Methods, systems and computer program products for determining the likelihood of presentation of neoantigens

Abstract

The present invention relates to computer-implemented methods, computer systems, and computer program products for determining the likelihood of presentation of neoantigens. Furthermore, the present invention relates to the use of the methods, systems, and / or computer program products for determining treatment for a subject.

Description

[Technical Field]

[0001] The present invention relates to a computer-implemented method, a computer system and a computer program product for determining the presentation likelihood of a neoantigen. [Background technology]

[0002] In addition to normal antigenic determinants, the surface of cancer cells likely presents neoantigens derived from aberrant genomic events and recognizable by T cells.

[0003] Neoantigens are newly formed antigens that have not previously been recognized by the immune system. Targeting these neoantigens has recently shown to be a very promising avenue for personalized medicine.

[0004] New technological developments have made it possible to increase the availability of mass spectrometry-derived lists of peptides that are actually bound to major histocompatibility complex (MHC) molecules on the cell surface. These lists are called "ligandomes." State-of-the-art neoantigen discovery methods start by generating a list of all potential neoantigens produced by cancer cells and rely on in silico prediction algorithms to extract the antigenic determinants most likely to be displayed on the surface of these cells, potentially eliciting an immune response.

[0005] WO 2017 106 638 describes a method for identifying one or more neoantigens from tumor cells of a subject that are likely to be presented on the tumor cell surface. Furthermore, this document discloses systems and methods for obtaining high-quality sequencing data from tumors and for identifying somatic alterations in polymorphic genomic data. Finally, WO '638 describes a specific cancer vaccine.

[0006] US 2019 0 311 781 describes methods for identifying peptides containing features associated with successful cellular processing, trafficking, and MHC presentation through the use of machine learning algorithms or statistical inference models. US 2018 0 085 447 describes methods for identifying immunogenic variant peptides with therapeutic utility as cancer vaccines. More specifically, methods for identifying T cell-activating neoantigenic determinants from all genetically modified proteins. These variant proteins contribute to neoantigenic determinants after proteolysis within antigen-presenting cells.

[0007] EP 3 256 853 describes a method for predicting T cell antigenic determinants useful for vaccination. In particular, this document relates to a method for predicting whether modifications in peptides or polypeptides, such as tumor-associated neoantigens, will be immunogenic and particularly useful for vaccination, or for predicting which of such modifications will be most immunogenic and particularly most useful for vaccination.

[0008] Several additional tools and approaches are available that address the same problem, such as NetMHCpan or MHCflurry. These approaches use methods to predict peptide binding affinity for a given HLA allele. Other approaches, such as EDGE or MARIA, also output presentation probabilities based on learning, but do not consider HLA sequences and do not encode HLA types as categorical variables.

[0009] Furthermore, the first prediction methods use the binding affinity of candidate neoantigens to MHC as an indication of the likelihood of presentation on the cell surface. However, these methods cannot model the entire cell surface presentation process, and therefore suffer from low positive predictive value. Furthermore, these methods cannot predict the likelihood of presentation of neoantigenic determinants for HLA molecules that are not included in the model training.

[0010] The present invention aims to provide a solution to at least some of the above-mentioned drawbacks as well as an improvement over the state of the art. Summary of the Invention

[0011] In a first aspect, the present invention relates to a computer-implemented method for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor of a subject according to claim 1.

[0012] In a second aspect, the present invention relates to a computer system for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor of a subject according to claim 12.

[0013] In a third aspect, the present invention relates to a computer program product for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor of a subject according to claim 13.

[0014] In a fourth aspect, the present invention relates to a use for determining the treatment of a subject according to claim 14.

[0015] The objective of the present invention is to predict the likelihood of presentation of variable length neoantigenic determinants on the surface of cancer cells given the set of HLA alleles expressed by said cells. For this purpose, a deep learning model is used.

[0016] The present invention is advantageous because it can predict the likelihood of presentation of a neoantigenic determinant for any HLA allele, even if the model has not been trained on that HLA allele.

