Identification, production, and use of neoantigens

Abstract

To provide an optimized approach for identifying and selecting neoantigens for personalized cancer vaccines.SOLUTION: Provided is a system and methods for determining the alleles, neoantigens, and vaccine composition as determined on the basis of an individual's tumor mutations. Also provided are systems and methods for obtaining high quality sequencing data from a tumor. Further, provided are systems and methods for identifying somatic changes in polymorphic genome data. Also further provided are systems and methods for selecting a subset of patients for treatment. A utility score indicating an estimated number of neoantigens presented on the surface of tumor cells is determined for each patient based on one or more neoantigen candidates identified for the patient. The subset of patients are selected based on the determined utility scores. The selected subset of patients can receive treatment, such as neoantigen vaccines or checkpoint inhibitor therapy. Still further provided are unique cancer vaccines.SELECTED DRAWING: Figure 13B

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of U.S. Provisional Application No. 62 / 517,786, filed June 9, 2017, which is incorporated by reference in its entirety. [Background technology]

[0002] background Therapeutic vaccines based on tumor-specific neoantigens hold great promise as the next generation of personalized cancer immunotherapy. 1~3 Cancers with high mutational burden, such as non-small cell lung cancer (NSCLC) and melanoma, are particularly promising targets for such therapies due to their relatively high likelihood of giving rise to neoantigens. 4,5 Early evidence suggests that neoantigen-based vaccination induces T cell responses. 6 , Neoantigen-targeted cell therapy can induce tumor regression in selected patients 7 It has been shown that both MHC class I and MHC class II influence T cell responses. 70~71 .

[0003] One question regarding neoantigen vaccine design is which of the many coding mutations present in the tumor of interest could give rise to the "best" therapeutic neoantigen (e.g., an antigen capable of eliciting antitumor immunity and causing tumor regression).

[0004] Early methods have been proposed that incorporate mutation-based analysis using next-generation sequencing, RNA gene expression, and prediction of MHC binding affinity of neoantigen peptides. 8However, these proposed methods involve many steps other than gene expression and MHC binding (e.g., TAP transport, proteasomal cleavage, MHC binding, transport of peptide-MHC complexes to the cell surface, and / or TCR recognition of MHC-I; endocytosis or autophagy, cleavage by extracellular or lysosomal proteases (e.g., cathepsins), competition with CLIP peptides for HLA binding catalyzed by HLA-DM, transport of peptide-MHC complexes to the cell surface, and / or TCR recognition of MHC-II). 9 The entire epitope generation process cannot be modeled, and therefore existing methods tend to suffer from low positive predictive value (PPV) (Figure 1A).

[0005] Indeed, analyses of peptides presented by tumor cells performed by several groups have shown that less than 5% of peptides predicted to be presented using gene expression and MHC binding affinity are found on MHC at the tumor surface. 10,11 (Figure 1B). This poor correlation between binding prediction and MHC presentation is further supported by the lack of improved prediction accuracy of binding-restricted neoantigens for checkpoint inhibitor response relative to number of mutations alone. 12 .

[0006] Such low positive predictive value (PPV) of existing methods for predicting presentation presents a problem in designing neoantigen-based vaccines. If a vaccine is designed using low PPV predictions, most patients will likely not receive therapeutic neoantigens, and even fewer will receive multiple neoantigens (even if all of the presented peptides are assumed to be immunogenic). Thus, neoantigen vaccination with current methods is unlikely to be successful in a significant number of tumor-bearing subjects (Figure 1C).

[0007] Furthermore, previous approaches have only used cis-acting mutations to generate candidate neoantigens, including splicing factor mutations that occur in multiple tumor types and lead to aberrant splicing of many genes. 13 , and additional sources of nascent ORFs, including mutations that create or remove protease cleavage sites, were not considered in most cases.

[0008] Finally, standard approaches to tumor genome and transcriptome analysis may miss somatic mutations that give rise to candidate neoantigens due to suboptimal conditions in library construction, exome and transcriptome capture, sequencing, or data analysis. Similarly, standard tumor analysis approaches may erroneously promote sequence artifacts or germline polymorphisms as neoantigens, which may lead to inefficient utilization of vaccine doses or risk of autoimmunity, respectively. Summary of the Invention

[0009] overview Optimized approaches to identify and select neoantigens for personalized cancer vaccines are disclosed herein. First, we address optimized tumor exome and transcriptome analysis approaches to identify neoantigen candidates using next generation sequencing (NGS). These methods build on standard approaches of tumor analysis by NGS so that the most sensitive and specific neoantigen candidates are developed across all classes of genomic alterations. Second, novel approaches to high PPV neoantigen selection are provided to overcome the specificity problem and make neoantigens developed for vaccine addition more likely to induce antitumor immunity. These approaches include, depending on the embodiment, trained statistical regression or nonlinear deep learning models that jointly model peptide-allele mapping, as well as motifs per allele for peptides of multiple lengths that share statistical potency across peptides of different lengths. In particular, nonlinear deep learning models can be designed and trained to treat different MHC alleles in the same cell as independent, thus resolving the problem with linear models where linear models interfere with each other. Finally, a further concern regarding the design and production of personalized neoantigen-based vaccines is addressed.

[0010] Also disclosed herein is a method for identifying a subset of patients suitable for treatment. At least one of tumor nucleotide sequencing data of the exome, transcriptome, or whole genome is obtained from tumor cells and normal cells of each patient. The tumor nucleotide sequencing data is used to obtain a peptide sequence for each of the set of neoantigens identified by comparing the nucleotide sequencing data from the tumor cells with the nucleotide sequencing data from the normal cells. The peptide sequence for each of the patient's neoantigens includes at least one change that makes it different from the corresponding wild-type parent peptide sequence identified from the patient's normal cells. The peptide sequence for each of the set of neoantigens is input into a machine-learned presentation model to generate a set of numerical presentation likelihoods for the set of neoantigens for each patient. Each presentation likelihood represents the likelihood that the corresponding neoantigen is presented by one or more MHC alleles on the surface of the patient's tumor cells. The set of presentation likelihoods is identified based at least on mass spectrometry data. One or more neoantigens are identified from the set of neoantigens of the patient. A utility score is determined for each patient, which indicates the estimated number of neoantigens presented on the surface of the patient's tumor cells, as determined by the corresponding likelihood of presentation of one or more neoantigens for the patient. A subset of patients is selected to receive treatment. Each patient within this subset of patients is associated with a utility score that meets predefined inclusion criteria. The selected subset of patients can be treated with treatment, such as a neoantigen vaccine or checkpoint inhibitor therapy. [The present invention 1001] 1. A method for identifying a subset of patients suitable for treatment, comprising: obtaining, for each patient, at least one of exome, transcriptome, or whole genome tumor nucleotide sequencing data from the patient's tumor cells and normal cells, wherein the tumor nucleotide sequencing data is used to obtain a peptide sequence for each of a set of neoantigens identified by comparing the nucleotide sequencing data from the tumor cells to the nucleotide sequencing data from the normal cells, wherein the peptide sequence of each neoantigen for the patient includes at least one alteration that makes it different from a corresponding wild-type parent peptide sequence identified from the patient's normal cells; generating, for each patient, a set of numerical presentation likelihoods for the set of neoantigens for the patient by inputting the peptide sequences of each of the set of neoantigens into a machine-learned presentation model, each presentation likelihood representing the likelihood that a corresponding neoantigen will be presented by one or more MHC alleles on the surface of tumor cells of the patient, the set of presentation likelihoods having been determined based at least on mass spectrometry data; identifying for each patient one or more neoantigens from said set of neoantigens for said patient; determining for each patient a utility score indicative of the estimated number of neoantigens presented on the surface of tumor cells of said patient as determined by the corresponding presentation likelihood for said one or more neoantigens for said patient; selecting a subset of patients suitable for treatment, wherein each patient in said subset of patients is associated with a utility score that meets predetermined inclusion criteria; The method comprising: [The present invention 1002] 1002. The method of claim 1001, wherein identifying said one or more neoantigens for said patient comprises selecting a subset of neoantigens in said set of neoantigens for said patient. [The present invention 1003] The method of claim 1002, wherein said subset of neoantigens are neoantigens having the highest presentation likelihood among said set of presentation likelihoods for said patient. [The present invention 1004] The method of claim 1001, further comprising treating each patient within the selected subset of patients with a corresponding neoantigen vaccine comprising at least one of the one or more neoantigens identified for said patient. [The present invention 1005] The method of claim 1001, further comprising identifying, for each patient within the selected subset of patients, one or more T cells or T cell receptors that are antigen-specific for at least one of the one or more neoantigens identified for the patient. [The present invention 1006] 1002. The method of claim 1001, wherein identifying one or more neoantigens for the patient comprises selecting the entire set of identified neoantigens for the patient. [The present invention 1007] The method of any one of claims 10 to 15, further comprising administering checkpoint inhibitor therapy to each patient within said selected subset of patients. [The present invention 1008] The method of the present invention 1001, wherein selecting a subset of patients suitable for treatment comprises selecting a subset of patients having a tumor mutational burden (TMB) higher than a minimum threshold, the TMB of a patient indicating the number of neoantigens in a set of neoantigens associated with that patient. [The present invention 1009] Selecting the subset of patients suitable for treatment is key. Selecting a subset of patients having a utility score above a minimum threshold The method of the present invention 1001, comprising: [The present invention 1010] 1002. The method of claim 1001, wherein said usefulness score is the sum of the likelihoods of presentation for each neoantigen within said identified subset of neoantigens for said patient. [The present invention 1011] 1002. The method of claim 1001, wherein said usefulness score is the probability that the number of presented neoantigens among said identified one or more neoantigens for said patient is above a minimum threshold. [The present invention 1012] The machine-learned presentation model is a label obtained by mass spectrometry to measure the presence of a peptide bound to at least one MHC allele identified as being present in at least one of the plurality of samples; a training peptide sequence including information on a set of amino acids constituting the training peptide sequence and positions of the amino acids within the training peptide sequence; at least one MHC allele associated with said training peptide sequence; a plurality of parameters determined based at least on the training data set, the plurality of parameters including: A function that represents the relationship between the peptide sequence and the likelihood of presentation based on the plurality of parameters. The method of the present invention 1001, comprising: [The present invention 1013] The training data set is (a) data relating to measurements of peptide-MHC binding affinity for at least one of the isolated peptides; and (b) data relating to a measurement of peptide-MHC binding stability for at least one of the isolated peptides; The method of the present invention 1012, further comprising at least one of the following: [The present invention 1014] The set of numerical likelihoods is (a) a C-terminal sequence adjacent to the neoantigen-encoding peptide sequence within the source protein sequence; and (b) an N-terminal sequence adjacent to the neoantigen-encoding peptide sequence within the source protein sequence; The method of the present invention 1001, further characterized by at least one of the following characteristics: [The present invention 1015] The method of claim 1001, wherein said set of presentation likelihoods is further identified by at least the expression level of said one or more MHC alleles in said subject, as measured by RNA-seq or mass spectrometry. [The present invention 1016] The set of presentation likelihoods is (a) predicted affinity between neoantigens within said set of neoantigens and said one or more MHC alleles; and (b) the predicted stability of the neoantigen-encoded peptide-MHC complex; The method of the present invention 1001, further characterized by at least one of the following characteristics: [The present invention 1017] inputting the peptide sequence into the machine-learned representation model; applying the machine learning presentation model to the peptide sequence of each neoantigen to generate, for each of the one or more MHC alleles, a dependency score indicative of whether the MHC allele presents the neoantigen based on a particular amino acid at a particular position of the peptide sequence. The method of the present invention 1001, comprising: [The present invention 1018] inputting the peptide sequence into the machine-learned representation model; transforming the dependency scores to generate, for each MHC allele, a corresponding per-allele likelihood that indicates the likelihood that the corresponding MHC allele presents the corresponding neoantigen; and combining said per-allele likelihoods to generate a presentation likelihood for said neoantigen. The method of the present invention 1017, comprising: [The present invention 1019] The method of claim 1018, wherein transforming said dependency score models presentation of said neoantigen as mutually exclusive across said one or more class MHC alleles. [The present invention 1020] inputting the peptide sequence into the machine-learned representation model; Transforming the combination of dependency scores to generate a representation likelihood. and transforming said combination of dependency scores models the presentation of said neoantigen as interfering between said one or more MHC alleles. [Brief description of the drawings]

[0011] These and other features, aspects and aspects of the present invention will become better understood with regard to the following description and accompanying drawings.

[0012] [Figure 1A] Current clinical approaches to neoantigen identification are presented. [Figure 1B] It is shown that less than 5% of the predicted binding peptides are displayed on tumor cells. [Figure 1C] 1 illustrates the impact of the specificity problem on neoantigen prediction. [Figure 1D] This shows that binding prediction is not sufficient to identify neoantigens. [Figure 1E] Probability of MHC-I presentation as a function of peptide length. [Figure 1F] 1 shows an exemplary peptide spectrum generated from a Promega dynamic range standard. [Figure 1G] We show how adding features increases the positive predictive value of the model. [Figure 2A] 1 is a schematic of an environment for identifying the likelihood of peptide presentation in a patient, according to one embodiment. [Figure 2B] A method for obtaining presentation information according to one embodiment is described below. FIG. 2B discloses SEQ ID NO:20. [Figure 2C] A method for obtaining presentation information according to one embodiment is illustrated in FIG. 2C, which discloses SEQ ID NOs: 3-8, respectively, in order of appearance. [Diagram 3] FIG. 1 is a high-level block diagram illustrating computer logic components of a presentation specific system according to one embodiment. [Figure 4]4 illustrates an exemplary set of training data, according to one embodiment. Figure 4 discloses, in order of appearance, "peptide sequences" as SEQ ID NOs: 10-13, and "C-terminal flanking sequences" as SEQ ID NOs: 15, 21-22, and 22, respectively. [Diagram 5] 1 illustrates an exemplary network model relating to MHC alleles. [Figure 6A] An exemplary network model NNH(·) shared by MHC alleles according to one embodiment [Figure 6B] FIG. 1 shows an exemplary network model NNH(·) shared by MHC alleles according to another embodiment. [Figure 7] 1 illustrates the generation of presentation likelihoods of peptides associated with MHC alleles using an exemplary network model. [Figure 8] 1 illustrates the generation of presentation likelihoods of peptides associated with MHC alleles using an exemplary network model. [Figure 9] 1 illustrates the generation of presentation likelihoods of peptides associated with MHC alleles using an exemplary network model. [Figure 10] 1 illustrates the generation of presentation likelihoods of peptides associated with MHC alleles using an exemplary network model. [Figure 11] 1 illustrates the generation of presentation likelihoods of peptides associated with MHC alleles using an exemplary network model. [Figure 12] 1 illustrates the generation of presentation likelihoods of peptides associated with MHC alleles using an exemplary network model. [Figure 13A] 1 shows the sample frequency distribution of tumor mutation burden in NSCLC patients. [Figure 13B] 1 shows the number of neoantigens presented in a simulated vaccine for patients selected based on the inclusion criterion of whether the patient met a minimum tumor mutational burden, according to one embodiment. [Figure 13C]FIG. 11 shows a comparison of the number of neoantigens presented in simulated vaccines between selected patients associated with a vaccine that includes a therapeutic subset identified based on a presented model according to one embodiment and selected patients associated with a vaccine that includes a therapeutic subset identified by a prior art model. [Figure 13D] 1 compares the number of presented neoantigens in simulated vaccines between selected patients associated with a vaccine containing therapeutic subsets identified based on a single allele-per-presentation model for HLA-A*02:01 and selected patients associated with a vaccine containing therapeutic subsets identified based on both an allele-per-presentation model for HLA-A*02:01 and HLA-B*07:02. According to one embodiment, the vaccine volume is set to v=20 epitopes. [Figure 13E] FIG. 11 compares the number of neoantigens presented in simulated vaccines between patients selected based on tumor mutation burden and patients selected by expected utility score, according to one embodiment. [Figure 14] An exemplary computer for implementing the entities illustrated in FIGS. 1 and 3 is described. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Detailed Description I. Definition In general, the terms used in the claims and the specification shall be interpreted as having their ordinary meanings as understood by those skilled in the art. Certain terms are defined below to provide further clarity. If there is a discrepancy between the ordinary meaning and a given definition, the given definition shall prevail.

[0014] As used herein, the term "antigen" refers to a substance that induces an immune response.

[0015] The term "neoantigen" as used herein refers to an antigen that has at least one change that makes the antigen different from the corresponding wild-type parent antigen, for example, due to a tumor cell mutation or a tumor cell-specific post-translational modification. A neoantigen may include a polypeptide sequence or a nucleotide sequence. A mutation may include a frameshift or non-frameshift insertion deletion (indel), a missense or nonsense substitution, a splice site change, a genomic rearrangement or gene fusion, or any genomic or expression change that results in a neo-ORF. A mutation may also include a splice variant. A tumor cell-specific post-translational modification may include aberrant phosphorylation. A tumor cell-specific post-translational modification may also include a splice antigen generated by the proteasome. See Liepe et al., A large fraction of HLAclass I ligands are proteasome-generated spliced ​​peptides; Science. 2016 Oct 21; 354(6310): 354-358.

[0016] As used herein, the term "tumor neoantigen" refers to a neoantigen that is present in tumor cells or tissues of a subject, but is not present in the corresponding normal cells or tissues of the subject.

[0017] As used herein, the term "neoantigen-based vaccine" refers to a vaccine construct that is based on one or more neoantigens, eg, multiple neoantigens.

[0018] As used herein, the term "candidate neoantigen" refers to a mutation or other abnormality that gives rise to a new sequence that may represent a neoantigen.

[0019] As used herein, the term "coding region" refers to the portion of a gene that encodes a protein.

[0020] As used herein, the term "coding mutation" refers to a mutation that occurs in a coding region.

[0021] As used herein, the term "ORF" means open reading frame.

[0022] As used herein, the term "neo-ORF" refers to a tumor-specific ORF that arises due to mutation or other abnormalities such as splicing.

[0023] As used herein, the term "missense mutation" is a mutation that results in the substitution of one amino acid for another.

[0024] As used herein, the term "nonsense mutation" is a mutation that results in the substitution of an amino acid for a stop codon.

[0025] As used herein, the term "frameshift mutation" is a mutation that causes an alteration in the frame of a protein.

[0026] As used herein, the term "insertion / deletion" refers to the insertion or deletion of one or more nucleic acids.

[0027] As used herein, the term "percent identity" in the context of two or more nucleic acid or polypeptide sequences refers to two or more sequences or subsequences in which a certain percentage of nucleotides or amino acid residues are the same when compared and aligned for maximum correspondence as measured using one of the sequence comparison algorithms described below (e.g., BLASTP and BLASTN, or other algorithms available to those of skill in the art) or by visual inspection. Depending on the application, the "percent identity" can exist over a region of the sequences being compared, e.g., over a functional domain, or over the full length of the two sequences being compared.

[0028] In sequence comparison, generally, one sequence serves as a reference sequence to which test sequence is compared.When using sequence comparison algorithm, test sequence and reference sequence are input into computer, subsequence coordinates are designated if necessary, and sequence algorithm program parameters are designated.The sequence comparison algorithm then calculates the sequence identity (%) of test sequence to reference sequence based on designated program parameters.Alternatively, sequence similarity or difference can be established by the combination of the presence or absence of a particular nucleotide at a selected sequence position (e.g., sequence motif) or an amino acid in a translated sequence.

[0029] Optimal alignment of sequences for comparison can be performed, for example, by the local homology algorithm of Smith & Waterman, Adv. Appl. Math. 2:482 (1981), by the homology alignment algorithm of Needleman & Wunsch, J. Mol. Biol. 48:443 (1970), by the search for similarity method of Pearson & Lipman, Proc. Nat'l. Acad. Sci. USA 85:2444 (1988), by computerized implementations of these algorithms (GAP, BESTFIT, FASTA, and TFASTA in the Wisconsin Genetics Software Package, Genetics Computer Group, 575 Science Dr., Madison, Wis.), or by visual inspection (see generally, Ausubel et al., infra).

[0030] One example of an algorithm that is suitable for determining percent sequence identity and percent 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.

[0031] As used herein, the term "non-stop or read-through" refers to a mutation that results in the removal of the natural stop codon.

[0032] As used herein, the term "epitope" refers to a specific portion of an antigen generally bound by an antibody or T-cell receptor.

[0033] As used herein, the term "immunogenic" refers to the ability to elicit an immune response, for example, via T cells, B cells, or both.

[0034] As used herein, the terms "HLA binding affinity", "MHC binding affinity" refer to the affinity of binding between a specific antigen and a specific MHC allele.

[0035] As used herein, the term "bait" refers to a nucleic acid probe used to enrich a specific sequence of DNA or RNA from a sample.

[0036] As used herein, the term "mutation" is a difference between the nucleic acid of a subject and a reference human genome used as a control.

[0037] As used herein, the term "variant calling" is the algorithmic determination, typically from sequencing, of the presence of a variant.

[0038] As used herein, the term "polymorphism" refers to a germline mutation, ie, a mutation found in all DNA-bearing cells of an individual.

[0039] As used herein, the term "somatic mutation" is a mutation that occurs in a non-germline cell of an individual.

[0040] As used herein, the term "allele" refers to one version of a gene or one version of a gene sequence or one version of a protein.

[0041] As used herein, the term "HLA type" refers to the complement of HLA gene alleles.

[0042] As used herein, the term "nonsense-mediated decay" or "NMD" refers to the degradation of an mRNA by a cell due to a premature stop codon.

[0043] As used herein, the term "truncal mutation" is a mutation that occurs early in the development of a tumor and is present in the majority of the cells of the tumor.

[0044] As used herein, the term "subclonal mutation" is a mutation that occurs late in the development of a tumor and is present in only a portion of the cells of the tumor.

[0045] As used herein, the term "exome" is the subset of the genome that encodes proteins. The exome can be the collection of exons of the genome.

[0046] As used herein, the term "logistic regression" is a regression model for binary data from statistics in which the logit of the probability that the dependent variable is equal to 1 is modeled as a linear function of the dependent variable.

[0047] As used herein, the term "neural network" is a machine learning model for classification or regression that consists of multiple layers of linear transformations followed by element-wise nonlinear transformations typically trained by stochastic gradient descent and backpropagation.

[0048] As used herein, the term "proteome" refers to the set of all proteins expressed and / or translated by a cell, a group of cells, or an individual.

[0049] As used herein, the term "peptidome" refers to the set of all peptides presented by MHC-I or MHC-II on the cell surface. Peptidome can also refer to the properties of a cell or a collection of cells (e.g., a tumor peptidome refers to the union of the peptidomes of all cells that comprise a tumor).

[0050] As used herein, the term "ELISPOT" refers to enzyme-linked immunosorbent spot assay, a common method for monitoring immune responses in humans and animals.

[0051] As used herein, the term "dextramer" is a dextran-based peptide-MHC multimer used for antigen-specific T cell staining in flow cytometry.

[0052] As used herein, the term "tolerance or immune tolerance" refers to a state of immune unresponsiveness to one or more antigens, eg, self-antigens.

[0053] As used herein, the term "central tolerance" is tolerance conferred in the thymus by either deleting autoreactive T cell clones or promoting the differentiation of autoreactive T cell clones into immunosuppressive regulatory T cells (Tregs).

[0054] As used herein, the term "peripheral tolerance" is tolerance conferred in the peripheral system by downregulating or anergizing autoreactive T cells that survived central tolerance or by promoting the differentiation of these T cells into Tregs.

[0055] The term "sample" can include a single cell, or multiple cells, or fragments of cells, or an aliquot of bodily fluid obtained from a subject by means including venipuncture, excretion, ejaculation, massage, biopsy, needle aspirate, lavage sample, scraping, surgical incision, or intervention, or other means known in the art.

[0056] The term "subject" includes cells, tissues, or organisms, either male or female, human or non-human, whether in vivo, ex vivo, or in vitro. The term subject includes mammals, including humans.

[0057] 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.

[0058] The term "clinical factor" refers to a measurement of a subject's condition, e.g., disease activity or severity. "Clinical factor" encompasses all markers of a subject's health status, including non-sample markers, and / or other characteristics of the subject, such as, but not limited to, age and sex. A clinical factor can be a score, value, or set of values ​​that can be obtained from the assessment of a subject or a sample (or a population of samples) from a subject under a given condition. A clinical factor can also be predicted by other parameters, such as markers and / or gene expression surrogates. Clinical factors can include tumor type, tumor subtype, and smoking history.

[0059] Abbreviations: MHC: major histocompatibility complex; HLA: human leukocyte antigen, or human MHC locus; NGS: next generation sequencing; PPV: positive predictive value; TSNA: tumor-specific neoantigens; FFPE: formalin-fixed paraffin-embedded; NMD: nonsense-mediated decay; NSCLC: non-small cell lung cancer; DC: dendritic cell.

[0060] Please note that 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.

[0061] Terms not directly defined herein should be understood to have the meanings generally associated with them as understood within the technical field of the present invention. Certain terms are discussed herein for the purpose of providing further guidance to the practitioner in describing the compositions, devices, methods, etc. of the present invention's embodiments, as well as their manufacture or use. It will be recognized that multiple ways of saying the same thing can be used. Thus, alternative terms and synonyms may be used for any one or more of the terms discussed herein. No weight should be placed on whether a term is detailed or discussed herein. Some synonyms or substitute methods, materials, etc. are provided. The description of one or several synonyms or equivalents does not exclude the use of other synonyms or equivalents unless expressly stated. The use of examples, including examples of terms, is for illustrative purposes only and does not limit the scope and meaning of the inventive embodiments herein.

