Methods and systems for targeting epitopes for neoantigen-based immunotherapy

文档序号：555563 发布日期：2021-05-14 浏览：12次中文

阅读说明：本技术 用于基于新抗原的免疫治疗的靶向抗原表位的方法和系统 (Methods and systems for targeting epitopes for neoantigen-based immunotherapy ) 是由布兰登·马龙尾上广祐吉原庆子于 2019-11-20 设计创作，主要内容包括：一种对源自新抗原的抗原表位作为个性化免疫治疗的靶标进行排序的方法包括基于癌症患者的患者数据收集候选抗原表位。为每个候选抗原表位计算评分集合,针对相应一个候选抗原表位的相应一个集合中的每个评分表示该相应一个候选抗原表位引发癌症患者体内的免疫应答的可能性的独立度量。将每个评分集合中的评分合并为针对每个候选抗原表位的单一评分。该针对候选抗原表位的单一评分在各种情况下反映引发患者体内的免疫应答的总体可能性。使用该单一评分对候选抗原表位排序以进行免疫治疗。(A method of ranking epitopes derived from a neoantigen as targets for personalized immunotherapy includes collecting candidate epitopes based on patient data of cancer patients. A set of scores is calculated for each candidate epitope, each score in a respective one of the set for a respective one of the candidate epitopes representing an independent measure of the likelihood that the respective one of the candidate epitopes elicits an immune response in a cancer patient. The scores in each score set are combined into a single score for each candidate epitope. This single score for a candidate epitope reflects in each case the overall likelihood of eliciting an immune response in the patient. Candidate epitopes are ranked for immunotherapy using the single score.)

1. A method of ranking epitopes derived from a neoantigen as targets for personalized immunotherapy, the method comprising:

collecting candidate epitopes based on patient data of a cancer patient;

calculating a set of scores for each candidate epitope, each score in a respective one of the set of respective one candidate epitopes representing an independent measure of the likelihood that the respective one candidate epitope will elicit an immune response in the cancer patient;

combining the scores in each score set into a single score for each candidate epitope, the single score for the candidate epitope reflecting in each case the overall likelihood of eliciting the immune response in the patient; and

ranking the candidate epitopes using the single score for immunotherapy.

2. The method of claim 1, wherein each set of scores comprises at least a first score and a second score, wherein the first score indicates likelihood of HLA binding as determined using cancer patient-specific human leukocyte antigen HLA alleles and the second score indicates T cell responses predicted using a T cell receptor TCR repertoire identified using cancer patient-specific healthy ribonucleic acid RNA sequence data.

3. The method of claim 2, wherein each set of scores further comprises a third score based on cancer patient-specific tumor RNA sequence data.

4. The method of claim 1, further comprising:

extracting experimental validation characteristics of the epitope and domain knowledge about the epitope; and

each epitope is embedded into a vector space based on the experimental validation properties and the domain knowledge.

5. The method of claim 4, wherein the candidate epitopes are ranked based on the unique score and the embeddings.

6. A method according to claim 5, wherein the ranking is performed in the order of the greatest weighted distance in the vector space, the weighted distances being determined in each case based on the Euclidean distance in the vector space multiplied by the single score of each respective one of the candidate epitopes such that the one candidate epitope with the greatest weighted distance from the origin of the vector space is ranked first and the second is the one candidate epitope with the greatest weighted difference from the first ranked epitope.

7. The method of claim 4, wherein embedding is performed using a representation learning embedding framework that uses an affinity graph in which nodes represent the epitopes and edges connect the epitopes having a similarity measure above a predetermined threshold, wherein attributes of the nodes include at least experimentally derived characteristics and the domain knowledge, and wherein an embedding function is learned for each attribute to map the attributes to a numerical vector.

8. The method of claim 4, wherein the embedding is performed by direct embedding in which at least experimentally derived characteristics and the domain knowledge are each embedded using a numerical vector concatenated together.

9. The method of claim 4, wherein the embedding comprises a vector representation of biochemical properties of the epitope.

10. The method of claim 4, wherein the embeddings comprise a vector representation of the amino acid sequence of the antigenic epitope.

11. A computer system for ranking epitopes derived from a neoantigen as targets for personalized immunotherapy, the computer system comprising a memory and one or more processors configured individually or in combination for performing a method comprising:

collecting candidate epitopes based on patient data of a cancer patient;

ranking the candidate epitopes using the single score for immunotherapy.

12. The computer system of claim 11, wherein each set of scores comprises at least a first score and a second score, wherein the first score indicates likelihood of HLA binding as determined using cancer patient-specific human leukocyte antigen HLA alleles and the second score indicates T cell responses predicted using a T cell receptor TCR repertoire identified using cancer patient-specific healthy ribonucleic acid RNA sequence data.

13. The computer system of claim 11, further configured for performing the steps of:

extracting experimental validation characteristics of the epitope and domain knowledge about the epitope; and

embedding each epitope into a vector space based on the experimentally validated properties and the domain knowledge,

wherein the candidate epitopes are ranked based on the unique scores and the insertions.

14. The computer system of claim 13, wherein the ranking is performed in order of greatest weighted distance in the vector space, the weighted distances being determined in each case based on euclidean distances in the vector space multiplied by the respective single scores for respective ones of the candidate epitopes such that the one candidate epitope with the greatest weighted distance from the origin of the vector space is ranked first and the second ranked one with the greatest weighted difference from the first ranked epitope.

15. A non-transitory computer-readable medium having instructions thereon, which when executed by one or more processors, alone or in combination and using memory, are for performing a method comprising:

collecting candidate epitopes based on patient data of a cancer patient;

ranking the candidate epitopes using the single score for immunotherapy.

Technical Field

The present invention relates to a computerized method and system for determining the likelihood that different neoantigens will elicit an immune response in a specific patient, particularly for neoantigen-based immunotherapy.

Background

Cancer cells typically include alterations in deoxyribonucleic acid (DNA), referred to as neoantigens, that are not present in normal, healthy cells. Because neoantigens are not present in healthy cells, they are interesting targets for cancer therapy. In immunotherapy, the goal is to stimulate the patient's immune system to attack and kill cancer cells. In neoantigen-based immunotherapy, the goal is to teach the immune system to specifically target neoantigens. Since healthy cells do not contain neoantigens, such treatments offer promise for avoiding off-target or autoimmune responses.

The embodiments of the present invention treat the neoantigen as an alteration of DNA that is transcribed into messenger ribonucleic acid (mRNA) and carries incorrect information according to the principles of molecular biology. These mrnas are then converted into malformed proteins. In other words, the neoantigen results in a peptide sequence (protein) with incorrect amino acids. These proteins are then processed by one of two antigen processing pathways: an endogenous processing pathway or an exogenous processing pathway. These pathways are discussed by Alberts, B. et al in Molecular Biology of the Cell, Garland Science (2002), the contents of which are incorporated herein by reference in their entirety.

In the endogenous processing pathway, proteins are retained within the cells that synthesize the proteins. Proteins are cleaved by the proteasome into small peptide sequences of about 9 amino acids (also referred to as epitopes). Some of these epitopes are then transported to the Endoplasmic Reticulum (ER) for processing. In the ER, some of the epitopes bind to major histocompatibility complex I protein (MHC-I). The epitope-MHC-I complex is presented on the cell surface. Thus, the cell is referred to as an Antigen Presenting Cell (APC). Finally, T cells with a cluster of differentiated 8 receptor proteins (CD8+) bind to this epitope-MHC-I complex. These CD8+ T cells (also known as cytotoxic T cells or CTCs) then induce the APC to initiate apoptosis, which generally means that the CTCs tell the APC to kill itself.

In the exogenous processing pathway, the malformed protein is first endocytosed from the extracellular environment into the endosome which will become the APC. In other words, the malformed protein is "taken up" into the cell. The protein is then degraded by proteases into antigenic epitopes in a manner similar to the endogenous processing pathway. Then, the epitope is bound to a major histocompatibility complex II protein (MHC-II), and the epitope-MHC-II complex is presented on the cell surface. The length of an epitope bound to MHC-II complex tends to be about 15 amino acids and is therefore somewhat longer than an epitope bound to MHC-I. Thus, the exogenous processing pathway also produces APCs. T cells with a cluster of differentiated 4 receptor proteins (CD4+) bind to epitope-MHC-II complexes. Unlike CTCs, CD4+ T cells release cytokines or signaling substances that activate B cells or CTCs. CD4+ T cells are often referred to as helper T cells because they activate other cells rather than acting directly.

The human MHC system is also known as the Human Leukocyte Antigen (HLA) system. Each person has three types of HLA-I genes, which are called HLA-A, HLA-B and HLA-C. In addition, for each of these genes, each person has two versions (one inherited from the mother and one from the father). Particular versions of these genes are called alleles. Thus, each person has up to six different HLA-I genes. Although these genes are similar in structure, the strength with which they bind epitopes is different. Furthermore, these genes are highly polymorphic, meaning that different people have different alleles.