[0017] Preferred embodiments of the present invention are discussed in claims 2 to 12 and throughout the description and examples. [Brief explanation of the drawings]

[0018] [Figure 1]Figure 1A shows precision-recall curves obtained as a result of testing a model according to the present invention on a test data set. A comparison of the performance of a model according to the present invention with prior art algorithms, the EDGE algorithm and the MHCflurry algorithm, when tested on the same test data set is shown. Figure 1B shows precision-recall curves obtained as a result of testing a model according to the present invention on a test data set. The predictive power of a model according to the present invention when tested on a new data set is shown. DETAILED DESCRIPTION OF THE INVENTION

[0019] In a first aspect, the present invention relates to a computer-implemented method for determining the likelihood of presentation of a set of neoantigens. In a second and third aspect, the present invention relates to a computer system and a computer program product. In a fourth aspect, the present invention relates to the use of any of the methods, systems, or products for determining treatment for a subject. In the following, the present invention will be described in detail and preferred embodiments will be discussed. Furthermore, the present invention will be illustrated by non-limiting examples.

[0020] Unless otherwise defined, all terms used in disclosing the present invention, including technical and scientific terms, have the meanings commonly understood by those skilled in the art to which the present invention belongs. By way of further guidance, definitions of terms used in the description are included to better understand the teachings of the present invention. The terms or definitions used herein are provided solely to aid in the understanding of the present invention.

[0021] As used herein, the following terms have the following meanings:

[0022] As used herein, the terms "indefinite article" and "definite article" refer to both singular and plural referents unless the context clearly dictates otherwise. By way of example, the term "a compartment" refers to one or more compartments.

[0023] As used herein, "comprise," "comprising," "comprises," and "comprised of" are synonymous with "include," "including," "includes," or "contain," "containing," or "contains," and are inclusive or open-ended terms specifying the presence of what follows, e.g., components, and do not exclude or preclude the presence of additional, unrecited components, features, elements, materials, or steps that are known in the art or disclosed therein.

[0024] The recitation of numerical ranges by endpoints includes all numbers and fractions subsumed within that range, as well as the recited endpoints. All percentages are understood to be percentages by weight unless otherwise defined or unless a different meaning is apparent to one of ordinary skill in the art from its use and in the context in which it is used. The terms "wt. %, "percent by weight", "% wt" or "wt%", as used herein and throughout this specification, refer to the relative weight of each component based on the total weight of the formulation, unless otherwise defined.

[0025] The term "one or more" or "at least one," e.g., one or more or at least one member of a group of members, will make itself clear by further example. This term specifically includes reference to any one of the aforementioned members, or any two or more of the aforementioned members, e.g., any ≧3, ≧4, ≧5, ≧6, or ≧7, etc. of the aforementioned members, and up to all of the aforementioned members.

[0026] Unless otherwise defined, all terms used in disclosing the present invention, including technical and scientific terms, have the meanings commonly understood by those skilled in the art to which the present invention belongs. By way of further guidance, definitions of terms used in the description are included to better understand the teachings of the present invention. The terms or definitions used herein are provided solely to aid in the understanding of the present invention.

[0027] References throughout this specification to "one embodiment" or "embodiment" mean that a particular feature, structure, or characteristic described in connection with an embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrase "in one embodiment" or "in one embodiment" in various places throughout this specification may, but do not necessarily, all refer to the same embodiment. Furthermore, particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments, as would be apparent to one of ordinary skill in the art from this disclosure. Furthermore, while some embodiments described herein include some features but not others included in other embodiments, combinations of features from different embodiments are within the scope of the present invention and are meant to form different embodiments, as would be understood by one of ordinary skill in the art. For example, in the following claims, any of the claimed embodiments may be used in any combination.

[0028] Furthermore, terms such as first, second, third, etc. in this specification and claims are used to distinguish between like elements and are not necessarily used to describe a sequential or chronological order, unless specified. Terms so used are interchangeable under appropriate circumstances, and it is understood that the embodiments of the invention described herein are capable of operating in other sequences than those described or illustrated herein.

