data 170A according to one embodiment is shown. Specifically, the first three data instances in the training data 170A indicate peptide presentation information obtained from a monoallelic cell line comprising alleles HLA-C*01:03 and three peptide sequences QCEIOWARE, FIEUHFWI, and FEWRHRJTRUJR. The fourth data instance in the training data 170A indicates peptide information obtained from a multiallelic cell line comprising alleles HLA-B*07:02, HLA-C*01:03, HLA-A*01:01, and a peptide sequence QIEJOEIJE. The first data instance indicates that the peptide sequence QCEIOWARE is not presented by the allele HLA-C*01:03. As discussed in the previous two paragraphs, the peptide sequences can be randomly generated by the data management module 312, or identified from the source protein presenting the peptides. Training data 170A also includes a binding affinity prediction value of 1000 nM for peptide sequence-allele pairs and a stability prediction value of 1 hour with a half-life. Training data 170A also includes allele non-interacting variables, such as the C-terminal flanking sequence of peptide FJELFISBOSJFIE, and 10 2 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. The training data 170A also includes the binding affinity prediction value and stability prediction value for each allele, as well as the C-terminal flanking sequence of the peptide and the mRNA quantitative measurement value of the peptide.

[0377] VII.C.2. Encoding module

[0378] 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 one-hot encodes sequences (e.g., peptide sequences or C-terminal flanking 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·ki 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 0 00 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 0 0 0 0 0 0 0 0 0 0 0 0 00 0]. C-terminal flanking sequence c i , and the protein sequences of the MHC alleles d h And other sequence data in the presentation information can be encoded in a manner similar to that described above.

[0379] 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 peptide sequence with the largest length in the training data 170. Therefore, when the peptide sequence with the largest length has k max When there are 20 amino acids, the encoding module 314 digitally represents each sequence as having (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 with 3 amino acids can be represented by a row vector with 105 elements: i=[1 0 0 0 00 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 0 1 0 000 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ]. i Or other sequence data can be encoded in a manner similar to that described above. Thus, the peptide sequence p i or c i Each independent variable or column of represents the presence of a specific amino acid at a specific position in the sequence.

[0380] Although the above method for encoding sequence data is described with reference to a sequence having an amino acid sequence, the method can be similarly extended to other types of sequence data, such as DNA or RNA sequence data.

[0381] The encoding module 314 also adds one or more MHC alleles a of the data instance i i 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-C*01:03}, the alleles HLA-B*07:02 and HLA-C*01:03 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 with 4 identified MHC allele types, the number of MHC allele types may actually be in the hundreds or thousands. As discussed previously, each data instance i typically contains at most 6 different peptide sequences p. i Related MHC allele types.

[0382] The encoding module 314 also adds a label y to each data instance i i Encoded as a binary variable with values ​​from the set {0,1}, where the value 1 indicates that the peptide x i By the relevant 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 these values ​​using various functions, such as a logarithmic function with a range of [-∞,∞] for ion currents between [0,∞].

[0383] The encoding module 314 can i A pair of alleles interacting with the associated MHC allele h h i is represented as a row vector in which the digital 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 Peptide p i and the predicted binding affinity for the relevant MHC allele h, and similarly for s h i Alternatively, one or more combinations of allele interaction variables may be stored individually (eg, in individual vector or matrix form).

[0384] In one example, the encoding module 314 generates an allele interaction variable x by incorporating a measured or predicted value of binding affinity into the allele interaction variable x. h i In the figure, the binding affinity information is shown.

[0385] In one example, the encoding module 314 generates a binding stability variable x by incorporating a measured or predicted value of binding stability into the allele interaction variable x. h i In the figure, the binding stability information is shown.

[0386] In one example, the encoding module 314 generates an association variable x by incorporating a measured or predicted value of the binding association rate into the allele interaction variable x. h i The binding association rate information is represented in .

[0387] In one example, encoding module 314 represents 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.

[0388] In one example, the encoding module 314 generates an MHC allele expression level by incorporating the allele interaction variable x based on RNA-seq into the allele interaction variable x. h i The RNA expression information of MHC alleles is represented in .

[0389] Similarly, the encoding module 314 can convert the allele non-interaction variable w i is represented as a row vector in which the numeric representations of the non-interacting allele 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 Except for 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 may be stored individually (eg, in individual vector or matrix form).

[0390] 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.

[0391] 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.

[0392] 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 activation of immunoproteasome is shown in the middle.

[0393] In one example, the encoding module 314 generates the allele-independent interaction variable w by incorporating the abundance of the source protein into 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 represented (quantified in FPKM, TPM by techniques such as RSEM).

[0394] 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 This probability is expressed in .

[0395] In one example, the encoding module 314 quantifies the expression level of the 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 evaluated by RNA-seq. The mean can be incorporated into the allele non-interaction variable w i middle.

[0396] 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 .

[0397] In one example, the encoding module 314 includes a measured or predicted TAP binding affinity value (e.g., in nanomolar concentration) in the allele non-interacting variable w i TAP binding affinity is indicated in the middle.

[0398] In one example, the encoding module 314 includes the TAP expression level measured using RNA-seq (and quantified in TPM using, for example, RSEM) in the allele non-interacting variable w i TAP expression levels are shown in the middle.

[0399] In one example, the encoding module 314 generates the allele non-interacting variable w i In the , tumor mutations are represented as vectors of indicator variables (i.e., if peptide p k From samples with KRAS G12D mutation, d k =1, otherwise 0).

[0400] In one example, the encoding module 314 represents germline polymorphisms in antigen presentation 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.

[0401] In one example, the encoding module 314 represents the tumor type as a one-hot encoded vector of length one according to an alphabet 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.

[0402] In one example, the encoding module 314 represents the MHC allele suffix by treating the HLA alleles with 4 digits with different suffixes. For example, for the purpose of this model, HLA-A*24:09N is considered to be a different allele from HLA-A*24:09. Alternatively, since HLA alleles ending with the N suffix are not expressed, the presentation probability of the MHC allele with the N suffix for all peptides can be set to zero.

[0403] 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 the tumor subtype (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.

[0404] 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 i Alternatively, smoking history can be encoded as a variable of length one one-hot encoding according to an alphabetical list of smoking severity. For example, smoking status can be rated on a 1-5 scale, where 1 indicates a non-smoker and 5 indicates 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 when the patient has a smoking history and the tumor type is lung tumor, and zero otherwise.

[0405] 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.

[0406] In one example, the encoding module 314 represents the expression level distribution of a specific gene or transcript of each gene or transcript in the human genome as a summary statistic (e.g., mean, median) of the expression level distribution by 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 variable w i In addition, peptide p in melanoma measured by TCGA k The average and / or median gene or transcript expression level of the source genes or transcripts.

[0407] In one example, the encoding module 314 represents the mutation type as a variable of length one one-hot encoding according to an 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.

[0408] In one example, the encoding module 314 generates the allele non-interacting variable w i In the example, the protein level features of the protein are represented as the annotation values ​​of the source protein (e.g., 5'UTR length). In another example, the encoding module 314 is generated by the non-interacting allele variable w i Indicator variables are included to indicate the source protein annotation at the residue level, i.e., if the peptide p k is equal to 1 if it overlaps with the helical motif and 0 otherwise, or if the peptide p k is completely contained within the helical motif is equal to 1. In another example, the peptide p k The feature of the proportion of residues contained in the helical motif annotation can be included in the allele non-interacting variable w i middle.

[0409] In one example, the encoding module 314 represents the type of protein or isoform in the human proteome as an indicator vector. 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 if yes, 0 otherwise.

[0410] The encoding module 314 can also convert the peptide p i and the variable z of the associated MHC allele h i The population set of is represented as a row vector, where the allele interaction variables x iand the allele non-interaction variable w i For example, the encoding module 314 may 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.

[0411] VIII. Training Module

[0412] 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 , 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.

[0413] VIII.A. Overview

[0414] The training module 316 constructs one or more presentation models based on the training data set stored in the memory 170 generated by 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 l(y i∈S ,u i∈S ; θ) represents the dependent variable y of 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 referred to throughout the rest of this specification, the loss function (y i∈S ,u i∈S ; θ) is the negative log-likelihood function provided by equation (1a):

[0415]

[0416] However, in practice, another loss function may be used. For example, when predicting mass spectrometry ion current, the loss function is the mean square loss provided by the following equation 1b:

[0417]

[0418] 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 θ) are minimized by various parameters of the parametric presentation model determined by a gradient-based numerical optimization algorithm, such as a batch gradient algorithm, a stochastic gradient algorithm, etc. Alternatively, the presentation model may 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 set of parameters.

[0419] VIII.B. Independent Allele Model

[0420] 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.

[0421] 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:

[0422] l(y i∈S ,u i∈S ; θ) = ∑ i∈S (y i log u i +(1-y i )log(1-u i )). (2)

[0423] The peptide sequence 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 a set of parameters θ for 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 hIt 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.

[0424] Correlation 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 relevance 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 relevance score of MHC allele h may have a higher value, while if presentation is not possible, it may have a lower value. The transformation function f(·) will be input, and more precisely, in this case will be composed of g h (x h k θ h ) is transformed into an appropriate value to indicate the peptide p k The probability that it will be presented by the MHC allele.

[0425] In a specific embodiment referred to throughout the rest 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:

[0426]

[0427] 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:

[0428] f(z)=tanh(z) (5)

[0429] Alternatively, when the predicted value of the mass spectrometer ion current exceeds the range [0,1], f(·) may be any function, such as an identity function, an exponential function, a logarithmic function, and the like.

[0430] Therefore, the correlation function g for MHC allele h can be h (·) applied to the peptide sequence p k The encoded form of the peptide sequence p is generated to produce the corresponding correlation score k The probability of an independent allele being presented by an MHC allele h. The correlation score can be transformed by a transformation function f(·) to produce a peptide sequence p kIndependent allele probabilities that will be presented by MHC allele h.

[0431] VIII.B.1 Correlation functions for allelic interaction variables

[0432] In a specific embodiment mentioned throughout the present invention, the correlation function g h (·) is the affine function given by:

[0433]

[0434] This function converts x h k Each allele interaction variable in θ is associated with the measured MHC allele h. h The corresponding parameters in are linearly combined.

[0435] In another specific embodiment mentioned throughout this specification, the correlation function g h (·) is the network function provided by:

[0436]

[0437] A network model NN with 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 nodes connected to the particular node weighted by the relevant parameters mapped by the activation function associated with the particular node. Network models are advantageous over affine functions because presentation models 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 this interaction affects peptide presentation.

[0438] Generally speaking, the network model NN h (·) Can be constructed as feedforward networks, such as artificial neural networks (ANN), convolutional neural networks (CNN), deep neural networks (DNN), and / or recurrent networks, such as long short-term memory networks (LSTM), bidirectional recurrent networks, deep bidirectional recurrent networks, etc.

[0439] In one example referred to 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.

[0440] Figure 5An example network model NN3(·) associated with an arbitrary MHC allele h=3 is shown. Figure 5 As shown in , the network model NN3(·) for 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 MHC allele h=3 k (1), x3 k (2) and x3 k (3) (including the encoded polypeptide sequence data and individual data instances of any other training data used) and outputs the value NN3(x3 k ).

[0441] 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 MHC allele h. In this example, the parameter set θ h can correspond to a set of parameters for this single network model, and therefore, the parameter set θ h May be common to all MHC alleles.

[0442] Fig. 6A Example network model NN showing sharing of MHC alleles h=1,2,…,m H (·).like Fig. 6A As shown in the figure, 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 ).

[0443] In yet another example, a single network model NN H (·) can be the allele interaction variable x for a given MHC allele h h k and the protein sequence encoded by h In this case, the network model outputs a relevance score. In this case, the parameter set θ h It can also correspond to a set of parameters of this single network model, and therefore, the parameter set θ hmay be common to all MHC alleles. Thus, in such an example, NN h (·) can be expressed as given the input [x h k d h ], the single network model NN H This type of network model is advantageous because the peptide presentation probability of unknown MHC alleles in the training data can only be predicted by identifying their protein sequences.

[0444] Figure 6B Example network model NN showing MHC allele sharing H (·).like Figure 6B As shown in the figure, the network model NN H (·) receives as input the allele interaction variable and protein sequence for MHC allele h=3 and outputs the correlation score NN3(x3) corresponding to MHC allele h=3 k ).

[0445] In yet another embodiment, the correlation function g h (·) can be expressed as:

[0446]

[0447] where g' h (x h k ;θ' h ) is a set of parameters θ' h , where the deviation parameter θ of a set of parameters of the allele interaction variables of the MHC alleles is h 0 represents the baseline presentation probability of MHC allele h.