[0062] All references, issued patents, and patent applications cited within the body of this specification are hereby incorporated by reference in their entirety for all purposes.

[0063] II. Methods for identifying neoantigens Disclosed herein is a method for identifying neoantigens from a subject's tumor that are likely to be presented on the cell surface of the tumor and / or likely to be immunogenic. By way of example, one such method includes obtaining at least one of tumor exome, transcriptome, or whole genome nucleotide sequencing data from the subject's tumor cells, wherein the tumor nucleotide sequencing data is used to obtain data representing each peptide sequence of a set of neoantigens, wherein the peptide sequence of each neoantigen includes at least one change that makes the peptide sequence different from a corresponding wild-type parent peptide sequence; inputting the peptide sequence of each neoantigen into one or more presentation models to generate a set of numerical likelihoods that each neoantigen will be presented by one or more MHC alleles on the tumor cell surface of the subject's tumor cells or by cells present within the tumor, wherein the set of numerical likelihoods has 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 numerical likelihoods to generate a set of selected neoantigens.

[0064] The presented model can include a statistical regression or machine learning (e.g., deep learning) model trained with a set of reference data (also referred to as a training data set) that includes a corresponding set of labels, the set of reference data being obtained from each of a plurality of separate subjects, some of which may have tumors, and the set of reference data includes at least one of the following: data representing an exome nucleotide sequence from tumor tissue, data representing an exome nucleotide sequence from normal tissue, data representing a transcriptome nucleotide sequence from tumor tissue, data representing a proteome sequence from tumor tissue, and data representing an MHC peptidome sequence from tumor tissue, and data representing an MHC peptidome sequence from normal tissue. The reference data can further include mass spectrometry data, sequencing data, RNA sequencing data, and proteomics data of synthetic proteins, normal and tumor human cell lines, and single allele cell lines engineered to express a given MHC allele that are subsequently exposed to fresh and frozen primary samples, as well as T cell assays (e.g., ELISPOT). In certain embodiments, the set of reference data includes each form of reference data.

[0065] The proposed model can include a set of features derived at least in part from a set of reference data, the set of features including at least one of an allele-dependent feature and an allele-independent feature. In certain embodiments, each feature is included.

[0066] The characteristics of dendritic cell presentation to naive T cells can include at least one of the above characteristics: the dose and type of antigen in the vaccine (e.g., peptide, mRNA, virus, etc.): (1) the route by which dendritic cells (DCs) take up the antigen type (e.g., endocytosis, micropinocytosis); and / or (2) the efficiency with which the antigen is taken up by DCs; the dose and type of adjuvant in the vaccine; the length of the vaccine antigen sequence; the number and site of vaccine administrations; baseline patient immune function (e.g., as measured by recent history of infection, blood counts, etc.). For RNA vaccines, (1) the turnover rate of the mRNA protein product in the dendritic cells, (2) the translation rate of the mRNA after uptake by the dendritic cells, as measured by in vitro or in vivo experiments, and / or (3) the number or rounds of translation of the mRNA after uptake by the dendritic cells, as measured by in vivo or in vitro experiments. Optionally, the presence of protease cleavage motifs in the peptide, giving additional weight to proteases typically expressed in dendritic cells (e.g., as measured by RNA-seq or mass spectrometry). The level of proteasome and immunoproteasome expression in typical activated dendritic cells (which may be measured by RNA-seq, mass spectrometry, immunohistochemistry, or other standard techniques). The level of expression of a particular MHC allele in the individual of interest (e.g., as measured by RNA-seq or mass spectrometry), optionally specifically measured in activated dendritic cells or other immune cells. The probability of peptide presentation by a particular MHC allele in other individuals expressing the particular MHC allele, optionally specifically measured in activated dendritic cells or other immune cells. The probability of peptide presentation by MHC alleles of the same molecular family (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other individuals, optionally specifically measured in activated dendritic cells or other immune cells.

[0067] The immune tolerance escape trait can include at least one of the following: direct measurement of the autopeptidome by protein mass spectrometry performed on one or several cell types; estimation of the autopeptidome by taking the union of all k-mer (e.g., 5-25) substrings of the self-protein; estimation of the autopeptidome using a model of presentation similar to the presentation model described above applied to all non-mutated self-proteins, optionally accounting for germline mutations.

[0068] Ranking can be performed using a plurality of neoantigens provided by at least one model based at least in part on numerical likelihood. After ranking, a selection can be performed to select a subset of the ranked neoantigens according to a selection criterion. After selection, the subset of ranked peptides can be provided as an output.

[0069] The set of selected neoantigens may be twenty in number.

[0070] The presentation model can represent the dependency between the presence of a pair of a particular one of the MHC alleles with a particular amino acid at a particular position in a peptide sequence, and the likelihood of presentation of such a peptide sequence containing a particular amino acid at a particular position on the tumor cell surface by a particular one of the paired MHC alleles.

[0071] The methods disclosed herein may also include applying one or more presentation models to a peptide sequence of a corresponding neoantigen to generate, for each of one or more MHC alleles, a dependency score indicative of whether the MHC allele presents the corresponding neoantigen based on at least the position of an amino acid in the peptide sequence of the corresponding neoantigen.

[0072] The methods disclosed herein may also include transforming the dependency scores to generate a corresponding per-allele likelihood for each MHC allele, the per-allele likelihood indicating the likelihood that the corresponding MHC allele presents the corresponding neoantigen; and combining the per-allele likelihoods to generate a numerical likelihood.

[0073] The step of transforming the dependency scores can model the presentation of peptide sequences of corresponding neoantigens as mutually exclusive.

[0074] The methods disclosed herein may also further include transforming the combination of dependency scores to generate a numerical likelihood.

[0075] Transforming the combination of dependency scores can model the presentation of the peptide sequences of the corresponding neoantigens as interference between MHC alleles.

[0076] The set of numerical likelihoods can be further identified by at least the allele non-interacting feature, and the methods disclosed herein may also include applying an allele non-interacting model of the one or more presentation models to the allele non-interacting feature to generate a dependency score for the allele non-interacting feature indicating whether the peptide sequence of the corresponding neoantigen is presented based on the allele non-interacting feature.

[0077] The methods disclosed herein may also include combining the dependency scores for each MHC allele with the dependency scores for the allele-non-interacting property in one or more MHC alleles; transforming the combined dependency scores for each MHC allele to generate a corresponding per-allele likelihood for the MHC allele, which indicates the likelihood that the corresponding MHC allele presents the corresponding neoantigen; and combining the per-allele likelihoods to generate a numerical likelihood.

[0078] The methods disclosed herein may also include generating a numerical likelihood by transforming a combination of the dependency scores for each of the MHC alleles and the dependency scores for the allele-non-interacting trait.

[0079] The set of numerical parameters for the proposed model can be trained based on a training dataset that includes at least a set of training peptide sequences identified as present in a plurality of samples and one or more MHC alleles associated with each training peptide sequence, where the training peptide sequences are identified by mass spectrometry of isolated peptides eluted from the MHC alleles from the plurality of samples.

[0080] The sample may also include a cell line engineered to express a single MHC class I or class II allele.

[0081] The sample may also include cell lines engineered to express multiple MHC class I or class II alleles.

[0082] Samples may also include human cell lines obtained or derived from multiple patients.

[0083] Samples may also include fresh or frozen tumor samples obtained from multiple patients.

[0084] Samples may also include fresh or frozen tissue samples obtained from multiple patients.

[0085] The sample may also contain peptides identified using a T cell assay.

[0086] The training dataset may further include data relating to the peptide abundance of the set of training peptides present in the sample; the peptide lengths of the set of training peptides in the sample.

[0087] The training dataset can be generated by comparing a set of training peptide sequences by alignment with a database containing a set of known protein sequences, where the set of training protein sequences is longer than and includes the training peptide sequences.

[0088] The training dataset may be generated based on performing, or having previously performed, nucleotide sequencing on the cell line to obtain at least one of exome, transcriptome, or whole genome sequencing data from the cell line, where the sequencing data includes at least one nucleotide sequence that includes the alteration.

[0089] The training dataset may be generated based on obtaining at least one of normal nucleotide sequencing data of the exome, transcriptome, or whole genome from a normal tissue sample.

[0090] The training dataset may further include data relating to proteomic sequences associated with the sample.

[0091] The training dataset may further comprise data relating to MHC peptidome sequences associated with the sample.

[0092] The training dataset may further include data relating to measurements of peptide-MHC binding affinity for at least one of the isolated peptides.

[0093] The training dataset may further include data relating to a measure of peptide-MHC binding stability for at least one of the isolated peptides.

[0094] The training dataset may further include data relating to the transcriptome associated with the sample.

[0095] The training dataset may further include data relating to the genome associated with the sample.

[0096] Training peptide sequences can be in the range of lengths k-mers, where k is between 8 and 15 inclusive for MHC class I, or between 6 and 30 inclusive for MHC class II.

[0097] The methods disclosed herein may also include encoding the peptide sequence using a one-hot encoding scheme.

[0098] The methods disclosed herein may also include encoding the training peptide sequences using a left-padded one-hot encoding scheme.

[0099] A method for treating a subject having a tumor, comprising carrying out the steps described in claim 1, and further comprising the steps of obtaining a tumor vaccine comprising a set of selected neoantigens, and administering the tumor vaccine to the subject.

[0100] Also disclosed herein is a method for producing a tumor vaccine, comprising: 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 a peptide sequence of each of a set of neoantigens, wherein the peptide sequence of each neoantigen comprises at least one mutation that makes the peptide sequence different from a corresponding wild-type parent peptide sequence; inputting the peptide sequence of each neoantigen into one or more presentation models to generate a set of numerical likelihoods that each of the neoantigens will be presented by one or more MHC alleles on the tumor cell surface of the tumor cells of the subject, wherein the set of numerical likelihoods have been identified based at least on received mass spectrometry data; selecting a subset of the set of neoantigens based on the set of numerical likelihoods to generate a set of selected neoantigens; and producing or having produced a tumor vaccine comprising the selected set of neoantigens.

[0101] Also provided herein is a tumor vaccine comprising a set of selected neoantigens selected by performing a method comprising: 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 a peptide sequence for each of a set of neoantigens, wherein the peptide sequence for each neoantigen comprises at least one mutation that renders the peptide sequence different from a corresponding wild-type parent peptide sequence; inputting the peptide sequence for each neoantigen into one or more presentation models to generate a set of numerical likelihoods that each of the neoantigens will be presented by one or more MHC alleles on the tumor cell surface of the tumor cells of the subject, wherein the set of numerical likelihoods have been identified based at least on received mass spectrometry data; selecting a subset of the set of neoantigens based on the set of numerical likelihoods to generate a set of selected neoantigens; and producing or having produced a tumor vaccine comprising the set of selected neoantigens.

[0102] The tumor vaccine may comprise one or more of a nucleotide sequence, a polypeptide sequence, RNA, DNA, a cell, a plasmid, or a vector.

[0103] A tumor vaccine may comprise one or more neoantigens displayed on the surface of tumor cells.

[0104] A tumor vaccine may comprise one or more neoantigens that are immunogenic in a subject.

[0105] A tumor vaccine may not include one or more neoantigens that induce an autoimmune response in a subject against normal tissue.

[0106] The tumor vaccine may include an adjuvant.

[0107] The tumor vaccine may include an excipient.

[0108] The methods disclosed herein may also include selecting neoantigens that have an increased likelihood of being presented on the tumor cell surface relative to neoantigens that are not selected based on the presentation model.

[0109] The methods disclosed herein may also include selecting neoantigens that have an increased likelihood of being able to induce a tumor-specific immune response in a subject against neoantigens that are not selected based on the presentation model.

[0110] The methods disclosed herein may also include selecting neoantigens that have an increased likelihood of being presented to naive T cells by professional antigen-presenting cells (APCs) relative to neoantigens not selected based on a presentation model, optionally where the APCs are dendritic cells (DCs).

[0111] The methods disclosed herein may also include selecting neoantigens that have a reduced likelihood of being inhibited by central or peripheral tolerance to neoantigens that are not selected based on the presentation model.

[0112] The methods disclosed herein may also include selecting neoantigens that have a reduced likelihood of being able to induce an autoimmune response against normal tissue in a subject relative to neoantigens that are not selected based on the presentation model.

[0113] Exome or transcriptome nucleotide sequencing data can be obtained by sequencing tumor tissue.

[0114] Sequencing may be next generation sequencing (NGS) or any massively parallel sequencing approach.

[0115] The set of numerical likelihoods can be further specified by at least MHC allele interaction properties including at least one of the following: predicted affinity of binding between the MHC allele and the neoantigen-encoded peptide; predicted stability of the neoantigen-encoded peptide-MHC complex; sequence and length of the neoantigen-encoded peptide; probability of presentation of a neoantigen-encoded peptide with similar sequence in cells from other individuals expressing the particular MHC allele as assessed by mass spectrometry proteomics or other means; expression level of the particular MHC allele in the subject of interest (e.g., as measured by RNA-seq or mass spectrometry); probability of presentation by the particular MHC allele in other distinct individuals expressing the particular MHC allele, independent of the overall neoantigen-encoded peptide sequence; probability of presentation by MHC alleles of the same molecular family (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other distinct subjects, independent of the overall neoantigen-encoded peptide sequence.

[0116] The set of numerical likelihoods is further specified by at least MHC allele non-interacting properties including at least one of the following: C- and N-terminal sequences flanking the neoantigen-encoded peptide within the source protein sequence; the presence of protease cleavage motifs within the neoantigen-encoded peptide, optionally weighted according to the expression of the corresponding protease in tumor cells (as measured by RNA-seq or mass spectrometry); turnover rate of the source protein measured in the appropriate cell type; measured by RNA-seq or proteomic mass spectrometry or predicted from annotation of germline or somatic splicing variants detected in DNA or RNA sequence data. the length of the source protein, possibly taking into account the particular splice variants ("isoforms") that are most highly expressed in the tumor cells; the level of expression of the proteasome, immunoproteasome, thymoproteasome, or other proteases in the tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry, or immunohistochemistry); the expression of the source gene of the neoantigen-encoded peptide (e.g., measured by RNA-seq or mass spectrometry); the typical tissue-specific expression of the source gene of the neoantigen-encoded peptide in different phases of the cell cycle; e.g., as determined by uniProt or PDB a comprehensive catalog of properties of the source protein and / or its domains, as can be found at http: / / www.rcsb.org / pdb / home / home.do; properties describing the nature of the domain of the source protein that contains the peptide, e.g., secondary or tertiary structure (e.g., alpha helix versus beta sheet); alternative splicing; the probability of presentation of the peptide derived from the source protein of the neoantigen-encoded peptide of interest in other separate subjects; the probability that the peptide will be undetected or over-represented by mass spectrometry due to technical bias; the expression of various gene modules / pathways (not necessarily including the source protein of the peptide) as measured by RNASeq, which informs about the status of the tumor cells, stroma, or tumor infiltrating lymphocytes (TILs); the copy number of the source gene of the neoantigen-encoded peptide in the tumor cells;the probability that the peptide will bind to TAP, or the measured or predicted binding affinity of the peptide to TAP; the expression level of TAP in the tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry, immunohistochemistry); the presence or absence of tumor mutations, including but not limited to: driver mutations in known cancer driver genes such as EGFR, KRAS, ALK, RET, ROS1, TP53, CDKN2A, CDKN2B, NTRK1, NTRK2, NTRK3, and genes encoding proteins involved in the antigen presentation machinery (e.g., B2M, HLA- A, HLA-B, HLA-C, TAP-1, TAP-2, TAPBP, CALR, CNX, ERP57, HLA-DM, HLA-DMA, HLA-DMB, HLA-DO, HLA-DOA, HLA-DOB, HLA-DP, HLA-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 a mutation in any of the genes encoding components of the proteasome or immunoproteasome. Peptides whose presentation depends on components of the antigen presentation machinery that carry loss-of-function mutations in the tumor have a low probability of presentation; the presence or absence of functional germline polymorphisms, including but not limited to: genes encoding proteins involved in antigen presentation machinery (e.g., B2M, HLA-A, HLA-B, HLA-C, TAP-1, TAP-2, TAPBP, CALR, CNX, ERP57, HLA-DM, HLA-DMA, HLA-DMB, HLA-DO, HLA-DOA, HLA-DOB, HLA-D ... polymorphisms in HLA-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 of the genes encoding components of the proteasome or immunoproteasome; tumor type (e.g., NSCLC, melanoma); clinical tumor subtype (e.g., squamous cell lung cancer vs. non-squamous); smoking history;Typical expression of the peptide's source gene in the relevant tumor type or clinical subtype, optionally stratified by driver mutations;

[0117] The at least one mutation may be a frameshift or non-frameshift indel, a missense or nonsense substitution, a splice site alteration, a genomic rearrangement or gene fusion, or any genomic or expression alteration that results in a nascent ORF.

[0118] The tumor cells can be 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.

[0119] 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.

[0120] At least one of the neoantigens in the set of selected neoantigens, when in polypeptide form, may comprise at least one of the following: a binding affinity to MHC with an IC50 value of less than 1000 nM; a length of 8-15, 8, 9, 10, 11, 12, 13, 14, or 15 amino acids for MHC class I polypeptides, or a length of 6-30, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 amino acids for MHC class II 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. In MHC class II, the presence of sequence motifs within or near the peptide that promote HLA binding catalyzed by cleavage by extracellular or lysosomal proteases (eg, cathepsins) or by HLA-DM.

[0121] Also disclosed herein is a method for generating a model for identifying one or more neoantigens likely to be presented on a tumor cell surface of a tumor cell, the method comprising: receiving mass spectrometry data comprising data relating to a plurality of isolated peptides eluted from major histocompatibility complexes (MHC) from a plurality of samples; obtaining a training dataset by at least identifying a set of training peptide sequences present in the samples and one or more MHCs associated with each training peptide sequence; and training a set of numerical parameters of a presentation model using the training dataset comprising the training peptide sequences, the presentation model providing a plurality of numerical likelihoods that a peptide sequence derived from a tumor cell will be presented by one or more MHC alleles on the tumor cell surface.

[0122] 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 of a peptide sequence having a particular amino acid at a particular position being presented by one of the MHC alleles on a tumor cell.

[0123] The sample may also include a cell line engineered to express a single MHC class I or class II allele.

[0124] The sample may also include cell lines engineered to express multiple MHC class I or class II alleles.

[0125] Samples may also include human cell lines obtained or derived from multiple patients.

[0126] Samples may also include fresh or frozen tumor samples obtained from multiple patients.

[0127] The sample may also contain peptides identified using a T cell assay.

[0128] The training dataset may further include data relating to the peptide abundances of the set of training peptides present in the sample; the peptide lengths of the set of training peptides in the sample.

[0129] The methods disclosed herein may also include obtaining a set of training protein sequences based on the training peptide sequences, the set of training peptide sequences being longer than and including the training peptide sequences, by comparing the set of training peptide sequences by alignment with a database comprising a set of known protein sequences.

[0130] The methods disclosed herein may also include performing or having previously 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, wherein the nucleotide sequencing data comprises at least one protein sequence that includes the mutation.

[0131] The methods disclosed herein may also include encoding the training peptide sequences using a one-hot encoding scheme.

[0132] The methods disclosed herein can also include obtaining at least one of exome, transcriptome, and whole genome normal nucleotide sequencing data from a normal tissue sample, and training a set of parameters of the proposed model using the normal nucleotide sequencing data.

[0133] The training dataset may further include data relating to proteomic sequences associated with the sample.

[0134] The training dataset may further comprise data relating to MHC peptidome sequences associated with the sample.

[0135] The training dataset may further include data relating to measurements of peptide-MHC binding affinity for at least one of the isolated peptides.

[0136] The training dataset may further include data relating to a measure of peptide-MHC binding stability for at least one of the isolated peptides.

[0137] The training dataset may further include data relating to the transcriptome associated with the sample.

[0138] The training dataset may further include data relating to the genome associated with the sample.

[0139] The methods disclosed herein may also include performing a logistic regression of the set of parameters.

[0140] Training peptide sequences can be in the range of lengths k-mers, where k is between 8 and 15 inclusive for MHC class I, or between 6 and 30 inclusive for MHC class II.

[0141] The methods disclosed herein may also include encoding the training peptide sequences using a left-padded one-hot encoding scheme.

[0142] The methods disclosed herein may also include determining values ​​for the set of parameters using a deep learning algorithm.

[0143] Disclosed herein is a method for identifying one or more neoantigens likely to be presented on the tumor cell surface of tumor cells, the method comprising the steps of receiving mass spectrometry data comprising data relating to a plurality of isolated peptides eluted from major histocompatibility complexes (MHC) derived from a plurality of fresh or frozen-like samples; obtaining a training dataset by identifying at least a set of training peptide sequences present in the tumor samples and 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 training a set of numerical parameters of a presentation model using the training protein sequences and the training peptide sequences, wherein the presentation model provides a plurality of numerical likelihoods that a peptide sequence derived from a tumor cell will be presented by one or more MHC alleles on the tumor cell surface.

[0144] The presentation model may represent the dependency between the presence of a pair of a particular one of the MHC alleles with a particular amino acid at a particular position in a peptide sequence, and the likelihood of presentation of such a peptide sequence comprising said particular amino acid at said particular position on the tumor cell surface by said particular one of the MHC alleles of said pair.

[0145] The methods disclosed herein may also include selecting a subset of neoantigens, each of which is selected because it has an increased likelihood of being presented on the cell surface of the tumor relative to one or more distinct tumor neoantigens.

[0146] The methods disclosed herein may also include selecting a subset of neoantigens, each of which has an increased likelihood of being able to induce a tumor-specific immune response in a subject against one or more distinct tumor neoantigens.

[0147] The methods disclosed herein may also include selecting a subset of neoantigens, each of which has an increased likelihood relative to one or more distinct tumor neoantigens that it can be presented to naive T cells by a professional antigen presenting cell (APC), optionally where the APC is a dendritic cell (DC).

[0148] The methods disclosed herein may also include selecting a subset of neoantigens, each of which has a reduced likelihood of being inhibited by central or peripheral tolerance to one or more distinct tumor neoantigens.

[0149] The methods disclosed herein may also include selecting a subset of neoantigens, each of which has a reduced likelihood of being able to induce an autoimmune response against normal tissue in a subject against one or more distinct tumor neoantigens.

[0150] The methods disclosed herein may also include selecting a subset of neoantigens, each of which has a reduced likelihood of being differentially post-translationally modified in tumor cells versus APCs, and optionally, the APCs are dendritic cells (DCs).

[0151] The practice of the methods herein will employ, unless otherwise indicated, conventional methods of protein chemistry, biochemistry, recombinant DNA technology, and pharmacology within the skill of the art. Such techniques are fully explained in the literature. See, e.g., TECreighton, Proteins: Structures and Molecular Properties (WH Freeman and Company, 1993); ALLehninger, Biochemistry (Worth Publishers, Inc., current addition); Sambrook, et al., Molecular Cloning: A Laboratory Manual (2nd Edition, 1989); Methods In Enzymology (S. Colowick and N. Kaplan eds., Academic Press, Inc.); Remington's Pharmaceutical Sciences, 18th Edition (Easton, Pennsylvania: Mack Publishing Company, 1990); Carey and Sundberg Advanced Organic Chemistry 3rd Ed. (Plenum Press) Vols A and B (1992).

[0152] A set of presentation likelihoods can also be generated based on the source genes of the set of neoantigens.

[0153] A set of presentation likelihoods can also be generated based on the source genes and source tissue type of the set of neoantigens.

[0154] The methods disclosed herein may include identifying a subset of patients suitable for treatment with a neoantigen vaccine, the steps including: obtaining, for each patient, at least one of exome, transcriptome, or whole genome tumor nucleotide sequencing data from the patient's tumor cells, wherein the tumor nucleotide sequencing data is used to obtain a peptide sequence for each of a set of neoantigens, wherein each neoantigen peptide sequence includes at least one alteration that makes it different from a corresponding wild-type parent peptide sequence; generating, for each patient, a set of numerical presentation likelihoods for the set of neoantigens for the patient by inputting the peptide sequences of each of the set of neoantigens into one or more presentation models, the set of presentation likelihoods representing the likelihood that each of the set of neoantigens will be presented by one or more MHC alleles on the surface of the patient's tumor cells, the set of presentation likelihoods having been determined based at least on the received mass spectrometry data; identifying for each patient a therapeutic subset of neoantigens from the set of neoantigens for the patient, the therapeutic subset corresponding to a predetermined number of neoantigens having the highest presentation likelihood within the set of presentation likelihoods generated for that patient; Selecting a subset of patients suitable for treatment with a neoantigen vaccine, wherein the selected subset of patients meets inclusion criteria based on a set of neoantigens obtained for each patient in the selected subset or based on tumor nucleotide sequencing data.

[0155] The methods disclosed herein may include treating each patient in a selected subset of patients with a corresponding neoantigen vaccine, where the neoantigen vaccine for the patient comprises a therapeutic subset identified by a set of presentation likelihoods for the patient.

[0156] The methods disclosed herein may include selecting a subset of patients having a tumor mutational burden (TMB) higher than a minimum threshold, where the TMB of a patient indicates the number of neoantigens in a set of neoantigens associated with that patient.

[0157] The methods disclosed herein may include identifying for each patient a utility score indicating a measure of the estimated number of neoantigens presented from a treatment subset of patients; and selecting a subset of patients having a utility score higher than a minimum threshold.

[0158] Presentation of neoantigens can be modeled as a Bernoulli random variable, and the utility score can represent the expected number of presented neoantigens in a therapeutic subset for a patient, and the utility score can be given by the sum of the likelihood of presentation for each neoantigen in the therapeutic subset of the patient.