For the HLA-II system, the situation is more complicated. Although there are also three types of HLA-II genes (which are called HLA-DR, HLA-DP, and HLA-DQ), they are each heterodimeric complexes formed by two polymorphic genes (called d-chain and β -chain, respectively). Also, for each of these genes, each person inherits two alleles (one from the mother and one from the father). Thus, each person has (up to) twelve different HLA-II complexes in total. As with HLA-I, different people have different alleles, and thousands of different combinations have been observed.

Disclosure of Invention

In embodiments, the present invention provides a method of ranking epitopes derived from a neoantigen as targets for personalized immunotherapy. Candidate epitopes are collected based on patient data of cancer patients. A set of scores is calculated for each candidate epitope, each score in a respective one of the set for a respective one of the candidate epitopes representing an independent measure of the likelihood that the respective one of the candidate epitopes elicits an immune response in a cancer patient. The scores in each score set are combined into a single score for each candidate epitope. This single score for a candidate epitope reflects in each case the overall likelihood of eliciting an immune response in the patient. Candidate epitopes are ranked for immunotherapy using the single score.

Drawings

The invention will be described in more detail below on the basis of exemplary drawings. The invention is not limited to the exemplary embodiments. In embodiments of the invention, all features described and/or illustrated herein may be used alone or in various combinations in combination. The features and advantages of various embodiments of the present invention will become apparent upon reading the following detailed description with reference to the accompanying drawings, in which:

fig. 1 is a schematic overview of a computer system and method for ranking and selecting target epitopes for immunotherapy;

FIG. 2 is a flow chart of a method for preparing a plasmid encoding a sequence carrying a patient-specific novel peptide; and

fig. 3 is a schematic overview of designing a neoepitope according to a somatic mutation type according to an embodiment of the present invention.

Detailed Description

Embodiments of the present invention provide a method and system for ranking or prioritizing neo-antigens or epitopes derived from neo-antigens as targets for immunotherapy based on their likelihood of eliciting an immune response in a particular patient (also referred to as neoepitopes). The method and system estimate a single personalized score for each epitope from various indices based on personal data. This score is then combined with domain knowledge to create a priority for the new antigen. In addition to being able to estimate with greater accuracy which epitopes will provide the best targets for a particular patient, embodiments of the present invention advantageously take into account the diversity of neoantigens to better identify the most promising targets.

Not all neo-epitopes are equally promising therapeutic targets. As mentioned above, the pathways by which neoantigens ultimately elicit an immune response are very complex, and they may fail at any step. For example, certain neoantigens produce epitopes that do not bind to the appropriate HLA complex present in a particular human, while other neoantigens may result in DNA that has not even been converted to a protein, and thus these pathways are not active from the outset. Thus, being able to rank or prioritize neoantigen epitopes based on their likelihood of eliciting an immune response can significantly increase the likelihood that neoantigen-based immunotherapy will be effective for a particular patient. In contrast to known methods, embodiments of the invention can produce more accurate predictions by explicitly combining known experimental results from similar epitopes to determine the ranking.

In the same or other embodiments, each score set comprises at least a first score and a second score, wherein the first score indicates likelihood of HLA binding determined using a cancer patient-specific human leukocyte antigen HLA allele and the second score indicates T cell response predicted using a T cell receptor TCR repertoire identified using cancer patient-specific healthy ribonucleic acid RNA sequence data.

In the same or other embodiments, each score set further comprises a third score based on cancer patient-specific tumor RNA sequence data.

In the same or other embodiments, the method further comprises: extracting experimental verification characteristics of the epitope and domain knowledge about the epitope; and embedding each epitope into a vector space based on the experimental validation properties and domain knowledge.

In the same or other embodiments, the candidate epitopes are ranked based on the unique score and the embedding.

In the same or other embodiments, the ranking is performed in order of greatest weighted distance in vector space, and in each case the weighted distances are determined based on the euclidean distance in vector space multiplied by the respective single score for the respective one of the candidate epitopes such that the one candidate epitope with the greatest weighted distance from the origin of vector space is ranked first and the second ranked one with the greatest weighted difference from the first ranked epitope.

In the same or other embodiments, the embedding is performed using a representation learning embedding framework that uses an affinity graph in which nodes represent epitopes and edges connect epitopes having a similarity measure above a predetermined threshold, wherein attributes of the nodes include at least experimentally derived characteristics and domain knowledge, and wherein an embedding function is learned for each attribute to map the attribute to a numerical vector. Alternatively, the embedding is performed by direct embedding, in which at least experimentally derived properties and domain knowledge are each embedded using a vector of numerical values concatenated together.

In the same or other embodiments, the intercalations comprise a vector representation of the biochemical properties of the antigenic epitope.

In the same or other embodiments, the insertion comprises a vector representation of the amino acid sequence of the epitope.

In another embodiment, the invention provides a computer system for ranking epitopes derived from a neoantigen as targets for personalized immunotherapy, the computer system comprising a memory and one or more processors configured, alone or in combination, for performing a method according to any of the embodiments above.

In the same or other embodiments, the computer system of claim 11, further configured for performing the steps of: extracting experimental verification characteristics of the epitope and domain knowledge about the epitope; and embedding each epitope into a vector space based on the experimental validation properties and domain knowledge, wherein candidate epitopes are ranked based on the unique scores and the embedding.

In yet another embodiment, the invention provides a non-transitory computer-readable medium having instructions thereon, which when executed by one or more processors, alone or in combination and using memory, are operable to perform a method according to any of the embodiments described above.

In yet another embodiment, the present invention provides a method of producing a novel peptide comprising: (a) a process of performing a method of ordering antigenic epitopes according to any of the embodiments described herein; and (b) a process of synthesizing a novel peptide identified by performing the method of ranking antigenic epitopes.

In yet another embodiment, the present invention provides a novel peptide obtained by the following process: a process of performing a method of ranking epitope according to any of the embodiments described herein and a process of synthesizing a new peptide identified by performing the method of ranking epitope.

In yet another embodiment, the present invention provides a pharmaceutical composition comprising a novel peptide identified by performing the method of ranking antigenic epitopes according to any of the embodiments described herein.

In yet another embodiment, the present invention provides a pharmaceutical composition for treating cancer, wherein the pharmaceutical composition comprises a novel peptide identified by performing the method of ranking antigenic epitopes according to any of the embodiments described herein.

In yet another embodiment, the present invention provides a method of treating cancer in a subject comprising administering to the subject a novel peptide identified by performing a method of ranking antigenic epitopes according to any of the embodiments described herein.

In a further embodiment, the invention provides the use of a novel peptide identified by performing a method of ranking antigenic epitopes according to any of the embodiments described herein for the manufacture of a medicament for the treatment of cancer.

Fig. 1 is an overview diagram of a method and system 10 for determining and prioritizing epitopes or neo-epitopes according to an exemplary embodiment, and shows several publicly available components to illustrate circumstances under which embodiments of the present invention may be used. The system 10 implements a method that includes three main stages:

1. generating a candidate epitope 26, wherein the candidate epitope 26 is identified based on whole exome sequencing (WXS) data 12 for the individual patient.

2. Candidate epitopes 26 are scored, wherein the evidence component assigns a score to each candidate epitope 26 independently. In a particularly advantageous embodiment of the invention, it is important that all or at least a part of these scores are based on personalized data.

3. The candidate epitopes 26 are ranked, with the personalized scores combined with historical data and domain knowledge, encoded in embedded form, and a final ranking 50 of the epitopes is constructed.

The various system components shown in fig. 1 for performing stages 1-3, as well as the components for embedding epitopes and HLA typing, may be a single server or memory accessible computer processor, or a plurality of different servers and/or memory accessible processors, each of which performs portions of stages 1-3, embedding and/or HLA typing.

Stage 1 is performed by a candidate epitope generator component 20 for generating candidate epitopes 26 and includes a somatic variant recognizer component 22 programmed to invoke somatic variants. The somatic variant identifier component 22 identifies the somatic variants as neoantigens from the WXS data 12. The somatic variant identifier 22 compares the patient's tumor and healthy WXS data 12 to determine variants that are present in the tumor sample but not in the healthy sample, and identifies these variants as either somatic variants or neoantigens. As a specific example, the Gene component analysis kit (GATK) developed by the Border Institute (Broad Institute) provides the best practices workflow for somatic short variant discovery (SNVs + Indels) that can be used to accomplish this step, which are commercially available on-line and incorporated herein by reference in their entirety.