[0448] In another embodiment, the deviation parameter θ h 0 can be common to the gene family of MHC allele h. In other words, the deviation 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 MHC allele h. For example, 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 parameters θ for each of these MHC alleles are h 0 Can be shared.

[0449] Returning to equation (2), as an example, using the affine correlation 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 by the following formula:

[0450]

[0451] 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.

[0452] As another embodiment, 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 by the following formula:

[0453]

[0454] where x3 k is the allele interaction variable identified for MHC allele h=3, and θ3 is the set of parameters of the network model NN3(·) determined to be associated with MHC allele h=3.

[0455] Figure 7 The peptide p associated with MHC allele h=3 is generated using the example network model NN3(·). 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 .

[0456] VIII.B.2. Independent Alleles with Allele Non-Interacting Variables

[0457] In one embodiment, the training module 316 incorporates allele non-interaction variables and uses the following formula to transform peptide p k The estimated probability of presentation u k Modeling:

[0458]

[0459] where w k Peptide p kThe non-interacting variable of the coded alleles, g w (·) is a set of parameters 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 parameter set θ for the non-interacting variables of the alleles w The value of θ can be obtained by h and θ w The loss function σ2 is minimized to determine σ2, where i is each instance in the subset S of training data 170 generated by cells expressing a single MHC allele.

[0460] Correlation function g w (w k θ w The output of ) represents the relevance 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 correlation 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 If the expression of β-catenin is related to the C-terminal flanking sequence of the protein, it may have a lower value.

[0461] According to equation (8), 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 correlation score of the allele interaction variable k The probability of an independent allele being presented by an MHC allele h. A function of the non-interacting variable g for the alleles w (·) is also applied to the encoded form of the allele-independent interacting variable to produce a correlation score for the allele-independent interacting 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 allele probabilities that will be presented by MHC allele h.

[0462] 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 wk Included in the predicted value. Therefore, the presentation probability can be obtained as follows:

[0463]

[0464] VIII.B.3 Correlation functions for non-interacting variables of alleles

[0465] Correlation function g with the interaction variable of the relevant allele h (·) Similarly, the correlation function g for the non-interacting variables of alleles 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.

[0466] Specifically, the correlation function g w (·) is the affine function given by:

[0467] g w (w k θ w )=w k ·θ w ,

[0468] This function converts the allele non-interaction variable w k With parameter set θ w The corresponding parameters in are linearly combined.

[0469] Correlation function g w (·) can also be a network function provided by:

[0470] g h (w k θ w )=NN w (w k θ w ),

[0471] This function is composed of a set of parameters θ w The network model NN with relevant parameters in w (·)express.

[0472] In another embodiment, the correlation function g of the allele non-interacting variable is w (·) can be provided by the following formula:

[0473]

[0474] where g' w (w k ;θ' w ) is an affine function with the set of allele non-interacting parameters θ'w The network function, m k Peptide p k The quantitative measurement of mRNA, h(·) is the function that transforms the quantitative measurement, and θ w m It is a parameter in the parameter set of non-interacting variables related to alleles, which is combined with the mRNA quantitative measurement value to generate a correlation score for the mRNA quantitative measurement value. In a specific embodiment mentioned throughout the rest of this specification, h(·) is a logarithmic function, but in fact, h(·) can be any of a variety of different functions.

[0475] In yet another example, the correlation function g for the correlation of the allele non-interacting variable w (·) can be provided by the following formula:

[0476]

[0477] where g' w (w k ;θ' w ) is an affine function with allele non-interacting parameter set θ' w Network functions, etc. k The above-described peptide p k The indicator vectors of proteins and isoforms, and θ w o is a set of parameters in the parameter set for allele non-interacting variables, which is combined with the indicator vector. In one variation, when o k The dimension and parameter set θ w o When it is significantly higher, the parameter regularization term can be used when determining the parameter value, such as Add to the loss function, where ||·|| represents L1 norm, L2 norm, combination, etc. The optimal value of the hyperparameter λ can be determined by appropriate methods.

[0478] Returning to equation (8), as an example, using the affine 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 the MHC allele h=3 can be generated by the following formula:

[0479]

[0480] where w k The identified peptide p k The alleles are non-interacting variables, and θw is the set of parameters for the measured allele non-interacting variables.

[0481] As another embodiment, 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 by the following formula:

[0482]

[0483] 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.

[0484] 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 the 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 .

[0485] VIII.C. Multi-allele model

[0486] The training module 316 can also construct a presentation model in a multi-allelic environment where two or more MHC alleles are present 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, cells expressing multiple MHC alleles, or a combination thereof.

[0487] VIII.C.1. Example 1: Maximum value for the independent allele model

[0488] 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 uk The presentation probability u of each MHC allele h in the set 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)-(11). Specifically, the probability u k It can be u k h∈H In one embodiment, as shown in equation (12), 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.

[0489]

[0490] VIII.C.2. Example 2.1: Function-of-Sums Model

[0491] In one embodiment, the training module 316 trains the peptide p by the following formula k The estimated probability of presentation u k Modeling:

[0492]

[0493] The element a h k For the peptide sequence p k The associated multiple MHC alleles H are 1, and x h k Indicates the encoded peptide p k and the allele interaction variable of the corresponding MHC allele. The parameter set θ for each MHC allele h h The value of θ can be obtained by h The correlation function g is determined by minimizing the loss function of , 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 can be expressed as the correlation function g introduced in Section VIII.B.1 above h any form of .

[0494] According to equation (13), the correlation 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 allele interaction variable kThe probability of presentation by one or more MHC alleles h. The scores for each MHC allele h are 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.

[0495] 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.

[0496] For example, when using the affine transformation function g h (·) Among the m = 4 different MHC alleles identified, peptide p k The probability of being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0497]

[0498] 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.

[0499] As another example, using the network transformation function g h (·), g w (·) Among the m = 4 different MHC alleles identified, peptide p k The probability of being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0500]

[0501] Where NN2(·), NN3(·) are the network models of the identified MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters of the measured MHC alleles h=2, h=3.

[0502] Fig. 9 The peptide p associated with MHC alleles h=2, h=3 is generated using the example network models NN2(·) and NN3(·). k The possibility of submission. Fig. 9 As shown in , the network model NN2(·) receives the allele interaction variable x2 for the MHC allele h=2. kAnd 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 .

[0503] VIII.C.3. Example 2.2: Functional model using the sum of allele non-interacting variables

[0504] In one embodiment, the training module 316 incorporates allele non-interaction variables and uses the following formula to transform peptide p k The estimated probability of presentation u k Modeling:

[0505]

[0506] where w k Indicates the encoded peptide p k Specifically, the parameter set θ for each MHC allele h h and the parameter set θ for the non-interacting variables of the alleles w The value of θ can be obtained by h and θ w The correlation function g is determined by minimizing the loss function of , 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 can be expressed as the correlation function g introduced in Section VIII.B.3 above w any form of .

[0507] 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 correlation score for 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 a correlation score for the allele non-interacting variables. The scores are combined and transformed by the transformation function f(·) to produce the peptide sequence p k The presentation probability will be presented by the MHC allele H.

[0508] 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.

[0509] 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 being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0510]

[0511] 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.

[0512] As another embodiment, using the network transformation function g h (·), g w (·) Among the m = 4 different MHC alleles identified, peptide p k The probability of being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0513]

[0514] 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.

[0515] Fig.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. Fig.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(x3k ). Network model NN w (·) Receive the 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 .

[0516] 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 (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:

[0517]

[0518] VIII.C.4. Example 3.1: Model using implicit independent allele likelihood

[0519] In another embodiment, the training module 316 trains the peptide p by the following formula: k The estimated probability of presentation u k Modeling:

[0520]

[0521] The element a h k For the peptide sequence p k The associated multiple MHC allele h∈H is 1, is the implicit independent allele presentation probability of MHC allele h, and the vector v is a vector in which element v h Corresponds to , 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 parameter set θ related to the implicit independent allele likelihood can be determined by minimizing the loss function with respect to θ, where i is each instance in a subset S of the training data 170 generated by cells expressing a single MHC allele and / or cells expressing multiple MHC alleles.

[0522] The presentation probability in the presentation model of equation (17) is changed by the respective peptide p kThe implicit independent allele presentation probability of the probability that will be presented by the individual MHC allele h The implicit independent allele probability differs from the independent allele presentation probability of Section VIII.B in that the parameters for the implicit independent allele probability can be learned from a multi-allelic environment, where the direct association between the presented peptide and the corresponding MHC allele is unknown except in a mono-allelic environment. 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 of MHC allele 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.

[0523] In a specific embodiment referred to throughout the rest of this specification, r(·) is a function with a range of [0,1]. For example, r(·) can be a clipping function:

[0524] r(z)=min(max(z,0),1),

[0525] 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:

[0526] r(z)=tanh(z).

[0527] VIII.C.5. Example 3.2: Sum-of-Functions Model

[0528] In a specific embodiment, s(·) is a summation function, and the presentation probability is obtained by summing the implicit independent allele presentation probabilities:

[0529]

[0530] In one embodiment, the implicit independent allele presentation probability of MHC allele h is given by:

[0531]

[0532] The probability of presentation is estimated by the following formula:

[0533]

[0534] According to equation (19), we can get the function g by 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 correlation score of the allele interaction variable k The probability of presentation by one or more MHC alleles H. Each correlation score is first transformed by a function f(·) to produce an implicit independent allele presentation probability Independent allele probability 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 correlation function g h can be expressed as the correlation function g introduced in Section VIII.B.1 above h any form of .

[0535] For example, when using the affine transformation function g h (·) Among the m = 4 different MHC alleles identified, peptide p k The probability of being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0536]

[0537] 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.

[0538] As another embodiment, using the network transformation function g h (·), g w (·) Among the m = 4 different MHC alleles identified, peptide p k The probability of being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0539]

[0540] Where NN2(·), NN3(·) are the network models of the identified MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters of the measured MHC alleles h=2, h=3.

[0541] Fig.11 The peptide p associated with MHC alleles h=2, h=3 is generated using the example network models NN2(·) and NN3(·). k The possibility of submission. Fig. 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 ). Each output is mapped by a function f(·) to produce an estimated presentation probability u k .

[0542] In another embodiment, when predicting the logarithm of the mass spectrometer ion current, r(·) is a logarithmic function and f(·) is an exponential function.

[0543] VIII.C.6. Example 3.3: Sum model using functions of allele non-interacting variables

[0544] In one embodiment, the implicit independent allele presentation probability of MHC allele h is given by:

[0545]

[0546] This results in the following presentation probability:

[0547]

[0548] To incorporate the effects of allele-noninteracting variables on peptide presentation.

[0549] According to equation (21), we can 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 correlation score for 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 relevance scores for the allele non-interacting variables. The scores for the allele non-interacting variables are combined with the individual relevance scores for the allele interacting variables. Each combined score is transformed by a function f(·) to produce an implicit independent allele presentation likelihood. The implicit likelihoods 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 correlation function g w can be expressed as the correlation function g introduced in Section VIII.B.3 abovew any form of .

[0550] 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 being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0551]

[0552] 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.

[0553] As another embodiment, using the network transformation function g h (·), g w (·) Among the m = 4 different MHC alleles identified, peptide p k The probability of being presented by MHC alleles h=2 and h=3 can be obtained by the following formula:

[0554]

[0555] 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.

[0556] Fig.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. Fig.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 the peptide p k The allele non-interacting variables and generate the 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(x3k ), and again compare this output with the same network model NN w The output NN of (·) w (w k ) and mapped by the function f(·). The two outputs are combined to produce the estimated presentation probability u k .

[0557] In another embodiment, the implicit independent allele presentation probability of MHC allele h is given by:

[0558]

[0559] The presentation probability is thus obtained by the following formula:

[0560]

[0561] VIII.C.7. Example 4: Second-Order Model

[0562] In one embodiment, s(·) is a second order function and the peptide p k The estimated probability of presentation u k It is obtained by the following formula:

[0563]

[0564] The elements is the implicit independent allele likelihood for MHC allele h. The value of the parameter set θ for implicit independent allele likelihood can be determined by minimizing a loss function with respect to θ, where i is each instance in a 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 likelihood can be in any of the forms shown in equations (18), (20), and (22) described above.

[0565] On the one hand, the model of equation (23) can imply the presence of peptide p k The possibility that both MHC alleles will be presented simultaneously, where the presentation of the two HLA alleles is statistically independent.