[0159] Presentation of neoantigens can also be modeled as a Poisson binomial random variable, and the utility score can be the probability that the number of presented neoantigens in the treatment subset for a patient is above a minimum threshold.

[0160] III. Identification of tumor-specific mutations in neoantigens Also disclosed herein are methods for identifying certain mutations (e.g., mutations or alleles present in cancer cells). In particular, these mutations may be present in the genome, transcriptome, proteome, or exome of cancer cells of a subject with cancer, but may not be present in normal tissues from the subject.

[0161] Genetic mutations in tumors can be considered useful for tumor immunological targeting if they cause changes in the amino acid sequence of a protein exclusively in tumors. Useful mutations include: (1) nonsynonymous 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 the 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 an intron in mature mRNA, thus resulting in a unique tumor-specific protein sequence; (4) chromosomal rearrangements (i.e., gene fusions) that result in chimeric proteins with tumor-specific sequences at the junction of two proteins; (5) frameshift mutations or deletions that result in new open reading frames with new tumor-specific protein sequences. Mutations can also include one or more of nonframeshift insertions, missense or nonsense substitutions, splice site changes, genomic rearrangements or gene fusions, or any genomic or expression changes that result in new ORFs.

[0162] For example, mutated peptides or mutated polypeptides resulting from splice site, frameshift, readthrough, or gene fusion mutations in tumor cells can be identified by sequencing DNA, RNA, or protein in tumor versus normal cells.

[0163] 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.

[0164] A variety of methods are available for detecting the presence of specific mutations or alleles in an individual's DNA or RNA. Advances in this field have provided accurate, easy, and inexpensive large-scale SNP genotyping. Several techniques have been described, including, for example, dynamic allele-specific hybridization (DASH), microplate array diagonal gel electrophoresis (MADGE), pyrosequencing, oligonucleotide-specific ligation, the TaqMan system, and various DNA "chip" technologies such as the Affymetrix SNP chip. These methods utilize the amplification of target gene regions, typically by PCR. Still other methods are based on the generation of small signal molecules by invasive cleavage followed by mass spectrometry, or on immobilized padlock probes and rolling circle amplification. Some of the methods known in the art for detecting specific mutations are summarized below.

[0165] PCR-based detection means can include multiplex amplification of multiple markers simultaneously. For example, it is well known in the art to select PCR primers to generate PCR products that do not overlap in size and can be analyzed simultaneously. Alternatively, it is possible to amplify different markers with primers that are differentially labeled and therefore can be differentially detected. Of course, hybridization-based detection means allows the differential detection of multiple PCR products in a sample. Other techniques that allow multiplex analysis of multiple markers are known in the art.

[0166] Several methods have been developed to facilitate the analysis of single nucleotide polymorphisms in genomic DNA or cellular RNA. For example, single nucleotide polymorphisms can be detected by using specialized exonuclease-resistant nucleotides, as disclosed, for example, in Mundy, CR (US Pat. No. 4,656,127). According to this method, a primer that is complementary to the allele sequence immediately 3' of the polymorphic site is hybridized to a target molecule obtained from a specific animal or human. If the polymorphic site on the target molecule contains a nucleotide that is complementary to the specific exonuclease-resistant nucleotide derivative present, the derivative is incorporated onto the end of the hybridized primer. Due to such incorporation, the primer becomes resistant to exonucleases, thereby allowing its detection. Since the identity of the exonuclease-resistant derivative of the sample is known, the knowledge that the primer has become resistant to exonucleases reveals that the nucleotide present at the polymorphic site of the target molecule is complementary to that of the nucleotide derivative used in the reaction. This method has the advantage that it does not require the determination of large amounts of exogenous sequence data.

[0167] Solution-based methods can be used to determine the identity of the nucleotide at a polymorphic site (Cohen, D. et al. (French Patent No. 2,650,840; PCT Application No. WO 91 / 02087)). As in the method of Mundy in U.S. Pat. No. 4,656,127, a primer is used that is complementary to the allelic sequence immediately 3' to the polymorphic site. This method uses a labeled dideoxynucleotide derivative that becomes incorporated onto the end of the primer if it is complementary to the nucleotide at the polymorphic site to determine the identity of the nucleotide at that site. An alternative method, known as Genetic Bit Analysis or GBA, has been described by Goelet, P. et al. (PCT Application No. 92 / 15712). The Goelet, P. et al. method uses a mixture of labeled terminators and a primer that is complementary to the sequence 3' to the polymorphic site. Goelet, P. et al. The method of Goelet, P. et al. uses a mixture of labeled terminators and a primer that is complementary to the sequence 3' of the polymorphic site. In contrast to the method of Cohen et al. (French Patent No. 2,650,840; PCT Application No. WO 91 / 02087), the method of Goelet, P. et al. can be a heterogeneous phase assay in which either the primer or the target molecule is immobilized on a solid phase.

[0168] Several primer-guided nucleotide incorporation procedures for assaying polymorphic sites in DNA have been described (Komher, J Set al., Nucl. Acids. Res. 17:7779-7784 (1989); Sokolov, B. P., Nucl. Acids Res. 18:3671 (1990); Syvanen, A.-C., et al., Genomics 8:684-692 (1990); Kuppuswamy, M Net al., Proc. Natl. Acad. Sci. (USA) 88:1143-1147 (1991); Prezant, T. R. et al., Hum. Mutat. 1:159-164 (1992); Ugozzoli, L. et al., GATA 9:107-112 (1992); Nyren, P. et al., GATA 9:107-112 (1992); al., Anal. Biochem. 208:171-175 (1993)). These methods differ from GBA in that they utilize the incorporation of labeled deoxynucleotides to discriminate between bases at the polymorphic site. In such formats, the signal is proportional to the number of deoxynucleotides incorporated, so that polymorphisms occurring in runs of the same nucleotides can result in a signal proportional to the length of the run (Syvanen, A.-C., et al., Amer. J. Hum. Genet. 52:46-59 (1993)).

[0169] Numerous initiatives obtain sequence information directly from millions of individual molecules of DNA or RNA in parallel. Real-time single molecule sequencing by synthesis techniques rely on the detection of fluorescent nucleotides as they are incorporated into nascent strands of DNA that are complementary to the template to be sequenced. In one method, oligonucleotides 30-50 bases in length are covalently anchored at their 5' ends to a glass coverslip. These anchored strands serve two functions. First, they act as capture sites for the target template strands when the template is constructed with a capture tail complementary to the surface-bound oligonucleotide. They also act as primers for template-directed primer extension, which forms the basis of sequence reading. The capture primers serve as fixed location sites for multiple cycles of synthesis, detection, and sequencing using chemical cleavage of the dye-linker to remove the dye. Each cycle consists of the addition of a polymerase / labeled nucleotide mix, rinsing, imaging, and cleavage of the dye. In an alternative method, the polymerase is modified with a fluorescent donor molecule and immobilized on a glass slide, while each nucleotide is color-coded with an acceptor fluorescent moiety attached to the γ-phosphate. As the nucleotide becomes incorporated into the new strand, the system detects the interaction between the fluorescently tagged polymerase and the fluorescently modified nucleotide. Other sequencing-by-synthesis techniques also exist.

[0170] Any suitable sequencing-by-synthesis platform can be used to identify mutations. As mentioned above, four major sequencing-by-synthesis platforms are currently available: the Genome Sequencer sold by Roche / 454 Life Sciences, the 1G Analyzer sold by Illumina / Solexa, the SOLiD system sold by Applied BioSystems, and the Heliscope system sold by Helicos Bioscience. Sequencing-by-synthesis platforms have also been described by Pacific BioSciences and VisiGen Biotechnologies. In some embodiments, the multiple nucleic acid molecules to be sequenced are bound to a support (e.g., a solid support). To immobilize the nucleic acid on the support, a capture sequence / universal priming site can be added to the 3' and / or 5' end of the template. The nucleic acid can be bound to the support by hybridizing the capture sequence to a complementary sequence covalently attached to the support. A capture sequence (also called a universal capture sequence) is a nucleic acid sequence complementary to a sequence attached to a support that can double as a universal primer.

[0171] As an alternative to capture sequences, members of a coupling pair (e.g., antibody / antigen, receptor / ligand, or an avidin-biotin pair, e.g., as described in U.S. Patent Application Publication No. 2006 / 0252077) can be linked to each fragment and captured onto a surface coated with the respective second member of the coupling pair.

[0172] Following capture, the sequence can be analyzed by single molecule detection / sequencing, e.g., as described in the Examples and in U.S. Pat. No. 7,283,337, including template-dependent sequencing by synthesis. In sequencing by synthesis, the surface-bound molecules are exposed to a multitude of labeled nucleotide triphosphates in the presence of a polymerase. The sequence of the template is determined by the order of labeled nucleotides that are incorporated into the 3' end of the growing strand. This can be done in real time and in a step-and-repeat mode. For real-time analysis, a different optical label can be incorporated for each nucleotide and multiple lasers can be utilized for stimulation of the incorporated nucleotides.

[0173] Sequencing can also include other massively parallel sequencing or next generation sequencing (NGS) techniques and platforms. Additional examples of massively parallel sequencing techniques and platforms are Illumina HiSeq or MiSeq, ThermoPGM or Proton, Pac Bio RS II or Sequel, Qiagen's Gene Reader, and Oxford Nanopore MinION. Additional similar current massively parallel sequencing technologies, and future generations of these technologies, can be used.

[0174] Any cell type or tissue can be utilized to obtain nucleic acid samples for use in the methods described herein. For example, DNA or RNA samples can be obtained from tumors or body fluids, such as blood obtained by known techniques (e.g., venipuncture), or saliva. Alternatively, nucleic acid testing can be performed on dry samples (e.g., hair or skin). In addition, a sample can be obtained from tumors for sequencing, and another sample can be obtained from normal tissues for sequencing, if the normal tissues are of the same tissue type as the tumor. A sample can be obtained from tumors for sequencing, and another sample can be obtained from normal tissues for sequencing, if the normal tissues are of a different tissue type from the tumor.

[0175] The tumor may include one or more 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.

[0176] Alternatively, protein mass spectrometry can be used to identify or demonstrate the presence of mutated peptides bound to MHC proteins on tumor cells. Peptides can be acid eluted from tumor cells or from HLA molecules immunoprecipitated from tumors and then identified using mass spectrometry.

[0177] IV. Neoantigens Neoantigens can comprise nucleotides or polynucleotides. For example, neoantigens can be RNA sequences that code for a polypeptide sequence. Neoantigens useful in vaccines can thus comprise nucleotide sequences or polypeptide sequences.

[0178] 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. Neoantigen peptides can be described in the context of their coding sequences when the neoantigens comprise nucleotide sequences (e.g., DNA or RNA) that encode the associated polypeptide sequence.

[0179] The one or more polypeptides encoded by the neoantigen nucleotide sequences can include at least one of the following: binding affinity to MHC with an IC50 value of less than 1000 nM, a length of 8-15, 8, 9, 10, 11, 12, 13, 14, or 15 amino acids for MHC class I peptides, the presence of sequence motifs within or near the peptide that promote proteasomal cleavage, and the presence of sequence motifs that promote TAP transport, a length of 6-30, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, or 30 amino acids for MHC class II peptides, and the presence of sequence motifs within or near the peptide that promote cleavage by extracellular or lysosomal proteases (e.g., cathepsins) or HLA binding catalyzed by HLA-DM.

[0180] One or more neoantigens can be present on the surface of the tumor.

[0181] One or more neoantigens can be immunogenic in a tumor-bearing subject, for example, capable of eliciting a T cell or B cell response in the subject.

[0182] One or more neoantigens that induce an autoimmune response in a subject can be eliminated from consideration in the context of generating a vaccine for a tumor-bearing subject.

[0183] The size of the at least one neoantigenic peptide molecule is 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 60, about 70, about 80, about 90, about 100, about 110, about 120 or more amino acid residues, and any range derivable therein. In a specific embodiment, the neoantigenic peptide molecule is 50 amino acids or less.

[0184] Neoantigenic peptides and polypeptides can be 15 residues or less in length for MHC class I, usually between about 8 and about 11 residues, particularly 9 or 10 residues; for MHC class II, they can be 6 to 30 residues in length.

[0185] If desired, longer peptides can be designed in several ways. In one example, where the presentation likelihood of a peptide on an HLA allele is predicted or known, the longer peptide can consist of either (1) an individual presented peptide with a 2-5 amino acid extension toward the N-terminus and C-terminus of each corresponding gene product; (2) a concatenation of some or all of the presented peptides, each with an extended sequence. In another example, where sequencing reveals long (longer than 10 residues) neo-epitope sequences present in the tumor (e.g., due to frameshifts, read-throughs, or inclusion of introns resulting in novel peptide sequences), the longer peptide will (3) consist of the entire stretch of novel tumor-specific amino acids, thus avoiding the need for computational or in vitro test-based selection of the strongest HLA-presented shorter peptide. In either example, the use of longer peptides may allow for endogenous processing by patient cells, resulting in more effective antigen presentation and induction of T cell responses.

[0186] The neoantigenic peptides and polypeptides can be presented on HLA proteins. In some embodiments, the neoantigenic peptides and polypeptides are presented on HLA proteins with stronger affinity than the wild-type peptides. In some embodiments, the neoantigenic peptides or polypeptides can have an IC50 of at least 5000nM, at least 1000nM, at least 500nM, at least 250nM, at least 200nM, at least 150nM, at least 100nM, at least 50nM, or less.

[0187] In some embodiments, the neoantigenic peptides and polypeptides do not induce an autoimmune response and / or do not cause immune tolerance when administered to a subject.

[0188] Also provided is a composition comprising at least two or more neoantigenic peptides. In some embodiments, the composition contains at least two different peptides. The at least two different peptides can be derived from the same polypeptide. Different polypeptides means that the peptides differ in length, amino acid sequence, or both. The peptides are derived from any polypeptide that is known or found to contain tumor-specific mutations. Suitable polypeptides from which neoantigenic peptides can be derived can be found, for example, in the COSMIC database. COSMIC manages comprehensive information on somatic mutations in human cancers. The peptides contain tumor-specific mutations. In some embodiments, the tumor-specific mutations are driver mutations for a particular cancer type.

[0189] Neoantigenic peptides and polypeptides with desired activities or properties can be modified to confer certain desired attributes, e.g., improved pharmacological characteristics, while at least retaining or enhancing substantially all of the biological activity of the unmodified peptides that bind to desired MHC molecules and activate appropriate T cells. By way of example, the neoantigenic peptides and polypeptides can be further subjected to various modifications, such as either conservative or non-conservative substitutions, which may provide certain advantages in their use, such as improved MHC binding, stability, or presentation. Conservative substitution means replacing an amino acid residue with another that is biologically and / or chemically similar, e.g., one hydrophobic residue with another hydrophobic residue, or one polar residue with another polar residue. Substitutions include combinations such as Gly, Ala; Val, Ile, Leu, Met; Asp, Glu; Asn, Gln; Ser, Thr; Lys, Arg; and Phe, Tyr. The effect of single amino acid substitutions may also be explored using D-amino acids. Such modifications can be carried out 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 & Young, Solid Phase Peptide Synthesis, (Rockford, Ill., Pierce), 2d Ed. (1984).

[0190] Modification of peptides and polypeptides with various amino acid mimetics or unnatural amino acids can be particularly useful in increasing the stability of peptides and polypeptides in vivo. Stability can be assayed in a number of ways. For example, peptidases and various biological media, such as human plasma and serum, have been used to test stability. See, for example, Verhoef et al., Eur. J. Drug Metab Pharmacokin. 11:291-302 (1986). Peptide half-life can be conveniently determined using a 25% human serum (v / v) assay. The protocol is generally as follows: Pooled human serum (type AB, non-heat inactivated) is delipidated by centrifugation before use. The serum is then diluted to 25% with RPMI tissue culture medium and used to test peptide stability. At predetermined time intervals, small aliquots of the reaction solution are removed and added to either 6% aqueous trichloroacetic acid or ethanol. The cloudy reaction sample is cooled (4° C.) for 15 minutes and then spun to precipitate the precipitated serum proteins. The presence of the peptide is then determined by reverse-phase HPLC using stability-specific chromatographic conditions.

[0191] Peptides and polypeptides can be modified to provide desirable attributes other than improved serum half-life. By way of example, the ability of a peptide to induce CTL activity can be enhanced by linkage to a sequence containing at least one epitope capable of inducing a T helper cell response. The immunogenic peptide / T helper conjugate can be linked by a spacer molecule. The spacer is typically composed of relatively small neutral molecules, such as amino acids or amino acid mimetics, that are substantially uncharged under physiological conditions. The spacer is typically selected, for example, from Ala, Gly, or other neutral spacers of non-polar amino acids or neutral polar amino acids. It will be understood that the spacer, if present, need not be composed of the same residues and thus may be a hetero- or homo-oligomer. If present, the spacer will usually be at least 1 or 2 residues, more usually 3-6 residues. Alternatively, the peptide can be linked to the T helper peptide without a spacer.

[0192] The neoantigenic peptide can be linked to the T helper peptide either directly or via a spacer at either the amino or carboxy terminus of the peptide. The amino terminus of either the neoantigenic peptide or the T helper peptide can be acylated. Exemplary T helper peptides include 830-843 of tetanus toxoid, 307-319 of influenza, and around 382-398 and 378-389 of malaria sporozoites.

[0193] Proteins or peptides can be produced by any technique known to those skilled in the art, including expressing the protein, polypeptide, or peptide through standard molecular biology techniques, isolating the protein or peptide from a natural source, or chemically synthesizing the protein or peptide. Nucleotide and protein, polypeptide, and peptide sequences corresponding to various genes have been previously disclosed and can be found in computerized databases known to those skilled in the art. One such database is the Genbank and GenPept databases of the National Center for Biotechnology Information, located at the National Institutes of Health website. The coding regions of known genes can be amplified and / or expressed using the techniques disclosed herein or as known to those skilled in the art. Alternatively, various commercial preparations of proteins, polypeptides, and peptides are known to those skilled in the art.

[0194] In a further embodiment, the neoantigen comprises a nucleic acid (e.g., polynucleotide) encoding a neoantigenic peptide or a portion thereof. The polynucleotide can be, for example, DNA, cDNA, PNA, CNA, RNA (e.g., mRNA), a polynucleotide having, for example, a phosphorothioate backbone, either single-stranded and / or double-stranded, or in a naturally occurring or stabilized form, or a combination thereof, and may or may not contain introns. Yet a further embodiment provides an expression vector capable of expressing the polypeptide or a portion thereof. Expression vectors for various cell types are well known in the art and can be selected without undue experimentation. Generally, the DNA is inserted into an expression vector, such as a plasmid, in the proper orientation and correct reading frame for expression. If necessary, the DNA can be linked to appropriate transcriptional and translational regulatory control nucleotide sequences recognized by the desired host, although such controls are generally available in the expression vector. The vector is then introduced into the host through standard techniques. Guidance can be found, for example, in Sambrook et al. (1989) Molecular Cloning, A Laboratory Manual, Cold Spring Harbor Laboratory, Cold Spring Harbor, NY.

[0195] IV. Vaccine Compositions Also disclosed herein are immunogenic compositions, e.g., vaccine compositions, that can generate a specific immune response, e.g., a tumor-specific immune response. Vaccine compositions typically include multiple neoantigens, e.g., selected using the methods described herein. Vaccine compositions can also be referred to as vaccines.

[0196] The vaccine can contain 1-30 peptides, 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 can include post-translational modifications. The vaccine may contain from 1 to 100 or more nucleotide sequences, 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, 101, 102, 103, 104, 105, 106, 107, 108, 109, 109, 109, 109, 102, 104, 105, 106, 107, 108, 109, 109, 109 It may contain 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 contains 1-30 neoantigen sequences, 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, 101, 102, 103, 104, 105, 106, 107, 108, 109, 109, 109, 101, 102, 103, 104, 105, 106, 107, 108, 109, 109, 110 It can contain 6, 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 neoantigen sequences, 6, 7, 8, 9, 10, 11, 12, 13, or 14 different neoantigen sequences, or 12, 13, or 14 different neoantigen sequences.

[0197] In one embodiment, the different peptides and / or polypeptides, or the nucleotide sequences encoding them, are selected such that the peptides and / or polypeptides are capable of binding to different MHC molecules, such as different MHC class I molecules and / or different MHC class II molecules. In some embodiments, a vaccine composition comprises coding sequences for peptides and / or polypeptides capable of binding to the most frequently occurring MHC class I molecules and / or MHC class II molecules. Thus, the vaccine composition can comprise different fragments capable of binding to at least two preferred, at least three preferred, or at least four preferred MHC class I molecules and / or MHC class II molecules.

[0198] The vaccine composition may be capable of generating a specific cytotoxic T cell response and / or a specific helper T cell response.

[0199] The vaccine composition may further comprise an adjuvant and / or a carrier. Examples of useful adjuvants and carriers are provided herein below. The composition may be combined with a carrier, such as, for example, a protein, or an antigen-presenting cell, such as, for example, a dendritic cell (DC), capable of presenting peptides to T cells.

[0200] An adjuvant is any substance whose incorporation into a vaccine composition increases or otherwise modifies the immune response to a neoantigen. The carrier can be a scaffold structure, such as a polypeptide or polysaccharide, to which the neoantigen can be bound. Optionally, the adjuvant is covalently or non-covalently conjugated.

[0201] The ability of adjuvant to increase immune response to antigen is typically manifested by a significant or substantial increase in immune-mediated reaction or a reduction in disease symptoms.For example, the increase in humoral immunity is typically manifested by a significant increase in the titer of antibody generated against antigen, and the increase in T cell activity is typically manifested in an increase in cell proliferation, or cellular cytotoxicity, or cytokine secretion.Adjuvant can also change immune response, for example, by changing a predominantly humoral or Th response to a predominantly cellular or Th response.

[0202] Suitable adjuvants include 1018 ISS, alum, aluminum salts, Amplivax, AS15, BCG, CP-870,893, CpG7909, CyaA, dSLIM, GM-CSF, IC30, IC31, Imiquimod, ImuFact IMP321, IS Patch, ISS, ISCOMATRIX, JuvImmune, LipoVac, MF59, monophosphoryl lipid A, Montanide IMS 1312, Montanide ISA206, Montanide ISA 50V, Montanide Adjuvants include, but are not limited to, 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 trap, R848, β-glucan, Pam3Cys, Aquila's QS21 stimulon (Aquila Biotech, Worcester, Mass., USA) derived from saponins, mycobacterial extracts and synthetic bacterial cell wall mimics, and other proprietary adjuvants such as Ribi's Detox.Quil or Superfos. Adjuvants such as incomplete Freund's or GM-CSF are useful. Several immunological adjuvants specific for dendritic cells (e.g., MF59) and their preparations have been previously described (Dupuis M, et al., Cell Immunol. 1998; 186(1):18-27; Allison AC; Dev Biol Stand. 1998; 92:3-11). Cytokines can also be used. Several cytokines have been directly linked to influencing dendritic cell migration to lymphoid tissues (e.g., TNF-α), accelerating dendritic cell maturation into efficient 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 immune adjuvants (e.g., IL-12) (Gabrilovich DI, et al., J Immunother Emphasis Tumor Immunol. 1996(6):414-418).

[0203] CpG immunostimulatory oligonucleotides have also been reported to enhance the effect of adjuvants in a vaccine setting. Other TLR binding molecules, such as RNA that binds to TLR 7, TLR 8, and / or TLR 9, may also be used.

[0204] Other examples of useful adjuvants include, but are not limited to, chemically modified CpG (e.g., CpR, Idera), Poly(I:C) (e.g., polyi: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 may act therapeutically and / or as adjuvants. The amounts and concentrations of adjuvants and additives can be readily determined by one of ordinary skill in the art without undue experimentation. Additional adjuvants include colony stimulating factors, such as granulocyte-macrophage colony stimulating factor (GM-CSF, sargramostim).

[0205] Vaccine compositions can include more than one different adjuvant. Additionally, therapeutic compositions can include any adjuvant material, including any of the above or combinations thereof. It is also contemplated that the vaccine and adjuvant can be administered together or separately in any suitable sequence.

[0206] The carrier (or excipient) can be present independent of the adjuvant. The function of the carrier can be, for example, to increase the activity or immunogenicity, to provide stability, to increase biological activity, or to increase serum half-life, particularly to increase the molecular weight of the variant. In addition, the carrier can help present the peptide to T cells. The carrier can be any suitable carrier known to those skilled in the art, for example, a protein or an antigen-presenting cell. The carrier protein can be, but is not limited to, a serum protein such as keyhole limpet hemocyanin, transferrin, bovine serum albumin, human serum albumin, thyroglobulin or ovalbumin, an immunoglobulin, or a hormone such as insulin, or palmitic acid. For human immunization, the carrier is generally a physiologically acceptable carrier that is acceptable and safe for humans. However, tetanus toxoid and / or diphtheria toxoid are suitable carriers. Alternatively, the carrier can be a dextran, for example, sepharose.

[0207] Cytotoxic T cells (CTLs) recognize antigens in the form of peptides bound to MHC molecules rather than intact foreign antigens themselves. The MHC molecules themselves are located on the cell surface of antigen-presenting cells. Thus, activation of CTLs is possible when a trimeric complex of peptide antigens, MHC molecules, and APCs is present. Correspondingly, it can enhance immune responses not only when peptides are used for activating CTLs, but also when APCs with the respective MHC molecules are added in addition. Thus, in some embodiments, the vaccine composition additionally contains at least one antigen-presenting cell.