The candidate epitopes 26 are then extracted in two steps by the candidate extraction component 24. First, the type of each identified somatic variant is annotated based on changes in the amino acid sequence of the protein coding region. For example, somatic variants may result in amino acid differences (missense mutations) or short insertions or deletions of amino acids at specific positions in tumor sequence data compared to healthy samples. Second, all possible 9-mers ("class I epitopes") and 15-mers ("class II epitopes") including the recognized somatic variants are generated as a set of candidate epitopes 26. As an example, a Variant Effect Predictor (VEP) tool from The Ensembl group may be used in conjunction with "The Ensembl variable Effect Predictor" of McLaren, W et al, Genome Biology, 6 months and 6 days; 17(1): 122(2016), which is available online and is incorporated herein by reference. 9-mers and 15-mers were generated by a sliding window strategy. For example, in the case of a missense mutation at position 10, a 9 mer based on positions 2 to 10 is generated; then, another 9-mer based on positions 3 to 11 is generated, and so on until a 9-mer based on positions 10 to 18 is generated. That is, all possible windows of size 9 (and 15), including mutations, are used to generate candidate epitopes. Similar methods were used to generate candidate epitopes from deletions and short insertions. In the case of long insertions (more than 9 amino acids), the sliding window method can also be applied, although it may only contain amino acids from somatic variants.

In the HLA typing component 28, the HLA-I allele of the patient is determined using the WXS data 12, while the HLA-II allele of the patient is determined using tumor RNA sequencing (RNA-seq) data 16. Both determinations can be made in accordance with standard practice. For example, a method described by Szolek, a. et al in "OptiType: the OptiType tool discussed in precision HLA typing from next-generation sequencing data ", Bioinformatics 30, pp.3310-3316(2014), and the Seq2HLA tool discussed by Boegel, S. et al in" HLA typing from RNA-sequence sequencing reads, "Genome Medicine, 4(2012), the contents of each of which are incorporated herein by reference, can be used to identify HLA-I alleles.

Stage 2 is performed by a candidate epitope scoring component 30 for scoring the candidate epitopes identified from stage 1, from which evidence component respective scores for ranking are calculated. According to an exemplary embodiment, three evidence components are specifically used, specifically HLA binding component 32, T cell response component 34 and RNA-seq expression component 36, although in other embodiments other evidence components may be used. These three components are substantially identical for class I and class II epitopes, although specific differences are discussed relatedly below. Predictive T cell response component 34 is a novel component introduced in embodiments of the present invention and provides the advantages discussed herein. HLA binding component 32 calculates a score for each candidate epitope for binding to each HLA allele identified in the patient. The HLA-binding module 32 considers only HLA-I alleles when predicting the binding score for class I epitopes (9-mers) and also considers only HLA-II alleles when predicting the binding score for class II epitopes (15-mers).

The current published and publicly available new antigen discovery and ranking procedures include prediction of HLA binding. Thus, HLA-binding component 32 simply behaves as a function (e.g., a machine learning model) that takes the epitope sequence and alleles as input and outputs a predicted binding score. The score may be the probability of binding or a number proportional to the biochemical binding affinity between the epitope and the allele. Published models for HLA binding can be used for this component, such as "High-order network and kernel methods for peptide-MHC binding reduction" by Kuksa, P. et al, Bioinformatics31, 3600-3607(2015), the contents of which are incorporated herein by reference in their entirety. Since this component takes into account patient-specific HLA alleles, the output can be considered a personalized score.

The T cell response component 34 calculates a score that represents the intensity or likelihood of the patient's immune response to the candidate epitope 26. In particular, in a first step, a pool of patient-specific T Cell Receptor (TCR) profiles is identified using healthy RNA-seq data 14, which can be enriched specifically for T cells, as discussed in the references mentioned below. This patient-specific information is then used to predict the T cell response for each candidate epitope 26. In particular, the T cell response was calculated as two independent scores. The first score calculates the likelihood of TCR and epitope-HLA binding, e.g., according to the pseudo-code below. The second score calculates the likelihood that an epitope will elicit a T cell response unrelated to binding (described below in the pseudo-code).

Calculation of T cell receptor, epitope-HLA binding affinities

● for each allele in the HLA allele set of the patient

For each candidate epitope e

■ for each T Cell Receptor (TCR) in the patient's TCR repertoire

● calculation of binding affinities of TCRs, epitopes, alleles, for example, as described in "A flex locking approach for prediction of T cell receptor-peptide-MHC complexes" of Pierce, B.G. et al, Protein Science22, 35-46(2013), the contents of which are incorporated below by reference

■ selection of maximum binding affinity for e

● Linear scaling of the maximum binding affinity of all candidate epitopes of a patient so that they are in the range of [0, 1]

To calculate the likelihood that an antigenic epitope will elicit a T cell response (independent of binding), historical data from immune response experiments are used to train supervised machine learning models, such as in vivo experiments in humans against antigenic epitopes whose source is not a neoantigen (e.g. viruses, see Dhanda, s.k. et al, "differentiating HLA CD4 Immunology in Human subjects", Frontiers in Immunology 9, 1369(2018)), or based on antigenic epitopes used in transgenic mice that have been genetically engineered to have Human MHC genes instead of MHC genes normally found in mice (see, e.g., "Properties of MHC Class I Presented Peptides which are used in Human mice", plodios 9 (2013)). The first model was trained to predict MHC-I responses in CD8+ T cells, while the second model was used to predict MHC-II responses in CD4+ T cells. Once trained, these models are used to predict the likelihood that each candidate epitope will elicit a response from the corresponding type of T cell.

The contents of the Scientific Reports 7(2017), an "Association of T-cell receptor reagent and cyclic extension in the technical T-cell lymphomic using RNA-seq data" by Gong, Q.et al, are incorporated herein by reference in their entirety, which indicates that RNA-seq is an effective tool for the evaluation of TCR libraries. The contents of "Effective screening of T cells recognizing neoantigens and confinement of T-cell receptor-engineered T cells" of Kato, T, et al, Oncotarget 9, 11009-11019(2018) are incorporated herein by reference in their entirety, which indicates that TCR profiling is an important consideration when T cells are engineered to target specific novel antigens for cell therapy. Furthermore, the conventional methods discussed in "a flexible locking approach for prediction of T cell receptor-peptide-MHC complexes", Protein Science22, 35-46(2013) of Pierce, b.g., et al, the contents of which are incorporated herein by reference in their entirety, have been shown to be predictive of binding of T cell receptors and epitope-HLA complexes. However, patient-specific TCR repertoires and patient-specific HLA alleles have not been considered in combination to prioritize neoantigens according to their likelihood of eliciting an immune response. The T cell response module is personalized as it takes into account the patient-specific TCR repertoire.

RNA-seq expression component 36 calculates scores based on RNA-seq expression of transcripts containing neoantigens in tumor samples (i.e., versions of genes that become RNA; see Alberts B. et al, incorporated herein by reference above). This can be performed using standard analytical procedures. It is derived directly from the patient's RNA and is therefore clearly personalized. For example, Conesa, Ana et al, "A surfey of best tasks for RNA-seq data analysis", Genome Biology, vol.17, 13.26, doi: 10.1186/s 13059-016-. The expression is given as a single number ("transcripts per million" or TPM), with a minimum of 0 and a theoretical maximum of 100 tens of thousands. In practice, a value of, for example, 100 is generally considered to be "high". According to an embodiment of the invention, the RNA-seq expression score is calculated by TPM-capping all transcripts to 100 and then dividing by 100. Thus, in this example, all transcripts had an RNA-seq expression score of 0 to 1. A score for the source transcript is assigned to each candidate epitope. In the case where the candidate epitope may be derived from multiple overlapping transcripts, it is preferred to create one copy of the epitope for each possible source transcript.

Three possible scoring factors are described herein for exemplary embodiments only. However, a wide variety of other epitope scoring factors have been proposed in the academic literature. For example, read depth and allele frequency of neoantigens within tumor RNA-seq data 16 is another common method for ranking candidate epitopes. In the case of read depth, the score for the candidate epitope is given as the number of RNA-seq reads that contain the somatic variant that led to the generation of the candidate epitope. As with RNA-seq expression, this number is capped at 100 and scaled between 0 and 1. For allele frequencies, two scores were calculated as the frequency of somatic variants (compared to a normal reference sequence) in the full exome sequencing or RNA-seq of the tumor sample. According to other embodiments of the invention, different scoring factors other than the three exemplary epitope scoring factors may alternatively or additionally be implemented.

In stage 3, where the candidate epitope ranker component 40 ranks the candidate epitopes 26, for each candidate epitope 26, a single personalized score is calculated for the likelihood that the corresponding candidate epitope 26 elicits an immune response using the personalized score calculator component 42 by combining the scores calculated by the evidence component. A supervised machine learning approach (preferably offline) is used to learn how to combine the scores.

When epitope-specific clinical or surrogate endpoints are available, such as Cancer Antigen (CA)125 levels in ovarian cancer blood or the time span of progression-free survival, these are first converted to appropriate representations to express whether an epitope elicits an immune response. For example, it can be determined that an epitope associated with a reduced CA 125 level in a patient has elicited a positive immune response, and thus the immune response is considered a binary variable. Alternatively, a reduced amount of CA 125 may be associated with each epitope, and thus, in this case, the immune response is a continuous variable. Such epitope-specific clinical or surrogate endpoints can be stored and retrieved from the clinical and surrogate endpoint database 47.