[0566] According to equation (23), the peptide sequence p k The presentation probability that a peptide will be presented by one or more MHC alleles H can be obtained by combining the implicit independent allele presentation probabilities and subtracting from the sum the probability that each pair of MHC alleles will simultaneously present a peptide p k The possibility to generate a peptide sequence p k Will be generated by the presentation probability presented by the MHC allele H.

[0567] For example, when using the affine transformation function g h (·) Among the m=4 different HLA alleles identified, peptide p k The probability of being presented by HLA alleles h=2, h=3 can be obtained from the following formula:

[0568]

[0569] 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.

[0570] As another embodiment, using the network transformation function g h (·), g w (·) Among the m=4 different HLA alleles identified, peptide p k The probability of being presented by HLA alleles h=2, h=3 can be obtained from the following formula:

[0571]

[0572] Where NN2(·), NN3(·) are the network models of the identified HLA alleles h=2, h=3, and θ2, θ3 are the sets of parameters of the determined HLA alleles h=2, h=3.

[0573] IX. Example 5: Prediction Module

[0574] The prediction module 320 receives the sequence data and selects candidate neoantigens in the sequence data using the presentation model. Specifically, the sequence data may be a DNA sequence, an RNA sequence, and / or a protein sequence extracted from a patient's tumor tissue cells. The prediction module 320 processes the sequence data into a plurality of peptide sequences p having 8-15 amino acids. k For example, the prediction module 320 can process the given sequence "IEFROEIFJEF" into three peptide sequences "IEFROEIFJ", "EFROEIFJE" and "FROEIFJEF" having 9 amino acids. In one embodiment, the prediction module 320 can identify the candidate neoantigen as a mutant peptide sequence by comparing the sequence data extracted from the patient's normal tissue cells with the sequence data extracted from the patient's tumor tissue cells to identify the portion containing one or more mutations.

[0575] The presentation module 320 applies one or more presentation models to the processed peptide sequences to estimate the presentation probability of these peptide sequences. Specifically, the prediction module 320 can select one or more candidate neoantigen peptide sequences that may be presented on tumor HLA molecules by applying the presentation model to the candidate neoantigens. In one embodiment, the presentation module 320 selects candidate neoantigen sequences whose estimated presentation probability exceeds a predetermined threshold. In another embodiment, the presentation module selects N candidate neoantigen sequences with the highest estimated presentation probability (where N is generally the maximum number of epitopes that can be delivered in a vaccine). A vaccine including candidate neoantigens selected for a given patient can be injected into the patient to induce an immune response.

[0576] X. Example 6: Experimental Results Demonstrating Performance of an Example Presentation Model

[0577] The effectiveness of the various presentation models described above are tested based on test data T, which is a subset of the training data 170 that is not used to train the presentation model or an independent data set from the training data 170 having variables and data structures similar to the training data 170.

[0578] Relevant metrics that indicate the performance of the presented model are:

[0579]

[0580] It indicates the ratio of the number of peptide instances correctly predicted to be presented on the relevant HLA allele to the number of peptide instances predicted to be presented on the HLA allele. i If it is greater than or equal to a given threshold t, then the peptide p in the test data T is predicted. i will be presented on one or more relevant HLA alleles. Another relevant metric indicating the performance of the presentation model is:

[0581]

[0582] It indicates the ratio of the number of peptide instances correctly predicted to be presented on the relevant HLA allele to the number of peptide instances known to be presented on the HLA allele. Another related measure indicating the performance of the presentation model is the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. ROC plots the recall rate against the false positive rate (FPR), which is provided by the following formula:

[0583]

[0584] Comparison of the performance of XA's mass spectrometry-based presentation model with existing technology models

[0585] Fig.13A The performance results of the exemplary presentation model described herein and the prior art model for predicting peptide presentation based on multi-allelic mass spectrometry data were compared. The results showed that the exemplary presentation model performed significantly better than the prior art model based on affinity and stability prediction values ​​in predicting peptide presentation.

[0586] Specifically, using the affine correlation function g h (·) and expit function f(·), Fig.13A The example presentation model shown as "MS" is the maximum of the independent allele presentation model shown in equation (12). The example presentation model is based on a subset of monoallelic HLA-A*02:01 mass spectrometry data from the IEDB dataset (data set "D1") (data can be found at http: / / www.iedb.org / doc / mhc_ligand_full.zip ) and a subset of monoallelic HLA-B*07:02 mass spectrometry data from the IEDB dataset (data set “D2”) (data can be found at http: / / www.iedb.org / doc / mhc_ ligand_full.zip ). All peptides from the source protein in the test set that contain the presented peptide are removed from the training data, so that the example presentation model does not simply record the sequence of the presented antigen.

[0587] Fig.13A The model shown as "affinity" in is a model similar to the prior art model that predicts peptide presentation based on affinity prediction values ​​NETMHCpan. Embodiments of NETMHCpan are provided in detail in http: / / www.cbs.dtu.dk / services / NetMHCpan / . Fig.13A The model shown as "Stability" in is a model similar to the prior art model, which predicts peptide presentation based on the stability prediction value NETMHCstab. The implementation of NETMHCstab is provided in detail at http: / / www.cbs.dtu.dk / services / NetMHCstab-1.0 / The test data is a subset of the multi-allelic JY cell line HLA-A*02:01 and HLA-B*07:02 mass spectrometry data from the Bassani-Sternberg dataset (dataset “D3”) (data can be found at www.ebi.ac.uk / pride / archive / projects / PXD000394 ). Error bars (indicated by solid lines) show 95% confidence intervals.

[0588] like Fig.13A As shown in the results, the example presentation model trained on mass spectrometry data has a significantly higher PPV value at 10% recall compared to the prior art models that predict peptide presentation based on MHC binding affinity prediction values ​​or MHC binding stability prediction values. Specifically, the PPV of the example presentation model is about 14% higher than that of the model based on affinity prediction values, and its PPV is about 12% higher than that of the model based on stability prediction values.

[0589] These results demonstrate that the example presentation model has significantly better performance than prior art models that predict peptide presentation based on MHC binding affinity or MHC binding stability predictions, even though the example model was not trained on protein sequences containing presenting peptides.

[0590] Comparison of XB's T cell epitope data-based presentation model performance with existing technology models

[0591] Fig. 13B The performance results of another example presentation model as described herein are compared with the prior art model for predicting peptide presentation based on T cell epitope data. T cell epitope data contains peptide sequences presented on the cell surface by MHC alleles and recognized by T cells. The results show that even though the example presentation model is trained based on mass spectrometry data, the performance of the example presentation model in predicting T cell epitopes is still significantly better than the prior art model based on affinity and stability prediction values. In other words, Fig. 13B The results indicate that the exemplary presentation model is not only superior to the prior art model in predicting peptide presentation based on mass spectrometry test data, but also significantly superior to the prior art model in predicting epitopes actually recognized by T cells. This indicates that the various presentation models provided herein can better identify antigens that may induce an immune response in the immune system.

[0592] Specifically, using the affine correlation function g h (·) and expit function f(·), Fig. 13B The example presentation model shown as "MS" is the independent allele presentation model shown in equation (2) trained on a subset of data set D1. All peptides in the test set that contain presentation peptides from the source protein are removed from the training data so that the presentation model does not simply record the sequence of the presented antigen.

[0593] These models were applied separately to test data, which is a subset of mass spectrometry data based on HLA-A*02:01 T cell epitope data (data set "D4") (data available at www.iedb.org / doc / tcell full v3.zip). Fig. 13B The model shown as "Affinity" in is a model similar to the prior art model that predicts peptide presentation based on affinity prediction values ​​NETMHCpan, while Fig. 13B The model shown as "Stability" in is a model similar to the prior art model that predicts peptide presentation based on stability prediction values ​​NETMHCstab. Error bars (indicated by solid lines) show 95% confidence intervals.

[0594] like Fig.13AAs shown in the results, the independent allele presentation model trained on mass spectrometry data has a significantly higher PPV value at 10% recall compared to the prior art models that predict peptide presentation based on MHC binding affinity or MHC binding stability prediction values, even though the presentation model is not trained on protein sequences containing the presented peptides. Specifically, the PPV of the independent allele presentation model is about 9% higher than that of the model based on affinity prediction values, and its PPV is about 8% higher than that of the model based on stability prediction values.

[0595] These results demonstrate that example presentation models trained on mass spectrometry data perform significantly better than state-of-the-art models in predicting epitopes recognized by T cells.

[0596] Comparison of the performance of different XC presentation models based on mass spectrometry data

[0597] Fig. 13C The performance results of the example function model of and (Equation (13)), the example function model of and (Equation (19)) and the example second-order model (Equation (23)) for predicting peptide presentation based on multi-allelic mass spectrometry data are compared. The results show that the performance of the function model of and and the second-order model is better than the function model of and. This is because the function model of and implies that alleles in a multi-allelic environment interfere with each other's peptide presentation, while in fact, peptide presentation is effectively independent.

[0598] To be precise, Fig. 13C The example presented model labeled “Sigmoid functions of and ” is to use the network correlation function g h The example model labeled “sum of sigmoid functions” is a function model that uses the network correlation function g h The example model labeled “Hyperbolic Tangent” is a network correlation function g h The example model labeled "second order" is a model that uses the network correlation function g h The second-order model in equation (23) of the implicit independent allele presentation probability form shown in equation (18) of the expit function f(·) and f(·). Each model is trained on a subset of data sets D1, D2 and D3. The example presentation model is applied to test data, which is a random subset of data set D3 that does not overlap with the training data.

[0599] like Fig. 13CAs shown in , the first column refers to the AUC of ROC when each presented model is applied to the test set, the second column refers to the negative logarithm of the likelihood loss, and the third column refers to the PPV at 10% recall. Fig. 13C As shown in , the performance of the presented models “sum of sigmoid functions”, “hyperbolic tangent”, and “second order” are roughly equivalent at about 15-16% PPV at 10% recall, while the performance of the model “sigmoid function of sum” is slightly lower at about 11%.

[0600] As discussed previously in Section VIII.C.4, the results show that the "sum of sigmoid models", "hyperbolic tangent", and "second order" presentation models have higher PPV values ​​than the "sum of sigmoid functions" model because these models correctly explain how each MHC allele presents peptides independently in a multi-allelic environment.

[0601] XD presents a comparison of model performance with and without training based on monoallelic mass spectrometry data. Comparison

[0602] Fig.13D The performance results of two example presentation models trained with and without monoallelic mass spectrometry data on predicting peptide presentation with multi-allelic mass spectrometry data were compared. The results indicate that the performance of the example presentation model trained without monoallelic data is comparable to that of the example presentation model trained with monoallelic data.

[0603] The example model for "Using A2 / B7 Monoallelic Data" is to use the network correlation function g h The "sum of sigmoid functions" in equation (19) of (·), expit function f(·) and identity function r(·) presents a model. The model is trained based on a subset of the dataset D3 and monoallelic mass spectrometry data of multiple MHC alleles from the IEDB database (data available at: http: / / www.iedb.org / doc / mhc_ligand_full.zip). The example model of "not using A2 / B7 monoallelic data" is the same model, but trained based on a subset of the multi-allele D3 dataset, without monoallelic mass spectrometry data of alleles HLA-A*02:01 and HLA-B*07:02, but using monoallelic mass spectrometry data of other alleles. In the multi-allele training data, cell line HCC1937 expresses HLA-B*07:02 but not HLA-A*02:01, while cell line HCT116 expresses HLA-A*02:01 but not HLA-B*07:02. The example presentation model is applied to the test data, which is a random subset of dataset D3 and does not overlap with the training data.

[0604] The "Correlation" column refers to the correlation between the actual markers indicating whether the peptide is presented on the corresponding allele in the test data and the predictive markers. Fig.13D As shown in , the performance of the implicit independent allele presentation probability prediction based on the MHC allele HLA-A*02:01 is significantly better than the performance of the prediction based on the single allele test data of the MHC allele HLA-A*02:01 but not the MHC allele HLA-B*07:02. The MHC allele HLA-B*07:02 showed similar results.

[0605] These results indicate that the implicit allele-independent presentation probability of the presentation model can correctly predict and distinguish binding motifs from individual MHC alleles, even though the direct association between these peptides and each individual MHC allele is not known in the training data.