[0208] Neoantigens may also be derived from viruses such as vaccinia, fowlpox, self-replicating alphaviruses, Maraba viruses, adenoviruses (see, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or lentiviruses, including but not limited to second, third, or hybrid second / third generation lentiviruses, and any generation of recombinant lentiviruses designed to target a specific cell type or receptor (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 The 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) can also be included in a viral vector-based vaccine platform. Depending on the packaging capacity of the viral vector-based vaccine platform described above, this approach can deliver one or more nucleotide sequences encoding one or more neoantigen peptides.The sequences may be flanked by non-mutated sequences, separated by linkers, or preceded by one or more sequences that target intracellular compartments (see, e.g., Gros et al., Prospective identification of neoantigen-specific lymphocytes in the peripheral blood of melanoma patients, Nat Med. (2016) 22(4): 433-8; Stronen et al., Targeting of cancer neoantigens with donor-derived T cell receptor repertoires, Science. (2016) 352(6291): 1337-41; Lu et al., Efficient identification of mutated cancer antigens recognized by T cells associated with durable tumor regressions, Clin Cancer Res. (2014) 20(13): 3401-10). Upon introduction into the host, the infected cells express the neoantigen, thereby eliciting a host immune (e.g., CTL) response against the peptide. Vaccinia vectors and methods useful in immunization protocols are described, for example, in U.S. Patent No. 4,722,848. Another vector is BCG (Bacille Calmette-Guerin). BCG vectors are described in Stover et al. (Nature 351:456-460 (1991)). A wide variety of other vaccine vectors useful for therapeutic administration or immunization of neoantigens, such as Salmonella typhi vectors, will be apparent to those skilled in the art from the description herein.

[0209] IV.A. Further Considerations in Vaccine Design and Manufacturing IV.A.1. Determination of a set of peptides covering all tumor subclones Truncal peptides, meaning those presented by all or most of the tumor subclones, are prioritized for inclusion in the vaccine. 53 Optionally, if there are no truncal peptides that are predicted to be highly presented and immunogenic, or if the number of truncal peptides that are predicted to be highly presented and immunogenic is small enough that additional non-truncal peptides can be included in the vaccine, additional peptides can be prioritized by estimating the number and identity of tumor subclones and selecting peptides to maximize the number of tumor subclones covered by the vaccine. 54 .

[0210] IV.A.2. Neoantigen prioritization After applying all of the above neoantigen filters, more candidate neoantigens may still be available for vaccine inclusion than the vaccine technology can accommodate. Additionally, there may remain uncertainties about various aspects of the neoantigen analysis, and trade-offs may exist between various properties of the candidate vaccine neoantigens. Therefore, instead of pre-determined filters at each stage of the selection process, an integral multidimensional model can be considered, in which the candidate neoantigens are placed in a space with at least the following axes, and the selection is optimized using an integral approach: 1. Risk of autoimmunity or tolerance (germline risk) (lower autoimmune risk is typically favorable) 2. Probability of sequencing artifacts (lower artifact probability is typically preferred) 3. Probability of immunogenicity (a higher probability of immunogenicity is typically preferred) 4. Probability of presentation (higher probabilities of presentation are typically preferred) 5. Gene Expression (higher expression is typically preferred) 6. HLA gene coverage (a higher number of HLA molecules involved in the presentation of a set of neoantigens may reduce the probability that a tumor will evade immune attack via downregulation or mutation of HLA molecules) 7. HLA class coverage (covering both HLA-I and HLA-II may increase the chance of a therapeutic response and decrease the chance of tumor immune evasion)

[0211] V. METHODS OF TREATMENT AND MANUFACTURING Also provided are methods for inducing a tumor-specific immune response in a subject, vaccinating against a tumor, and treating and / or alleviating symptoms of cancer in a subject by administering to the subject one or more neoantigens, such as multiple neoantigens identified using the methods disclosed herein.

[0212] In some embodiments, the subject has been diagnosed with cancer or is at risk of developing cancer.The subject can be a human, a dog, a cat, a horse, or any animal in which a tumor-specific immune response is desired.The tumor can be any solid tumor, such as breast, ovarian, prostate, lung, kidney, stomach, colon, testicular, head and neck, pancreatic, brain, melanoma, and other tissue organ tumors, as well as hematological tumors, such as lymphomas and leukemias, including acute myeloid leukemia, chronic myeloid leukemia, chronic lymphocytic leukemia, T-cell lymphocytic leukemia, and B-cell lymphoma.

[0213] The neoantigen can be administered in an amount sufficient to induce a CTL response.

[0214] The neoantigens can be administered alone or in combination with other therapeutic agents, such as chemotherapeutic agents, radiation, or immunotherapy. Any suitable therapeutic treatment for the particular cancer can be administered.

[0215] In addition, the subject can be further administered with an anti-immunosuppressive / immunostimulatory substance, such as a checkpoint inhibitor.For example, the subject can be further administered with an anti-CTLA antibody or anti-PD-1 or anti-PD-L1.The blockade of CTLA-4 or PD-L1 with an antibody can enhance the immune response against cancerous cells in the patient.In particular, CTLA-4 blockade has been shown to be effective when vaccination protocols are adopted.

[0216] The optimal amount of each neoantigen to be included in the vaccine composition and the optimal dosing regimen can be determined. For example, the neoantigen or variants thereof can be prepared for intravenous (iv), subcutaneous (sc), intradermal (id), intraperitoneal (ip), or intramuscular (im) injection. Methods of injection include sc, id, ip, im, and iv. Methods of DNA or RNA injection include id, im, sc, ip, and iv. Other methods of administration of the vaccine composition are known to those skilled in the art.

[0217] Vaccines can be edited such that the selection, number, and / or amount of neoantigens present in the composition are tissue, cancer, and / or patient specific. As an example, the exact selection of peptides can be guided by the expression pattern of the parent protein in a given tissue. Selection can depend on the specific type of cancer, the state of the disease, earlier treatment regimes, the immune status of the patient, and of course the HLA haplotype of the patient. Furthermore, vaccines can contain components that are personalized according to the personal needs of a particular patient. Examples include altering the selection of neoantigens according to the expression of neoantigens in a particular patient, or adjusting for secondary treatments after a first round or scheme of treatment.

[0218] For compositions to be used as vaccines for cancer, neoantigens with similar normal self-peptides that are abundantly expressed in normal tissues can be avoided or present in low amounts in the compositions described herein. On the other hand, if a patient's tumor is known to express a high amount of a particular neoantigen, the respective pharmaceutical composition for the treatment of this cancer can be present in high amount and / or can include more than one neoantigen specific for this particular neoantigen or pathway of this neoantigen.

[0219] Compositions containing neoantigens can be administered to individuals already suffering from cancer. In therapeutic applications, the compositions are administered to patients in an amount sufficient to elicit an effective CTL response against tumor antigens and to cure or at least partially halt symptoms and / or complications. An amount adequate to achieve this is defined as a "therapeutically effective dose". Amounts effective for this use will depend, for example, on the composition, the mode of administration, the stage and severity of the disease being treated, the weight and general health of the patient, and the judgment of the prescribing physician. It should be kept in mind that the compositions can generally be used in severe disease states, i.e., life-threatening or potentially life-threatening situations, particularly when the cancer has metastasized. In such instances, it may be possible, and the treating physician may feel desirable, to administer substantial excesses of these compositions, taking into account the minimization of foreign material and the relatively non-toxic nature of the neoantigens.

[0220] For therapeutic use, administration can begin upon detection or surgical removal of the tumor, followed by boosting doses until at least symptoms are substantially abated, and for a period thereafter.

[0221] Pharmaceutical compositions for therapeutic treatment (e.g., vaccine compositions) are intended for parenteral, topical, nasal, oral, or local administration. Pharmaceutical compositions can be administered parenterally, e.g., intravenously, subcutaneously, intradermally, or intramuscularly. The compositions can be administered at the site of surgical resection to induce a local immune response against the tumor. Disclosed herein are compositions for parenteral administration that include a solution of a neoantigen, the vaccine composition being dissolved or suspended in an acceptable carrier, e.g., an aqueous carrier. A variety of aqueous carriers can be used, e.g., water, buffered water, 0.9% saline, 0.3% glycine, hyaluronic acid, and the like. These compositions can be sterilized by conventional, well-known sterilization techniques, or can be sterile filtered. The resulting aqueous solutions can be packaged for use as is, or lyophilized, the lyophilized preparation being combined with a sterile solution prior to administration. The compositions may contain pharma- ceutically acceptable auxiliary substances required to approximate physiological conditions, such as pH adjusting and buffering agents, isotonicity agents, wetting agents, and the like, for example, sodium acetate, sodium lactate, sodium chloride, potassium chloride, calcium chloride, sorbitan monolaurate, triethanolamine oleate, and the like.

[0222] Neoantigens can also be administered via liposomes, which target them to specific cellular tissues, such as lymphoid tissues. Liposomes are also useful for increasing half-life. Liposomes include emulsions, foams, micelles, insoluble monolayers, liquid crystals, phospholipid dispersions, lamellar layers, and the like. In these preparations, the neoantigen to be delivered is incorporated as part of the liposome, either alone or with a molecule that binds to a receptor that is predominant among lymphoid cells, such as a monoclonal antibody that binds to the CD45 antigen, or with other therapeutic or immunogenic compositions. Thus, liposomes filled with the desired neoantigen can be directed to the site of lymphoid cells, where they then deliver the selected therapeutic / immunogenic composition. Liposomes can generally be formed from standard vesicle-forming lipids, including neutral and negatively charged phospholipids and sterols, such as cholesterol. The choice of lipid is generally guided by considerations, for example, of liposome size, acid lability, and stability of the liposomes in the bloodstream. A variety of methods are available for preparing 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.

[0223] For targeting to immune cells, the ligand to be incorporated into the liposomes can include, for example, an antibody or fragment thereof specific for a cell surface determinant of the desired immune system cell. The liposome suspension can be administered intravenously, topically, locally, etc., in doses that vary according to, among other things, the mode of administration, the peptide being delivered, and the stage of the disease being treated.

[0224] For therapeutic or immunization purposes, the peptides described herein, and optionally nucleic acids encoding one or more of the peptides, can also be administered to a patient. Numerous methods are conveniently used to deliver nucleic acids to a patient. For example, the nucleic acid can be delivered directly as "naked DNA". This approach is described, for example, in Wolff et al., Science 247:1465-1468 (1990), and in U.S. Pat. Nos. 5,580,859 and 5,589,466. Nucleic acids can also be administered using ballistic delivery, as described, for example, in U.S. Pat. No. 5,204,253. Particles consisting simply of DNA can be administered. Alternatively, DNA can be attached to particles, such as gold particles. Approaches for delivering nucleic acid sequences include viral vectors, mRNA vectors, and DNA vectors, with or without electroporation.

[0225] Nucleic acids can also be delivered by complexing with cationic compounds, such as cationic lipids.Lipid-mediated gene delivery methods are described, for example, in 9618372 WOAWO 96 / 18372; 9324640 WOAWO 93 / 24640; Mannino & Gould-Fogerite, BioTechniques 6(7): 682-691 (1988); U.S. Patent No. 5,279,833 Rose, U.S. Patent No. 5,279,833; 9106309 WOAWO 91 / 06309; and Felgner et al., Proc.Natl.Acad.Sci.USA 84: 7413-7414 (1987).

[0226] Neoantigens can also be derived from viruses such as vaccinia, fowlpox, self-replicating alphaviruses, Maraba viruses, adenoviruses (see, e.g., Tatsis et al., Adenoviruses, Molecular Therapy (2004) 10, 616-629), or lentiviruses, including but not limited to second, third, or hybrid second / third generation lentiviruses, and recombinant lentiviruses of any generation designed to target a specific cell type or receptor (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:basicto 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 The ubiquitin C promoter, Nucl.AcidsRes.(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) can also be included in a viral vector-based vaccine platform. Depending on the packaging capacity of the viral vector-based vaccine platform described above, this approach can deliver one or more nucleotide sequences encoding one or more neoantigen peptides.The sequences may be flanked by non-mutated sequences, separated by linkers, or preceded by one or more sequences that target intracellular compartments (see, e.g., Gros et al., Prospective identification of neoantigen-specific lymphocytes in the peripheral blood of melanoma patients, Nat Med. (2016) 22 (4):433-8; Stronen et al., Targeting of cancer neoantigens with donor-derived T cell receptor repertoires, Science. (2016) 352 (6291):1337-41; Lu et al., Efficient identification of mutated cancer antigens recognized by T cells associated with durable tumor regressions, Clin Cancer Res. (2014) 20(13):3401-10). Upon introduction into the host, the infected cells express the neoantigen, thereby eliciting a host immune (e.g., CTL) response against the peptide. Vaccinia vectors and methods useful in immunization protocols are described, for example, in U.S. Patent No. 4,722,848. Another vector is BCG (Bacillus Calmette-Guerin). BCG vectors are described in Stover et al. (Nature 351:456-460 (1991)). A wide variety of other vaccine vectors useful for therapeutic administration or immunization of neoantigens, such as Salmonella typhi vectors, will be apparent to those skilled in the art from the description herein.

[0227] The means of administering the nucleic acid uses a minigene construct that encodes one or more epitopes. To generate a DNA sequence (minigene) encoding the selected CTL epitopes for expression in human cells, the amino acid sequence of the epitope is reverse translated. A human codon usage table is used to guide the codon selection for each amino acid. The DNA sequences encoding these epitopes are directly adjacent to generate a continuous polypeptide sequence. Additional elements can be incorporated into the minigene design to optimize expression and / or immunogenicity. Examples of amino acid sequences that can be reverse translated and included in the minigene sequence include helper T lymphocyte 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 that encode the plus and minus strands of the minigene. Overlapping oligonucleotides (30-100 bases long) are synthesized, phosphorylated, purified, and annealed under appropriate conditions using well-known techniques. The ends of the oligonucleotides are ligated using T4 DNA ligase. This synthetic minigene encoding the CTL epitope polypeptide can then be cloned into a desired expression vector.

[0228] Purified plasmid DNA can be prepared for injection using a variety of formulations. The simplest of these is the reconstitution of lyophilized DNA in sterile phosphate-buffered saline (PBS). A variety of methods have been described, and new techniques may become available. As mentioned above, nucleic acids are conveniently formulated with cationic lipids. In addition, glycolipids, fusogenic liposomes, peptides, and compounds collectively referred to as protective, interactive, non-condensing (PINC) can also be complexed with purified plasmid DNA to affect variables such as stability, intramuscular distribution, or transport to specific organs or cell types.

[0229] Also disclosed herein is a method of producing a tumor vaccine comprising performing the steps of the methods disclosed herein; and producing a tumor vaccine comprising multiple neoantigens or a subset of multiple neoantigens.

[0230] The neoantigens disclosed herein can be produced using methods known in the art. For example, a method for producing a neoantigen or vector disclosed herein (e.g., a vector comprising at least one sequence encoding one or more neoantigens) can include culturing a host cell under conditions suitable for expressing the neoantigen or vector, the host cell comprising at least one polynucleotide encoding the neoantigen or vector, and purifying the neoantigen or vector. Standard purification methods include chromatographic, electrophoretic, immunological, precipitation, dialysis, filtration, concentration, and chromatofocusing techniques.

[0231] The host cell can include Chinese Hamster Ovary (CHO) cells, NS0 cells, yeast, or HEK293 cells. The host cell can be transformed with one or more polynucleotides comprising at least one nucleic acid sequence encoding a neoantigen or vector disclosed herein, and optionally, the isolated polynucleotide further comprises a promoter sequence operably linked to the at least one nucleic acid sequence encoding the neoantigen or vector. In certain embodiments, the isolated polynucleotide can be a cDNA.

[0232] VI. Identification of neoantigens VI.A. Identification of candidate neoantigens A methodology for NGS analysis of tumor and normal exomes and transcriptomes in the specific space of neoantigens is described and applied. 6,14,15The examples below consider certain optimizations for greater sensitivity and specificity for identifying neoantigens in a clinical setting. These optimizations can be grouped into two areas: those related to laboratory processes and those related to NGS data analysis.

[0233] VI.A.1. Laboratory Process Optimization The process improvements presented herein build on the concepts developed for reliable assessment of cancer driver genes in targeted cancer panels. 16 This addresses the challenges in high-precision neoantigen discovery from low tumor content and small volume clinical specimens by expanding the full-exome and full-transcriptome settings required for neoantigen identification. Specifically, these improvements include: 1. Targeting deep (greater than 500x) unique average coverage across the tumor exome to detect mutations present at low mutant allele frequency due to either low tumor content or subclonal status. 2. Less than 5% of bases are covered by less than 100x to minimize missing potential neoantigens, e.g. a. Use of DNA-based capture probes with individual probe QC 17 b. Inclusion of additional baits for areas that are not sufficiently covered 3. Targeting uniform coverage across the normal exome with less than 5% of bases covered under 20x to minimize the chance of potential neoantigens remaining unclassified for somatic / germline status (and therefore not usable as TSNAs). 4. To minimize the total amount of sequencing required, sequence capture probes are designed only for the coding regions of genes, since non-coding RNA cannot give rise to neoantigens. Additional optimizations include: a. Supplemental probes for HLA genes that are GC-rich and not well captured by standard exome sequencing18 . b. Elimination of genes predicted to produce few or no candidate neoantigens due to factors such as poor expression, suboptimal digestion by the proteasome, or atypical sequence characteristics. 5. Tumor RNA is also sequenced at high depth (>100M reads) to allow for mutation detection, quantification of gene and splice variant ("isoform") expression, and fusion detection. RNA from FFPE samples is sequenced using probe-based enrichment with the same or similar probes used to capture the exome in DNA. 19 It is extracted using

[0234] VI.A.2. Optimizing NGS Data Analysis Analytical method improvements address the suboptimal sensitivity and specificity of common research variant calling approaches and specifically allow for customization relevant for identification of neoantigens in the clinical setting. These include: 1. Use of the HG38 reference human genome or later versions for alignment, as it contains multiple MHC region assemblies that better reflect population polymorphism, as opposed to earlier genome releases. 2. Various programs 5 Overcoming the limitations of single mutation callers 20 by merging results from. a. Single nucleotide mutations and indels are detected in tumor DNA, tumor RNA, and normal DNA with a range of tools including: Strelka 21 and Mutec t22 programs based on comparison of tumor and normal DNA, such as; 23 , UNCeqR, and other programs that incorporate tumor DNA, tumor RNA, and normal DNA. b. Indels are identified in Strelka and ABRA 24 This is determined by a program that performs local reassembly, such as C. Structural rearrangements are Pindel 25 or Breakseq26 It is determined using specialized tools such as 3. To detect and prevent sample swapping, mutation calls from samples for the same patient are compared at a selected number of polymorphic sites. 4. Extensive filtering of artificial calls is performed, for example by: a. Removal of mutations found in normal DNA, potentially with relaxed detection parameters in the case of low coverage, and permissive proximity criteria in the case of indels. b. Removal of mutations due to poor mapping quality or poor base quality 27 . c. Removal of mutations resulting from re-emerging sequencing artifacts, even if not observed in the corresponding normal 27 Examples include mutations that are detected primarily on one strand. d. Removal of mutations detected in a set of unrelated controls 27 . 5.seq2HLA 28 , ATHLETES 29 or Optitype, and also combining exome and RNA sequencing data 28 , accurate HLA calling from normal exomes. Additional potential optimizations include the adoption of dedicated assays for HLA typing, such as long-read DNA sequencing. 30 or adaptation of methods for linking RNA fragments to maintain continuity 31 Includes. 6. Robust detection of nascent ORFs arising from tumor-specific splice variants in CLASS 32 , Bayesembler 33 , StringTie 34 This is done by assembling transcripts from RNA-seq data using Cufflinks, or a similar program in its reference-guided mode (i.e., using known transcript structures rather than attempting to recreate the entire transcripts from each experiment). 35Although commonly used for this purpose, it frequently produces an incredibly large number of splice variants, many of which are much shorter than the full-length gene, and may not be able to recover a simple positive control, SpliceR, which can cloning the coding sequence and potentially nonsense-mediated decay mechanisms that reintroduce mutant sequences. 36 and MAMBA 37 Gene expression is determined using tools such as Cufflinks 35 or Express (Roberts and Pachter, 2013). Wild-type and mutant-specific expression counts and / or relative levels are determined using ASE. 38 or HTSeq 39 Potential filtering steps include: a. Removal of candidate nascent ORFs that are thought to be poorly expressed. b. Removal of candidate nascent ORFs predicted to undergo nonsense-mediated decay (NMD). 7. Candidate neoantigens observed only in RNA (e.g., neo-ORFs) that cannot be directly verified as tumor-specific are classified as likely to be tumor-specific according to additional parameters, for example, by considering the following: a. Presence of cis-acting frameshift or splice site mutation support in tumor DNA only. b. The presence of confirmed trans-acting mutations in splicing factors in tumor DNA only. As an example, the gene that exhibited the most differential splicing in three independently published experiments with R625 mutant SF3B1 was 40 The second experiment examined uveal melanoma cell lines. 41 , and a third study looked at breast cancer patients. 42 Nevertheless, there was a match. c. For novel splicing isoforms, the presence of confirmatory “novel” splice-junction reads in the RNASeq data. d. For a de novo rearrangement, the presence of conclusive evidence of exon-proximal reads in tumor DNA that are not present in normal DNA. e.GTEx 43 etc. from the gene expression compendium (i.e., making germline origin less likely). 8. Complementing reference genome alignment-based analyses by comparing tumor and normal reads (or k-mers derived from such reads) of assembled DNA to directly avoid alignment and annotation-based errors and artifacts (e.g., for somatic mutations occurring near germline mutations or repeat context indels).

[0235] In samples with polyadenylated RNA, the presence of viral and microbial RNA in the RNA-seq data will be assessed using RNA CoMPASS44 or similar methods towards identifying additional factors that may predict patient response.

[0236] VI.B. Isolation and Detection of HLA Peptides Isolation of HLA peptide molecules was performed using classical immunoprecipitation (IP) techniques after lysis and solubilization of tissue samples. 55~58 The clarified lysates were used for HLA-specific IP.

[0237] Immunoprecipitation was performed using antibodies coupled to beads, where the antibodies are specific for HLA molecules. For pan-class I HLA immunoprecipitation, a pan-class I CR antibody is used, and for class II HLA-DR, an HLA-DR antibody is used. The antibodies are covalently attached to the NHS-Sepharose beads during overnight incubation. After covalent attachment, the beads were washed and aliquoted for IP. 59、60Immunoprecipitation can also be performed using antibodies that are not covalently attached to beads. Typically, this is done using sepharose or magnetic beads coated with protein A and / or protein G to retain the antibodies on the column. Below are some antibodies that can be used to selectively enrich for MHC / peptide complexes. TIFF0007651623000001.tif46160

[0238] The clarified tissue lysate is added to the antibody beads for immunoprecipitation. After immunoprecipitation, the beads are removed from the lysate and the lysate is stored for additional experiments, including additional IPs. The IP beads are washed to remove non-specific binding and the HLA / peptide complexes are 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 in some cases stored at -20°C prior to MS analysis.

[0239] The dried peptides are reconstituted in HPLC buffer suitable for reversed-phase chromatography and loaded onto a C-18 microcapillary HPLC column for gradient elution in a Fusion Lumos mass spectrometer (Thermo). MS1 spectra of peptide mass / charge (m / z) are collected at high resolution in an Orbitrap detector, and then MS2 low-resolution scans are collected in an ion trap detector after HCD fragmentation of selected ions. Additionally, MS2 spectra can be acquired using either CID or ETD fragmentation methods, or any combination of the three techniques to obtain a larger amino acid coverage of the peptide. MS2 spectra can also be measured with high resolution mass accuracy in an Orbitrap detector.

[0240] The MS2 spectra from each analysis were analyzed using Comet 61、62 The peptides were searched against a protein database using Percolator63~65 Further sequencing was performed using PEAKS studio (Bioinformatics Solutions Inc.) and other search engines, or spectral matching and de novo sequencing were performed. 75 Sequencing methods including the following can be used.

[0241] VI.B.1. Study of MS detection limits for comprehensive HLA peptide sequencing Using the peptide YVYVADVAAK (SEQ ID NO:1), what the limit of detection was was determined using various 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) was in the attomolar range (10 -18 ), a dynamic range spanning five orders of magnitude, and a signal-to-noise ratio in the low femtomole range (10 -15 ) appears to be sufficient for sequencing.

[0242] TIFF0007651623000002.tif61128

[0243] VII. Presented Model VII.A. System Overview 2A is an overview of an environment 100 for identifying the likelihood of peptide presentation in a patient, according to one embodiment. The environment 100 provides a context for implementing a presentation identification system 160, which itself includes a presentation information store 165.

[0244] The presentation identification system 160 is a computer model or a system embodied in a computational system such as that discussed below with respect to FIG. 14 that receives a peptide sequence associated with a set of MHC alleles and determines the likelihood that the peptide sequence will be presented by one or more of the set of associated MHC alleles. The presentation identification system 160 can be applied to both class I and class II MHC alleles. This is useful in a variety of contexts. An example of one specific application of the presentation identification system 160 is to receive a nucleotide sequence of a candidate neoantigen associated with a set of MHC alleles from tumor cells of a patient 110 and determine the likelihood that the candidate neoantigen will be presented by one or more of the associated MHC alleles of the tumor and / or will induce an immunogenic response in the immune system of the patient 110. Those candidate neoantigens that have a high likelihood as determined by the system 160 can be selected for inclusion in the vaccine 118 such that an anti-tumor immune response can be elicited from the immune system of the patient 110 that provided the tumor cells.