[0029] In a first aspect, the present invention relates to a computer-implemented method for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor in a subject. The method preferably comprises obtaining at least one of exome or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from tumor cells and normal cells associated with the tumor in the subject. The method preferably further comprises obtaining a set of aberrant genomic events associated with the tumor by comparing the exome and / or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from the tumor cells to the exome and / or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from the normal cells. The method preferably further comprises obtaining data representing the peptide sequence of each of the set of neoantigens identified based at least in part on the set of aberrant events. The peptide sequence of each neoantigen comprises at least one modification that distinguishes it from a corresponding wild-type peptide sequence identified from normal cells of the subject. The method preferably further comprises obtaining data representing HLA peptide sequences from tumor cells based on tumor exome and / or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data. The method preferably further comprises training a deep learning model with a training dataset including a positive dataset. The positive dataset comprises a plurality of input-output pairs, each pair comprising, as input, an item of an antigenic determinant sequence. The antigenic determinant sequences are identified or inferred from surface-bound or secreted HLA / peptide complexes encoded by the corresponding HLA allele expressed by the training cells. Each pair further comprises, as output, an item of an alpha chain peptide sequence encoded by the corresponding HLA allele. The method preferably further comprises determining the likelihood of presentation of each of a set of neoantigens for the HLA peptide sequence by the trained model.

[0030] In a second aspect, the present invention relates to a computer system for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor in a subject, the computer system being configured to perform the computer-implemented method according to the first aspect of the invention.

[0031] In a third aspect, the present invention relates to a computer program product for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor in a subject, the computer program product comprising instructions that, when the computer program product is executed by a computer, cause the computer to carry out a method according to the first aspect of the invention.

[0032] In a fourth aspect, the present invention relates to the use of a method according to the first aspect of the invention, and / or a computer system according to the second aspect of the invention, and / or a computer program product according to the third aspect of the invention, for determining a treatment for a subject.

[0033] The present invention provides a computer-implemented method, a computer system, and a computer program product for determining the likelihood of presentation of neoantigens by tumor cells of a tumor in a subject, and the use of any of the methods, systems, or products for determining a treatment for the subject. Those skilled in the art will understand that the method may be implemented in a computer program product and executed using a computer system. It will also be apparent to those skilled in the art that the likelihood of presentation of a set of neoantigens may be used to determine a treatment for the subject. Therefore, the following will address the four aspects of the present invention together.

[0034] "Subject" as used herein refers to the term known in the state of the art and should preferably be understood as a human or animal body, most preferably a human body. "Animal" as used herein preferably refers to vertebrates, more preferably birds and mammals, even more preferably mammals. "Subject in need thereof" as used herein should be understood as a subject who will benefit from treatment.

[0035] A simple embodiment of the present invention preferably provides for obtaining at least one of exome or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from tumor cells and normal cells associated with said tumor of a subject. A simple embodiment preferably further provides for obtaining a set of aberrant genomic events associated with said tumor by comparing the exome and / or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from tumor cells to the exome and / or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from normal cells. It will be apparent that the exome, whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data are each compared with the data type of the corresponding nucleotide sequencing data.

[0036] As used herein, "neoantigenic determinant" refers to a term known in the art and should be understood as a class of major histocompatibility complex (MHC)-binding peptides that preferably arise from tumor-specific variants. These peptides represent antigenic determinants of neoantigens. Neoantigenic determinants can be recognized by the immune system as targets for T cells and induce immune responses against cancer.

[0037] As used herein, "neoantigen" refers to a term known in the art and should preferably be understood as an antigen that has at least one modification that distinguishes it from its most closely related wild-type antigen, i.e., the corresponding wild-type sequence, for example, through tumor cell mutations, tumor cell-specific post-translational modifications, fusions, transposable element insertions, alternative splicing events, or any method of modification known to those skilled in the art. Furthermore, neoantigens may or may not comprise a polypeptide or nucleotide sequence.

[0038] Preferably, the set of abnormal genomic events comprises one or more of single nucleotide polymorphisms (SNPs), indel mutations, gene fusions, chromosomal rearrangements (such as inversions, translocations, duplications, or chronotropy), transposable element insertions, or alternative splicing events. In the present context, the term "indel" is understood as a molecular biology term for the insertion or deletion of one or more nucleic acids in the genome of an organism. Furthermore, in the present context, the term "SNP" or "single nucleotide polymorphism" refers to a single base substitution occurring at a specific position in the genome of an organism.

[0039] The present invention may or may not use input peptide or neo-antigenic determinant sequences generated by a neo-antigenic determinant discovery pipeline that starts with raw sequencing data from a subject, preferably a patient. This raw sequencing data includes at least tumor DNA, preferably tumor DNA generated by a biopsy. Preferably, this raw data further includes tumor RNA, more preferably tumor RNA generated by a biopsy. Preferably, this raw data further includes normal DNA generated from a subject specimen, preferably a blood specimen. Preferably, this raw data further includes normal RNA generated from a subject specimen, preferably a blood specimen.