[0606] Comparison of independent allele prediction performance of XE without training based on single-allelic mass spectrometry data

[0607] Fig.13E Display based on Fig.13D Monoallelic mass spectrometry data for the alleles HLA-A*02:01 and HLA-B*07:02 retained in the analysis shown in, Fig.13D The performance of the example models "without A2 / B7 monoallelic data" and "with A2 / B7 monoallelic data" shown in . The results indicate that even though the example presentation model was trained without monoallelic mass spectrometry data for these two alleles, the model was still able to learn the binding motifs for each MHC allele.

[0608] like Fig.13E As shown in , "A2 model predicts B7" indicates the performance of the model when predicting peptide presentation for monoallelic HLA-B*07:02 data based on implicit independent allele presentation likelihood estimates for the relevant MHC allele HLA-A*02:01. Similarly, "A2 model predicts A2" indicates the performance of the model when predicting peptide presentation for monoallelic HLA-A*02:01 data based on implicit independent allele presentation likelihood estimates for the relevant MHC allele HLA-A*02:01. "B7 model predicts B7" indicates the performance of the model when predicting peptide presentation for monoallelic HLA-B*07:02 data based on implicit independent allele presentation likelihood estimates for the relevant MHC allele HLA-B*07:02. "B7 Model Predicts A2" indicates the performance of the model when predicting peptide presentation for the monoallelic HLA-A*02:01 data based on implicit independent allele presentation probability estimates for the MHC allele HLA-B*07:02.

[0609] like Fig.13EAs shown in , the predictive power of the implicit independent allele probability for the HLA allele is significantly higher than that for the predetermined allele and significantly lower than that for the other HLA alleles. Fig.13D Similar to the results shown in , the example presentation model correctly learned to distinguish peptide presentation for individual alleles HLA-A*02:01 and HLA-B*07:02, even though a direct association between peptide presentation and these alleles was not present in the multi-allele training data.

[0610] Frequently occurring anchor residues in XF independent allele predictions match known canonical anchor motifs

[0611] Fig.13F Show in Fig.13D The common anchor residues at positions 2 and 9 among the 9-mers predicted by the "without A2 / B7 monoallele data" example model shown in . If the estimated probability is greater than 5%, these peptides are predicted to be presented. The results show that the most common anchor residues in the peptides presented on the MHC alleles HLA-A*02:01 and HLA-B*07:02 identified match the anchor motifs previously known for these MHC alleles. This indicates that, as expected, the example presentation model correctly learns peptide binding based on the specific amino acid positions of the peptide sequence.

[0612] like Fig.13F As shown in , it is known that the amino acid L / M at position 2 and the amino acid V / L at position 9 are typical anchor residue motifs of HLA-A*02:01 (eg https: / / link.springer.com / article / 10.1186 / 1745-7580-4-2 4), and it is known that amino acid P at position 2 and amino acid L / V at position 9 are typical anchor residue motifs for HLA-B*07: 02. The most common anchor residue motifs at positions 2 and 9 of the peptide identified by the model match the typical anchor residue motifs known for the two HLA alleles.

[0613] Comparison of performance of XG presented models with and without allele non-interaction variables

[0614] Figure 13G The performance results between the example presentation model incorporating C-terminal and N-terminal flanking sequences as allele interaction variables and the example presentation model incorporating C-terminal and N-terminal flanking sequences as allele non-interaction variables were compared. The results show that incorporating C-terminal and N-terminal flanking sequences as allele non-interaction variables significantly improves model performance. More specifically, it is very meaningful to identify and model the features that are common to all different MHC alleles that are suitable for peptide presentation, thereby making the statistical strength of these allele non-interaction variables common to all MHC alleles to improve presentation model performance.

[0615] An example "allele interaction" model is a sum model using the implicit independent allele presentation probability form in Eq. (22), incorporating the C-terminal and N-terminal flanking sequences as allele interaction variables, and utilizing the network correlation function g h (·) and the expit function f(·). An example "allele non-interaction" model is the sum model of the functions shown in Eq. (21), which incorporates the C-terminal and N-terminal flanking sequences as allele non-interaction variables and utilizes the network correlation function g h (·) and expit function f(·). Allele non-interacting variables are obtained through independent network correlation function g w (·) Modeling. Both models were trained based on a subset of the dataset D3 and monoallelic mass spectrometry data of multiple MHC alleles from the IEDB database (data available at: http: / / www.iedb.org / doc / mhc_ligand_full.zip). The presented models were applied to the test dataset, which was a random subset of the dataset D3 that did not overlap with the training data.

[0616] like Figure 13G As shown in , incorporating C-terminal and N-terminal flanking sequences as allele-independent interactions in the example presentation model improved the PPV value by approximately 3% relative to modeling them as allele-independent interactions. In general, this is because the "allele-independent interactions" example presentation model is able to model the effects with minimal increase in computational power while sharing the statistical strength of allele-independent interactions across all MHC alleles by utilizing independent network correlation functions.

[0617] Correlation between XH-presenting peptides and mRNA quantification

[0618] Fig.13H The correlation between the percentage of gene presented peptides obtained by mRNA quantification based on mass spectrometry data of tumor cells is shown. The results show that there is a strong correlation between mRNA expression and peptide presentation.

[0619] To be precise, Figure 13G The horizontal axis in indicates the mRNA expression level expressed in quartiles of transcripts per million reads (TPM). Figure 13G The vertical axis in indicates the percentage of epitopes presented by genes in the corresponding mRNA expression level quartile. Each solid line is a curve related to two measurements obtained from tumor samples, which are associated with the corresponding mass spectrometry data and mRNA expression measurements. Figure 13GAs shown in , there is a strong positive correlation between mRNA expression and the percentage of peptides in the corresponding gene. Specifically, genes in the top quartile of RNA expression are more than 20 times more likely to present peptides than those in the bottom quartile. In addition, genes that were not detected by RNA essentially did not present peptides.

[0620] The results indicate that the performance of the presentation model can be substantially improved by incorporating mRNA quantification measurements, as these measurements favorably predict peptide presentation.

[0621] Comparison of XI presentation model performance with incorporation of RNA quantification data

[0622] Fig.13I The performance of two example presentation models is shown, one of which was trained on mass spectrometry tumor cell data and the other incorporated mRNA quantification data and mass spectrometry tumor cell data. Fig.13H As expected, the results indicate that, since mRNA expression is a strong indicator of peptide presentation, performance is significantly improved by incorporating mRNA quantification measurements in the exemplary presentation model.

[0623] "MHCflurry + RNA filter" is a model similar to the current state-of-the-art model for predicting peptide presentation based on affinity predictions. It is implemented using MHCflurry along with a standard gene expression filter that removes all peptides less than 3.2 FPKM in the protein using mRNA quantitative measurements. Implementation details of MHCflurry are provided in https: / / github.com / hammerlab / mhcflurry / ,and http: / / biorxiv.org / content / early / 2016 / 05 / 22 / 054775 The “Example Model, No RNA” model is a network correlation function g h (·), network correlation function g w The “sum of sigmoid functions” example presentation model shown in equation (21) of the expit function f(·) and the network correlation function g. w (·) C-terminal flanking sequences were incorporated as allele-non-interacting variables.

[0624] The “Example Model, with RNA” model is based on the network correlation function g h (·), the network correlation function g of mRNA quantitative data is incorporated into equation (10) through a logarithmic function w (·), and the expit function f(·) in equation (19) shows the “sum of sigmoid functions” presenting the model. The “example model, with RNA” model is represented by the network correlation function “g w (·) C-terminal flanking sequences were incorporated as allele-independent interacting variables and into mRNA quantification measurements via a logarithmic function.

[0625] Each model was trained based on monoallelic mass spectrometry data from the IEDB dataset, 7 cell lines from the Bassani-Sternberg dataset, and 20 mass spectrometry tumor samples. Each model was applied to a test set consisting of 5,000 provided proteins from 7 tumor samples, which constituted 9,830 presented peptides from a total of 52,156,840 peptides.

[0626] like Fig.13I As shown in the first two bars of , the PPV value of the "Example Model, No RNA" model at 20% recall is 21%, while the PPV value of the prior art model is about 3%. This indicates an initial performance improvement of 18% in PPV value, even without incorporating mRNA quantification measurements. Fig.13I As shown in the third bar graph of , the “Example Model, with RNA” model that incorporates mRNA quantitation data into the presentation model shows a PPV value of approximately 30%, an increase of nearly 10% in performance compared to the Example Presentation Model that does not utilize mRNA quantitation measurements.

[0627] Therefore, the results indicate that, as Fig.13H As expected from the findings in , mRNA expression is actually a powerful predictor of peptide prediction, which can significantly improve the performance of presentation models with little increase in computational complexity.

[0628] Example of parameters related to MHC allele HLA-C*16:04 determined by XJ

[0629] Fig.13J The results of the comparison of peptide lengths were compared. Fig.13I The described "Example Model, with RNA" presentation model produces results that compare favorably with those predicted by prior art models that do not take length into account when predicting peptide presentation. The results indicate that Fig.13I The “Example Model, with RNA” example presents a model that captures the variation in likelihood between peptides of different lengths.

[0630] The horizontal axis indicates peptide samples with lengths of 8, 9, 10, and 11. The vertical axis indicates the probability of peptide presentation depending on the peptide length. The curve "actual test data probability" shows the proportion of presented peptides in the sample test data set depending on the peptide length. The presentation probability varies with the peptide length. For example, Fig.13JAs shown in , a 10mer peptide with a typical HLA-A2 L / V anchor motif has a presentation probability that is approximately 3 times lower than a 9mer peptide with the same anchor residue. The curve "Model ignoring length" indicates the measured values ​​predicted when a prior art model that ignores peptide length is applied to the same test data set for presentation prediction. These models can be versions of NetMHC prior to version 4.0, versions of NetMHCpan prior to version 3.0, and MHCflurry, which do not take into account the variation of peptide presentation with peptide length. Fig.13J As shown in , the proportion of presented peptides is constant across different values ​​of peptide length, indicating that these models will not be able to capture the variation in peptide presentation with length. The curve "Gritstone, with RNA" indicates the measurements generated by the "Gritstone, with RNA" presentation model. Fig.13J As shown in , the measurements produced by the "Gritstone, with RNA" model closely follow the measurements shown in "Actual Test Data Probability" and correctly account for the different extents of peptide presentation at lengths 8, 9, 10, and 11.

[0631] Thus, the results show that the exemplary presentation model as presented herein not only produces improved predictions for 9mer peptides, but also improves predictions for peptides of other lengths between 8-15, which account for up to 40% of the presented peptides in class I HLA alleles.

[0632] Example of parameters related to MHC allele HLA-C*16:04 determined by XK

[0633] Shown below is a set of parameters determined for a variation of the independent allele presentation model (Equation (2)) for the MHC allele HLA-C*16:04 represented by h:

[0634]

[0635] where relu(·) is the rectified linear unit (RELU) function, and W h 1 、b h 1 , W h 2 and b h 2 is the set of parameters θ of the model that are determined. The allele interaction variable x h k It is composed of a peptide sequence. h 1 The dimension of b is (231×256), h 1 The dimension of W is (1×256), h 2 The dimension of b is (256×1), and bh 2 is a scalar. For the purpose of verification, b h 1 、b h 2 , W h 1 and W h 2 The values ​​are listed below.