[0245] The presentation identification system 160 determines the presentation likelihood through one or more presentation models. Specifically, the presentation model generates a likelihood that a given peptide sequence is presented for a set of related MHC alleles, and the likelihood is generated based on the presentation information stored in the storage device 165. For example, the presentation model may generate a likelihood that a peptide sequence "YVYVADVAAK (SEQ ID NO:1)" is presented for a set of alleles HLA-A*02:01, HLA-A*03:01, HLA-B*07:02, HLA-B*08:03, HLA-C*01:04 on the cell surface of the sample. The presentation information 165 contains information about whether these peptides bind to various types of MHC alleles such that the peptides are presented by the MHC alleles, which is determined in the model according to the position of the amino acid in the peptide sequence. Based on the presentation information 165, the presentation model can predict whether an unrecognized peptide sequence is presented in conjunction with a related set of MHC alleles. As stated above, the presentation model can be applied to both class I and class II MHC alleles.

[0246] VII.B. Presentation information 2 illustrates a method of obtaining presentation information according to one embodiment. Presentation information 165 includes two general categories of information: allele interaction information and allele non-interaction information. Allele interaction information includes information that influences presentation of a peptide sequence that is dependent on the type of MHC allele. Allele non-interaction information includes information that influences presentation of a peptide sequence that is independent of the type of MHC allele.

[0247] VII.B.1. Allelic interaction information The allele interaction information primarily includes identified peptide sequences that are known to be presented by one or more identified MHC molecules from humans, mice, etc. Of note, this may or may not include data obtained from tumor samples. The presented peptide sequences may be identified from cells expressing a single MHC allele. In this example, the presented peptide sequences are generally collected from a single allele cell line engineered to express a predetermined MHC allele and then exposed to a synthetic protein. The peptides presented on the MHC alleles are isolated by techniques such as acid elution and identified by mass spectrometry. Figure 2B shows an exemplary peptide presented on a predetermined MHC allele HLA-DRB1*12:01. An example of this is shown in TIFF0007651623000003.tif5128, which was isolated and identified by mass spectrometry. In this situation, the direct association between the presented peptide and the MHC protein to which it bound is conclusively known, since the peptide is identified through cells engineered to express a single, predefined MHC protein.

[0248] The presented peptide sequences may also be collected from cells expressing multiple MHC alleles. Typically in humans, six different types of MHC-I molecules and up to 12 different types of MHC-II molecules are expressed on cells. Such presented peptide sequences may be identified from multi-allelic cell lines that have been engineered to express multiple predetermined MHC alleles. Such presented peptide sequences may also be identified from tissue samples, either normal tissue samples or tumor tissue samples. In this particular example, MHC molecules can be immunoprecipitated from normal or tumor tissue. Peptides presented on multiple MHC alleles can similarly be isolated by techniques such as acid elution and identified by mass spectrometry. Figure 2C shows six exemplary peptides. An example of this is shown in TIFF0007651623000004.tif17166, presented on identified class I MHC alleles HLA-A*01:01, HLA-A*02:01, HLA-B*07:02, HLA-B*08:01, and class II MHC alleles HLA-DRB1*10:01, HLA-DRB1:11:01, isolated and identified by mass spectrometry. In contrast to monoallelic cell lines, the bound peptide is isolated from the MHC molecule prior to identification, so the direct association between the presented peptide and the MHC protein to which it is bound may be unknown.

[0249] The allele interaction information can also include mass spectrometry ion currents, which depend on both the concentration of peptide-MHC molecule complexes and the ionization efficiency of peptides. Ionization efficiency varies from peptide to peptide in a sequence-dependent manner. In general, ionization efficiency varies from peptide to peptide over approximately two orders of magnitude, while the concentration of peptide-MHC complexes varies over a larger range.

[0250] Allele interaction information can also include measurements or predictions of binding affinity between a given MHC allele and a given peptide. One or more affinity models can generate such predictions (72,73,74). For example, returning to the example shown in FIG. 1D, the representation 165 may include a combination of peptide YEMFNDKSF (SEQ ID NO:3) and class I allele HLA-A * 01:01 may include a predicted binding affinity value of 1000 nM between the peptides. Few peptides with IC50>1000 nm are presented by the MHC, with lower IC50 values ​​increasing the probability of presentation. The predicted binding affinity between TIFF0007651623000005.tif4128 and class II allele HLA-DRB1:11:01 may be included.

[0251] The allele interaction information can also include measurements or predictions of the stability of MHC complexes. 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 number on tumor cells and on antigen-presenting cells that encounter vaccine antigens. For example, returning to the example shown in FIG. 2C, the presentation information 165 can include a half-life stability prediction of 1 hour for the class I molecule HLA-A*01:01. The presentation information 165 can also include a half-life stability prediction of the class II molecule HLA-DRB1:11:01.

[0252] Allele interaction information can also include measured or predicted rates of peptide-MHC complex formation. Complexes that form at a faster rate are more likely to be presented at high concentrations on the cell surface.

[0253] The allele interaction information can also include the sequence and length of the peptide. MHC class I molecules typically prefer to present peptides with a length of 8-15 peptides. 60-80% of the presented peptides have a length of 9. MHC class II molecules generally tend to present peptides with a length of 6-30 peptides.

[0254] Allele interaction information can also include the presence of kinase sequence motifs on the neoantigen-encoded peptide, and the presence or absence of specific post-translational modifications on the neoantigen-encoded peptide. The presence of a kinase motif influences the probability of post-translational modifications that may enhance or interfere with MHC binding.

[0255] Allelic interaction information can also include expression or activity levels of proteins, e.g., kinases, involved in post-translational modification processes (as measured or predicted by RNA-seq, mass spectrometry, or other methods).

[0256] Allelic interaction information can also include the probability of presentation of peptides with similar sequences in cells from other individuals expressing particular MHC alleles, as assessed by mass spectrometry proteomics or other means.

[0257] Allelic interaction information can also include the expression levels of particular MHC alleles 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.

[0258] Allelic interaction information can also include the overall neoantigen-encoded peptide sequence-independent probability of presentation by a particular MHC allele in other individuals expressing that particular MHC allele.

[0259] Allele interaction information can also include the overall peptide sequence-independent probability of presentation by MHC alleles of the same family of molecules (e.g., HLA-A, HLA-B, HLA-C, HLA-DQ, HLA-DR, HLA-DP) in other individuals. For example, HLA-C molecules are typically expressed at lower levels than HLA-A or HLA-B molecules, and therefore presentation of peptides by HLA-C is a priori less likely than presentation by HLA-A or HLA-B II. As another example, HLA-DP is generally expressed at lower levels than HLA-DR or HLA-DQ, so presentation of peptides by HLA-DP is presumed to be less likely than presentation by HLA-DR or HLA-DQ.

[0260] The allele interaction information can also include the protein sequence of a particular MHC allele.

[0261] Any of the MHC allele non-interacting information listed in the section below can also be modeled as MHC allele interacting information.

[0262] VII.B.2. Allelic non-interaction information The allele non-interacting information can include the C-terminal sequence adjacent to the neoantigen-encoded peptide in its source protein sequence. In MHC-I, the C-terminal flanking sequence can 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 surface of the cell. As a result, the MHC molecule does not receive any information about the C-terminal flanking sequence, and therefore the effect of the C-terminal flanking sequence cannot vary depending on the MHC allele type. For example, returning to the example shown in FIG. 2C, the presentation information 165 includes the C-terminal flanking sequence of the presented peptide FJIEJFOESS (SEQ ID NO:5) identified from the source protein of the peptide. TIFF0007651623000006.tif5128 may be included.

[0263] Allele non-interaction information can also include mRNA quantification measurements. For example, mRNA quantification data can be obtained for the same samples that provide mass spectrometry training data. As described later with respect to Figure 13H, RNA expression has been identified as a strong predictor of peptide presentation. In one embodiment, mRNA quantification measurements are identified from the software tool RSEM. A detailed implementation 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).

[0264] The allele-non-interacting information can also include N-terminal sequences adjacent to the peptide within the source protein sequence.

[0265] The allele non-interaction information can also include the source gene of the peptide sequence. The source gene can be defined as the Ensembl protein family of the peptide sequence. In another example, the source gene can be defined as the source DNA or source RNA of the peptide sequence. The source gene can be represented, for example, as a string of nucleotides that encodes a protein, or alternatively, in a more categorical form based on a named set of known DNA or RNA sequences that are known to encode a particular protein. In another example, the allele non-interaction information can also include the source transcript or isoform of the peptide sequence, or a set of potential source transcripts or isoforms, extracted from databases such as Ensembl or RefSeq.

[0266] The allele non-interaction information can also include the tissue type, cell type, or tumor type of the cell from which the peptide sequence is derived.

[0267] The allele non-interaction information can also include the presence of protease cleavage motifs in peptides, which are optionally weighted according to the expression of the corresponding proteases in tumor cells (as measured by RNA-seq or mass spectrometry). Peptides containing protease cleavage motifs are more easily degraded by proteases and therefore less stable in cells, and therefore less likely to be presented.

[0268] Allele non-interaction information can also include the turnover rate of the source protein when measured in the appropriate cell type. A faster turnover rate (i.e., a lower half-life) increases the probability of presentation, but this feature has low predictive power when measured in dissimilar cell types.

[0269] The allelic non-interaction information can also include the length of the source protein, optionally taking into account the specific splice variants ("isoforms") that are most highly expressed in tumor cells, as measured by RNA-seq or proteomic mass spectrometry, or as predicted from annotation of germline or somatic splicing mutations detected in DNA or RNA sequence data.

[0270] Allele non-interaction information can also include the expression level of proteasome, immunoproteasome, thymoproteasome, or other proteases in tumor cells (which can be measured by RNA-seq, proteome mass spectrometry, or immunohistochemistry). Different proteasomes have different cleavage site preferences. More weight is given to the cleavage preference of each type of proteasome in proportion to its expression level.

[0271] Allele-free interaction information can also include the expression of the peptide's source gene (e.g., as measured by RNA-seq or mass spectrometry). Possible optimizations include adjusting the measured expression to account for the presence of stromal cells and tumor-infiltrating lymphocytes in the tumor sample. Peptides from genes with higher expression are more likely to be presented. Peptides from genes with undetectable levels of expression can be eliminated from consideration.

[0272] Allelic non-interaction information can also include the probability that the source mRNA of the neoantigen-encoding peptide will be subject to nonsense-mediated decay as predicted by models of nonsense-mediated decay, e.g., the model from Rivas et al., Science 2015.

[0273] Allele-free interaction information can also include typical tissue-specific expression of the peptide's source gene during various stages of the cell cycle. Genes that are expressed at low levels overall (as measured by RNA-seq or sample analysis proteomics) but are known to be expressed at high levels during specific stages of the cell cycle are more likely to produce peptides that are presented than genes that are stably expressed at very low levels.

[0274] The allelic non-interacting information can also include a comprehensive catalog of properties of the source protein, such as those provided in uniProt or the PDB http: / / www.rcsb.org / pdb / home / home.do. These properties can include, among others, the secondary and tertiary structure of the protein, the subcellular localization,11 and Gene Ontology (GO) terms. Specifically, this information can contain annotations that operate at the level of the protein, such as the 5'UTR length, and annotations that operate at the level of specific residues, such as the helix motif at residues 300-310. These properties can also include turn motifs, sheet motifs, and disordered residues.

[0275] Allelic non-interacting information can also include features describing the nature of the domain of the source protein that contains the peptide, such as secondary or tertiary structure (eg, alpha helices versus beta sheets); alternative splicing.

[0276] The allelic non-interacting information can also include features describing the presence or absence of presentation hotspots at the peptide's position in its source protein.

[0277] Allelic non-interaction information can also include the probability of presentation of peptides derived from the source protein of the peptide in question in other individuals (after adjusting for the expression levels of the source protein in those individuals and the effects of the various HLA types of those individuals).

[0278] Allelic non-interaction information can also include the probability that a peptide will be undetected or over-represented by mass spectrometry due to technical bias.

[0279] Expression of various gene modules / pathways (not necessarily containing the source protein of the peptide) as measured by gene expression assays such as RNASeq, microarrays, targeted panels such as Nanostring, etc., or single / multiple genes representing gene modules measured by assays such as RT-PCR, that inform on the status of tumor cells, stroma, or tumor infiltrating lymphocytes (TILs).

[0280] Allele non-interaction information can also include the copy number of the peptide's source gene in the tumor cell. For example, a peptide derived from a gene that is subject to homozygous deletion in the tumor cell can be assigned a presentation probability of zero.

[0281] The allele non-interaction information can also include the probability that the peptide will bind to TAP, or the binding affinity of the peptide to TAP, measured or predicted. Peptides that are more likely to bind to TAP, or that bind with higher affinity to TAP, are more likely to be presented by MHC-I.

[0282] Allele-free information can also include the expression level of TAP in tumor cells (which can be measured by RNA-seq, proteomic mass spectrometry, immunohistochemistry). At MHC-I, higher TAP expression levels increase the probability of presentation of all peptides.

[0283] Allelic non-interaction information can also include the presence or absence of tumor mutations, including but not limited to: i.Driver mutations in known cancer driver genes such as EGFR, KRAS, ALK, RET, ROS1, TP53, CDKN2A, CDKN2B, NTRK1, NTRK2, and NTRK3. ii. In genes encoding proteins involved in 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 of the genes encoding components of the proteasome or immunoproteasome). Peptides whose presentation relies on components of the antigen presentation machinery that are affected by loss-of-function mutations in the tumor have a reduced probability of presentation.

[0284] Presence or absence of functional germline polymorphisms, including but not limited to: i. In genes encoding proteins involved in 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 of the genes encoding components of the proteasome or immunoproteasome).

[0285] The allelic non-interaction information can also include tumor type (eg, NSCLC, melanoma).

[0286] The allele non-interaction information can also include the known functionality of the HLA allele, as reflected, for example, by the HLA allele suffix. For example, the N suffix in the allele name HLA-A*24:09N indicates a null allele that is not expressed and is therefore unlikely to present an epitope; the complete HLA allele suffix nomenclature is described in https: / / www.ebi.ac.uk / ipd / imgt / hla / nomenclature / suffixes.html.

[0287] Allelic non-interaction information can also include clinical tumor subtype (eg, squamous cell lung cancer vs. non-squamous).

[0288] The allele non-interaction information can also include smoking history.

[0289] Allele non-interaction information can also include a history of sunburn, sun exposure, or exposure to other mutagens.

[0290] The allelic non-interaction information can also include regional expression of the peptide's source gene in the relevant tumor type or clinical subtype, optionally stratified by driver mutations. Genes that are typically expressed at high levels in the relevant tumor type are more likely to be represented.

[0291] The allelic non-interaction information can also include the frequency of a mutation in all tumors, or in tumors of the same type, or in tumors from individuals with at least one shared MHC allele, or in tumors of the same type in individuals with at least one shared MHC allele.

[0292] In the example of a mutated tumor-specific peptide, the list of characteristics used to predict the probability of presentation can also include the mutation annotation (e.g., missense, read-through, frameshift, fusion, etc.) or whether the mutation is predicted to result in nonsense-mediated decay (NMD). For example, a peptide derived from a protein segment that is not translated in tumor cells due to a homozygous premature termination mutation can be assigned a presentation probability of zero. NMD results in a decrease in mRNA translation, which reduces the probability of presentation.

[0293] VII.C. Presentation Identification System 3 is a high-level block diagram illustrating computer logic components of a presentation identification system 160, according to one embodiment. In this exemplary 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 comprises a training data store 170 and a presentation model store 175. Some embodiments of the model management system 160 have different modules than those described herein. Similarly, functionality may be distributed among the modules in a manner different than that described herein.

[0294] VII.C.1. Data Management Module The data management module 312 generates sets of training data 170 from the representation information 165. Each training data set contains a number of data examples, each of which includes at least a representation of a peptide sequence p i and the peptide sequence p i Combined with one or more relevant MHC alleles i and the dependent variable y , which represents information that the presentation identification system 160 is interested in predicting new values ​​of the independent variables. i and the independent variable z i Contains a set of:

[0295] In one particular implementation that will be mentioned throughout the remainder of this specification, the dependent variable y i is peptide p i but one or more associated MHC alleles i However, in other implementations, the dependent variable y i is a function of the presentation identification system 160 for determining the independent variable z i It will be appreciated that the dependent variable y may represent any other type of information that one is interested in predicting. For example, in another implementation, the dependent variable y i may also be a numerical value indicating the mass analysis ion current determined for the example data.

[0296] Peptide sequence p for data example i i is k i is a sequence of k amino acids, i can vary within a range among data examples i. For example, the range can be 8 to 15 for MHC class I, or 6 to 30 for MHC class II. In one specific implementation of the system 160, all peptide sequences p in the training data set are i can have the same length, e.g., 9. The number of amino acids in a peptide sequence can vary depending on the type of MHC allele (e.g., MHC allele in humans, etc.). MHC allele a for data example i i is the peptide sequence p i Indicates whether it exists in combination with

[0297] The data management module 312 also calculates the peptide sequences p i and bound MHC allele a i Together with the binding affinity b i and stability i For example, the training data 170 may include a set of allele interaction variables, such as a predicted value of i And, a i The predicted binding affinity b between each of the bound MHC molecules shown in iAs another example, the training data 170 may include a i The predicted stability value s for each of the MHC alleles shown in i may contain

[0298] The data management module 312 also receives the peptide sequence p i Along with allele-free interaction variables such as C-terminal flanking sequences and mRNA quantification measurements, i It may also include.

[0299] The data management module 312 also identifies peptide sequences that are not presented by MHC alleles to generate the training data 170. Generally, this involves identifying a "longer" sequence of the source protein that contains the peptide sequence to be presented prior to presentation. If the presentation information contains an engineered cell line, the data management module 312 identifies a set of peptide sequences in the synthetic protein to which the cell was exposed that were not presented on the MHC alleles of the cell. If the presentation information contains a tissue sample, the data management module 312 identifies the source protein from which the presented peptide sequence originated and identifies a set of peptide sequences in the source protein that were not presented on the MHC alleles of the tissue sample cells.

[0300] The data management module 312 also artificially generates peptides with random sequences of amino acids and identifies the generated sequences as peptides that are not presented on MHC alleles. This can be accomplished by randomly generating peptide sequences, allowing the data management module 312 to easily generate a large amount of synthetic data for peptides that are not presented on MHC alleles. In practice, because a small percentage of peptide sequences are presented by MHC alleles, it is highly likely that synthetically generated peptide sequences are not presented by MHC alleles, even if they are included in proteins processed by the cell.

[0301] 4 illustrates an exemplary set of training data 170A, according to one embodiment. Specifically, the first three data examples in training data 170A are a monoallelic cell line containing the allele HLA-C*01:03, as well as three peptide sequences: 1 shows peptide presentation information from TIFF0007651623000007.tif9128. The fourth data example in training data 170A shows peptide information from a multi-allelic cell line including alleles HLA-B*07:02, HLA-C*01:03, HLA-A*01:01, and peptide sequence QIEJOEIJE (SEQ ID NO:13). The first data example shows that peptide sequence QCEIOWARE (SEQ ID NO:14) was not presented by allele HLA-DRB3:01:01. As discussed in the previous two paragraphs, the negatively labeled peptide sequences may be randomly generated by data management module 312 or may be identified from the source protein of the presented peptide. Training data 170A also includes predicted binding affinity of 1000 nM and predicted half-life stability of 1 hour for the peptide sequence-allele pair. The training data 170A also contains the C-terminal flanking sequence of peptide FJELFISBOSJFIE (SEQ ID NO:15), and 2 It also includes allele non-interacting variables, such as mRNA quantification measurements of TPM. A fourth data example shows that the peptide sequence QIEJOEIJE (SEQ ID NO:13) was presented by one of the alleles HLA-B*07:02, HLA-C*01:03, or HLA-A*01:01. Training data 170A also includes predicted binding affinity and stability values ​​for each of the alleles, as well as the C-terminal flanking sequences of the peptide and mRNA quantification measurements for the peptide.

[0302] VII.C.2. Coding Module The encoding module 314 encodes the information contained in the training data 170 into a numerical representation that can be used to generate one or more representation models. In one implementation, the encoding module 314 one-hot encodes sequences (e.g., peptide sequences or C-terminal flanking sequences) for a predetermined 20-letter amino acid alphabet. Specifically, k i A peptide sequence p having amino acids i is 20·k i p i 20·(j-1)+1 ,p i 20·(j-1)+2 ,...,p i 20·j A single element in has a value of 1. The remaining elements have a value of 0. As an 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 3 amino acid peptide sequence EAF of data example i is a 60 element row vector The C-terminal flanking sequence c i , and the protein sequence for the MHC allele d h , and other sequence data in the presentation information can similarly be coded as above.

[0303] If the training data 170 contains sequences of amino acids of different lengths, the encoding module 314 may further encode the peptides into vectors of equivalent length by adding PAD characters to extend the predetermined alphabet. For example, this may be done by left padding the peptide sequence with PAD characters until the length of the peptide sequence reaches the peptide sequence with the longest length in the training data 170. Thus, if the peptide sequence with the longest length is k 最大 If the sequence has 20 amino acids, the encoding module 314 encodes each sequence as (20+1) k 最大We express it numerically as a row vector of k elements. For example, consider 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 最大 For a maximum amino acid length of =5, the same exemplary peptide sequence EAF of 3 amino acids is represented as a 105-element row vector The C-terminal flanking sequence c i or other sequence data can be similarly encoded as above. Thus, the peptide sequence p i or c i Each argument or row in represents the occurrence of a particular amino acid at a particular position in the sequence.

[0304] Although the above method for encoding sequence data has been described with respect to sequences having amino acid sequences, the method can be extended to other types of sequence data as well, such as, for example, DNA or RNA sequence data.

[0305] The encoding module 314 also encodes one or more MHC alleles a for data instance i. i into a row vector of m elements, where each element h=1,2,...,m corresponds to a unique specified MHC allele. The element corresponding to the MHC allele specified for data instance i has a value of 1. The remaining elements have a value of 0. As an example, the alleles HLA-B*07:02 and HLA-DRB1*10:01 for data instance i corresponding to a multi-allelic cell line among m=4 unique specified MHC allele types {HLA-A*01:01, HLA-C*01:08, HLA-B*07:02, HLA-DRB1*10:01} are encoded into a four-element row vector a i = [0 0 1 1], and a3 i =1 and a4 i = 1. An example with four identified MHC allele types is described herein, but the number of MHC allele types can in fact be hundreds or thousands. As mentioned above, each data instance i typically contains a peptide sequence pi Contains up to six different MHC allele types associated with

[0306] The encoding module 314 also generates a label y for each data instance i. i We code,x,as a binary variable with values ​​from the set {0,1}, where a value of 1 indicates that the peptide x i However, the associated MHC allele a i A value of 0 indicates that the peptide was presented by one of the peptides x i However, the associated MHC allele a i The dependent variable y i If represents the mass analysis ion current, the encoding module 314 may additionally scale the value using various functions, such as a log function having a range of [-∞, ∞] for ion current values ​​between [0, ∞].

[0307] The coding module 314 encodes the peptide p i and the allele interaction variable x for the associated MHC allele h h i A pair of alleles x may be represented as a row vector in which the numerical representations of the allele interaction variables are concatenated one after the other. For example, the encoding module 314 may h i [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 is the predicted binding affinity for peptide p and associated MHC allele h, and similarly, s h i is for stability. Alternatively, one or more combinations of allele interaction variables may be stored individually (e.g., as individual vectors or matrices).

[0308] In one example, the encoding module 314 encodes the measured or predicted values ​​for binding affinity as a function of the allele interaction variable x h i The binding affinity information is expressed by incorporating

[0309] In one example, the encoding module 314 encodes the measured or predicted values ​​for binding stability as allele interaction variables x h i By incorporating it into

[0310] In one example, the encoding module 314 encodes the measured or predicted values ​​for the binding on-rate as a function of the allele interaction variable x h i The binding on-rate information is represented by incorporating

[0311] In one example, for peptides presented by class I MHC molecules, the encoding module 314 encodes the peptide length in a vector TIFF0007651623000010.tif5132 (However, TIFF0007651623000011.tif3128 is the index function, L k is peptide p k The vector T k Let x be the allele interaction variable h i In another example, for peptides presented by class II MHC molecules, the encoding module 314 can include the peptide length in the vector TIFF0007651623000012.tif19161 (However, TIFF0007651623000013.tif3128 is the index function, L k is peptide p k The vector T k Let x be the allele interaction variable hi can be included in

[0312] In one example, the encoding module 314 represents the RNA expression information of MHC alleles by incorporating the RNA-seq based expression levels of the MHC alleles into an allele interaction variable xhi.

[0313] Similarly, the encoding module 314 encodes the allele non-interacting variable w i can be represented as a row vector in which the numerical representations of the allele-noninteraction variables are concatenated one after the other. For example, w i is [c i ] or [c i m i w i ], or a row vector equivalent to w i is the C-terminal adjacent sequence of peptide pi and the mRNA quantification measurement m i Alternatively, one or more combinations of allele-non-interacting variables may be stored individually (e.g., as individual vectors or matrices).

[0314] In one example, the encoding module 314 encodes the turnover rate or half-life as a function of the allele non-interacting variable w i represents the turnover rate of the source protein for a peptide sequence.

[0315] In one example, the encoding module 314 encodes the protein length in terms of the allele non-interacting variable w i represents the length of the source protein or isoform.

[0316] In one example, the encoding module 314 i , β2 i , β5 i The mean expression of immunoproteasome-specific proteasome subunits, including the subunits, was calculated using the allele-free interaction variable w iIncorporation into the IL-1 protein indicates activation of the immunoproteasome.