In the case where no endpoint is available, then a proxy endpoint is designed. These endpoints may be based on other experimental data that are not clinical or surrogate endpoints. Alternatively, they may be determined by computer simulation, or may be selected manually.

In either case, any suitable, state-of-the-art supervised machine learning model can be trained to predict the selected endpoint (i.e., immune response) based on the scores from the evidence component. In particular, the clinical and surrogate (or surrogate) endpoint databases 47 are used to calculate the score for all epitopes. The supervised machine learning model is then trained to predict known endpoints in the database. Where a linear model is chosen, the result of learning will be an appropriate weight for each score to best predict the endpoint. If other model classes are selected, such as random forests or neural networks, the exact interpretation of the learned models may be less clear. However, the results in either case are machine learning models that take as input the score of the epitope and predict the selected endpoint (i.e., immune response). The same model is then used to predict the immune response to epitopes with unknown endpoints.

In embedding all epitope components 46, a "position" or embedding is calculated for each candidate epitope 26 within the vector space. The insertion may involve sequence similarity, biochemical characteristics, known experimental results, domain knowledge, and other characteristics of each epitope 26. The embed all epitopes component 46 has access to a physical memory database containing such information, such as a historical epitope experiment results database 48 and a domain knowledge database 49. Two examples, referred to as "direct embedding" and "representation learning embedding" are given here to illustrate how these features can be incorporated into embedding. These are merely illustrative examples.

For sequence similarity in direct intercalation, each epitope is represented as a unique heat-encoding vector based on its sequence. For example, only a small subset of amino acids is considered: r, K, D, E, one-hot encoding would use (1, 0,0, 0) for R, 0, 1, 0,0 for K, etc. The epitope is then represented as a concatenation of each of its amino acids, as shown in this example:

REDD：R(1，0，0，0)；E(0，0，0，1)；D(0，0，1，0)；D(0，0，1，0)：(1，0，0，0，0，0，0，1，0，0，1，0，0，0，1，0)

for biochemical characteristics in direct intercalation, intercalation in which 4-mers are embedded in a 12-dimensional space based on the charge, polarity and hydrophobicity of each amino acid in an epitope can be utilized. In this example, amino acids are assumed to have the following properties, which are also available on-line:

charged: r, K, D, E

Polar: q, N, H, S, T, Y, C, W

Hydrophobic: a, I, L, M, F, V, P, G

Thus, one can choose to intercalate each charged amino acid as (1, 0, 0), each polar amino acid as (0, 1, 0), and each hydrophobic amino acid as (0, 0, 1). These embeddings are not exclusive and other embeddings schemes may be used, such as 1 for charging, 2 for polarity, 3 for hydrophobicity.

Using the selected embedding scheme, the following examples are provided:

MSDE：M(0，0，1)；S(0，1，0)；D(1，0，0)；E(1，0，0)：(0，0，1，0，1，0，1，0，0，1，0，0)

RKAD：R(1，0，0)；K(1，0，0)；A(0，0，1)；D(1，0，0)：(1，0，0，1，0，0，0，0，1，1，0，0)

WILD：(0，1，0，0，0，1，0，0，1，1，0，0)

these insertions are not relevant for a particular patient and can be considered to represent "background knowledge" about the epitope.

In some cases, known experimental results are available for specific epitopes, and they can be used in direct intercalation. For example, the binding affinity of a particular epitope for a particular HLA-I or HLA-II allele can be known. This information is directly embedded using a numerical vector containing appropriate values. When the corresponding experimental results for a particular epitope are not known, this value is considered "missing". These missing values can then be resolved using standard machine learning techniques for processing the missing values.

With the domain knowledge in direct intercalation, in many cases additional information about a particular epitope may be known. For example, it may arise from mutations documented in a single nucleotide polymorphism database (dbSNP, available on-line) that includes the clinical significance (e.g., "benign" or "potentially pathogenic") of mutations for many diseases (having an identifier such as "RCV 000302825.1"), or the epitope may be due to a change in a known tumor-associated gene. This information is captured using suitable data representations such as text packets of text data or indicators of binary data (e.g., whether the epitope is due to a change in a known tumor-associated gene), and the like, and pre-processing.

By combining the above vectors into a single vector, the final direct insertion is found for each epitope. For example, each individual vector is concatenated to form one large vector, e.g., as described below. In doing so, the binary/classification value may be considered a normal value of 0 or 1, depending on the embodiment. According to another embodiment, discussed further below, a more complex approach may be used.

Simplified examples of the domain knowledge of known epitopes and the direct insertions that occur when the values differ from the domain knowledge are shown below. The standard machine learning method or "no change" for creating direct embedding is given in parentheses.

Epitope 1

Epipote _ sequence (not used for embedding): AGTW

Sequence _ biochemical _ properties (unchanged): [0,0,1,0,0,1,0,1,0,0,1,0]

HLA _ a 0201_ binding (unchanged): 5.3

HLA _ B2705 _ binding (unchanged): is there a

HLA _ DRB1 × 1201_ binding (unchanged): 3.2

dbSNP _ RCV000302825.1_ clinical _ signficce (unique thermal coding): is there a ([0,0,0,0])

dbSNP _ RCV000587704.1_ clinical _ signficce (unique thermal coding): benign ([1, 0,0, 0])

-Oncogene (one-hot coding): no ([1, 0])

Gene _ description (Standard Natural language preprocessing, successor frequency-inverse document frequency): component of ribosome (a large ribonucleoprotein complex responsible for protein synthesis in cells) ([0.2, 0,0, 0.1, 0.1, 0.5, 0.3, 0.6])

-direct embedding: catenate ([0, 0, 1, 0,0, 1, 0, 1, 0,0, 1, 0], [5.3,

epitope 2

Epipote _ sequence (not used for embedding): PLKK

Sequence _ biochemical _ properties (unchanged): [0,0,1,0,0,1,1,0,0,1,0,0]

HLA _ a 0201_ binding (unchanged): is there a

HLA _ B2705 _ binding (unchanged): 6.5

HLA _ DRB1 × 1201_ binding (unchanged): is there a

dbSNP _ RCV000302825.1_ clinical _ signficce (unique thermal coding): likely-benign ([0, 1, 0, 0])

dbSNP _ RCV000587704.1_ clinical _ signficce (unique thermal coding): benign ([1, 0,0, 0])

-Oncogene (one-hot coding): yes ([0, 1])

Gene _ description (Standard Natural language preprocessing, successor frequency-inverse document frequency): mediate the regulation of alternative splicing of pre-mRNA. Binding to and modulating the choice of splice sites in pre-mRNA ([0, 0, 0.3, 0.8, 0, 0.1, 0.2, 0])

-direct embedding: catenate ([0, 0, 1, 0,0, 1, 1, 0,0, 1, 0, 0], [

Epitope 3

Epipote _ sequence (not used for embedding): RMI

Sequence _ biochemical _ properties (unchanged): [1,0,0,0,0,1,0,0,1]

HLA _ a 0201_ binding (unchanged): 2.3

HLA _ B2705 _ binding (unchanged): 5.9

HLA _ DRB1 × 1201_ binding (unchanged): 6.1

dbSNP _ RCV000302825.1_ clinical _ signficce (unique thermal coding): likely-pathetic ([0, 0, 1, 0])

dbSNP _ RCV000587704.1_ clinical _ signficce (unique thermal coding): pathological ([0, 0,0, 1])

-Oncogene (one-hot coding): yes ([0, 1])

Gene _ description (Standard Natural language preprocessing, successor frequency-inverse document frequency): transcription regulators are critical for the development and suppressive function of regulatory T cells (tregs). Plays a crucial role in maintaining the homeostasis of the immune system by achieving a complete suppressive function and stability of the Treg lineage ([0.9, 0.7, 0.8, 0,0, 0,0, 0.9])

-direct embedding: catenate ([1, 0,0, 0,0, 1, 0,0, 1], [2.3, 5.9, 6.1], [0, 0, 1, 0], [0, 0,0, 1], [0, 1], [0.9, 0.7, 0.8, 0,0, 0,0, 0, 0.9])

Epitope 4

Epipote _ sequence (not used for embedding): TAG

Sequence _ biochemical _ properties (unchanged): [0,1,0,0,0,1,0,0,1]

HLA _ a 0201_ binding (unchanged): 6.1

HLA _ B2705 _ binding (unchanged): is there a

HLA _ DRB1 × 1201_ binding (unchanged): 2.1

dbSNP _ RCV000302825.1_ clinical _ signficce (unique thermal coding): is there a ([0,0,0,0])

dbSNP _ RCV000587704.1_ clinical _ signficce (unique thermal coding): is there a ([0,0,0,0])

-Oncogene (one-hot coding): is there a ([0,0])

Gene _ description (Standard Natural language preprocessing, successor frequency-inverse document frequency): is there a ([0,0,0,0,0,0,0,0])

-direct embedding: catenate ([0, l, 0,0, 0, 1, 0,0, 1], [6.1,

in these examples, the order of the clinically significant indicators in the one-hot code is: [ benign, possibly pathogenic, pathogenic ]. For the binary variable "Oncogene" ("is this somatic variant occurring in a known Oncogene: [ NO, YES ]. For "Gene _ descriptions," eight terms are retained after standard pre-processing, so the length of the vector is 8. For these variables, a vector of all zeros is used to indicate missing values. Other strategies may also be used, such as additional indicator fields (e.g., [ no, yes, deletion ] for oncogenes). The number of deletions (e.g., known binding affinities) is retained as "deletions" using a standard representation (e.g., "not a number").