[0040] "Specimen" as used herein refers to the term known in the state of the art and should be understood as a single cell, or multiple cells, or fragments of cells, or an aliquot of bodily fluid, preferably obtained from a subject by means including venipuncture, excretion, ejaculation, massage, biopsy, needle aspiration, lavage, scraping, surgical incision or intervention, or any other means known in the art.

[0041] The neoantigenic determinant discovery pipeline outputs a list of all genomic and transcriptomic alteration events occurring within a tumor. These "aberrant genomic events" include novel transposable element insertion events, novel RNA isoforms, novel gene fusions, novel RNA editing events, and novel nucleotide-based post-translational modification events of produced proteins. In addition, it detects single nucleotide polymorphisms (SNPs) and indels (focal insertion or deletion mutations) at both the RNA and DNA levels, and compares the results from both analyses to produce a list of high-confidence SNPs and indels.

[0042] According to a preferred embodiment, a confidence score is associated with each of the sets of aberrant genomic events based at least in part on the number of sequencing reads in the sequencing data supporting each associated aberrant genomic event. Preferably, the confidence score is further based at least in part on the pervasiveness in the genome of the sequencing data supporting each associated aberrant genomic event. A preferred embodiment further includes obtaining a subset of aberrant genomic events by comparing the confidence score of each aberrant genomic event in the set of aberrant genomic events with a threshold. If the associated confidence score exceeds the threshold, the event is added to the subset. A set of neoantigens identified at least in part based on the set of aberrant events is identified according to this preferred embodiment at least in part based on the subset of aberrant events. Events with high confidence scores display a high number of sequencing reads, are pervasive in the genome, and are therefore selected for further investigation. As a result, performance is improved.

[0043] It should be noted that the present invention will not function if the input sequence contains non-standard amino acids. In the context of this specification, the term "non-standard amino acid" is understood as a non-standard or non-encoded amino acid that is not naturally encoded or found in the genetic code of any organism.

[0044] A simple embodiment of the present invention preferably provides for obtaining data representing peptide sequences of HLAs from tumor cells based on tumor exome and / or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data. Thus, the HLA composition of a tumor biopsy is assessed using the same genomic data used to identify the set of neoantigens. Preferably, the present invention provides for obtaining data representing peptide sequences of each of a set of HLAs based on tumor exome and / or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from tumor cells.

[0045] As used herein, "human leukocyte antigen (HLA)" refers to a term known in the state of the art and should preferably be understood as the gene complex encoding "major histocompatibility complex (MHC)" proteins in humans. These cell surface proteins are involved in regulating the immune system in humans. HLA genes are highly polymorphic, i.e., can have different alleles, thereby allowing HLA genes to fine-tune the adaptive immune system of a subject. In the context of this specification, the term "HLA binding affinity" or "MHC binding affinity" is understood as the affinity of binding between a specific antigen and a specific MHC allele. In the context of this specification, the term "HLA type" is understood as the complement of HLA gene alleles.

[0046] A simple embodiment of the present invention preferably provides for training a deep learning model with a training data set. The training data set preferably includes a positive data set. The positive data set preferably includes a plurality of input-output pairs. Each pair preferably includes, as input, an item of antigenic determinant sequence. The antigenic determinant sequence is preferably identified or inferred from surface-bound or secreted HLA / peptide complexes encoded by corresponding HLA alleles expressed by the training cells. Each pair preferably further includes, as output, an item of alpha chain peptide sequence encoded by the corresponding HLA allele.

[0047] As used herein, "training cells" should preferably be understood as cells from which specimens are derived. Such specimens are used to obtain the input and output of the input-output pairs in the positive dataset. The training cells may or may not be cells obtained from a monoallelic cell line, such as a human cell line, or cells obtained from a multiallelic tissue, such as a human tissue.

[0048] According to a most preferred embodiment, each positive input consists of a sequence of an 8-15 amino acid antigenic determinant shown to be present on the cell surface in a given dataset, and each associated positive output consists of the concatenated amino acid sequence, up to 71 amino acids, of the alpha chains of HLA alleles expressed by cells within the same dataset.