[0317] In one example, the encoding module 314 may encode the RNA-seq abundance of a peptide (quantified in units of FPKM, TPM by techniques such as RSEM) or a source protein of a gene or transcript of the peptide, and correlate the abundance of the source protein with an allele-free interaction variable w i This is expressed by incorporating it into

[0318] In one example, the encoding module 314 calculates the probability that the transcript from which the peptide originates will undergo nonsense-mediated decay (NMD), for example as estimated by the model in Rivas et al. Science, 2015, and calculates this probability as a function of an allele-noninteraction variable w i This is expressed by incorporating it into

[0319] In one example, the encoding module 314 represents the activation status of a gene module or pathway assessed via RNA-seq, for example, by quantifying the expression of genes in the pathway, e.g., in units of TPM using RSEM, for each of the genes in the pathway, and then computing a summary statistic, e.g., the mean, across the genes in the pathway. The mean is calculated using the allele non-interaction variable w i can be incorporated into.

[0320] In one example, the encoding module 314 encodes the copy number of the source gene by dividing the copy number by the allele non-interacting variable w i This is expressed by incorporating it into

[0321] In one example, the encoding module 314 encodes a measured or predicted TAP binding affinity (e.g., in nanomolar units) relative to the allele-non-interacting variable w i The TAP binding affinity is expressed by including

[0322] In one example, the encoding module 314 encodes the TAP expression levels measured by RNA-seq (and quantified, for example, by RSEM in units of TPM) into the allele-non-interacting variable w i The expression level of TAP is represented by including

[0323] In one example, the encoding module 314 encodes the tumor mutation as a non-allele interaction variable w i A vector of indicator variables in (i.e., peptide p k d if derived from a sample with a KRAS G12D mutation k = 1, otherwise 0).

[0324] In one example, the encoding module 314 encodes germline polymorphisms in antigen presenting genes as vectors of indicator variables (i.e., peptide p k d if derived from a sample with a specific germline polymorphism in TAP k = 1). We use these indicator variables as the allele non-interaction variables w i can be included in

[0325] In one example, the encoding module 314 represents the tumor types as a one-hot coded vector of length 1 over an alphabet of tumor types (e.g., NSCLC, melanoma, colon cancer, etc.). These one-hot coded variables are then combined into a set of allele-non-interaction variables w i can be included in

[0326] In one example, the encoding module 314 represents MHC allele suffixes by processing four-digit HLA alleles with various suffixes. For example, HLA-A*24:09N is considered a different allele than HLA-A*24:09 for the purposes of the model. Alternatively, the probability of presentation by MHC alleles with an N suffix can be set to zero for all peptides, since HLA alleles ending with an N suffix are not expressed.

[0327] In one example, the encoding module 314 represents the tumor subtype as a one-hot coded vector of length 1 for an alphabet of tumor subtypes (e.g., lung adenocarcinoma, lung squamous cell carcinoma, etc.). These one-hot coded variables are combined with the allele non-interaction variables w i can be included in

[0328] In one example, the encoding module 314 encodes smoking history as a function of the allele non-interaction variable w i A binary indicator variable (d if the patient has a smoking history) can be included in k = 1, otherwise 0). Alternatively, smoking history can be coded as a one-hot coded variable of length 1 for the alphabet of smoking severity. For example, smoking status can be assessed on a scale of 1 to 5, with 1 indicating non-smoker and 5 indicating current heavy smoker. Because smoking history is primarily relevant for lung tumors, when training models for multiple tumor types, this variable can also be defined as being equal to 1 if the patient has a history of smoking and the tumor type is lung tumor, and zero otherwise.

[0329] In one example, the encoding module 314 encodes sunburn history as a function of the allele non-interacting variable w i A binary indicator variable (d if the patient has a history of severe sunburn) can be included in k = 1, otherwise 0). Because severe sunburn is primarily associated with melanoma, when training models for multiple tumor types, this variable can also be defined as equal to 1 if the patient has a history of severe sunburn and the tumor type is melanoma, and zero otherwise.

[0330] In one example, the coding module 314 represents the distribution of expression levels of a particular gene or transcript for each gene or transcript in the human genome as a summary statistic (e.g., mean, median) of the distribution of expression levels by using a reference database such as TCGA. Specifically, the expression level of peptide p in samples with tumor type melanoma is expressed as k Regarding peptide p k The measured gene or transcript expression levels of the genes or transcripts of origin of alleles are compared with the allele-non-interacting variable w i Not only can it be included in the peptide p in melanoma as measured by TCGA, k The mean and / or median gene or transcript expression of the genes or transcripts of a given source may also be included.

[0331] In one example, the encoding module 314 represents the mutation types as one-hot coded variables of length 1 for an alphabet of mutation types (e.g., missense, frameshift, NMD-induced, etc.). These one-hot coded variables are grouped together into a set of allele-non-interacting variables w i can be included in

[0332] In one example, the encoding module 314 encodes the protein-level characteristics of the protein as values ​​of the source protein annotation (e.g., 5'UTR length) and the allele-non-interacting variable w i In another example, the coding module 314 encodes the peptide p i The residue-level annotation of the source protein for peptide p i is equal to 1 if it overlaps with the helical motif, and 0 otherwise; or i The allele non-interaction variable w i represents an indicator variable equal to 1 if p is completely contained within the helical motif. i The property that represents the proportion of residues in the allele-free variable w ican be included in

[0333] In one example, the encoding module 314 encodes the types of proteins or isoforms in the human proteome into an index vector o having a length equal to the number of proteins or isoforms in the human proteome. k and the corresponding element o k i is peptide p k is 1 if comes from protein i, and 0 otherwise.

[0334] In one example, the coding module 314 encodes the peptide p i Source gene G = gene(p i ) as a categorical variable with L possible categories, where L denotes an upper bound on the number of subscripted source genes 1, 2, ..., L.

[0335] In one example, the coding module 314 encodes the peptide p i T = tissue type, cell type, tumor type, or tumor histology type (p i ) as a categorical variable with M possible categories (where M denotes an upper limit on the number of subscripted types 1, 2, ..., M). Tissue types can include, for example, lung tissue, cardiac tissue, intestinal tissue, neural tissue, etc. Cell types can include, for example, dendritic cells, macrophages, CD4 T cells, etc. Lung adenocarcinoma, lung squamous cell carcinoma, melanoma, non-Hodgkin's lymphoma, etc.

[0336] The coding module 314 also encodes the peptide p i and the variable z for the associated MHC allele h i The entire set of alleles is treated as the interaction variable x i and the allele non-interaction variable w i For example, the encoding module 314 may represent z h i [xh i w i ] or [w i x h i ] can be represented as a row vector equivalent to

[0337] VIII. Training Module The training module 316 constructs one or more presentation models that generate a likelihood that a peptide sequence will be presented by an MHC allele associated with the peptide sequence. k and peptide sequence p k MHC alleles associated with a k Given a set of peptide sequences, each proposed model k However, the associated MHC allele a k The estimate u, which indicates the likelihood that one or more of k Generate.

[0338] VIII.A. Overview The training module 316 constructs one or more representation models based on a training data set stored in storage 170, which is generated from the representation information stored in 165. Generally, regardless of the specific type of representation model, all representation models capture the dependencies between independent and dependent variables in the training data 170 such that a loss function is minimized. Specifically, the loss function TIFF0007651623000014.tif5128 is a graph of the dependent variable y for one or more data examples S in the training data 170. i∈S and the estimated likelihood u for a data example S generated by the proposed model. i∈S In one particular implementation that is referred to throughout the remainder of this specification, the loss function TIFF0007651623000015.tif4128 is the negative log-likelihood function given by equation (1a) as follows: However, in practice, another loss function may be used. For example, if a prediction is made for mass spectrometry ion current, the loss function is the mean squared loss given by Equation 1b as follows: TIFF0007651623000017.tif11128

[0339] The proposed model may be a parametric model, in which one or more parameters θ mathematically specify the dependency between the independent and dependent variables. Typically, the loss function The various parameters of the proposed model of parametric type that minimizes TIFF0007651623000018.tif4128 are determined through a gradient-based numerical optimization algorithm, such as a batch gradient algorithm, a stochastic gradient algorithm, etc. Alternatively, the proposed model can be a non-parametric model, where the model structure is determined from training data 170 and is not strictly based on a fixed set of parameters.

[0340] VIII.B. Allele-specific model The training module 316 may build a presentation model to predict the presentation likelihood of a peptide on a per-allele basis. In this example, the training module 316 may train the presentation model based on example data S in the training data 170 generated from cells expressing a single MHC allele.

[0341] In one implementation, the training module 316 includes: TIFF0007651623000019.tif7128 gives the estimated presentation likelihood u of peptide pk for a particular allele h k where the peptide sequence x h k is peptide p k and the coded allele interaction variable for the corresponding MHC allele h, where f(·) is an arbitrary function, which for convenience of description will be referred to as a transformation function throughout this specification. h(·) is an arbitrary function, which for ease of description will be referred to throughout this specification as the dependence function, and the parameter θ determined for the MHC allele h h Based on the set of allele interaction variables x h k Generate a dependency score for each MHC allele h using the parameters θ h The set of values ​​is θ h where i is each example in a subset S of training data 170 generated from cells expressing a single MHC allele h.

[0342] Dependency function g h (x h k ;θ h ) is the output of the MHC allele h that has at least one allele interaction characteristic x h k Based on, and in particular on, the peptide p k The dependency score for MHC allele h indicates whether MHC allele h presents the corresponding neoantigen based on the amino acid position of the peptide sequence of p. k The transformation function f(·) transforms the input, more specifically, g in this example. h (x h k ;θ h ) is used to calculate the dependency score for peptide p k into an appropriate value indicating the likelihood that a gene will be presented by an MHC allele.

[0343] In one particular implementation mentioned throughout the remainder of this specification, f(·) is a function with range in [0,1] for the appropriate domain range. In one example, f(·) is This is the expit function given by TIFF0007651623000020.tif10128. As another example, f(·) also holds for values ​​in the domain z greater than or equal to 0. It can also be the hyperbolic tangent function given by TIFF0007651623000021.tif5128. Alternatively, if the prediction is made for mass analysis ion currents with values ​​outside the range [0,1], f(·) can be any function, for example the identity function, the exponential function, the log function, etc.

[0344] Therefore, the peptide sequence p k The allele-specific likelihood that ���� will be presented by MHC allele h is the dependence function g h (·) is the peptide sequence p k to generate a corresponding dependency score. k may be transformed by a transformation function f(·) to generate the per-allele likelihood that h will be presented by MHC allele h.

[0345] VIII.B.1 Dependence Functions for Allelic Interaction Variables In one particular implementation mentioned throughout this specification, the dependency function g h (·) is x h k Each allele interaction variable in is expressed as a function of the parameter θ determined for the relevant MHC allele h. h with the corresponding parameters in the set This is an affine function given by TIFF0007651623000022.tif6128.

[0346] In another specific implementation mentioned throughout this specification, the dependency function g h (·) is a network model NN with a set of nodes arranged in one or more layers. h (·), The network function is given by TIFF0007651623000023.tif6128. The nodes are connected with the parameters θh A node may be connected to other nodes through connections, each having an associated parameter in a set of activation functions. The value at one particular node may be expressed as the sum of the values ​​of the nodes connected to the particular node, weighted by the associated parameters mapped by the activation function associated with the particular node. In contrast to affine functions, network models are advantageous because the presentation model can incorporate nonlinearity and process data with amino acid sequences of different lengths. Specifically, through nonlinear modeling, the network model can capture the interactions between amino acids at different positions in a peptide sequence and how this interaction affects peptide presentation.

[0347] In general, the network model NN h (·) may be structured 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, and deep bidirectional recurrent networks.

[0348] In one example, which will be mentioned throughout the remainder of this specification, each MHC allele in h=1, 2, ..., m is associated with a separate network model, NN h (·) denotes the output from the network model related to MHC allele h.

[0349] FIG. 5 illustrates an exemplary network model NN3(·) associated with an arbitrary MHC allele h=3. As shown in FIG. 5, 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(·) includes three allele interaction variables x3 for MHC allele h=3. k (1), x3 k (2) and x3 k Receive input values ​​for (3) (individual data examples, including encoded polypeptide sequence data and any other training data used) and generate the value NN3(x3 k ) The network function may include one or more network models, each of which takes different allele interaction variables as input.

[0350] In another example, the identified MHC alleles h=1,2,...,m are modeled using a single network model NN H (·), and NN h (·) denotes one or more outputs of a single network model associated with MHC allele h. In such an example, the parameters θ h may correspond to a set of parameters for a single network model, and thus the parameters θ h A set of can be shared by all MHC alleles.

[0351] FIG. 6A shows an exemplary network model NN shared by MHC alleles h=1, 2, ..., m. H As shown in FIG. 6A, the network model NN H (·) contains m output nodes, each corresponding to an MHC allele. The network model NN3(·) has an allele interaction variable x3 for MHC allele h=3. k , and the value NN3(x3k ) to output m values.

[0352] In yet another example, a single network model NN H (·) is the allele interaction variable x for MHC allele h h k and the encoded protein sequence d h In such an example, the parameter θ h The set of parameters θ may again correspond to the set of parameters for a single network model, and thus the parameters θ h The set of MHC alleles can be shared by all MHC alleles. Thus, in such an example, NNh(·) is a single network model with inputs [x h k d h ], a single network model NN H Such a network model is advantageous because it can correctly predict peptide presentation probabilities for MHC alleles that were unknown in the training data simply by identifying their protein sequences.

[0353] FIG. 6B shows an exemplary network model NN shared by MHC alleles. H As shown in FIG. 6B, the network model NN H (·) takes as input the allele interaction variables and protein sequence of MHC allele h=3 and calculates the dependency score NN3 (x3 k ) to output.

[0354] In yet another example, the dependency function g h (·)teeth, TIFF0007651623000024.tif6128, where g' h (x h k ;θ' h) is an affine function, network function, etc., with a set of parameters θ′h, which represents the baseline probability of presentation for MHC allele h, and a bias parameter θ h 0 This is accompanied by:

[0355] In another implementation, the bias parameter θ h 0 may be shared according to the gene family of the MHC allele h. That is, the bias parameter θ for the MHC allele h h 0 is θ 遺伝子(h) 0 where gene (h) is the gene family of MHC allele h. For example, class I MHC alleles HLA-A*02:01, HLA-A*02:02, and HLA-A*02:03 may be assigned to the gene family "HLA-A", and the bias parameter θ for each of these MHC alleles may be h 0 As another example, the class II MHC alleles HLA-DRB1:10:01, HLA-DRB1:11:01, and HLA-DRB3:01:01 may be assigned to the gene family "HLA-DRB" and the bias parameters θ for each of these MHC alleles may be shared. h 0 can be shared.

[0356] As an example, going back to equation (2), the affine dependency function g h (·) was used to identify peptide p by MHC allele h = 3 among m = 4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000025.tif6128, where x3k is the allele interaction variable identified for MHC allele h=3 and θ3 is the set of parameters determined for MHC allele h=3 through loss function minimization.

[0357] As another example, we consider the MHC allele h = 3-dependent peptide p ... k The likelihood that will be presented is TIFF0007651623000026.tif6128 can be generated by k are the allele interaction variables identified for MHC allele h = 3, and θ3 is the set of parameters determined for the network model NN3(·) associated with MHC allele h = 3.

[0358] FIG. 7 shows the correlation coefficients of peptide p associated with MHC allele h=3 using the exemplary network model NN3(·). k As shown in Figure 7, the network model NN3(·) is a set of allele interaction variables x3 for MHC allele h=3. k Receives and outputs NN3(x3 k The output is then mapped by a function f(·) to produce an estimated presentation likelihood u k Generate.

[0359] VIII.B.2. Per-allele with non-allele interaction variables In one implementation, the training module 316 incorporates allele non-interacting variables to: Peptide p by TIFF0007651623000027.tif7131 k We model the estimated presentation likelihood uk of k is peptide p k means the coded allele non-interaction variable for g w (·) is the parameter θ determined for the allele non-interaction variable w Based on the set of allele non-interaction variables w k Specifically, the parameter θ h and the parameters θ for the allele-noninteracting variables w The set of values ​​of θ h and θw where i is each example in a subset S of training data 170 generated from cells expressing a single MHC allele.

[0360] Dependency function g w (w k ;θ w ) output is the percentage of peptides p10 ... k For example, the dependency score for an allele-non-interacting variable is k The C-terminal flanking sequences and peptide p, which are known to positively affect the presentation of k If peptide p is bound, it may have a high value k The C-terminal flanking sequences and peptide p, which are known to negatively affect the presentation of k If bound, it may have a low value.

[0361] According to equation (8), the peptide sequence p k The allele-specific likelihood that is presented by MHC allele h is the function g h (·) is the peptide sequence p k to generate the corresponding dependency scores for the allele interaction variables. w (·) is also applied to the coded version of the allele-non-interacting variables to generate a dependency score for the allele-non-interacting variables. Both scores are combined, and the combined score is the probability of a peptide sequence p being associated with MHC allele h. k are transformed by a transformation function f(·) to produce the allele-specific likelihoods that will be presented.

[0362] Alternatively, the training module 316 may include a training module 316 that uses the allele non-interaction variable w k Let x be the allele interaction variable. hk We may include the allele non-interaction variable w in the prediction by adding The image may be given by TIFF0007651623000028.tif7128.

[0363] VIII.B.3 Dependence Functions for Allelic Non-Interacting Variables Dependence function g for allelic interaction variables h Similarly to (·), the dependence function g w (·) is an affine function, or a separate network model for the allele non-interaction variables w k It may be a network function related to

[0364] Specifically, the dependency function g w (·) is w k The allele-free interaction variables in are the parameters θ w with the corresponding parameters in the set This is an affine function given by TIFF0007651623000029.tif6128.

[0365] Dependency function g w (·) also represents the parameter θ w The network model NN has relevant parameters in the set w (·), TIFF0007651623000030.tif6128. The network function may include one or more network models, each taking different allele-non-interacting variables as input.

[0366] In another example, the dependence function g for the allele-non-interacting variables w (·)teeth, TIFF0007651623000031.tif6128, where g' w (w k ;θ' w) is the allele non-interaction parameter θ' w are affine functions, network functions, etc., with a set of m k is peptide p k is the mRNA quantification measurement for θ(·), h(·) is a function that transforms the quantification measurement, and θ w m is a parameter in the set of parameters for the allele-non-interacting variables that is combined with the mRNA quantification measurement to generate a dependency score for the mRNA quantification measurement. In one particular embodiment mentioned throughout the remainder of this specification, h(·) is a log function, although in practice h(·) can be any one of a variety of different functions.

[0367] In yet another example, the dependence function g for the allele-non-interacting variables w (·)teeth, TIFF0007651623000032.tif6128, where g' w (w k ;θ' w ) is the allele non-interaction parameter θ' w affine functions, network functions, etc. with a set of k is peptide p k is the indicator vector described in Section VII.C.2, which represents proteins and isoforms in the human proteome, and w o is the set of parameters in the set of parameters for the allele non-interaction variables that are combined with the indicator vector. k and parameter θ w o If the dimension of the set of is significantly higher, TIFF0007651623000033.tif5128( A parameter regularization term such as L1 norm, L2 norm, combination, etc. can be added to the loss function when determining the value of the parameter. The optimal value of the hyperparameter λ can be determined through an appropriate method.

[0368] In yet another example, the dependence function g for the allele-non-interacting variables w (·)teeth, TIFF0007651623000035.tif14131, where g' w (w k ;θ' w ) is the allele non-interaction parameter θ' w affine functions, network functions, etc. with a set of TIFF0007651623000036.tif5128 is peptide p k is an indicator function that is equal to 1 if θ is derived from the source gene l as described above for allele-free interaction variables, and w l is a parameter that indicates the "antigenicity" of the source gene l. In one variation, L is sufficiently large, so that the number of parameters θ w l=1, 2,...,L If is large enough, Parameter regularization term such as TIFF0007651623000037.tif5128 (where TIFF0007651623000038.tif4128 can add L1 norm, L2 norm, combination, etc. to the loss function when determining the value of the parameter. The optimal value of the hyperparameter λ can be determined by an appropriate method.

[0369] In yet another example, the dependence function g for the allele-non-interacting variables w (·)teeth, TIFF0007651623000039.tif14158, where g' w (w k ;θ' w) is the allele non-interaction parameter θ' w affine functions, network functions, etc. with a set of TIFF0007651623000040.tif5128 is peptide p as described above for allele non-interaction variables. k is derived from source gene l, and peptide p k is an indicator function that is equal to 1 if originates from tissue type m, and θ w lm is a parameter indicative of the antigenicity of the combination of source gene l and tissue type m. In particular, the antigenicity of gene l in tissue type m may indicate the residual tendency of cells of tissue type m to present peptides derived from gene l after adjustment for RNA expression and peptide sequence context.

[0370] In one variation, L or M is sufficiently large, so that the number of parameters θ w lm=1, 2,...,LM If is large enough, Parameter regularization term such as TIFF0007651623000041.tif5128 (where TIFF0007651623000042.tif4128 can be added to the loss function when determining the value of the parameter (such as L1 norm, L2 norm, combination, etc.). The optimal value of the hyperparameter λ can be determined by an appropriate method. In another variation, a parameter regularization term can be added to the loss function when determining the value of the parameter so that the coefficients for the same source gene do not differ significantly between tissue types. For example, a penalty term such as: TIFF0007651623000043.tif19128 (in the formula, TIFF0007651623000044.tif5128 is the average antigenicity across tissue types for source gene l) can add a penalty to the standard deviation of antigenicity across different tissue types in the loss function.

[0371] In practice, the dependence function g for the allele-non-interacting variables can be calculated by combining any of the additional terms in equations (10), (11), (12a) and (12b). w For example, the term h(·) representing the mRNA quantification measurement in equation (10) and the term representing the antigenicity of the source gene in equation (12) can be added together along with any other affine or network functions to generate a dependency function for the allele non-interacting variables.

[0372] As an example, going back to equation (8), the affine transformation function g h (·), g w (·) was used to identify peptide p by MHC allele h = 3 among m = 4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000045.tif6128, where w k is peptide p k are allele-free interaction variables identified for θ w is the set of parameters determined for the allele non-interacting variables.

[0373] As another example, the network transformation function g h (·), g w (·) was used to identify peptide p by MHC allele h = 3 among m = 4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000046.tif6128, where w k is peptide p k are allele interaction variables identified for θ w is the set of parameters determined for the allele non-interacting variables.

[0374] FIG. 8 shows exemplary network models NN3(·) and NN w Peptide p associated with MHC allele h=3 using (·) kAs shown in Figure 8, the network model NN3(·) is a set of allele interaction variables x3 for MHC allele h=3. k Receives and outputs NN3(x3 k ) is generated. w (·) is peptide p k The allele non-interaction variable w k Receives and outputs NN w (w k The outputs are then combined and mapped by a function f(·) to produce an estimated presentation likelihood u k Generate.

[0375] VIII.C. Multi-Allele Models The training module 316 may also build a presentation model for predicting the presentation likelihood of a peptide in a multi-allelic setting where more than one MHC allele is present. In this example, the training module 316 may train the presentation model based on example data S in the training data 170 generated from cells expressing a single MHC allele, cells expressing multiple MHC alleles, or a combination thereof. EXAMPLES

[0376] VIII.C.1. Example 1: Maximum per Allele Model In one implementation, the training module 316 trains peptides p associated with a set of MHC alleles H. k The estimated presentation likelihood u k is the presentation likelihood determined for each of the MHC alleles h in set H determined based on cells expressing a single allele, as described above in conjunction with equations (2)-(11). We model the likelihood of the presentation as a function of TIFF0007651623000047.tif4128. Specifically, k teeth, The function can be any function of TIFF0007651623000048.tif4128. In one implementation, the function is a maximum function, as shown in equation (12), where the presented likelihood u kcan be determined as the maximum of the presentation likelihood for each MHC allele h in set H. TIFF0007651623000049.tif6128

[0377] VIII.C.2. Example 2.1: Sum Function Model In one implementation, the training module 316 trains peptides p k The estimated presentation likelihood u k of, Modeled by TIFF0007651623000050.tif14128, where element a h k is the peptide sequence p k 1 for multiple MHC alleles H associated with x h k is peptide p k and the coded allele interaction variables for the corresponding MHC alleles. h The set of values ​​is θ h The dependence function g can be determined by minimizing a loss function for each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. h is the dependency function g introduced above in Section VIII.B.1. h The composition may be in any of the following forms:

[0378] According to equation (13), the peptide sequence p k The likelihood that a gene will be presented by one or more MHC alleles h is given by the dependence function g h (·) for each of the MHC alleles H, k to generate a corresponding score for the allele interaction variables. The scores for each MHC allele h are combined to generate a corresponding score for the peptide sequence p k is transformed by a transformation function f(·) to produce the presentation likelihood that α will be presented by the set of MHC alleles H.

[0379] The model presented in equation (13) is that for each peptide p k It differs from the allele-by-allele model of equation (2) in that the number of associated alleles for a can be greater than 1. In other words, h k More than one element in the peptide sequence p k can have a value of 1 for multiple MHC alleles H associated with

[0380] As an example, the affine transformation function g h (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000051.tif6128 can be generated by k , x3 k are the allele interaction variables identified for MHC alleles h=2, h=3, and θ2, θ3 are the set of parameters determined for MHC alleles h=2, h=3.

[0381] As another example, the network transformation function g h (·), g w (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000052.tif6128, where NN2(·), NN3(·) are the network models specified for MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters determined for MHC alleles h=2, h=3.