As an alternative to the direct embedding method described above, a more complex representation learning embedding method may be used. In a preferred embodiment, the Embedded Propagation (EP) framework discussed in Garc, a-Duran, A. et al, "Learning Graph reproduction with Embedding Propagation," Advances in Neural Information Processing Systems 30(2017), the contents of which are incorporated herein by reference in their entirety, is used. The EP takes as input an affinity graph, where nodes represent entities and edges connect similar entities. A set of attributes is provided for each node. For some nodes, some attributes may be missing. EP learns an embedding function that maps attributes to a numerical vector. Different functions are learned for each attribute, and different types of functions may be learned for different attribute types. And learning the parameters of the function to make the numerical vectors of the adjacent nodes in the graph similar.

In this setup, each node in the EP map corresponds to an epitope, and edges connect similar epitopes. As an example of similarity, all epitopes with similarity above a certain threshold can be linked according to a sequence similarity measure (e.g., the levenstein distance). The attributes of each node may be, for example, the biochemical characteristics, experimental results, and domain knowledge described above. This representation learning method has not been used previously in this situation.

Since the intercalation is only used to calculate the distance, according to an embodiment, a kernel can also be defined on the epitope without intercalation. However, the embedding-based approach is the preferred approach because studies have shown that in many circumstances the embedding-based approach is superior to the kernel-based approach, e.g., "Learning computational Networks" by Niepert, M.et al, Proceedings of the 33rd International Conference on Machine Learning (2016), the contents of which are incorporated herein by reference in their entirety.

In the rank epitopes component 44, the candidate epitopes 26 are ranked based on their personalization scores and embeddings. The rank epitope component 44 calls candidate epitopes 26 into a final rank 50 such that the highly ranked epitopes are both likely to induce an immune response and in different parts of the intercalation space. Furthermore, the ranking 50 remains diversified in case some top-ranked epitopes have to be discarded for technical reasons.

According to the examples, the maximum weighted distance between a single epitope and the set of "currently" selected epitopes is used. This is considered to be the maximum euclidean distance between a single epitope and any epitope in the "current" set multiplied by the score of the single epitope. This means that an epitope with a low score will always have a low "maximum weighted distance" regardless of its position of intercalation. On the other hand, epitopes with high scores but similar insertion positions will have relatively low distances. Thus, the process also encourages diversity by selecting large scoring epitopes that are distant from each other in the intercalation space.

Candidate epitopes 26 may be ranked using the following algorithm:

1. the epitope with the greatest weighted distance from the origin (0, 0.,) position is selected as the first-ranked epitope.

2. And selecting the epitope with the largest weighted distance from the epitope ranked at the first position as the epitope ranked at the second position.

3. The epitope having the largest weighted distance from both the epitope ranked first and the epitope ranked second is selected as the epitope ranked third.

This process continues until all epitopes are ordered.

Embodiments of the present invention provide the following improvements:

1) a single score reflecting the likelihood of eliciting an immune response is calculated for each epitope. The single score is a combined score of a collection of individual, independent scores reflecting individualized data, including a patient-specific T cell receptor repertoire and HLA alleles.

2) Based on their experimental validation properties and domain knowledge, representation learning is used to embed epitopes into vector space.

3) Antigen epitopes were ordered by combining scores, embedding positions and sequence diversity. In particular, this means that an order is created rather than selecting a subset of epitopes. Furthermore, the ranking is based on both immune response likelihood and diversity, not just on the likelihood of a response.

According to an embodiment of the invention, a method of prioritizing antigenic epitopes based on the likelihood that they are derived from a neoantigen to elicit an immune response, comprises the steps of:

1) extracting epitope characteristics verified by experiments;

2) extracting domain knowledge about epitopes;

3) embedding all antigen epitopes into a vector space based on experimental verification characteristics;

4) collecting a candidate epitope set;

5) calculating for each epitope a set of scores each giving an independent measure of the likelihood that the epitope will elicit an immune response;

6) pooling the score set for each epitope into a single score that reflects the overall likelihood that the epitope will elicit an immune response; and

7) epitopes are ranked based on their immune response potential, intercalation and sequence diversity.

Steps 1) -3) may be performed off-line and steps 4) -7) may be performed on-line.

Modular scoring methods according to embodiments of the present invention advantageously allow for natural incorporation of antigenic epitope immunogenicity. To date, all described neoantigen selection protocols have only regarded HLA binding as the "end point" for selection of neoantigens. For example, "MuPeXI: the prediction of neo-epitopes from the molecule sequencing data ", Cancer Immunology, Immunology 66, 1123-1130(2017) does not include any description of T cell responses. Us patent No. 10,055,540 explicitly states that its method predicts the likelihood of presentation on HLA alleles. Rubinsteyn, a. et al, "Vaxrank: vaccine Peptide Selection ", J.computational Pipeline for the PGV-001Neoantigen Vaccine Trial, Frontiers in Immunology 8(2018), describes the final ordering of candidate Vaccine peptides based on predicted MHC binding and expression. "pVAC-Seq" of Hundal, J.: a Genome-bound in silico aproach to identifying molecular neo-antibodies ", Genome Medicine 8(2016) is intended to" predict high affinity peptides binding to HLA class I molecules ". However, their prediction does not include immunogenicity. U.S. published patent application No. 2016/0069895 describes a peptide screening platform. U.S. published patent application No. 2017/0224799 describes a method of predicting the immunogenicity of peptides using the conformational stability of epitopes in MHC protein binding grooves. Therefore, they only consider the stability of epitope and MHC binding. Therefore, they also do not include immunogenicity in their prediction. The contents of each of the foregoing publications are hereby incorporated by reference in their entirety.

Further, the system according to the embodiment of the present invention allows arbitrary functional comments to be included. For example, neoantigens derived from DNA changes in regions known to be associated with cancer may be prioritized. Thus, the system may naturally prioritize the "drive mutations". Also, the prior published methods do not contain this type of domain knowledge.

The intercalation model advantageously allows for the direct incorporation of experimental evidence for an epitope of an antigen (when it is known). In contrast, existing methods only indirectly include this evidence through a trained machine learning model (see, NetMHCPan by Bjerregaard et al and Rubinsteyn et al; NetMHC by Hundal et al; and the custom neural network model in U.S. Pat. No. 10,055,540).

Moreover, the method according to embodiments of the present invention is advantageously equally applicable to the intrinsic pathway and the extrinsic pathway. Previous approaches have focused only on the intrinsic pathway. Although U.S. patent No. 10,055,540 mentions the use of HLA-II binding epitopes in its training set, it is predicted to be directed only to HLA-I binding (i.e., the intrinsic pathway).

For the proposed prioritization scheme, there should be a set of candidate epitopes available. "signals of biological processes in human cancer" by Alexandrov et al, Nature 2013, 500, 415-421(2013) indicate that mutations associated with certain forms of cancer (e.g., hairy cell astrocytomas and acute lymphocytic leukemia) are rare, in which case there may not be enough candidate epitopes to prioritize at present.

The determination of the weight of an individual score in a combined score depends on the outcome selected, and different clinical trials will typically consider different outcomes. For example, CA 125 levels in the blood are a common quantitative endpoint for some types of cancer, particularly ovarian cancer. Other tests may take into account progression free survival time or other results. Since these are different types of numbers on different scales, they may affect the importance of the different scores whose weights are taken into account.

Embodiments of the invention may be used to select epitopes for use in vaccine-based immunotherapy and/or to provide "neo-antigen discovery as a service".

Figure 2 illustrates a method for predicting, ranking and selecting target neoepitopes to reach a metastatic sequence 120 for a particular patient 100. Several steps S1-S5 are performed, starting with the collection of patient samples (tumor tissue and normal tissue) in step S1 until the transfer sequence 120 is designed for generating a plasmid carrying the sequence encoding of the patient-specific neo-peptide (or "patient-specific plasmid") after step S5.