[0636] b h 1 :

[0637] [-1.25887644-0.84448904-1.71391594-1.38860381-1.15529644-2.42168117-1.98687959-0.8488462-1.6607399-1.12956274-2.08983159-0.53710765-0.49313864-1.15045631-0.48757577-0.55577797-0.31031775-0.70026076-1.04614675-1.3850919-1.82895124-1.15633833-1.2794342-2.40924239-1.38725305-1.98276925-1.45160246-0.81624526-0.59441668-1.86188185-0.83900708-1.33677316-1.68388879-1.1717515-1.27319682-1.05872869-1.34773123-1.61631954-0.82552254-1.64032412-1.26584375-0.56390315-1.00684011-1.16896808-2.27648878-0.66619354-2.08268309-0.90704113-1.17440355-1.06266129-1.14868402-0.98686731-1.1801722-1.41404176-0.92722374-0.61310995-1.69453728-1.26077592-1.80019915-0.55121636-1.42537642-2.1196804-1.32291269-1.30997157-0.8173914-1.23448598-0.92230183-1.04590237-0.68719882-0.99965096-1.85294855-0.67477435-1.31483507-1.21778536-1.18890083-0.80372357-1.90210617-1.85402811-1.05570829-1.345541-0.68872309-1.91153145-1.15126705-0.9407478-1.01246905-1.36009204-1.15897787-1.38631177-0.87737125-0.81053489-0.5077101-0.36021063-1.06372654-1.03310466-1.15494275-2.19268966-0.80285913-0.79190463-0.93620723-1.30614185-1.85568225-2.38452196-1.67495275-0.70025575-0.98393178-1.83200908-1.07078218-0.98362595-0.71674275-2.04092884-0.81029147-0.92036235-1.75404763-1.45618188-0.87118417-0.54004192-1.65154266-1.28891158-1.98474801-1.20875657-1.51076365-1.44827867-1.60024345-1.5518645-1.22961164-1.33044851-1.78681803-1.60184741-1.70247972-1.48592603-1.19963896-2.06061363-1.59096014-1.18787074-1.19529891-0.61437321-0.92168951-0.91861475-1.55775297-1.57438934-1.76918828-2.1351223-1.28325438-1.17691207-0.9070273-0.92114311-1.37144518-1.05090868-1.00942171-1.12824321-0.94788575-1.27639067-1.39774501-1.06832922-1.03379047-1.49880815-1.23712206-1.15053117-1.34290957-1.66009867-0.92846054-0.97625399-0.77744013-1.93288314-1.11750412-0.97511715-0.74683744-1.13343358-1.20803583-1.436432-1.50354922-1.00136673-2.01420403-1.2733649-0.92629886-1.29439116-1.64390576-1.23867738-0.88478297-0.85971212-1.85386622-1.53856933-2.06025767-1.88411045-1.49906313-1.32475007-1.117347-1.68465662-0.69363773-1.5894047-1.97358358-2.59916759-1.35396743-2.30461049-0.99292755-1.27674675-0.86507457-0.94777668-1.35009873-1.47957981-0.66513908-1.06404948-1.06712294-1.72618425-1.63874471-1.46202075-0.89948207-1.20573676-1.24084079-1.19317305-0.93100727-1.02617919-0.76241934-1.87542462-1.03179228-1.45539415-1.36384034-0.96758604-0.86005205-0.80283695-1.1137774-1.53601909-1.13732922-1.12731791-1.23490679-1.33152759-1.02586186-1.44582832-1.92057085-1.30637109-2.32667851-1.57416546-1.30796921-0.70605123-2.00818539-1.46872818-0.92903972-1.57068658-1.23099935-1.05193675-1.35168147-0.52318192-0.8894254-1.69644272-1.18933356-1.11279356-1.06844199-1.36696768-1.31789732-1.63881576-0.56565195-0.52174371-0.93638539-2.07027268-1.60358965-1.22863901].

[0638] b h 2 :

[0639] -2.88706302643

[0640] W h 2 :

[0641] [-1.35359335-0.578529-0.13680433 0.17503858-0.20488028-0.54827738-0.2597248-1.73813021-1.02167261 0.09387285-0.63083726-0.40662372-0.47041351-0.32108155 0.11540248-0.10392507-0.27698821 0.03336413-0.72273839-0.53308272-0.25452602-0.68693435-0.11388378-1.02962244-1.22807848-0.56124699-0.56379735-0.18465492 0.14469336-0.41948733-1.01185989-1.1557889-0.70585167-0.54171222-0.14044708-0.4110463-0.90434784-0.61688143-0.99438584-1.55196273-0.94430661-0.39086425-0.99821037 0.33283517-0.69542056 0.14739829-0.47792649 0.27000278-0.64110023-1.90988696-0.37607646 0.11039938-0.7492047-0.17413628-1.15171742-0.68284678-0.06158932-1.04064727-0.7101987-0.86431879-0.67057228-0.682955680.08244683-0.86153275-0.31266105-0.21919173-0.79513979 0.17517358-0.29372135-1.68132675-0.6964252-0.479926850.00768639-0.3944906-0.99549603-1.29167581-0.82665157-0.661383750.14107071-0.89010292-0.49536058-0.90695-0.84300721-0.85446638-0.99457145-0.84283727-1.26285946-0.7482127-1.32341206-0.14178833-0.47957143 0.22300801 0.22044657-0.07665028 0.19888243-0.686866880.090933250.20991775-0.47505447-1.29607451-0.79738855-0.6163758-0.25245398-0.24964713-0.70785236-0.14511365 0.23526534-1.37460887-0.42035979-0.76954895 0.01340491-0.23353948-1.45486987-2.102533820.31412357 0.05441735-1.16719246-0.53678679-1.21121192 0.36880198-1.7501055-1.08759594-1.03163946-0.87725466-1.04990077-0.23935798-0.70022678-0.30579087-1.51719499-0.05505106-0.30621436-0.37509263-0.35138479 0.08471824-0.30281609-0.40951991-0.88454992 0.045063550.12505099-0.79208314-0.98382056-0.73998731-0.68265402-0.30925721-0.30487028-0.18522757-0.51589108-0.14071934-0.58638161-0.37126878-0.36587363-1.98553813-0.52241606-0.33231446-1.05311215-1.14458430.1126269-0.18052928-0.87758267-1.06622291-1.76154435-0.16173303-1.38007092-0.67007738-0.95604581-0.71263856 0.04040499-0.847063240.3145974-0.89040732-0.75084466-2.24529719-0.90706474 0.08718969-2.26292086-0.43557408-0.16575792-1.28786123-0.76409894-1.2387414-1.00480986-1.59728515-1.76675069-0.45098865 0.37770402-2.141566510.17127-1.07672346-0.63889885-1.85108674 0.09201332-0.64665085-1.47701621-0.27428123 0.11656716-0.71997839 0.25774828-0.65556616-0.12599011-1.19412673 0.05263189-0.250155930.15155405-0.71724749-1.8973484-0.01961765-0.85265714-0.54098094-1.15418613-0.55546618-0.95463139-1.48755825-1.03432381-0.82399344-0.91085857 0.0421642-0.55967122-0.70061207 0.24029407-0.97962326-0.75661993-0.6151405-1.64357328-1.16723275-0.04848668-0.62112832-0.91200793-0.5882951-0.50034207 0.01658128-0.58651829-0.12295453-0.13033544 0.07790214-0.678482-0.62363708 0.08507859-0.94984204 0.00797078-1.039689780.01618595-0.11407378 0.2762318-0.7622599-1.18186867-0.714379310.25219718-1.02577722-0.3844451-0.210338-0.34110329-1.00708425-0.23432316-0.29608929 0.1013524-0.40958044].

[0642] XI. Sample Computer

[0643] Fig.14An example computer 1400 for implementing the entities shown in Figures 1 and 3 is shown. The computer 1400 includes at least one processor 1402 coupled to a chipset 1404. The chipset 1404 includes a memory controller hub 1420 and an input / output (I / O) controller hub 1422. The memory 1406 and the graphics adapter 1412 are coupled to the memory controller hub 1420, and the display 1418 is coupled to the graphics adapter 1412. The storage device 1408, the input device 1414, and the network adapter 1416 are coupled to the I / O controller hub 1422. Other embodiments of the computer 1400 have different architectures.

[0644] Storage device 1408 is a non-transitory computer-readable storage medium, such as a hard drive, a compact disk read-only memory (CD-ROM), a DVD, or a solid-state memory device. Memory 1406 stores instructions and data used by processor 1402. Input interface 1414 is a touch screen interface, a mouse, a trackball or other type of pointing device, a keyboard, or some combination thereof, and is used to enter data into computer 1400. In some embodiments, computer 1400 can be configured to receive input (e.g., commands) from input interface 1414 through gestures of a user. Graphics adapter 1412 displays images and other information on display 1418. Network adapter 1416 couples computer 1400 to one or more computer networks.

[0645] The computer 1400 is adapted to execute computer program modules to provide the functions described herein. As used herein, the term "module" refers to computer program logic for providing a specified function. Therefore, a module can be implemented in hardware, firmware and / or software. In one embodiment, the program module is stored on a storage device 1408, loaded into the memory 1406 and executed by the processor 1402.

[0646] The type of computer 1400 used by the entity of Figure 1 can vary depending on the implementation and the processing power required by the entity. For example, the presentation authentication system 160 can be run in a single computer 1400 or in multiple computers 1400 that communicate with each other through a network, such as in a server farm. The computer 1400 may lack some of the components described above, such as a graphics adapter 1412 and a display 1418.

[0647] The present invention also includes the following embodiments:

[0648] Embodiment 1. A method for identifying one or more new antigens that may be presented on the surface of tumor cells from a subject, comprising the following steps:

[0649] Obtaining at least one of exome, transcriptome, or whole genome tumor nucleotide sequencing data from a tumor cell of the subject, wherein the tumor nucleotide sequencing data is used to obtain data representing a peptide sequence of each neo-antigen in the set of neo-antigens, and wherein the peptide sequence of each neo-antigen comprises at least one change that makes it different from a corresponding wild-type parent peptide sequence;

[0650] inputting the peptide sequence of each neo-antigen into one or more presentation models to generate a set of digital probabilities that each of the neo-antigens is presented by one or more MHC alleles on the tumor cell surface of the tumor cell of the subject, the set of digital probabilities having been identified based at least on the received mass spectrometry data; and

[0651] Based on the set of numerical possibilities, a subset of the set of neo-antigens is selected to produce a set of selected neo-antigens.

[0652] Embodiment 2. A method according to embodiment 1, wherein the number of selected neoantigens in the collection is 20.

[0653] Embodiment 3. The method according to any one of embodiments 1 to 2, wherein the presentation model represents the dependencies between:

[0654] The presence of a specific allele in the MHC allele and a specific amino acid at a specific position in the peptide sequence; and

[0655] The likelihood that such a peptide sequence comprising said particular amino acid at said position is presented on the surface of said tumor cell by a particular allele of said pair of MHC alleles.

[0656] Embodiment 4. The method according to any one of embodiments 1 to 3, wherein inputting the peptide sequence comprises:

[0657] The one or more presentation models are applied to the peptide sequence of the corresponding neoantigen to generate a dependency score for each of the one or more MHC alleles based at least on the amino acid position of the peptide sequence of the corresponding neoantigen, wherein the dependency score indicates whether the MHC allele will present the corresponding neoantigen.

[0658] Embodiment 5. The method according to embodiment 4, further comprising:

[0659] transforming the dependency scores to obtain a corresponding independent allele likelihood for each MHC allele, thereby indicating the likelihood that the corresponding MHC allele will present the corresponding neoantigen; and

[0660] The independent allele probabilities are combined to generate the numerical likelihood.

[0661] Embodiment 6. A method according to embodiment 5, wherein the transformation of the dependency score models the presentation of the peptide sequences of the corresponding neoantigens as mutually exclusive.

[0662] Embodiment 7. The method according to any one of embodiments 4 to 6, further comprising:

[0663] Combinations of the dependency scores are transformed to produce the numerical likelihood.

[0664] Embodiment 8. A method according to embodiment 7, wherein the combination of transforming the dependency scores models the presentation of the peptide sequence of the corresponding neoantigen as the presence of interference between MHC alleles.

[0665] Embodiment 9. The method of any one of embodiments 4 to 8, wherein the set of numerical possibilities is further identified by at least one allele non-interaction feature, and further comprising:

[0666] Alleles that do not interact with one of the one or more presentation models are 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.

[0667] Embodiment 10. The method according to embodiment 9, further comprising:

[0668] 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;

[0669] transforming the combined dependency scores for each MHC allele to generate a corresponding independent allele likelihood for the MHC allele, thereby indicating a likelihood that the corresponding MHC allele will present the corresponding neoantigen; and

[0670] The independent allele probabilities are combined to produce the numerical likelihood.

[0671] Embodiment 11. The method according to any one of embodiments 9 to 10, further comprising:

[0672] The combination of the dependency scores for each of the MHC alleles and the dependency scores for the allele non-interacting features is transformed to produce the numerical likelihood.

[0673] Embodiment 12. A method according to any one of embodiments 1 to 11, wherein the set of digital parameters of the presentation model is trained based on a training data set, the training data set comprising at least one set of training peptide sequences identified as being present in multiple samples and one or more MHC alleles associated with each training peptide sequence, wherein the training peptide sequences are identified by mass spectrometry of separated peptides eluted from MHC alleles derived from the multiple samples.

[0674] Embodiment 13. A method according to Embodiment 12, wherein the training data set also includes data on the mRNA expression levels of the tumor cells.

[0675] Embodiment 14. A method according to any one of Embodiments 12 to 13, wherein the sample comprises a cell line engineered to express a single MHC class I or class II allele.

[0676] Embodiment 15. A method according to any one of Embodiments 12 to 14, wherein the sample comprises a cell line engineered to express multiple class I or class II MHC alleles.