[0382] FIG. 9 shows the correlation between peptide p associated with MHC alleles h=2 and h=3 using exemplary network models NN2(·) and NN3(·). kAs shown in Figure 9, the network model NN2(·) is a set of allele interaction variables x2 for MHC allele h=2. k Receives and outputs NN2(x2 k ) and the network model NN3(·) is the allele interaction variable x3 for MHC allele h=3. k Receives and outputs NN3(x3 k The outputs are then combined and mapped by a function f(·) to produce an estimated presentation likelihood u k Generate.

[0383] VIII.C.3. Example 2.2: Sum Function Models with Allelic Non-Interacting Variables In one implementation, the training module 316 incorporates allele non-interacting variables to: Peptide p by TIFF0007651623000053.tif14142 k The estimated presentation likelihood u k where w k is peptide p k Specifically, the parameters θ for each MHC allele h h and the parameters θ for the allele-noninteracting variables w The set of values ​​of θ h and θ w The dependence function g can be determined by minimizing a loss function for each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. w is the dependency function g introduced above in Section VIII.B.3. w The composition may be in any of the following forms:

[0384] Thus, according to equation (14), one or more MHC alleles H can bind to a peptide sequence p k The likelihood of being presented is the function g h (·) for each of the MHC alleles H, kto generate corresponding dependency scores for allele interaction variables for each MHC allele h. w (·) is also applied to the coded versions of the allele-non-interacting variables to generate a dependency score for the allele-non-interacting variables. The scores are combined and the combined score is used to estimate the association between the peptide sequence p and the MHC allele H. k is transformed by a transformation function f(·) to produce the presentation likelihood that

[0385] In the model presented in equation (14), each peptide p k The number of related alleles for a can be greater than 1. h k More than one element in the peptide sequence p k can have a value of 1 for multiple MHC alleles H associated with

[0386] As an example, the affine transformation function g h (·), g w (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000054.tif6128, where w k is peptide p k are allele-free interaction variables identified for θ w is the set of parameters determined for the allele non-interacting variables.

[0387] As another example, the network transformation function g h (·), g w (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000055.tif6128, where w k is peptide p k are allele interaction variables identified for θ w is the set of parameters determined for the allele non-interacting variables.

[0388] FIG. 10 shows exemplary network models NN2(·), NN3(·), and NN w Peptide p associated with MHC alleles h=2 and h=3 using (·) k As shown in FIG. 10, the network model NN2(·) generates the allele interaction variable x2 for the MHC allele h=2. k Receives and outputs NN2(x2 k The network model NN3(·) generates the allele interaction variable x3 for MHC allele h=3. k Receives and outputs NN3(x3 k ) is generated. w (·) is peptide p k The allele non-interaction variable w k Receives and outputs NN w (w k The outputs are then combined and mapped by a function f(·) to produce an estimated presentation likelihood u k Generate.

[0389] Alternatively, the training module 316 may use the allele non-interacting variable w k Let x be the allele interaction variable. h k By adding to the allele non-interaction variable w k Thus, the presentation likelihood may include The image can be given by TIFF0007651623000056.tif14134.

[0390] VIII.C.4. Example 3.1: Model with Implicit Per-Allele Likelihood In another implementation, the training module 316 trains peptide p k The estimated presentation likelihood u k of, Modeled by TIFF0007651623000057.tif7142, where element a h k is the peptide sequence p k Multiple MHC alleles associated with 1 for TIFF0007651623000058.tif4128, and u' k h is the implicit allele-wise presentation likelihood for MHC allele h, and vector v has elements v h But, a h k ·u' k h where v is a vector corresponding to, s(·) is a function that maps the elements of v, and r(·) is a clipping function that clips the values ​​of the input into a predetermined range. As described in more detail below, s(·) may be a summation function or a quadratic function, although it will be appreciated that in other embodiments s(·) may be any function, such as a maximum function. A set of values ​​for the parameters θ for the implicit per-allele likelihood can be determined by minimizing a loss function with respect to θ, where i are each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles.

[0391] The presentation likelihood in the presentation model of equation (17) is the likelihood that each individual MHC allele h will present a peptide p k The implicit allele presentation likelihood u' corresponds to the likelihood that k h The implicit per-allele likelihood differs from the per-allele presentation likelihood of Section VIII.B in that parameters for the implicit per-allele likelihood can be learned from a multi-allelic setting, in addition to a single-allelic setting, where the direct association between the presented peptide and the corresponding MHC allele is unknown. Thus, in the multi-allelic setting, the presentation model is kNot only can we estimate whether peptide p is presented by the set of MHC alleles H as a whole, but also which MHC alleles h are presented by peptide p k The individual likelihoods that TIFF0007651623000059.tif4128 can also be provided. The advantage of this is that the presented model can generate implicit likelihoods without training data for cells expressing a single MHC allele.

[0392] In one particular implementation that will be mentioned throughout the remainder of this specification, r(·) is a function with range [0,1]. For example, r(·) is the clip function: r(z)=min(max(z,0),1) and the minimum value between z and 1 is the presented likelihood u k In another implementation, r(·) is chosen as r(z)=tanh(z) where the domain z is 0 or greater.

[0393] VIII.C.5. Example 3.2: Sum of Functions Model In one particular implementation, s(·) is a summation function, and the presentation likelihood is given by summing the implicit per-allele presentation likelihoods. TIFF0007651623000060.tif16128

[0394] In one implementation, the implicit per-allele presentation likelihood for MHC allele h is defined as: The proposed likelihood is generated by TIFF0007651623000061.tif7128. As estimated by TIFF0007651623000062.tif14129.

[0395] According to equation (19), one or more MHC alleles H can bind to a peptide sequence p k The likelihood of being presented is the function g h(·) for each of the MHC alleles H, k to generate the corresponding dependency scores for the allele interaction variables. Each dependency score can be generated by first applying k h The allele likelihood u' is transformed by the function f(·) to generate k h are combined and a clipping function is applied to the combined likelihood to clip the values ​​into the range [0,1] to produce a peptide sequence p k A presentation likelihood can be generated that g will be presented by a set of MHC alleles H. h is the dependency function g introduced above in Section VIII.B.1. h The composition may be in any of the following forms:

[0396] As an example, the affine transformation function g h (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000063.tif7128 can be generated by k , x3 k are the allele interaction variables identified for MHC alleles h=2, h=3, and θ2, θ3 are the set of parameters determined for MHC alleles h=2, h=3.

[0397] As another example, the network transformation function g h (·), g w (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000064.tif7128, where NN2(·), NN3(·) are the network models specified for MHC alleles h=2, h=3, and θ2, θ3 are the sets of parameters determined for MHC alleles h=2, h=3.

[0398] FIG. 11 shows the correlation between peptide p associated with MHC alleles h=2 and h=3 using exemplary network models NN2(·) and NN3(·). k As shown in Figure 9, the network model NN2(·) is a set of allele interaction variables x2 for MHC allele h=2. k Receives and outputs NN2(x2 k ) and the network model NN3(·) is the allele interaction variable x3 for MHC allele h=3. k Receives and outputs NN3(x3 k Each output is then mapped by a function f(·) and combined to produce an estimated presentation likelihood u k Generate.

[0399] In another implementation, if the prediction is made in terms of the log of the mass analysis ion current, then r(·) is a log function and f(·) is an exponential function.

[0400] VIII.C.6. Example 3.3: Sum of Functions Model with Allelic Non-Interacting Variables In one implementation, the implicit per-allele presentation likelihood for MHC allele h is defined as: The proposed likelihood is generated by TIFF0007651623000065.tif7128. Incorporates the influence of allele-non-interacting variables into peptide presentation as generated by TIFF0007651623000066.tif14131.

[0401] According to equation (21), one or more MHC alleles H can bind to a peptide sequence p k The likelihood of being presented is the function g h(·) for each of the MHC alleles H, k to generate corresponding dependency scores for allele interaction variables for each MHC allele h. w (·) is also applied to the coded version of the allele non-interaction variables to generate a dependency score for the allele non-interaction variables. The scores of the allele non-interaction variables are combined with each of the dependency scores of the allele interaction variables. Each of the combined scores is transformed by a function f(·) to generate an implicit per-allele presentation likelihood. The implicit likelihoods are combined and a clipping function is applied to the combined output to clip the values ​​into the range [0,1] to estimate the likelihood of presentation of the peptide sequence p by the MHC allele H. k A likelihood of being proposed may be generated. w is the dependency function g introduced above in Section VIII.B.3. w The composition may be in any of the following forms:

[0402] As an example, the affine transformation function g h (·), g w (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000067.tif7128, where w k is peptide p k are allele-free interaction variables identified for θ w is the set of parameters determined for the allele non-interacting variables.

[0403] As another example, the network transformation function g h (·), g w (·) was used to identify peptide p by MHC alleles h=2 and h=3 among m=4 different identified MHC alleles. k The likelihood that will be presented is TIFF0007651623000068.tif7139, where w k is peptide p k are allele interaction variables identified for θ w is the set of parameters determined for the allele non-interacting variables.

[0404] FIG. 12 shows exemplary network models NN2(·), NN3(·), and NN w Peptide p associated with MHC alleles h=2 and h=3 using (·) k As shown in FIG. 12, the network model NN2(·) is a model of the MHC allele h=2 with an allele interaction variable x2 k Receives and outputs NN2(x2 k ) is generated. w (·) is peptide p k The allele non-interaction variable w k Receives and outputs NN w (w k The outputs are combined and mapped by a function f(·). The network model NN3(·) generates a set of allele interaction variables x3 for MHC allele h=3. k Receives and outputs NN3(x3 k ) is generated, which is also the same network model NN w (·) Output NN w (w k ) and mapped by a function f(·). Both outputs are combined to give the estimated presentation likelihood u k Generate.

[0405] In another implementation, the implicit per-allele presentation likelihood for MHC allele h can be expressed as: The proposed likelihood is generated by TIFF0007651623000069.tif7128. It will be generated by TIFF0007651623000070.tif14128.

[0406] VIII.C.7. Example 4: Quadratic Model In one implementation, s(·) is a quadratic function, and the peptide p k The estimated presentation likelihood u k teeth, TIFF0007651623000071.tif14137, where element u' k h is the implicit per-allele presentation likelihood for MHC allele h. A set of values ​​for parameters θ for the implicit per-allele likelihood can be determined by minimizing a loss function with respect to θ, where i is each example in a subset S of training data 170 generated from cells expressing a single MHC allele and / or cells expressing multiple MHC alleles. The implicit per-allele presentation likelihood can be of any of the forms shown in equations (18), (20), and (22) above.

[0407] In one embodiment, the model of equation (23) is k However, the possibility exists that a given antigen may be presented simultaneously by two MHC alleles, which may imply that presentation by the two HLA alleles is statistically independent.

[0408] According to equation (23), one or more MHC alleles H can bind to a peptide sequence p k The presentation likelihood that p will be presented is calculated by combining the implicit per-allele presentation likelihoods and the likelihood that p will be presented by the MHC allele H. k Each pair of MHC alleles is presented with a peptide p such that it generates a presentation likelihood that k can be generated by subtracting from the sum the likelihood that

[0409] As an example, the affine transformation function g h (·) was used to identify peptide p by HLA alleles h=2 and h=3 among m=4 different identified HLA alleles. k The likelihood that will be presented is TIFF0007651623000072.tif6128 can be generated byk , x3 k are the allele interaction variables specified for HLA alleles h=2, h=3, and θ2, θ3 are the set of parameters determined for HLA alleles h=2, h=3.

[0410] As another example, the network transformation function g h (·), g w (·) was used to identify peptide p by HLA alleles h=2 and h=3 among m=4 different identified HLA alleles. k The likelihood that will be presented is TIFF0007651623000073.tif6137, where NN2(·), NN3(·) are the network models specified for HLA alleles h=2, h=3, and θ2, θ3 are the sets of parameters determined for HLA alleles h=2, h=3.

[0411] IX. Example 5: Prediction Module The prediction module 320 receives the sequence data and selects candidate neoantigens in the sequence data using the proposed model. Specifically, the sequence data may be DNA sequences, RNA sequences, and / or protein sequences extracted from the patient's tumor tissue cells. The prediction module 320 converts the sequence data into a plurality of peptide sequences p having 8-15 amino acids for MHC-I or 6-30 amino acids for MHC-II. k For example, the prediction module 320 processes a given sequence TIFF0007651623000074.tif4128, three peptide sequences each with nine amino acids TIFF0007651623000075.tif9128. In one embodiment, the prediction module 320 can identify candidate neoantigens that are mutated peptide sequences by comparing sequence data extracted from normal tissue cells of a patient with sequence data extracted from tumor tissue cells of the patient to identify portions that have one or more mutations.

[0412] The prediction module 320 applies one or more of the presentation models to the processed peptide sequences to estimate the presentation likelihood of the peptide sequences. Specifically, the prediction module 320 can select one or more candidate neoantigen peptide sequences that are likely to be presented on tumor HLA molecules by applying the presentation model to the candidate neoantigens. In one implementation, the prediction module 320 selects the candidate neoantigen sequences that have an estimated presentation likelihood above a pre-determined threshold. In another implementation, the presentation model selects the v candidate neoantigen sequences with the highest estimated presentation likelihood (v is generally the maximum number of epitopes that can be delivered in a vaccine). A vaccine containing the selected candidate neoantigens for a given patient can be injected into the patient to induce an immune response.

[0413] X. Example 6: Patient Selection Module The patient selection module 324 selects a subset of patients for vaccine treatment based on whether the patient meets the inclusion criteria. In one embodiment, the inclusion criteria is determined based on the likelihood of presentation of the neoantigen candidates of the patient generated by the presentation model. By adjusting the inclusion criteria, the patient selection module 324 can adjust the number of patients who receive the vaccine based on the likelihood of presentation of the neoantigen candidates of the patient. Specifically, strict inclusion criteria may result in a smaller number of patients treated with the vaccine, but a higher proportion of vaccine-treated patients who receive effective treatment (e.g., one or more tumor-specific neoantigens (TSNAs) are delivered). In contrast, loose inclusion criteria may result in a larger number of patients treated with the vaccine, but a lower proportion of vaccine-treated patients who receive effective treatment. The patient selection module 324 modifies the inclusion criteria based on a desired balance between a target proportion of patients who receive the vaccine and a proportion of patients who receive effective treatment as a result of the vaccine treatment.

[0414] In one embodiment, a patient is associated with a corresponding therapeutic subset of v neoantigen candidates that can potentially be included in a personalized vaccine for the patient with a vaccine volume v. In one embodiment, the therapeutic subset for a patient is the neoantigen candidate with the highest presentation likelihood as determined by the presentation model. For example, if a vaccine may include v=20 epitopes, the vaccine may include a therapeutic subset for each patient with the highest presentation likelihood as determined by the presentation model. However, it is recognized that in other embodiments, the therapeutic subset for a patient may be determined based on other methods. For example, the therapeutic subset for a patient may be randomly selected from the set of neoantigen candidates for that patient, or may be determined in part based on a combination of factors including prior art models of binding affinity or stability of peptide sequences, or presentation likelihood obtained from a presentation model and affinity or stability information for those peptide sequences.

[0415] In one embodiment, the patient selection module 324 determines that a patient meets the inclusion criteria if the patient's tumor mutation burden is equal to or higher than a minimum mutation burden. A patient's tumor mutation burden (TMB) indicates the total number of nonsynonymous mutations in the tumor exome. In one embodiment, the patient selection module 324 selects a patient suitable for vaccine treatment if the absolute number of the patient's TMB is equal to or higher than a predefined threshold. In another implementation, the patient selection module 324 selects a patient suitable for vaccine treatment if the patient's TMB is within a threshold percentile between the TMBs determined for the set of patients.

[0416] In another embodiment, the patient selection module 324 determines that a patient meets the inclusion criteria if the patient's utility score based on the patient's therapeutic subset is equal to or higher than a minimum utility score. In one embodiment, the utility score is a measure of the estimated number of presented antigens from the therapeutic subset.

[0417] The expected number of presented antigens can be predicted by modeling the presentation of neoantigens as random variables with one or more probability distributions. In one implementation, the utility score for patient i is the expected number of presented neoantigen candidates from the treatment subset, or a specific function thereof. As an example, the presentation of each neoantigen can be modeled as a Bernoulli random variable, where the probability of presentation (success) is given by the presentation likelihood of the neoantigen candidate. In particular, the presentation likelihood u of each neoantigen candidate with the highest presentation likelihood u i1 , u i2 , …, u iv Possessing v types of neoantigen candidates p i1 , p i2 , …, p iv Treatment subset of S i Regarding neoantigen candidate p ij The presentation of random variable A ij where: The expected number of neoantigens presented is given by the sum of the likelihoods of presentation of each neoantigen candidate. In other words, the utility score for patient i is expressed as: The patient selection module 324 selects a subset of patients having a utility score equal to or greater than the minimum utility for the vaccine treatment.

[0418] In another implementation, the utility score for patient i is the probability that at least a threshold number of neoantigens k are presented. In one example, a therapeutic subset of neoantigen candidates S i The number of presented antigens in is modeled as a Poisson binomial random variable where the probability of presentation (success) is given by the presentation likelihood of each of the epitopes. In detail, the number of presented antigens for patient i is determined by the random variable N i can be given by: TIFF0007651623000078.tif14128In the formula, PBD(·) denotes the Poisson binomial distribution. The probability that at least a threshold number of neoantigens k are presented is iis given by the operation of the probability that k is equal to or greater than k. In other words, the utility score for patient i is expressed as: The patient selection module 324 selects a subset of patients having a utility score equal to or greater than the minimum utility for the vaccine treatment.

[0419] In another embodiment, the number of neoantigens in the therapeutic subset for a patient need not be limited to the vaccine volume v, and the patient selection module 324 can select the patient using a utility score determined based on any set of candidate neoantigens for that patient. For example, the utility score can be determined based on all mutations or candidate neoantigens identified for that patient. The utility score can be generated, for example, using the methods described in conjunction with equations (24)-(27), where v is a variable v(i) that depends on patient i, and indicates the total number of mutations or candidate neoantigens identified for that patient.

[0420] In another implementation, the utility score for patient i is calculated based on a therapeutic subset S of neoantigen candidates that have a binding affinity or predicted binding affinity below a fixed threshold (e.g., 500 nM) for one or more of the patient's HLA alleles. i The number of neoantigens in the nucleus is the number of neoantigens in the nucleus. In one example, the fixed threshold is in the range of 1000 nM to 10 nM. In some cases, the utility score may count only neoantigens detected as expressed by RNA-seq.

[0421] In another implementation, the utility score for patient i is determined by dividing the therapeutic subset S of candidate neoantigens for which the binding affinity for one or more HLA alleles for that patient is equal to or less than a threshold percentile of the binding affinity of a random peptide for that HLA allele. iThe threshold percentile is the number of neoantigens in the 10th percentile. In one example, the threshold percentile is between the 10th percentile and the 0.1th percentile. In some cases, the utility score may only count neoantigens detected as expressed by RNA-seq.

[0422] It will be appreciated that the example utility scores described with respect to equations (25) and (27) are merely illustrative, and that the patient selection module 324 may use other statistics or probability distributions to generate the utility scores.

[0423] XI. Example 7: Neoantigen Loading for Immune Checkpoint Inhibitor Therapy and Other Immunotherapies The patient selection module 324 can also use the utility score defined in section X above to select patients for immune checkpoint inhibitor therapy (e.g., PD-1, CTLA4) or any other immunotherapy for which neoantigen load may be associated with efficacy, including immunostimulatory agents, immune stimulatory molecule agonists (e.g., CD40), oncolytic viruses (e.g., T-VEC), neoantigen or other cancer antigen-containing therapeutic vaccines, neoantigen or other cancer antigen-targeted adoptive cell therapy, tumor microenvironment modulators (e.g., TGFβ), or any combination of these with immune checkpoint inhibitors.

[0424] For example, in some embodiments, the immunostimulatory agent is an agent that blocks signaling of an inhibitory receptor or its ligand on an immune cell. In some embodiments, the inhibitory receptor or ligand is selected from CTLA-4, PD-1, PD-L1, LAG-3, Tim3, TIGIT, neuritin, BTLA, KIR, and combinations thereof. In some aspects, the agent is selected from an anti-PD-1 antibody (e.g., pembrolizumab or nivolumab), an anti-PD-L1 antibody (e.g., atezolizumab), an anti-CTLA-4 antibody (e.g., ipilimumab), and combinations thereof. In some aspects, the agent is pembrolizumab. In some aspects, the agent is nivolumab. In some aspects, the agent is atezolizumab.

[0425] In some embodiments, the therapeutic agent is an agent that inhibits the interaction of PD-1 and PD-L1. In some aspects, the additional agent that inhibits the interaction of PD-1 and PD-L1 is selected from an antibody, a peptidomimetic, and a small molecule. In some aspects, the additional agent that inhibits the interaction of PD-1 and PD-L1 is selected from pembrolizumab, nivolumab, atezolizumab, avelumab, durvalumab, BMS-936559, sulfamonomethoxine 1, and sulfamethizole 2. In some embodiments, the additional agent that inhibits the interaction of PD-1 and PD-L1 is any therapeutic agent known in the art having such activity, for example, as described in Weinmann et al., Chem Med Chem, 2016, 14:1576 (DOI:10.1002 / cmdc.201500566), which is incorporated by reference in its entirety.

[0426] In some embodiments, the immunostimulatory agent is an agonist of a costimulatory receptor of an immune cell. In some aspects, the costimulatory receptor is selected from OX40, ICOS, CD27, CD28, 4-1BB, and CD40. In some embodiments, the agonist is an antibody.

[0427] In some embodiments, the immunostimulatory agent is a cytokine, hi some aspects, the cytokine is selected from IL-2, IL-5, IL-7, IL-12, IL-15, IL-21, and combinations thereof.

[0428] In some embodiments, the immunostimulatory agent is an oncolytic virus. In some aspects, the oncolytic virus is selected from herpes simplex virus, vesicular stomatitis virus, adenovirus, Newcastle disease virus, vaccinia virus, and Maraba virus.

[0429] In some embodiments, the immunostimulatory agent is a T cell bearing a chimeric antigen receptor (CAR-T cell). In some embodiments, the immunostimulatory agent is a bispecific or multispecific T cell directed antibody. In some embodiments, the immunostimulatory agent is an anti-TGF-β antibody. In some embodiments, the immunostimulatory agent is a TGF-β trap.

[0430] In some embodiments, the therapeutic agent is a vaccine against a tumor antigen. The vaccine can be targeted to any suitable antigen, provided that the antigen is present in the tumor treated by the methods provided herein. In some aspects, the tumor antigen is a tumor antigen that is overexpressed relative to its expression level in normal tissue. In some aspects, the tumor antigen is selected from a cancer testis antigen, a differentiation antigen, NY-ESO-1, MAGE-A1, MART, and combinations thereof. In some embodiments, the therapeutic agent is a vaccine against one or more neoantigens. The neoantigens in the vaccine can be identified by the methods provided herein.

[0431] In particular, the patient selection module 324 determines a neoantigen load indicating the total expected number of presented neoantigens for each patient. Checkpoint inhibitor therapy can be administered to patients with a neoantigen load that meets the inclusion criteria. For example, such therapy can be administered to patients with a neoantigen load above a predefined threshold. In one embodiment, the neoantigen load is a utility score as set forth in Section XI, where v is the total number of mutations or candidate neoantigens identified for the patient, rather than a subset of candidate antigens for that patient.

[0432] A high neoantigen load in a particular tumor relative to the median may indicate that a subject with that tumor is more likely to benefit from treatment with a checkpoint inhibitor, such as anti-CTLA4, anti-PD1, and / or anti-PDL1. For example, neoantigen load may be a better indicator of efficacy with checkpoint inhibitors compared to mutational load, since neoantigens are generally presented on the surface of tumor cells and are more likely to be recognized by T cells with higher activity against tumors following checkpoint inhibitor therapy.

[0433] In another embodiment, the patient selection module 324 may use a utility score generated from a combination of one or more of the following characteristics: predicted HLA class I neoantigen load, predicted HLA class II neoantigen load, and tumor mutation load. The predicted HLA class I neoantigen load for a patient is the neoantigen load for the patient's set of class I HLA alleles, indicating the total expected number of neoantigens presented on the patient's class I HLA alleles. The predicted HLA class II neoantigen load for a patient is the neoantigen load for the patient's set of class II HLA alleles, indicating the total expected number of neoantigens presented on the patient's class II HLA alleles. For example, the utility score may be calculated as f(class I neoantigen load, class II neoantigen load, tumor mutation load; b), where f(·) is a function parameterized by the set of machine-learned parameters b. The set of machine-learned parameters b may be tumor type dependent (e.g., b may be different for melanoma and non-small cell lung cancer).

[0434] In another embodiment, the patient selection module 324 can use a utility score that incorporates information about immunogenic tumor antigens other than neoantigens. Examples of immunogenic tumor antigens other than neoantigens include cancer-germline antigens (CGAs, e.g., MAGEA3), differentiation antigens (e.g., tyrosinase), and antigens overexpressed in tumors (e.g., CEA). The expression levels of these antigens can be determined using at least tumor RNA sequencing data, and the expected number of HLA class I or class II epitopes from these genes that are presented by the HLA alleles of the patient's tumor can be determined by applying a presentation model to each peptide from the set of tumor antigens using the RNA sequencing data for each gene. These presentation likelihoods can be incorporated into a utility score that is calculated as f(class I neoantigen load, class II neoantigen load, tumor mutation load, class I non-neoantigen tumor antigen load, class II non-neoantigen tumor antigen load; b), where f(·) is a function parameterized by the set of machine-learned parameters b. The set of machine-learned parameters b may depend on the tumor type (e.g., b may be different for melanoma and non-small cell lung cancer).