In step S1, a patient tumor sample is obtained immediately after surgery. A portion of the sample was removed for Formalin Fixation and Paraffin Embedding (FFPE), and another piece of tissue was immediately frozen. Blood samples were collected as whole blood in PAXgene tubes or as Ficoll gradient isolated Peripheral Blood Mononuclear Cells (PBMC). The samples were stored at-80 ℃ (or in nitrogen vapor) until sequence analysis was performed.

In step S2, Whole Exome Sequencing (WES) is preferably performed in a qualifying laboratory. Genomic DNA from tumor and peripheral samples was sheared, end-repaired, ligated with barcoded ILLUMINA sequencing adaptors, amplified, and size selected. Preferably frozen tumor tissue is used. When not available, FFPE tumor samples were used for WES. In this example, a Nextera Rapid Capture Exome v1.2 decoy pool from ILLUMINA or equivalent was used to target exomes. This capture method covers approximately 37.7Mb of exon regions, which include all coding regions of the RefSeqGene database at NCBI (see O' Leary et al, "Reference sequence (RefSeq) database at NCBI: current status, taxonic expansion, and functional organization," Nucleic Acids Res 44: D733-45 (2016)). The resulting library was then qPCR quantified, pooled and sequenced using an ILLUMINA sequencer to paired-end reads of at least 2x75bp to obtain a fastq file.

For constructing RNA sequencing libraries, preferably RNA is extracted from frozen samples, or FFPE samples when frozen material is not available. The TruSeq RNA Access Library Prep kit from ILLUMINA or equivalent was used to prepare the RNA-Seq Library. Total RNA concentration was quantified and normalized prior to library preparation. Using the TruSeq RNA Access Library Prep kit or equivalent, a chain cDNA Library is prepared, which is then hybridized with a collection of DNA oligonucleotide probes to enrich for mRNA transcript fragments of the Library. Transcriptome Capture targets 21,415 genes, accounting for 98.3% of the RefSeq Exome (same decoy pool as Rapid Capture Exome). Each sequencing run was performed with a read length of at least 2x50bp paired ends.

In step S3, a somatic mutation is identified. For each patient, candidate epitopes were identified by variant calling from tumor and normal whole exome sequencing reads. Tumor and normal whole exome sequencing reads were trimmed and filtered using trimmatic, a flexible read pruning tool for sequence data. After quality control, they were aligned to the human GRCh38 reference genome using BWA-MEM. The alignment files were processed according to GATK best practices. For generating candidates, all possible peptide sequences, including mutated amino acids, were generated based on 9-mer or 15-mer window sizes.

To determine the patient HLA genotype, whole exome sequencing reads were trimmed using razer s3 and aligned to the IMGT/HLA database. HLA class I alleles were identified using OptiType. After filtering the low quality reads using Flexbar, the tumor RNA-seq reads were trimmed and the ribosomal RNA reads filtered out using bowtie 2. After quality control, they were aligned to the IMGT/HLA database using bowtie. HLA class II alleles were recognized using seq2 HLA.

Further in step S3, the candidate epitopes are subjected to immunogenicity scoring and ranking. The relevance of candidate epitopes to vaccine design is scored and ranked using a series of biological and biochemical factors that drive their relevance to tumor-specific immune targets. These factors include binding affinity to the patient's HLA, similarity to epitopes known to be immunogenic, expression level at the transcriptional level, frequency of mutations, homology to normal human sequences, homology to viral proteins, and the likelihood that a given sequence will be processed by intracellular machinery for presentation. As in the above embodiments, these factors are considered in the scoring through the calculation of several evidence components that define an index that reflects each of these factors. The evidence component (examples of which are described above and below) is used to derive an overall score and ranking for each candidate epitope. The evidence components for class I and class II epitopes may generally be the same, although specific differences are indicated below when relevant. The evidence component is a computer processing component specifically configured to receive its respective input (preferably from a memory or database) and output a respective score.

For HLA binding affinity, a high performance machine learning algorithm based on a higher order kernel support vector machine was trained using a binding affinity proprietary database (which was measured using a laboratory in vitro assay). Briefly, peptide binding to HLA class I molecules was measured by a stabilization assay using a TAP deficient tumor cell line. This allows for accurate measurement of binding affinity, enabling better prediction. In this example, analysis of HLA-a 02: 01 study of binding peptides and using widely available T2 cells, several cell lines suitable for analysis of other HLA class I allele molecules were also generated. C1R cells (ATCC, Manassas, VA) expressing neither HLA-A nor B molecules were transfected with different HLA-A genes of interest. The criprpr/Cas 9 system was then used to remove the transporter associated with The Antigen Processing (TAP) gene, which resulted in the appearance of a large number of "empty HLA molecules" on the cell surface. A monoclonal antibody (mAb) was developed to detect peptide-loaded HLA-a molecules, which discriminates between HLA-a allele molecules in most peptide promiscuous forms. By using a TAP deficient cell line and the mAb, binding affinity can be measured with high accuracy. To analyze HLA class II molecules, a method of measuring binding of peptides to HLA class II molecules on the cell surface of live Antigen Presenting Cells (APCs) was developed. The method is characterized first by measuring the binding of an 11-mer peptide with diamino acid extensions at the N and C termini to protect the peptide from degradation by cell-associated peptidases. This ensures the accuracy of the peptide concentration during the binding assay. In public databases, most HLA class II binding data is obtained using longer peptides, which obscures the exact sequence information of peptides in direct contact with HLA class II molecules. Second, binding of the peptide to HLA class II molecules on living cells is facilitated at ph6.0 and by the addition of parachlorophenol, a hydrogen bond exchanger that allows for efficient peptide loading, thereby accurately measuring binding affinity. This method takes advantage of the natural mechanism of antigen presentation and, unlike other methods using affinity purified HLA class II, it does not use detergents that may affect peptide binding. These models predict values proportional to binding affinity; although the range of scores varies, the scores of typical "strong binders" are in the range of [5, 7 ].

To determine similarity to epitopes with known immunogenicity, the evidence component uses a deep Convolutional Neural Network (CNN) to score the likelihood that a candidate epitope will elicit a T cell response in an in vitro immunogenicity assay. Instead of learning any insertions for each amino acid, known biochemical properties (e.g., polarity and hydrophobicity) and evolutionary features (BLOSUM62 mutation values) were used. Public CD4 or CD8 immune response data from the Immune Epitope Database (IEDB) were used to train these models. This score is always in the range of [0, 1] due to the model's prediction probability.

For RNA expression, FPKM (million fragments per kilobase) values were extracted from RNA-seq read files. The RNA-seq values were converted to the range of [0, 1] by first capping all FPKM values to 100 (i.e., setting the FPKM values above 100 to 100). Epitopes derived from transcripts estimated to have FPKM less than 1 are filtered out. To generate RNA expression scores, these values were then scaled linearly from [0, 1 ].

For DNA, RNA allele frequencies, the evidence component gives the frequency of mutations or indels in WES or RNA sequencing of tumor samples, respectively. Therefore, it is always in the range of [0, 1 ].

For RNA allele depth, the evidence component gives the number of RNA sequencing reads that include mutations or insertion deletions responsible for the antigenic epitope. The count is clipped to 100 and scaled linearly from [0, 1 ]. The epitope without any support for RNA sequencing was filtered.

For human sequence homology, the evidence component compares the epitope sequence with its closest homolog in the human proteome. In particular, the Basic Local Alignment Search Tool (BLAST) database was constructed using the human proteome (Ensemble, GRCh38, version 90). Then, a BLAST search is performed and for each hit, a normalized block substitution matrix (BLOSUM) similarity is calculated, ranging from 0 (completely different sequences) to 1 (completely identical sequences). The score from this component is taken as (1-similarity). For example, in the case where the epitope actually appears elsewhere in the human proteome, the score is 0.

With respect to homology to viral sequences, the evidence component compares the antigenic epitope sequence with its closest homolog in the viral proteome, since viral proteins are more likely to elicit an immune response. It is similar to the human homology component. BLAST databases were constructed using non-redundant viral protein sequences from RefSeq version 91. Performing a search using the same parameters as the human homology component and likewise finding the most similar match; however, in this case, the similarity is used as a score. Thus, epitopes similar to viral sequences have a higher score. The score ranged from [0, 1], where 1 represents a perfect match in the viral sequence.

For intracellular processing, the evidence component predicts the likelihood that a particular epitope will undergo intracellular processing (proteasomal cleavage, TAP binding and transport) and will be available for presentation by the corresponding HLA molecule (score in the range of [0, 1 ]). Training a gradient boosting tree to predict this; they used the same inputs for each epitope as predicted for T cell responses and the corresponding HLA molecule pseudo-sequences. The model is trained using "positives" based on public mass spectral data, while "negatives" for training and testing are also based on available public data.

The evidence component is preferably weighted. The weight or relative importance of each evidence component was determined using public ex vivo T cell response data. In particular, the above values are calculated for each epitope that has been tested in ex vivo experiments. The linear model was then trained to predict the observed T cell response. The coefficients learned in the linear model are taken as the weight of each component.