[0677] Embodiment 16. A method according to any one of Embodiments 12 to 15, wherein the sample comprises human cell lines obtained or derived from multiple patients.

[0678] Embodiment 17. A method according to any one of Embodiments 12 to 16, wherein the sample comprises fresh or frozen tumor samples obtained from multiple patients.

[0679] Embodiment 18. A method according to any one of Embodiments 12 to 17, wherein the sample comprises fresh or frozen tissue samples obtained from multiple patients.

[0680] Embodiment 19. A method according to any one of Embodiments 12 to 18, wherein the sample comprises peptides identified using a T cell assay.

[0681] Embodiment 20. The method according to any one of embodiments 12 to 19, wherein the training dataset further comprises data related to:

[0682] the abundance of peptides of the training peptide set present in the sample;

[0683] The peptide lengths of the training peptide set present in the sample.

[0684] Embodiment 21. A method according to any one of embodiments 12 to 20, wherein the training data set is generated by comparing a set of training peptide sequences with a database containing a set of known protein sequences via alignment, wherein the set of training protein sequences is longer than and includes the training peptide sequences.

[0685] Embodiment 22. A method according to any one of embodiments 12 to 21, wherein the training data set is generated based on performing or having performed mass spectrometry on a cell line to obtain at least one of exome, transcriptome or whole genome sequencing data from the cell line, and the peptide sequencing data includes at least one protein sequence containing changes.

[0686] Embodiment 23. A method according to any one of embodiments 12 to 22, wherein the training data set is generated based on obtaining at least one of exome, transcriptome or whole genome normal nucleotide sequencing data from a normal tissue sample.

[0687] Embodiment 24. A method according to any one of embodiments 12 to 23, wherein the training dataset further comprises data related to proteomic sequences associated with the sample.

[0688] Embodiment 25. A method according to any one of embodiments 12 to 24, wherein the training dataset further comprises data related to MHC peptidomimetic sequences associated with the sample.

[0689] Embodiment 26. A method according to any one of Embodiments 12 to 25, wherein the training dataset further comprises data related to peptide-MHC binding affinity measurements of at least one of the isolated peptides.

[0690] Embodiment 27. A method according to any one of Embodiments 12 to 26, wherein the training dataset further comprises data related to peptide-MHC binding stability measurements of at least one of the isolated peptides.

[0691] Embodiment 28. A method according to any one of Embodiments 12 to 27, wherein the training dataset further comprises data related to the transcriptome associated with the sample.

[0692] Embodiment 29. A method according to any one of embodiments 12 to 28, wherein the training data set also includes data related to the genome associated with the sample.

[0693] Embodiment 30. A method according to any one of embodiments 12 to 29, wherein the length of the training peptide sequence is in the range of k-mers, where k is between 8-15 (including the endpoints).

[0694] Embodiment 31. A method according to any one of embodiments 12 to 30, further comprising encoding the peptide sequence using a one-hot encoding scheme.

[0695] Embodiment 32. A method according to embodiment 31, which also includes encoding the training peptide sequence using a left-filled one-hot encoding scheme.

[0696] Embodiment 33. A method for treating a subject having a tumor, comprising performing any one of the steps described in Embodiments 1 to 32, and further comprising obtaining a tumor vaccine comprising a collection of selected neoantigens, and administering the tumor vaccine to the subject.

[0697] Embodiment 34. A method for manufacturing a tumor vaccine, comprising performing any one of the steps described in embodiments 1 to 33, and further comprising producing or having produced a tumor vaccine comprising the collection of selected new antigens.

[0698] Embodiment 35. A tumor vaccine comprising a collection of selected neoantigens according to any one of Embodiments 1 to 32, wherein the collection of neoantigens is selected by performing the method according to any one of Embodiments 1 to 32.

[0699] Embodiment 36. A vaccine according to embodiment 35, wherein the tumor vaccine comprises one or more of a nucleotide sequence, a polypeptide sequence, RNA, DNA, a cell, a plasmid or a vector.

[0700] Embodiment 37. The vaccine of any one of Embodiments 35 to 36, wherein the tumor vaccine comprises one or more neoantigens presented on the surface of tumor cells.

[0701] Embodiment 38. The vaccine of any one of Embodiments 35 to 37, wherein the tumor vaccine comprises one or more neoantigens that are immunogenic in a subject.

[0702] Embodiment 39. The vaccine of any one of Embodiments 35 to 38, wherein the tumor vaccine does not comprise one or more neoantigens that induce an autoimmune response against normal tissues in the subject.

[0703] Embodiment 40. The vaccine of any one of Embodiments 35 to 39, wherein the tumor vaccine further comprises an adjuvant.

[0704] Embodiment 41. The vaccine of any one of Embodiments 35 to 40, wherein the tumor vaccine further comprises an excipient.

[0705] Embodiment 42. A method according to any one of Embodiments 1 to 41, wherein selecting the set of selected neo-antigens comprises selecting neo-antigens that have an increased likelihood of being presented on the surface of the tumor cells relative to unselected neo-antigens based on the presentation model.

[0706] Embodiment 43. A method according to any one of Embodiments 1 to 42, wherein selecting the set of selected neoantigens comprises selecting neoantigens that have an increased likelihood of inducing a tumor-specific immune response in the subject relative to unselected neoantigens based on the presentation model.

[0707] Embodiment 44. A method according to any one of Embodiments 1 to 43, wherein selecting the set of selected neo-antigens comprises selecting neo-antigens that have an increased likelihood of being presented to natural 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).

[0708] Embodiment 45. A method according to any one of Embodiments 1 to 44, wherein selecting the set of selected neoantigens comprises selecting neoantigens that have a reduced likelihood of undergoing central or peripheral tolerance suppression relative to unselected neoantigens based on the presentation model.

[0709] Embodiment 46. A method according to any one of Embodiments 1 to 45, 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 presentation model.

[0710] Embodiment 47. A method according to any one of Embodiments 1 to 46, wherein the exome or transcriptome nucleotide sequencing data is obtained by sequencing the tumor tissue.

[0711] Embodiment 48. A method according to any one of embodiments 1 to 47, wherein the sequencing is next generation sequencing (NGS) or any massively parallel sequencing method.

[0712] Embodiment 49. The method of any one of Embodiments 1 to 48, wherein the set of numerical possibilities is further identified by at least an MHC allele interaction feature, the feature comprising at least one of:

[0713] a. predicted affinity of the MHC allele for binding to the peptide encoded by the neoantigen;

[0714] b. predicted stability of the neoantigen encoding peptide-MHC complex;

[0715] c. the sequence and length of the neoantigen encoding peptide;

[0716] d. the probability of presentation of a neoantigen-encoding peptide of similar sequence in cells from other individuals expressing a particular MHC allele, as assessed by mass spectrometry proteomics or other means;

[0717] e. the expression level of said particular MHC allele in the relevant subject (e.g., as measured by RNA-seq or mass spectrometry);

[0718] f. the probability of presentation by the specific MHC allele in other different subjects expressing the specific MHC allele independent of the overall neoantigen encoding peptide sequence;

[0719] g. The probability of presentation by MHC alleles in the same family of molecules (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other different subjects independent of the overall neoantigen encoding peptide sequence.

[0720] Embodiment 50. The method of any one of Embodiments 1 to 49, wherein the set of numerical possibilities is further identified by at least an MHC allele non-interaction feature, the feature comprising at least one of:

[0721] a. The C-terminal and N-terminal sequences of the new antigen encoding peptide are flanked within its source protein sequence;

[0722] b. the presence of a protease cleavage motif in the neoantigen encoding peptide, optionally weighted according to the expression of the corresponding protease in the tumor cells (as measured by RNA-seq or mass spectrometry);

[0723] c. the turnover rate of the source protein as measured in an appropriate cell type;

[0724] d. the length of the source protein as measured by RNA-seq or proteomic mass spectrometry, or as predicted based on annotations 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 the tumor cells;

[0725] e. the expression level of proteasome, immunoproteasome, thymic proteasome or other proteases in the tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry or immunohistochemistry);

[0726] f. expression of the source gene encoding the neoantigen peptide (e.g., measured by RNA-seq or mass spectrometry);

[0727] g. Typical tissue-specific expression of the source gene encoding the neoantigen peptide during various stages of the cell cycle;

[0728] h. a comprehensive catalog of features of the source protein and / or its domains, as can be found, for example, in uniProt or PDB http: / / www.rcsb.org / pdb / home / home.do;

[0729] i. Features that describe the properties of the domain of the source protein containing the peptide, for example: secondary or tertiary structure (e.g., alpha helix versus beta sheet); alternative splicing;

[0730] j. The presentation probability of peptides from the source protein of the related neoantigen encoding peptides in other different subjects;

[0731] k. The probability that the peptide is not detected or over-represented by mass spectrometry due to technical bias;

[0732] 1. Expression of various gene modules / pathways that provide information about the status of the tumor cells, stroma or tumor infiltrating lymphocytes (TILs) measured by RNASeq (without the need for the source protein containing the peptide);

[0733] m. the copy number of the source gene encoding the neoantigen peptide in the tumor cell;

[0734] n. a probability of the peptide binding to the TAP or a measured value or predicted value of the binding affinity of the peptide to the TAP;

[0735] o. The expression level of TAP in the tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry, immunohistochemistry);

[0736] p. The presence or absence of tumor mutations, including but not limited to:

[0737] i. Driver mutations in known cancer driver genes, such as EGFR, KRAS, ALK, RET, ROS1, TP53, CDKN2A, CDKN2B, NTRK1, NTRK2, and NTRK3

[0738] 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;

[0739] q. The presence or absence of functional germline polymorphisms, including but not limited to:

[0740] 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-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)

[0741] r. Tumor type (e.g., NSCLC, melanoma);

[0742] s. Clinical tumor subtype (e.g., squamous versus non-squamous lung cancer);

[0743] t. Smoking history;

[0744] u. Typical expression of the source gene of the peptide in relevant tumor types or clinical subtypes, optionally stratified by driver gene mutations.

[0745] Embodiment 51. The method of any one of Embodiments 1 to 50, wherein the at least one mutation is a frameshift or non-frameshift indel, a missense or nonsense substitution, a splice site change, a genomic rearrangement or a gene fusion, or any genomic or expression change that produces a neoORF.

[0746] Embodiment 52. A method according to any one of Embodiments 1 to 51, wherein the tumor cells are selected from the group consisting of lung cancer, melanoma, breast cancer, ovarian cancer, prostate cancer, kidney cancer, gastric 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.

[0747] Embodiment 53. A method according to any one of Embodiments 1 to 52, further comprising obtaining a tumor vaccine comprising the selected set of neoantigens or a subset thereof, and optionally further comprising administering the tumor vaccine to the subject.

[0748] Embodiment 54. A method according to any one of embodiments 1 to 53, wherein at least one neoantigen in the selected collection of neoantigens comprises at least one of the following when in polypeptide form: a binding affinity to MHC with an IC50 value of less than 1000 nM; a length of 8-15, i.e., 8, 9, 10, 11, 12, 13, 14 or 15 amino acids for class 1 MHC polypeptides; the presence of a sequence motif within or near the polypeptide in the parent protein sequence that promotes proteasomal cleavage; and the presence of a sequence motif that promotes TAP transport.

[0749] Embodiment 55. A method of generating a model for identifying one or more neoantigens that may be presented on the tumor cell surface of a tumor cell, comprising performing the following steps:

[0750] receiving mass spectrometry data comprising data associated with a plurality of separated peptides eluted from a major histocompatibility complex (MHC) derived from a plurality of samples;

[0751] obtaining a training data set by at least identifying a set of training peptide sequences present in the sample and one or more MHCs associated with each training peptide sequence;

[0752] The training data set comprising the training peptide sequences is used to train a set of numerical parameters of a presentation model, the presentation model providing a plurality of numerical probabilities that a peptide sequence from the tumor cell is presented on the surface of the tumor cell by one or more MHC alleles.

[0753] Embodiment 56. A method according to embodiment 55, wherein the presentation model represents the dependency between:

[0754] The presence of a specific amino acid at a specific position in the peptide sequence; and

[0755] The likelihood that a peptide sequence containing said specific amino acid at said specific position will be presented on said tumor cell by one of said MHC alleles.

[0756] Embodiment 57. A method according to any one of Embodiments 55 to 56, wherein the sample comprises a cell line engineered to express a single MHC class I or class II allele.

[0757] Embodiment 58. A method according to any one of Embodiments 55 to 57, wherein the sample comprises a cell line engineered to express multiple class I or class II MHC alleles.