[0435] A higher usefulness score indicates that the tumor presents more HLA epitopes that are recognized as foreign or non-self by the patient's immune system. Patients with tumors that present more non-self HLA epitopes may be more likely to respond to checkpoint inhibitors or other immunotherapy, since these tumors are more likely to be recognized by T cells with higher activity against tumors after immunotherapy.

[0436] The utility score described in section X above can also be adapted to select patients for treatment with adoptive cell therapy (e.g., expanded TIL, CAR-T, or engineered TCR) by using f(class I neoantigen burden, class II neoantigen burden, tumor mutation burden, class I non-neoantigen tumor antigen burden, class II non-neoantigen tumor antigen burden) (wherein class I and class II neoantigens and non-neoantigens are only considered as present or predicted to be present in the adoptive immunotherapy). For example, in the case of engineering a TCR therapy against a single neoantigen epitope, f can be reduced to the likelihood of presentation of that single epitope.

[0437] XII. Example 8: Experimental Results Showing Exemplary Patient Selection Performance The validity of the patient selection described in Section X is verified by performing patient selection on a set of simulated patients, each associated with a test set of simulated neoantigen candidates, for which a subset of the simulated neoantigens are known to be represented in the mass spectrometry data. Specifically, each simulated neoantigen candidate in the test set is associated with a label indicating whether the neoantigen is represented in the mass spectrometry data set of multi-allelic JY cell line HLA-A*02:01 and HLA-B*07:02 from the Bassani-Sternberg dataset (dataset "D1") (data can be found at www.ebi.ac.uk / pride / archive / projects / PXD0000394). As described in more detail below in conjunction with Figure 13A, a number of neoantigen candidates for the simulated patients are sampled from the human proteome based on known frequency distributions of mutation burden in non-small cell lung cancer (NSCLC) patients.

[0438] Per-allele presentation models for the same HLA alleles are trained using a training set that is a subset of the mass spectrometry data for the single alleles HLA-A*02:01 and HLA-B*07:02 from the IEDB dataset (dataset "D2") (data can be found at http: / / www.iedb.org / doc / mhc_ligand_full.zip). In particular, the presentation model for each allele is trained using the network dependency function g h (·) and g w The presentation model for the HLA-A*02:01 allele generates the presentation likelihood that a particular peptide will be presented on the HLA-A*02:01 allele, given the peptide sequence as the allele interaction variable and the N- and C-terminal flanking sequences as the allele non-interaction variables. The presentation model for the HLA-B*07:02 allele generates the presentation likelihood that a particular peptide will be presented on the HLA-B*07:02 allele, given the peptide sequence as the allele interaction variable and the N- and C-terminal flanking sequences as the allele non-interaction variables.

[0439] As disclosed in the following examples with reference to Figures 13A-13G, different therapeutic subsets for patients are identified based on the predictions by applying different models, such as a presentation model trained on peptide binding prediction and a prior art model, to a test set of neoantigen candidates for each simulated patient. Patients who meet the inclusion criteria for vaccine treatment are selected and associated with a personalized vaccine that contains epitopes in the therapeutic subset of patients. The size of the therapeutic subsets varies according to different vaccine doses. No overlap is introduced between the training set used to train the presentation model and the test set of simulated neoantigen candidates.

[0440] In the following example, the proportion of selected patients with at least a certain number of presented neoantigens among the epitopes contained in the vaccine is analyzed. This statistic indicates the effectiveness of the simulated vaccine in delivering potential neoantigens to induce an immune response in patients. In particular, the simulated neoantigens in a test set are presented if the neoantigen is presented in the mass spectrometry dataset D2. A high proportion of patients with presented neoantigens indicates the likelihood of successful treatment with the neoantigen vaccine by inducing an immune response.

[0441] XII.A. Example 8A: Frequency Distribution of Tumor Mutation Burden in NSCLC Cancer Patients FIG. 13A shows the sample frequency distribution of mutation burden in NSCLC patients. Mutation burden and mutations in different tumor types, including NSCLC, can be found, for example, in the Cancer Genome Atlas (TCGA) (https: / / cancergenome.nih.gov). The X-axis represents the number of nonsynonymous mutations for each patient, and the Y-axis represents the proportion of sample patients with a particular number of nonsynonymous mutations. The sample frequency distribution in FIG. 13A shows a range of 3 to 1786 mutations, with 30% of patients having fewer than 100 mutations. Although not shown in FIG. 13A, studies have shown that mutation burden is higher in smokers compared to nonsmokers, and that mutation burden can be a strong indicator of neoantigen burden in patients.

[0442] As introduced at the beginning of Section XI above, each of the simulated patient populations is associated with a test set of neoantigen candidates. Each patient's test set is determined by selecting the mutation load m from the frequency distribution shown in Figure 13A for each patient. iA test set of candidate neoantigen sequences is generated for patient i by identifying each (8, 9, 10, 11)-mer peptide sequence across the mutations in the 21-mer. Each candidate neoantigen is associated with a label indicating whether the candidate neoantigen sequence is present in the mass spectrometry D1 dataset. For example, candidate neoantigen sequences present in dataset D1 can be associated with the label "1" and sequences not present in dataset D1 can be associated with the label "0". As described in more detail below, Figures 13B-13G show experimental results of patient selection based on the presented neoantigens of patients in the test set.

[0443] XII.B. Example 8B: Proportion of Selected Patients with Neoantigen Presenting Based on Inclusion Criteria for Tumor Mutational Burden Figure 13B shows the number of neoantigens presented in the simulated vaccine for patients selected based on the inclusion criteria of whether the patient met the minimum tumor mutational burden. The proportion of selected patients with at least a certain number of presented neoantigens in the corresponding study is identified.

[0444] In FIG. 13B, the x-axis shows the proportion of patients excluded from vaccine treatment based on tumor mutation burden, as indicated by the label "Minimum Number of Mutations." For example, the data points at "Minimum Number of Mutations" 200 indicate that the patient selection module 324 selected only a subset of simulated patients with a tumor mutation burden of at least 200 mutations. As another example, the data points at "Minimum Number of Mutations" 300 indicate that the patient selection module 324 selected a lower proportion of simulated patients with at least 300 mutations. The y-axis shows the proportion of selected patients associated with at least a certain number of presented neoantigens in the test set that does not have vaccine volume v. In particular, the top plot shows the proportion of selected patients presenting at least one neoantigen, the middle plot shows the proportion of selected patients presenting at least two antigens, and the bottom plot shows the proportion of selected patients presenting at least three antigens.

[0445] As shown in Figure 13B, the proportion of selected patients with presented neoantigens significantly increased with increasing tumor mutation burden, indicating that tumor mutation burden as an inclusion criterion can be effective in selecting patients in whom neoantigen vaccines are more likely to induce effective immune responses.

[0446] XII.C. Example 8C: Comparison of neoantigen presentation in vaccines identified by presentation models and prior art models Figure 13C compares the number of presented neoantigens in simulated vaccines between selected patients associated with a vaccine containing a therapeutic subset identified based on the presented model and selected patients associated with a vaccine containing a therapeutic subset identified by a prior art model. The plot on the left assumes a limiting vaccine volume of v=10, and the plot on the right assumes a limiting vaccine volume of v=20. Patients are selected based on a utility score indicating the expected number of presented neoantigens.

[0447] In FIG. 13C, the solid lines show patients associated with a vaccine that includes therapeutic subsets identified based on presentation models for alleles HLA-A*02:01 and HLA-B*07:02. Therapeutic subsets for each patient are identified by applying each of the presentation models to sequences in the test set and identifying the v neoantigen candidates with the highest presentation likelihood. The dotted lines show patients associated with a vaccine that includes therapeutic subsets identified based on the prior art model NETMHCpan for single allele HLA-A*02:01. Implementation details for NETMHCpan are provided at http: / / www.cbs.dtu.dk / services / NetMHCpan. Therapeutic subsets for each patient are identified by applying the NETMHCpan model to sequences in the test set and identifying the v neoantigen candidates with the highest estimated binding affinity. The x-axis of both graphs shows the proportion of patients excluded from vaccine treatment based on an expected utility score, which indicates the expected number of presented neoantigens in the therapeutic subsets identified based on the presentation models. The expected utility score is determined as described in Section X in relation to equation (25). The y-axis represents the proportion of selected patients presenting at least a particular number of neoantigens (1, 2, or 3 neoantigens) contained in the vaccine.

[0448] As shown in Figure 13C, a significantly higher proportion of patients associated with a vaccine that includes a therapeutic subset based on the presentation model receive a vaccine that includes the presented neoantigen than patients associated with a vaccine that includes a therapeutic subset based on the prior art model. For example, as shown in the graph on the right, 80% of selected patients associated with a vaccine based on the presentation model receive at least one presented neoantigen in the vaccine, compared to only 40% of selected patients associated with a vaccine based on the prior art model. These results indicate that the presentation model described herein is effective in selecting neoantigen candidates for a vaccine that are likely to induce an immune response to treat tumors.

[0449] XII.D. Example 8D: Effect of HLA coverage on neoantigen presentation of vaccines identified by presentation model Figure 13D compares the number of presented neoantigens in simulated vaccines between selected patients associated with a vaccine containing a therapeutic subset identified based on a single allele-per-presentation model for HLA-A*02:01 and selected patients associated with a vaccine containing a therapeutic subset identified based on both HLA-A*02:01 and HLA-B*07:02 allele-per-presentation models. Vaccine volume is set to v=20 epitopes. For each experiment, patients are selected based on expected utility scores determined based on different therapeutic subsets.

[0450] In FIG. 13D, the solid line shows patients associated with a vaccine that includes a therapeutic subset based on both presentation models for HLA alleles HLA-A*02:01 and HLA-B*07:02. The therapeutic subset for each patient is identified by applying each of the presentation models to the sequences in the test set and identifying the v neoantigen candidates with the highest presentation likelihood. The dotted line shows patients associated with a vaccine that includes a therapeutic subset based on a single presentation model for HLA allele HLA-A*02:01. The therapeutic subset for each patient is identified by applying a presentation model for only a single HLA allele to the sequences in the test set and identifying the v neoantigen candidates with the highest presentation likelihood. In the solid line plot, the x-axis shows the proportion of patients excluded from vaccine treatment based on expected utility scores for the therapeutic subsets identified by both presentation models. In the dotted line plot, the x-axis shows the proportion of patients excluded from vaccine treatment based on expected utility scores for the therapeutic subsets identified by a single presentation model. The y-axis indicates the proportion of selected patients presenting at least a particular number of neoantigens (1, 2, or 3 neoantigens).

[0451] As shown in Figure 13D, patients associated with vaccines containing therapeutic subsets defined by presentation models for both HLA alleles presented neoantigens at a significantly higher rate than patients associated with vaccines containing therapeutic subsets defined by a single presentation model. These results demonstrate the importance of establishing presentation models with high HLA coverage.

[0452] XII.E. Example 8E: Comparison of Neoantigen Presentation in Patients Selected by Tumor Mutational Burden and Expected Number of Presented Neoantigens Figure 13E compares the number of neoantigens presented in simulated vaccines between patients selected based on tumor mutation burden and patients selected by expected utility score, which is determined based on the therapeutic subset identified by the presentation model with a size of v=20 epitopes.

[0453] In FIG. 13E, the solid lines indicate patients selected based on expected utility scores associated with a vaccine that includes the therapeutic subset identified by the proposed model. The therapeutic subset for each patient is identified by applying each of the proposed models to the sequences in the test set and identifying the v=20 neoantigen candidates with the highest likelihood of presentation. The therapeutic utility score is determined based on the likelihood of presentation of the therapeutic subset identified in section X based on equation (25). The dotted lines indicate patients selected based on tumor mutation burden associated with a vaccine that includes the therapeutic subset identified by the proposed model. The x-axis indicates the proportion of patients excluded from vaccine treatment based on expected utility scores in the solid plot and the proportion of patients excluded based on tumor mutation burden in the dotted plot. The y-axis indicates the proportion of selected patients who will receive a vaccine that includes at least a certain number of presented neoantigens (1, 2, or 3 neoantigens).

[0454] As shown in Figure 13E, patients selected based on expected utility score are administered a vaccine containing a higher proportion of presented neoantigens than patients selected based on tumor mutation load. However, patients selected based on tumor mutation load are administered a vaccine containing a higher proportion of presented neoantigens than patients not selected. Thus, although tumor mutation load is an effective patient selection criterion for effective neoantigen vaccine therapy, expected utility score is more effective.

[0455] XIII. Exemplary Computer Figure 14 illustrates an exemplary computer 1400 for implementing the entities shown in Figures 1 and 3. 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. A memory 1406 and a graphics adapter 1412 are coupled to the memory controller hub 1420, and a display 1418 is coupled to the graphics adapter 1412. A storage device 1408, an input device 1414, and a network adapter 1416 are coupled to the I / O controller hub 1422. Other embodiments of the computer 1400 have different architectures.

[0456] The storage device 1408 is a non-transitory computer-readable storage medium, such as a hard drive, a compact disc read-only memory (CD-ROM), a DVD, or a solid-state memory device. The memory 1406 holds instructions and data used by the processor 1402. The input interface 1414 is a touch screen interface, a mouse, a trackball, or other type of pointing device, a keyboard, or some combination thereof, used to input data into the computer 1400. In some embodiments, the computer 1400 may be configured to receive input (e.g., commands) from the input interface 1414 via gestures from a user. The graphics adapter 1412 displays images and other information on the display 1418. The network adapter 1416 couples the computer 1400 to one or more computer networks.

[0457] The computer 1400 is adapted to execute computer program modules for providing the functionality described herein. As used herein, the term "module" refers to computer program logic used to provide a particular functionality. Thus, a module may be implemented in hardware, firmware, and / or software. In one embodiment, the program modules are stored in the storage device 1408, loaded into the memory 1406, and executed by the processor 1402.

[0458] 1 can vary depending on the implementation and processing power required by the entity. For example, the presentation specification system 160 can run on a single computer 1400 or on multiple computers 1400 that communicate with each other over 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.

[0459] References TIFF0007651623000080.tif223158TIFF0007651623000081.tif238159TIFF0007651623 000082.tif238159TIFF0007651623000083.tif233160TIFF0007651623000084.tif52153

[0460] Sequence information SEQUENCE LISTING <110> GRITSTONE BIO, INC. <120> NEOANTIGEN IDENTIFICATION, MANUFACTURE, AND USE <150> US 62 / 517,786 <151> 2017-06-09 <160> twenty two <170> PatentIn version 3.5 <210> 1 <211> 10 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <400> 1 Tyr Val Tyr Val Ala Asp Val Ala Ala Lys 1 5 10 <210> 2 <211> 17 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <400> 2 Tyr Glu Met Phe Asn Asp Lys Ser Gln Arg Ala Pro Asp Asp Lys Met 1 5 10 15 Phe <210> 3 <211> 9 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <400> 3 Tyr Glu Met Phe Asn Asp Lys Ser Phe 1 5 <210> 4 <211> 11 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (3)..(3) <223> Pyrrolysine <220> <221> MOD_RES <222> (11)..(11) <223> Ile or Leu <400> 4 His Arg Xaa Glu Ile Phe Ser His Asp Phe Xaa 1 5 10 <210> 5 <211> 10 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (2)..(2) <223> Ile or Leu <220> <221> MOD_RES <222> (5)..(5) <223> Ile or Leu <220> <221> MOD_RES <222> (7)..(7) <223> Pyrrolysine <400> 5 Phe Xaa Ile Glu Xaa Phe Xaa Glu Ser Ser 1 5 10 <210> 6 <211> 10 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (4)..(4) <223> Pyrrolysine <400> 6 Asn Glu Ile Xaa Arg Glu Ile Arg Glu Ile 1 5 10 <210> 7 <211> 27 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (1)..(1) <223> Ile or Leu <220> <221> MOD_RES <222> (11)..(11) <223> Ile or Leu <220> <221> MOD_RES <222> (15)..(15) <223> Selenocysteine <220> <221> MOD_RES <222> (21)..(21) <223> Ile or Leu <220> <221> MOD_RES <222> (27)..(27) <223> Ile or Leu <400> 7 Xaa Phe Lys Ser Ile Phe Glu Met Met Ser Xaa Asp Ser Xaa Island 1 5 10 15 Phe Leu Lys Ser Xaa Phe Ile Glu Ile Phe Xaa 20 25 <210> 8 <211> 13 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (11)..(11) <223> Pyrrolysine <400> 8 Lys Asn Phe Leu Glu Asn Phe Ile Glu Ser Xaa Phe Ile 1 5 10 <210> 9 <211> 15 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (2)…(2) <223> Pyrrolysine <220> <221> MOD_RES <222> (14)..(14) <223> Ile or Leu <400> 9 Phe Xaa Glu Ile Phe Asn Asp Lys Ser Leu Asp Lys Phe Xaa Ile 1 5 10 15 <210> 10 <211> 16 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (5)..(5) <223> Pyrrolysine <220> <221> MOD_RES <222> (16)..(16) <223> Ile or Leu <400> 10 Gln Cys Glu Ile Xaa Trp Ala Arg Glu Phe Leu Lys Glu Ile Gly Xaa 1 5 10 15 <210> 11 <211> 8 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (4)..(4) <223> Selenocysteine <400> 11 Phe Ile Glu Xaa His Phe Trp Ile 1 5 <210> 12 <211> 12 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (7)..(7) <223> Ile or Leu <220> <221> MOD_RES <222> (10)..(10) <223> Selenocysteine <220> <221> MOD_RES <222> (11)..(11) <223> Ile or Leu <400> 12 Phe Glu Trp Arg His Arg Xaa Thr Arg Xaa Xaa Arg 1 5 10 <210> 13 <211> 9 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (4)..(4) <223> Ile or Leu <220> <221> MOD_RES <222> (5)..(5) <223> Pyrrolysine <220> <221> MOD_RES <222> (8)..(8) <223> Ile or Leu <400> 13 Gln Ile Glu Xaa Xaa Glu Ile Xaa Glu 1 5 <210> 14 <211> 9 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (5)..(5) <223> Pyrrolysine <400> 14 Gln Cys Glu Ile Xaa Trp Ala Arg Glu 1 5 <210> 15 <211> 14 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (2)..(2) <223> Ile or Leu <220> <221> MOD_RES <222> (9)..(9) <223> Pyrrolysine <220> <221> MOD_RES <222> (11)..(11) <223> Ile or Leu <400> 15 Phe Xaa Glu Leu Phe Ile Ser Asx Xaa Ser Xaa Phe Ile Glu 1 5 10 <210> 16 <211> 11 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (5)..(5) <223> Pyrrolysine <220> <221> MOD_RES <222> (9)..(9) <223> Ile or Leu <400> 16 Ile Glu Phe Arg Xaa Glu Ile Phe Xaa Glu Phe 1 5 10 <210> 17 <211> 9 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (5)..(5) <223> Pyrrolysine <220> <221> MOD_RES <222> (9)..(9) <223> Ile or Leu <400> 17 Ile Glu Phe Arg Xaa Glu Ile Phe Xaa 1 5 <210> 18 <211> 9 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (4)..(4) <223> Pyrrolysine <220> <221> MOD_RES <222> (8)..(8) <223> Ile or Leu <400> 18 Glu Phe Arg Xaa Glu Ile Phe Xaa Glu 1 5 <210> 19 <211> 9 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (3)..(3) <223> Pyrrolysine <220> <221> MOD_RES <222> (7)..(7) <223> Ile or Leu <400> 19 Phe Arg Xaa Glu Ile Phe Xaa Glu Phe 1 5 <210> 20 <211> 18 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <400> 20 Tyr Glu Met Phe Asn Asp Lys Ser Phe Gln Arg Ala Pro Asp Asp Lys 1 5 10 15 Met Phe <210> 21 <211> 9 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (6)..(6) <223> Selenocysteine <220> <221> MOD_RES <222> (7)..(8) <223> Pyrrolysine <400> 21 Phe Glu Gly Arg Lys Xaa Xaa Xaa Ile 1 5 <210> 22 <211> 14 <212> PRT <213> Artificial Sequence <220> <223> Description of Artificial Sequence: Synthetic peptide <220> <221> MOD_RES <222> (2)..(2) <223> Ile or Leu <220> <221> MOD_RES <222> (5)..(5) <223> Pyrrolysine <220> <221> MOD_RES <222> (7)..(7) <223> Ile or Leu <220> <221> MOD_RES <222> (8)..(8) <223> Pyrrolysine <220> <221> MOD_RES <222> (10)..(10) <223> Ile or Leu <220> <221> MOD_RES <222> (14)..(14) <223> Pyrrolysine <400> 22 Pro Half Phe Island Half Glu Half Island Half Gly Glu Island 1 5 10

Claims

1. 1. A method for identifying a subset of patients suitable for treatment, comprising: obtaining, for each patient, at least one of exome, transcriptome, or whole genome tumor nucleotide sequencing data from the patient's tumor cells and normal cells, wherein the tumor nucleotide sequencing data is used to obtain a peptide sequence for each of a set of neoantigens identified by comparing the nucleotide sequencing data from the tumor cells to the nucleotide sequencing data from the normal cells, wherein the peptide sequence of each neoantigen for the patient includes at least one alteration that makes it different from a corresponding wild-type parent peptide sequence identified from the patient's normal cells; generating for each patient a set of numerical presentation likelihoods for the set of neoantigens for the patient by inputting the peptide sequence of each of the set of neoantigens into a neural network presentation model, each presentation likelihood representing the likelihood that a corresponding neoantigen will be presented by one or more MHC alleles on the surface of tumor cells of the patient, wherein the neural network presentation model: a label obtained by mass spectrometry to measure the presence of a peptide bound to at least one MHC allele identified as being present in at least one of the plurality of samples; a training peptide sequence; at least one MHC allele associated with said training peptide sequence; A plurality of parameters determined based on at least the training data set, the parameters including: A function that represents the relationship between the peptide sequence and the likelihood of presentation based on the plurality of parameters. said generating comprising: identifying for each patient one or more neoantigens from the set of neoantigens of the patient, wherein identifying the one or more neoantigens for the patient comprises selecting a subset of neoantigens in the set of neoantigens for the patient, the subset of neoantigens being neoantigens having a highest presentation likelihood among the set of presentation likelihoods for the patient; determining for each patient a utility score indicative of the estimated number of neoantigens presented on the surface of tumor cells of said patient as determined by the corresponding presentation likelihood for said one or more neoantigens for said patient; selecting a subset of patients suitable for treatment, wherein each patient in said subset of patients is associated with a utility score that meets predetermined inclusion criteria; The method comprising:

2. 2. The method of claim 1, further comprising identifying, for each patient in the selected subset of patients, one or more T cells or T cell receptors that are antigen-specific for at least one of the one or more neoantigens identified for the patient.

3. 2. The method of claim 1, wherein identifying one or more neoantigens for the patient comprises selecting the entire set of identified neoantigens for the patient.

4. 2. The method of claim 1, wherein selecting a subset of patients suitable for treatment comprises selecting a subset of patients having a tumor mutational burden (TMB) higher than a minimum threshold, the TMB of a patient indicating the number of neoantigens in a set of neoantigens associated with that patient.

5. Selecting the subset of patients suitable for treatment is key. Selecting a subset of patients having a utility score above a minimum threshold The method of claim 1 , comprising:

6. 2. The method of claim 1, wherein the utility score is the sum of the presentation likelihoods for each neoantigen within the identified subset of neoantigens for the patient.

7. The method of claim 1 , wherein the utility score is the probability that the number of presented neoantigens among the one or more identified neoantigens for the patient is above a minimum threshold.

8. The training data set is (a) data relating to measurements of peptide-MHC binding affinity for at least one of the isolated peptides; and (b) data relating to measurements of peptide-MHC binding stability for at least one of the isolated peptides; The method of claim 7 , further comprising at least one of:

9. The set of numerical likelihoods is (a) a C-terminal sequence adjacent to the peptide sequence of each neoantigen; and (b) the N-terminal sequence adjacent to the peptide sequence of each neoantigen; The method of claim 1 , further characterized by characteristics including at least one of:

10. 2. The method of claim 1, wherein the set of presentation likelihoods is further identified by at least the expression level of the one or more MHC alleles of the patient as measured by RNA-seq or mass spectrometry.

11. The set of presentation likelihoods is (a) the predicted affinity between neoantigens within said set of neoantigens and said one or more MHC alleles; and (b) the predicted stability of neoantigens and the one or more MHC alleles within said set of neoantigens. The method of claim 1 , further characterized by characteristics including at least one of:

12. inputting said peptide sequence into said neural network representation model; applying the neural network presentation model to the peptide sequence of each neoantigen to generate, for each of the one or more MHC alleles, a dependency score indicative of whether the MHC allele presents the neoantigen based on a particular amino acid at a particular position of the peptide sequence. The method of claim 1 , comprising:

13. inputting said peptide sequence into said neural network representation model; transforming the dependency scores to generate, for each MHC allele, a corresponding per-allele likelihood that indicates the likelihood that the corresponding MHC allele presents the corresponding neoantigen; and combining said per-allele likelihoods to generate a presentation likelihood for said neoantigen. The method of claim 12, comprising:

14. 14. The method of claim 13, wherein transforming the dependency score models presentation of the neoantigen as mutually exclusive across the one or more class MHC alleles.

15. inputting said peptide sequence into said neural network representation model; Transforming the combination of dependency scores to generate a representation likelihood. and transforming the combination of dependency scores models presentation of the neoantigen as interfering between the one or more MHC alleles.

16. The method of claim 1, wherein the training peptide sequences are 6 to 30 amino acids in length.

17. The method of claim 1, wherein the training peptide sequences are 8 to 15 amino acids in length.

18. The method of claim 1 , wherein one or more of the training peptide sequences are artificially generated.

Images