Finally, in step S3, the final ranking of the epitopes is based on three elements: a score merge component (which is a weighted merge of the evidence components described above), an epitope embedder component (a position in a multidimensional space), and an epitope order component that merges scores and positions.

By combining scores from all evidence components, a single score for each candidate epitope is calculated. An important advantage of the system according to embodiments of the present invention stems from the use of HLA binding affinity datasets. The single score is computed as a weighted sum of all the components described above and or a different combination of evidence components used in other embodiments.

The epitope intercalating components of this embodiment can be similar to those discussed in the above embodiments. The "position" within the intercalation space is calculated for each epitope. As a simple example, one can consider the insertion of a 9-mer into a 27-dimensional space (27-dimensional with 3 properties of 9 amino acids) based on the charge, polarity and hydrophobicity of each amino acid in the epitope. These insertions are not relevant for a particular patient and they can be considered to represent "background knowledge" about the epitope. For example, the EP algorithm described above may be used to learn the embedding location. EP consists of two stages: an off-line learning phase (using known experimental results) and an on-line embedding phase (in which the positions of new candidate epitopes are determined). In the learning phase, the EP takes as input a map linking epitopes based on sequence similarity and all known characteristics of these epitopes (e.g. known HLA binding affinity data, presentation in mass spectral data, and information such as the genetic ontology conditions of the gene from which the epitope is derived). Then, the EP trains the neural network to map epitopes that are close in the map and have similar properties to close positions in the embedding space. Again, this is done during the off-line learning phase and no information about candidate epitopes is used. During the online intercalation phase, the location of each candidate epitope is determined. First, for each candidate epitope, its neighbors in the training graph are determined based on sequence similarity. The trained neural network is then used to determine the location of the candidate epitope in the intercalation space.

Candidate epitopes are ranked based on their above-described patient-specific scores (from the score pool component implemented as described above) and "diversity". The aim is to order the epitopes so that highly ordered epitopes are both likely to induce an immune response and are diverse. Furthermore, the ranking is designed to remain diversified in case some of the top-ranked epitopes are unusable due to problems with synthesis etc. First, the patient-specific location of each candidate epitope is determined by multiplying the score of each candidate epitope by its location. Thus, for example, all candidate epitopes with scores close to 0 will be close together, while candidate epitopes with large scores will be far apart. The epitopes are then ranked using an iterative process. The candidate epitope with the highest score was selected as the first ranked epitope. Then, the candidate epitope farthest from the epitope ranked first is recognized and taken as the epitope ranked second. The third ranked candidate epitope is the candidate epitope that is most distant from both of the first two epitopes. For example, this process continues until all candidate epitopes in the first 30 positions are ranked.

Advantageously, candidate epitopes with high scores but similar positions will have relatively low distances; thus, the process also encourages diversity. In other words, the method will select epitopes that are distant from each other in the intercalation space with a large score.

To allow presentation of the neoepitopes identified as described above to a broad spectrum of immune responses, the neopeptides were designed in step S4 by extending the predicted 9-mer neoepitopes from the mutation site in each direction to cover the 15-mer window. The resulting novel peptides depend on the type of mutation that results in the generation of the epitope. Various scenario scenarios are shown in fig. 3. The general rule for the design of the neopeptides is then defined as an extension of up to 14 residues upstream and downstream, respectively, of the first and last mutation positions that are part of the predicted neoepitope.

In FIG. 3, M represents a mutation, M1 represents the first mutation in an epitope, Mn represents the last mutation in an epitope (1 < n.ltoreq.9), Δ represents a deletion event, SI represents a short insertion in the structure (1 < M < 9), LI represents a long insertion in the structure (> 9), and FS represents a frameshift. The detected mutations driven by the epitopes cannot exceed 9 (9-mer epitopes).

In the design of the neopeptide fusion transfer sequences, the eligibility of the sequenced neopeptide as part of the expression cassette encoding it is fused depends on various criteria including sequence homology and biochemical properties that may affect the production of the recombinant vector, such as hydrophobicity and protein characteristics associated with hydrophobicity (e.g., propensity to form transmembrane domains). In step S5, based on the above characteristics, a custom tool is used to design an optimized expression cassette. The tool detects and discards any new peptide combinations that may lead to incorrect protein fusion. Any new peptide candidate that has had a forbidden feature or induced a highly hydrophobic fusion protein will be automatically disqualified and replaced by the next candidate (if any) in the initial list. The resulting expression cassette is then embedded in the transfer sequences required for plasmid production.

Each of the novel peptides of the present invention can be synthesized using techniques known to those skilled in the art. For example, it can be artificially synthesized by a solid phase method (e.g., Fmoc method or tBoc method) or a liquid phase method. The desired peptide can also be produced by expressing a polynucleotide encoding the novel peptide of the present invention or a recombinant vector containing the polynucleotide. The novel peptides thus obtained can each be verified using techniques known to those skilled in the art. For example, it can be verified using edman degradation or mass spectrometry.

Briefly, synthesis of peptides by using a solid phase synthesis method involves first attaching the protected C-terminal amino acid of the peptide to a resin. After attachment, the resin is filtered, washed and the protecting group (e.g., t-butyloxycarbonyl) on the alpha amino group of the C-terminal amino acid is removed. Of course, the protecting group must be removed without breaking the bond between the amino acid and the resin. Then, the penultimate C-terminal protected amino acid is coupled to the resulting resin peptide. The coupling is achieved by forming an amide bond between the free carboxyl group of the second amino acid and the amino group of the first amino acid attached to the resin. This sequence of events is repeated with consecutive amino acids until all the amino acids of the peptide are attached to the resin. Finally, the protected peptide is stripped from the resin and the protecting group is removed to obtain the desired peptide. The cleavage techniques used to separate the peptide from the resin and remove the protecting groups depend on the choice of resin and protecting group and are known to those familiar with the art of peptide synthesis.

According to one embodiment, the neopeptide is obtained by a process of performing a method according to any of the above embodiments and a process of generating the neopeptide identified by performing the method.

Although other methods exist to determine the epitope to be targeted, these methods all have significant disadvantages and do not provide the improvements described above. For example, all possible epitopes can be experimentally verified, thereby avoiding the necessity of ranking. However, this is a time and cost prohibitive solution and is therefore not a viable solution. As another example, a set of epitopes can be computationally selected based on a set of hard filters. However, the filter will need to be designed manually and it is not clear how to handle the situation when many epitopes pass through all filters or none. As yet another example, the set of epitopes may be manually selected by an expert. However, there is already literature (e.g., "NetMHCpan-4.0: Improved Peptide-MHC Class I Interaction precursors Integrated excited Ligand and Peptide Binding Affinity Data" by Jurtz, V, Journal of Immunology 199, 3360-. Thus, this approach is likely to result in many false positives. Furthermore, some tumor samples may result in thousands of epitope candidates, making manual ranking or selection impractical in these cases. Epitopes can also be ranked based solely on their predicted HLA binding affinity. However, gross, A. et al, "therapeutic identification of neoantigen-specific lymphocytes in the periphytol bulk of mammalian properties", Nature Medicine 22, pp.433-438(2016) have indicated that many epitopes with high predicted HLA binding affinity fail to elicit an immune response. Thus, this approach is likely to result in many false positives.

In any of the embodiments described herein, the ranked candidate epitopes are preferably used for the treatment of a particular patient by targeting the epitopes according to their ranking in immunotherapy.

Thus, in a further embodiment, the invention provides the use of a novel peptide identified by performing a method of ranking antigenic epitopes according to any of the embodiments described herein, in one or more embodiments also provided is a pharmaceutical composition for the treatment of cancer.

The pharmaceutical composition for treating or preventing cancer according to one or more embodiments of the present invention comprises at least one novel peptide of the present invention as an active ingredient. The novel peptide of the present invention induces Cytotoxic T Lymphocytes (CTLs) by being presented on antigen presenting cells, and the induced CTLs damage cancer cells. Therefore, the active ingredient of the pharmaceutical composition of the present invention is not limited to the novel peptide of the present invention, but may be a component capable of inducing CTLs specifically by the novel peptide directly or indirectly, for example, the active ingredient may also be a polynucleotide encoding the novel peptide or a vector comprising such a polynucleotide, or mRNA encoding the novel peptide, or an antigen-presenting cell presenting a complex of the novel peptide and an HLA molecule on the surface or an exosome secreted from the antigen-presenting cell, or a combination thereof. Examples of antigen presenting cells used include macrophages and dendritic cells. However, dendritic cells having high CTL inducibility are preferably used. Any other ingredient known for use in the treatment of cancer, such as chemokines, cytokines, tumor necrosis factors, and chemotherapeutic agents, may be included in the pharmaceutical compositions of the present invention.