[0758] Embodiment 59. A method according to any one of Embodiments 55 to 58, wherein the sample comprises human cell lines obtained or derived from multiple patients.

[0759] Embodiment 60. A method according to any one of Embodiments 55 to 59, wherein the sample comprises fresh or frozen tumor samples obtained from multiple patients.

[0760] Embodiment 61. A method according to any one of Embodiments 55 to 60, wherein the sample comprises peptides identified using a T cell assay.

[0761] Embodiment 62. The method according to any one of embodiments 55 to 61, wherein the training data set further includes data related to:

[0762] the abundance of peptides of the training peptide set present in the sample;

[0763] The peptide lengths of the training peptide set present in the sample.

[0764] Embodiment 63. The method of any one of Embodiments 55 to 62, wherein obtaining the training data set comprises:

[0765] Based on the training peptide sequence, a set of training protein sequences is obtained by comparing the set of training peptide sequences with a database comprising a set of known protein sequences through alignment, wherein the set of training protein sequences is longer than the training peptide sequence and includes the training peptide sequence.

[0766] Embodiment 64. The method of any one of Embodiments 55 to 63, wherein obtaining the training data set comprises:

[0767] The cell line is or has been subjected to mass spectrometry to obtain at least one of exome, transcriptome, or whole genome nucleotide sequencing data from the cell line, the nucleotide sequencing data comprising at least one protein sequence comprising a mutation.

[0768] Embodiment 65. The method of any one of Embodiments 55 to 64, wherein training the set of parameters of the presentation model comprises:

[0769] The training peptide sequences were encoded using a one-hot encoding scheme.

[0770] Embodiment 66. The method according to any one of embodiments 55 to 65, further comprising:

[0771] Obtaining at least one of exome, transcriptome and whole genome normal nucleotide sequencing data from a normal tissue sample; and

[0772] The normal nucleotide sequencing data is used to train a set of parameters for the presentation model.

[0773] Embodiment 67. A method according to any one of Embodiments 55 to 66, wherein the training dataset further comprises data related to proteomic sequences associated with the sample.

[0774] Embodiment 68. A method according to any one of Embodiments 55 to 67, wherein the training dataset further comprises data related to MHC peptidomimetic sequences associated with the sample.

[0775] Embodiment 69. A method according to any one of Embodiments 55 to 68, wherein the training data set further comprises data related to peptide-MHC binding affinity measurements of at least one of the isolated peptides.

[0776] Embodiment 70. A method according to any one of Embodiments 55 to 69, wherein the training data set further comprises data related to peptide-MHC binding stability measurements of at least one of the isolated peptides.

[0777] Embodiment 71. A method according to any one of Embodiments 55 to 70, wherein the training dataset further comprises data related to the transcriptome associated with the sample.

[0778] Embodiment 72. A method according to any one of embodiments 55 to 71, wherein the training data set also includes data related to the genome associated with the sample.

[0779] Embodiment 73. The method of any one of Embodiments 55 to 72, wherein training the set of digital parameters further comprises:

[0780] A logistic regression is performed on the set of parameters.

[0781] Embodiment 74. A method according to any one of embodiments 55 to 73, wherein the length of the training peptide sequence is in the range of k-mers, where k is between 8-15 (including the endpoints).

[0782] Embodiment 75. The method of any one of Embodiments 55 to 74, wherein training the set of digital parameters of the rendering model comprises:

[0783] The training peptide sequences were encoded using a left-filled one-hot encoding scheme.

[0784] Embodiment 76. The method of any one of Embodiments 55 to 75, wherein training the set of digital parameters further comprises:

[0785] The values ​​of the set of parameters are determined using a deep learning algorithm.

[0786] Embodiment 77. A method of generating a model for identifying one or more neoantigens that may be presented on the tumor cell surface of a tumor cell, comprising performing the following steps:

[0787] receiving mass spectrometry data comprising data associated with a plurality of isolated peptides eluted from a major histocompatibility complex (MHC) derived from a plurality of fresh or frozen tumor samples;

[0788] obtaining a training data set by at least identifying a set of training peptide sequences present in the sample and presented on one or more MHC alleles associated with each training peptide sequence;

[0789] Based on the training peptide sequence, obtaining a training protein sequence set; and

[0790] The training protein sequences and the training peptide sequences are used to train a set of numerical parameters for a presentation model that provides a plurality of numerical probabilities that a peptide sequence from the tumor cell is presented on the surface of the tumor cell by one or more MHC alleles.

[0791] Embodiment 78. The method of embodiment 77, wherein the presentation model represents the dependency between:

[0792] The presence of a specific allele in the MHC allele and a specific amino acid at a specific position in the peptide sequence; and

[0793] The likelihood that such a peptide sequence comprising said particular amino acid at said position is presented on the surface of said tumor cell by a particular allele of said pair of MHC alleles.

[0794] References

[0795] 1.Desrichard,A.,Snyder,A.&Chan,T.A.Cancer Neoantigens andApplications for Immunotherapy.Clin.Cancer Res.Off.J.Am.Assoc.Cancer Res.(2015).doi:10.1158 / 1078-0432.CCR-14-3175

[0796] 2.Schumacher,T.N.&Schreiber,R.D.Neoantigens in cancerimmunotherapy.Science 348,69–74(2015).

[0797] 3.Gubin,M.M.,Artyomov,M.N.,Mardis,E.R.&Schreiber,R.D.Tumorneoantigens:building a framework for personalized cancerimmunotherapy.J.Clin.Invest.125,3413–3421(2015).

[0798] 4.Rizvi,N.A.et al.Cancer immunology.Mutational landscape determinessensitivity to PD-1blockade in non-small cell lung cancer.Science 348,124–128(2015).

[0799] 5.Snyder,A.et al.Genetic basis for clinical response to CTLA-4blockade in melanoma.N.Engl.J.Med.371,2189–2199(2014).

[0800] 6.Carreno,B.M.et al.Cancer immunotherapy.A dendritic cell vaccineincreases the breadth and diversity of melanoma neoantigen-specific Tcells.Science 348,803–808(2015).

[0801] 7.Tran,E.et al.Cancer immunotherapy based on mutation-specific CD4+Tcells in a patient with epithelial cancer.Science 344,641–645(2014).

[0802] 8.Hacohen,N.&Wu,C.J.-Y.United States Patent Application:0110293637-COMPOSITIONS AND METHODS OF IDENTIFYING TUMOR SPECIFIC NEOANTIGENS.(A1).at<http: / / appft1.uspto.gov / netacgi / nph-Parser?Sect1=PTO1&Sect2=HITOFF&d=PG01&p=1&u= / netahtml / PTO / srchnum.html&r=1&f=G&l=50&s1=20110293637.PGNR.>

[0803] 9.Lundegaard,C.,Hoof,I.,Lund,O.&Nielsen,M.State of the art andchallenges in sequence based T-cell epitope prediction.Immunome Res.6 Suppl2,S3(2010).

[0804] 10.Yadav,M.et al.Predicting immunogenic tumour mutations by combiningmass spectrometry and exome sequencing.Nature 515,572–576(2014).

[0805] 11.Bassani-Sternberg,M.,Pletscher-Frankild,S.,Jensen,L.J.&Mann,M.Massspectrometry of human leukocyte antigen class I peptidomes reveals strongeffects of protein abundance and turnover on antigen presentation.Mol.Cell.Proteomics MCP 14,658–673(2015).

[0806] 12.Van Allen,E.M.et al.Genomic correlates of response to CTLA-4blockade in metastatic melanoma.Science 350,207–211(2015).

[0807] 13.Yoshida,K.&Ogawa,S.Splicing factor mutations and cancer.WileyInterdiscip.Rev.RNA 5,445–459(2014).

[0808] 14.Cancer Genome Atlas Research Network.Comprehensive molecularprofiling of lung adenocarcinoma.Nature 511,543–550(2014).

[0809] 15.Rajasagi,M.et al.Systematic identification of personal tumor-specific neoantigens in chronic lymphocytic leukemia.Blood 124,453–462(2014).

[0810] 16.Downing,S.R.et al.United States Patent Application:0120208706-OPTIMIZATION OF MULTIGENE ANALYSIS OF TUMOR SAMPLES.(A1).at<http: / / appft1.uspto.gov / netacgi / nph-Parser?Sect1=PTO1&Sect2=HITOFF&d=PG01&p=1&u= / netahtml / PTO / srchnum.html&r=1&f=G&l=50&s1=20120208706.PGNR.>

[0811] 17.Target Capture for NextGen Sequencing-IDT.at<http: / / www.idtdna.com / pages / products / nextgen / target-capture>

[0812] 18.Shukla,S.A.et al.Comprehensive analysis of cancer-associatedsomatic mutations in class I HLA genes.Nat.Biotechnol.33,1152–1158(2015).

[0813] 19.Cieslik,M.et al.The use of exome capture RNA-seq for highlydegraded RNA with application to clinical cancer sequencing.Genome Res.25,1372–1381(2015).

[0814] 20.Bodini,M.et al.The hidden genomic landscape of acute myeloidleukemia:subclonal structure revealed by undetected mutations.Blood 125,600–605(2015).

[0815] 21.Saunders,C.T.et al.Strelka:accurate somatic small-variant callingfrom sequenced tumor-normal sample pairs.Bioinforma.Oxf.Engl.28,1811–1817(2012).

[0816] 22.Cibulskis,K.et al.Sensitive detection of somatic point mutationsin impure and heterogeneous cancer samples.Nat.Biotechnol.31,213–219(2013).

[0817] 23.Wilkerson,M.D.et al.Integrated RNA and DNA sequencing improvesmutation detection in low purity tumors.Nucleic Acids Res.42,e107(2014).

[0818] 24.Mose,L.E.,Wilkerson,M.D.,Hayes,D.N.,Perou,C.M.&Parker,J.S.ABRA:improved coding indel detection via assembly-based realignment.Bioinforma.Oxf.Engl.30,2813–2815(2014).

[0819] 25.Ye,K.,Schulz,M.H.,Long,Q.,Apweiler,R.&Ning,Z.Pindel:a patterngrowth approach to detect break points of large deletions and medium sizedinsertions from paired-end short reads.Bioinforma.Oxf.Engl.25,2865–2871(2009).

[0820] 26.Lam,H.Y.K.et al.Nucleotide-resolution analysis of structuralvariants using BreakSeq and a breakpoint library.Nat.Biotechnol.28,47–55(2010).

[0821] 27.Frampton,G.M.et al.Development and validation of a clinical cancergenomic profiling test based on massively parallel DNAsequencing.Nat.Biotechnol.31,1023–1031(2013).

[0822] 28.Boegel,S.et al.HLA typing from RNA-Seq sequence reads.GenomeMed.4,102(2012).

[0823] 29.Liu,C.et al.ATHLATES:accurate typing of human leukocyte antigenthrough exome sequencing.Nucleic Acids Res.41,e142(2013).

[0824] 30.Mayor,N.P.et al.HLA Typing for the Next Generation.PloS One 10,e0127153(2015).

[0825] 31.Roy,C.K.,Olson,S.,Graveley,B.R.,Zamore,P.D.&Moore,M.J.Assessinglong-distance RNA sequence connectivity via RNA-templated DNA-DNAligation.eLife 4,(2015).

[0826] 32.Song,L.&Florea,L.CLASS:constrained transcript assembly of RNA-seqreads.BMC Bioinformatics 14 Suppl 5,S14(2013).

[0827] 33.Maretty,L.,Sibbesen,J.A.&Krogh,A.Bayesian transcriptomeassembly.Genome Biol.15,501(2014).

[0828] 34.Pertea,M.et al.StringTie enables improved reconstruction of atranscriptome from RNA-seq reads.Nat.Biotechnol.33,290–295(2015).

[0829] 35.Roberts,A.,Pimentel,H.,Trapnell,C.&Pachter,L.Identification ofnovel transcripts in annotated genomes using RNA-Seq.Bioinforma.Oxf.Engl.(2011).doi:10.1093 / bioinformatics / btr355

[0830] 36.Vitting-Seerup,K.,Porse,B.T.,Sandelin,A.&Waage,J.spliceR:an Rpackage for classification of alternative splicing and prediction of codingpotential from RNA-seq data.BMC Bioinformatics 15,81(2014).

[0831] 37.Rivas,M.A.et al.Human genomics.Effect of predicted protein-truncating genetic variants on the human transcriptome.Science 348,666–669(2015).