The pharmaceutical compositions of the present invention are believed to be useful for killing cancer cells by, for example, but not limited to, the following mechanisms of action. As such, a pharmaceutical composition for treating cancer is disclosed, wherein the pharmaceutical composition comprises a novel peptide identified by performing a method of ranking antigenic epitopes according to any of the embodiments described herein. Administration of the pharmaceutical composition of the present invention to a specific cancer patient allows the novel peptide in the pharmaceutical composition to be presented in a state where it is bound to an HLA molecule on the surface of an antigen presenting cell. When a novel peptide on such antigen presenting cells is identified, CTLs are activated, proliferated, and circulate systemically. When the neopeptide-specific CTL enters cancer tissues, it recognizes the same neopeptide derived from a specific cancer antigen, and naturally binds to HLA molecules present on the surface of cancer cells to kill the cancer cells. This effect contributes to the treatment of cancer. Thus, in yet another embodiment, the invention relates to a method of treating cancer in a subject in need thereof.

The pharmaceutical composition of the present invention can be used not only for treating cancer but also for preventing cancer. For example, administration of the pharmaceutical composition of the present invention to healthy humans induces CTLs, and the induced cytotoxic T cells remain in the body, and thus, when a specific cancer cell appears, the cancer cell can be damaged. Similarly, the composition may be administered to a human after treatment of cancer to prevent recurrence of the cancer. In both cases, the pharmaceutical composition is a vaccine composition.

In this specification, the term "cancer" is used in its broadest sense. Examples of cancer include, but are not limited to, astrocytoma, oligodendroglioma, meningioma, neurofibroma, glioblastoma, ependymoma, schwannoma, neurofibrosarcoma, neuroblastoma, pituitary tumor (e.g., pituitary adenoma), medulloblastoma, melanoma, brain tumor, prostate cancer, head and neck cancer, esophageal cancer, kidney cancer, renal cell carcinoma, pancreatic cancer, breast cancer, lung cancer, colon cancer, colorectal cancer, gastric cancer, skin cancer, ovarian cancer, bladder cancer, fibrosarcoma, squamous cell carcinoma, neuroectodermal tumors, thyroid tumor, lymphoma, leukemia, multiple myeloma, hepatocellular carcinoma, mesothelioma, and epidermoid carcinoma.

The pharmaceutical composition of the present invention may be dissolved in an aqueous solvent, configured in the form of a pharmaceutically acceptable salt, and administered to a patient. Examples of such pharmaceutically acceptable salt forms include forms buffered at physiological pH in the form of a physiologically acceptable water-soluble salt, such as a salt of sodium, potassium, magnesium or calcium. In addition to the water-soluble solvent, a water-insoluble solvent may be used; examples of such water-insoluble solvents include alcohols such as ethanol and propylene glycol.

Formulations containing the pharmaceutical compositions of this embodiment may contain agents for various purposes; examples of such agents include preservatives and buffers. Examples of preservatives include sodium bisulfite, sodium bisulfate, sodium thiosulfate, benzalkonium chloride, chlorobutanol, thimerosal, phenylmercuric acetate, phenylmercuric nitrate, methylparaben, polyvinyl alcohol, phenylethyl alcohol, ammonia, dithiothreitol, and β -mercaptoethanol. Examples of buffering agents include sodium carbonate, sodium borate, sodium phosphate, sodium acetate, and sodium bicarbonate. These agents may be present in an amount capable of maintaining the pH of the system at 2 to 9, preferably 4 to 8.

The dosage form of the pharmaceutical composition of the present invention is not particularly limited; however, when it is used in the form of a vaccine, examples of the dosage form thereof include injections (intramuscular, subcutaneous and intradermal), oral agents and nasal drops. When the pharmaceutical composition of the invention is in the form of a vaccine, it may be a mixed cocktail vaccine comprising a plurality of active ingredients. For example, such vaccines may comprise any two or more novel peptides of the invention, or a plurality of active ingredients by combination with other active ingredients.

The vaccine of the present invention may be a vaccine comprising an inert ingredient which is an ingredient other than the pharmaceutical composition, is not active per se and has the effect of further enhancing the utility of the pharmaceutical composition as a vaccine. Examples of inert ingredients include adjuvants and toxoids. Examples of adjuvants include, but are not limited to, precipitation-type adjuvants (e.g., aluminum hydroxide, aluminum phosphate, and calcium phosphate) and oily adjuvants (e.g., freund's complete adjuvant and freund's incomplete adjuvant).

When present in the form of a vaccine, the pharmaceutical composition of the present invention is preferably administered into the body orally or by injection or infusion (e.g., intradermal, subcutaneous or intramuscular administration) or by dermal administration or by inhalation through the mucosa of the nose, pharynx, etc. Its single dose can be set between a dose that can significantly induce cytotoxic T cells and a dose where a large number of non-cancerous cells suffer damage.

The pharmaceutical compositions of the present invention are not only designed for administration to the human body, but are also intended for in vitro use. More specifically, the pharmaceutical composition of the present invention can be used for the purpose of stimulating antigen-presenting cells in vitro or ex vivo to increase their CTL inducing activity. For example, in the case of using the pharmaceutical composition of the present invention for dendritic cell therapy of cancer, the composition may be previously contacted with antigen-presenting cells (e.g., dendritic cells) derived from a patient in need of cancer treatment or prevention, and then the antigen-presenting cells may be administered to the patient by returning them to the patient. The peptide contained in the pharmaceutical composition can be introduced into antigen-presenting cells, for example, by lipofection or injection. When a polynucleotide encoding a peptide of the present invention is used in such applications, the polynucleotide may be introduced into an antigen presenting cell by techniques known in the art. For example, antigen-presenting cells derived from a patient can be transformed in vitro using a polynucleotide of interest or a vector encoding the polynucleotide by lipofection, electroporation, microinjection, cell fusion, DEAE-dextran, calcium phosphate method, or the like.

The invention includes a method of treating cancer by administering a medicament according to the invention in a therapeutically effective dose. The therapeutically effective dose can be appropriately determined by those skilled in the art, for example, according to the symptoms, age, sex, body weight and sensitivity difference of patients, administration method, administration interval and formulation type.

The novel peptides of the invention are not only designed for administration to the human body, but are also intended for in vitro use. More specifically, the novel peptides of the present invention can be used for the purpose of stimulating antigen-presenting cells in vitro or ex vivo to increase their CTL-inducing activity. For example, in the case where the novel peptide of the present invention is used for dendritic cell therapy, the novel peptide may be previously contacted with antigen-presenting cells (e.g., dendritic cells) derived from a patient in need of immune induction, and then the antigen-presenting cells may be administered to the patient by returning them to the patient. The novel peptide can be introduced into antigen-presenting cells, for example, by transfection via liposomes (lipofection) or injection. When a polynucleotide encoding the novel peptide of the present invention is used in such applications, the polynucleotide may be introduced into antigen-presenting cells by techniques known in the art. For example, antigen-presenting cells derived from a patient can be transformed in vitro by lipofection, electroporation, microinjection, cell fusion, DEAE-dextran, calcium phosphate method, or the like, using a polynucleotide of interest or a vector expressing the polynucleotide.

As used herein, "immune induction" refers to induction of an immune response, for example, increasing CTL-inducing activity of antigen presenting cells, and further increasing cytotoxic activity of CTLs against cancer cells. As used herein, "CTL induction" refers to induction or proliferation of CTLs that specifically discriminate specific antigens, or differentiation of naive T cells into effector cells having the ability (cytotoxic activity) to kill target cells (e.g., cancer cells), and/or increase the cytotoxic activity of CTLs by presenting the peptides of the present invention on the surface of antigen-presenting cells in vitro or in vivo.

While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. It will be understood that changes and modifications may be effected by one of ordinary skill in the art within the scope of the appended claims. In particular, the invention covers other embodiments having any combination of features from the different embodiments described above and below. Moreover, the statements herein reciting the present disclosure refer to embodiments of the disclosure, and not necessarily all embodiments.

The terms used in the claims should be construed with the broadest reasonable interpretation consistent with the foregoing description. For example, use of the article "a" or "the" in introducing an element is not to be construed as excluding a plurality of the elements. Likewise, references to "or" should be construed as inclusive, such that references to "a or B" do not exclude "a and B," unless it is clear from the context or the foregoing description that only one of a and B is intended. Further, reference to "at least one of A, B and C" should be interpreted as one or more of a group of elements consisting of A, B and C, and should not be interpreted as requiring at least one of each listed element A, B and C, whether A, B and C are related as a category or otherwise. Furthermore, references to "A, B and/or C" or "A, B or at least one of C" should be interpreted as including any singular entity (e.g., a), any subset of the listed elements (e.g., a and B), or the entire list of elements A, B and C.

26页详细技术资料下载

上一篇：一种医用注射器针头装配设备

下一篇：提高患者生活质量、参与度和持续性的疾病因子的早期反馈

Methods and systems for targeting epitopes for neoantigen-based immunotherapy

相关技术

网友询问留言