[0832] 38.Skelly,D.A.,Johansson,M.,Madeoy,J.,Wakefield,J.&Akey,J.M.Apowerful and flexible statistical framework for testing hypotheses of allele-specific gene expression from RNA-seq data.Genome Res.21,1728–1737(2011).

[0833] 39.Anders,S.,Pyl,P.T.&Huber,W.HTSeq--a Python framework to work withhigh-throughput sequencing data.Bioinforma.Oxf.Engl.31,166–169(2015).

[0834] 40.Furney,S.J.et al.SF3B1 mutations are associated with alternativesplicing in uveal melanoma.Cancer Discov.(2013).doi:10.1158 / 2159-8290.CD-13-0330

[0835] 41.Zhou,Q.et al.A chemical genetics approach for the functionalassessment of novel cancer genes.Cancer Res.(2015).doi:10.1158 / 0008-5472.CAN-14-2930

[0836] 42.Maguire,S.L.et al.SF3B1 mutations constitute a novel therapeutictarget in breast cancer.J.Pathol.235,571–580(2015).

[0837] 43.Carithers,L.J.et al.A Novel Approach to High-Quality PostmortemTissue Procurement:The GTEx Project.Biopreservation Biobanking 13,311–319(2015).

[0838] 44.Xu,G.et al.RNA CoMPASS:a dual approach for pathogen and hosttranscriptome analysis of RNA-seq datasets.PloS One 9,e89445(2014).

[0839] 45.Andreatta,M.&Nielsen,M.Gapped sequence alignment using artificialneural networks:application to the MHC class I system.Bioinforma.Oxf.Engl.(2015).doi:10.1093 / bioinformatics / btv639

[0840] 46. K.W.,Rasmussen,M.,Buus,S.&Nielsen,M.NetMHCstab-predictingstability of peptide-MHC-I complexes;impacts for cytotoxic T lymphocyteepitope discovery.Immunology 141,18–26(2014).

[0841] 47.Larsen,M.V.et al.An integrative approach to CTL epitopeprediction:a combined algorithm integrating MHC class I binding,TAP transportefficiency,and proteasomal cleavage predictions.Eur.J.Immunol.35,2295–2303(2005).

[0842] 48.Nielsen,M.,Lundegaard,C.,Lund,O.& C.The role of theproteasome in generating cytotoxic T-cell epitopes:insights obtained fromimproved predictions of proteasomal cleavage.Immunogenetics 57,33–41(2005).

[0843] 49.Boisvert,F.-M.et al.A Quantitative Spatial Proteomics Analysis ofProteome Turnover in Human Cells.Mol.Cell.Proteomics 11,M111.011429–M111.011429(2012).

[0844] 50.Duan,F.et al.Genomic and bioinformatic profiling of mutationalneoepitopes reveals new rules to predict anticancerimmunogenicity.J.Exp.Med.211,2231–2248(2014).

[0845] 51.Janeway’s Immunobiology:9780815345312:Medicine&Health ScienceBooks@Amazon.com.at<http: / / www.amazon.com / Janeways-Immunobiology-Kenneth-Murphy / dp / 0815345313>

[0846] 52.Calis,J.J.A.et al.Properties of MHC Class I Presented PeptidesThat Enhance Immunogenicity.PLoS Comput.Biol.9,e1003266(2013).

[0847] 53.Zhang,J.et al.Intratumor heterogeneity in localized lungadenocarcinomas delineated by multiregion sequencing.Science 346,256–259(2014)

[0848] 54.Walter,M.J.et al.Clonal architecture of secondary acute myeloidleukemia.N.Engl.J.Med.366,1090–1098(2012).

[0849] 55.Hunt DF,Henderson RA,Shabanowitz J,Sakaguchi K,Michel H,Sevilir N,Cox AL,Appella E,Engelhard VH.Characterization of peptides bound to the classI MHC molecule HLA-A2.1 by mass spectrometry.Science 1992.255:1261-1263.

[0850] 56.Zarling AL,Polefrone JM,Evans AM,Mikesh LM,Shabanowitz J,Lewis ST,Engelhard VH,Hunt DF.Identification of class I MHC-associated phosphopeptidesas targets for cancer immunotherapy.Proc Natl Acad Sci U S A.2006 Oct 3;103(40):14889-94.

[0851] 57.Bassani-Sternberg M,Pletscher-Frankild S,Jensen LJ,Mann M.Massspectrometry of human leukocyte antigen class I peptidomes reveals strongeffects of protein abundance and turnover on antigen presentation.Mol CellProteomics.2015 Mar;14(3):658-73.doi:10.1074 / mcp.M114.042812.

[0852] 58.Abelin JG,Trantham PD,Penny SA,Patterson AM,Ward ST,Hildebrand WH,Cobbold M,Bai DL,Shabanowitz J,Hunt DF.Complementary IMAC enrichment methodsfor HLA-associated phosphopeptide identification by mass spectrometry.NatProtoc.2015 Sep;10(9):1308-18.doi:10.1038 / nprot.2015.086.Epub 2015 Aug 6

[0853] 59.Barnstable CJ,Bodmer WF,Brown G,Galfre G,Milstein C,Williams AF,Ziegler A.Production of monoclonal antibodies to group A erythrocytes,HLA andother human cell surface antigens-new tools for genetic analysis.Cell.1978May;14(1):9-20.

[0854] 60.Goldman JM,Hibbin J,Kearney L,Orchard K,Th'ng KH.HLA-DR monoclonalantibodies inhibit the proliferation of normal and chronic granulocyticleukaemia myeloid progenitor cells.Br J Haematol.1982 Nov;52(3):411-20.

[0855] 61.Eng JK,Jahan TA,Hoopmann MR.Comet:an open-source MS / MS sequencedatabase search tool.Proteomics.2013 Jan;13(1):22-4.doi:10.1002 / pmic.201200439.Epub 2012 Dec 4.

[0856] 62.Eng JK,Hoopmann MR,Jahan TA,Egertson JD,Noble WS,MacCoss MJ.Adeeper look into Comet--implementation and features.J Am Soc MassSpectrom.2015 Nov;26(11):1865-74.doi:10.1007 / s13361-015-1179-x.Epub 2015 Jun27.

[0857] 63.Lukas Jesse Canterbury,Jason Weston,William Stafford Noble andMichael J.MacCoss.Semi-supervised learning for peptide identification fromshotgun proteomics datasets.Nature Methods 4:923–925,November 2007

[0858] 64.Lukas John D.Storey,Michael J.MacCoss and William StaffordNoble.Assigning confidence measures to peptides identified by tandem massspectrometry.Journal of Proteome Research,7(1):29-34,January 2008

[0859] 65.Lukas John D.Storey and William Stafford Noble.Nonparametricestimation of posterior error probabilities associated with peptidesidentified by tandem mass spectrometry.Bioinformatics,24(16):i42-i48,August2008

[0860] 66.Bo Li and Colin N.Dewey.RSEM:accurate transcript quantificationfrom RNA-Seq data with or without a referenfe genome.BMC Bioinformatics,12:323,August 2011

[0861] 67.Hillary Pearson,Tariq Daouda,Diana Paola Granados,Chantal Durette,Eric Bonneil,Mathieu Courcelles,Anja Rodenbrock,Jean-Philippe Laverdure,Caroline Sylvie Mader,Sébastien Lemieux,Pierre Thibault,and ClaudePerreault.MHC class I-associated peptides derive from selective regions ofthe human genome.The Journal of Clinical Investigation,2016,

[0862] 68.Juliane Liepe,Fabio Marino,John Sidney,Anita Jeko,DanielE.Bunting,Alessandro Sette,Peter M.Kloetzel,Michael P.H.Stumpf,AlbertJ.R.Heck,Michele Mishto.A large fraction of HLA class I ligands areproteasome-generated spliced peptides.Science,21,October 2016.

[0863]

[0864]

[0865]

[0866]

[0867]

[0868]

[0869]

[0870]

[0871]

[0872]

[0873]

[0874]

[0875]

[0876]

[0877]

[0878]

[0879]

[0880]

[0881]

[0882]

[0883]

[0884]

[0885]

[0886]

[0887]

[0888]

[0889]

[0890]

[0891]

[0892]

[0893]

[0894]

[0895]

[0896]

[0897]

[0898]

[0899]

[0900]

[0901]

[0902]

[0903]

[0904]

[0905]

[0906]

[0907]

[0908]

[0909]

[0910]

[0911]

[0912]

[0913]

[0914]

[0915]

[0916]

[0917]

[0918]

[0919]

[0920]

[0921]

[0922]

[0923]

[0924]

[0925]

[0926]

[0927]

[0928]

[0929]

[0930]

[0931]

[0932]

[0933]

[0934]

[0935]

[0936]

[0937]

[0938]

[0939]

[0940]

[0941]

[0942]

[0943]

[0944]

[0945]

[0946]

[0947]

[0948]

[0949]

[0950]

[0951]

[0952]

[0953]

[0954]

[0955]

[0956]

[0957]

[0958]

[0959]

[0960]

Claims

1. A method for identifying one or more new antigens that may be presented on the surface of tumor cells from a subject, comprising the following steps: Obtaining at least one of exome, transcriptome, or whole genome tumor nucleotide sequencing data from a tumor cell of the subject, wherein the tumor nucleotide sequencing data is used to obtain data representing a peptide sequence of each neo-antigen in the set of neo-antigens, and wherein the peptide sequence of each neo-antigen comprises at least one change that makes it different from a corresponding wild-type parent peptide sequence; inputting the peptide sequence of each neo-antigen into one or more presentation models to generate a set of digital likelihoods that each of the neo-antigens is presented by one or more MHC alleles on the tumor cell surface of the tumor cell of the subject, the set of digital likelihoods having been identified based at least on the received mass spectrometry data; and Based on the set of numerical possibilities, a subset of the set of neo-antigens is selected to produce a set of selected neo-antigens.

2. The method of claim 1, wherein the number of selected neoantigens in the collection is 20.

3. The method according to any one of claims 1 to 2, wherein the presentation model represents the dependencies between: The presence of a specific allele in the MHC allele and a specific amino acid at a specific position in the peptide sequence; and The likelihood that such a peptide sequence comprising said particular amino acid at said position is presented on the surface of said tumor cell by a particular allele of said pair of MHC alleles.

4. The method according to any one of claims 1 to 3, wherein inputting the peptide sequence comprises: The one or more presentation models are applied to the peptide sequence of the corresponding neoantigen to generate a dependency score for each of the one or more MHC alleles based at least on the amino acid position of the peptide sequence of the corresponding neoantigen, wherein the dependency score indicates whether the MHC allele will present the corresponding neoantigen.

5. The method according to claim 4, further comprising: transforming the dependency scores to obtain 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 numerical likelihood.

6. A method for treating a subject having a tumor, comprising performing any one of the steps of claims 1 to 5, and further comprising obtaining a tumor vaccine comprising a collection of selected neoantigens, and administering the tumor vaccine to the subject.

7. A method for producing a tumor vaccine, comprising performing any one of the steps of claims 1 to 6, and further comprising producing or having produced a tumor vaccine comprising the collection of selected new antigens.

8. A tumor vaccine comprising a collection of selected neoantigens according to any one of claims 1 to 5, wherein the collection of neoantigens is selected by performing the method according to any one of claims 1 to 5.

9. A method of generating a model for identifying one or more neoantigens that may be presented on the tumor cell surface of a tumor cell, comprising performing the following steps: receiving mass spectrometry data comprising data associated with a plurality of separated peptides eluted from a major histocompatibility complex (MHC) derived from a plurality of samples; obtaining a training data set by at least identifying a set of training peptide sequences present in the sample and one or more MHCs associated with each training peptide sequence; The training data set comprising the training peptide sequences is used to train a set of numerical parameters of a presentation model, the presentation model providing a plurality of numerical probabilities that a peptide sequence from the tumor cell is presented on the surface of the tumor cell by one or more MHC alleles.

10. A method of generating a model for identifying one or more neoantigens that may be presented on the tumor cell surface of a tumor cell, comprising performing the following steps: receiving mass spectrometry data comprising data associated with a plurality of isolated peptides eluted from a major histocompatibility complex (MHC) derived from a plurality of fresh or frozen tumor samples; obtaining a training data set by at least identifying a set of training peptide sequences present in the sample and presented on one or more MHC alleles associated with each training peptide sequence; Based on the training peptide sequence, obtaining a training protein sequence set; and The training protein sequences and the training peptide sequences are used to train a set of numerical parameters for a presentation model that provides a plurality of numerical probabilities that a peptide sequence from the tumor cell is presented on the surface of the tumor cell by one or more MHC alleles.

Images