[0049] According to a preferred embodiment, the antigenic determinant sequence of the input of each input-output pair of the positive data set is obtained by mass spectrometry. In another or alternative embodiment, the peptide sequence of the alpha chain encoded by the corresponding HLA allele of the output of each input-output pair of the positive data set is obtained by mass spectrometry.

[0050] In one embodiment of the present invention, positive input-output pairs can be assigned different weights, preferably depending on their frequency of occurrence in the mass spectrometry data used to construct the positive training set. The weights modify the influence of the pairs on the training of the deep learning model. Larger weights will result in greater adjustment of parameters associated with the deep learning model when training the model using such input-output pairs, as further described below.

[0051] According to another preferred embodiment, the training dataset for training the deep learning model further comprises a negative dataset. The negative dataset preferably comprises a plurality of input-output pairs. Each pair preferably comprises, as input, an entry of a peptide sequence. The peptide sequence is preferably a random sequence of the human proteome. Each pair preferably further comprises, as output, a peptide sequence encoded from a random HLA allele.

[0052] According to a most preferred embodiment, each positive input is a random sequence from the human proteome that is not present in any ligandome dataset. The input is a random sequence of 8 to 15 amino acids. Each associated output is the concatenation of the alpha chain sequences of a random set of HLA alleles present in the positive dataset.

[0053] "Proteome" as used herein refers to a term known in the state of the art and should preferably be understood as the entire set of proteins expressed or capable of being expressed by a genome, cell, tissue, or organism at a given time. It is the set of proteins expressed in a given type of cell or organism at a given time under defined conditions. "Proteomics" is the study of the proteome.

[0054] Preferably, a portion, preferably a majority, and more preferably both the positive and negative data sets of the input-output pairs of the positive data set are used to train the deep learning model. Preferably, a portion, preferably a minority, and more preferably both the positive and negative data sets of the input-output pairs of the positive data set are used to validate the trained deep learning model.

[0055] The ratio between the number of positive and negative input-output pairs for training a deep learning model may or may not be changed, and said ratio is an important parameter for training the model.

[0056] The ratio between the number of positive and negative input-output pairs for validating a deep learning model may or may not be changed. The ratio is an important parameter for model validation.

[0057] According to preferred embodiments, the positive dataset comprises a mono-allelic dataset and a multi-allelic dataset. The mono-allelic dataset preferably comprises input-output pairs obtained from training cells from a mono-allelic cell line. The multi-allelic dataset preferably comprises input-output pairs obtained from training cells from a multi-allelic tissue. The training cells obtained from a mono-allelic cell line are preferably cells obtained from a mono-allelic human cell line. The training cells obtained from a multi-allelic tissue are preferably cells obtained from a human tissue. The multi-allelic human tissue may or may not be healthy or cancerous.

[0058] "Mono-allele" as used herein refers to the term known in the state of the art and should preferably be understood as a situation where only one allele occurs at a site or locus in a population.

[0059] "Multiple alleles" as used herein refers to a term known in the art and should preferably be understood as a situation in which many alleles occur. Polymorphisms are "multiple alleles", also called "polyallelic".

[0060] According to a preferred embodiment, training the deep learning model comprises two or more training cycles. Each training cycle preferably comprises a plurality of training steps. Each training step preferably comprises processing a pair of a plurality of input-output pairs. Preferably, one of said two or more training cycles comprises training the deep learning model on a single-allele dataset. Preferably, one of said two or more training cycles comprises training the deep learning model on both a single-allele dataset and a multi-allelic dataset.

[0061] According to a further preferred embodiment, the present invention provides three or more training cycles. One of the three or more training cycles is a supervised learning period, during which the model is trained on both mono-allelic and bi-allelic datasets to predict the complete sequence of amino acids represented by a particular set of alleles. One of the three or more training cycles is a burn-in period, during which only samples derived from the mono-allelic dataset are used for the model to learn a particular peptide-HLA relationship. One of the three or more cycles is a generalization period, during which the bi-allelic dataset is used to generalize the model, thereby learning patient data.

[0062] According to a preferred embodiment, the antigenic determinant sequences of the inputs of each input-output pair of the positive data set are obtained by mass spectrometry. New technological developments have made it possible to increase the availability of mass spectrometry-derived lists of peptides that are actually bound to MHC molecules on the cell surface. These lists are called "ligandomes." In the context of this specification, the term "ligandome" is understood as the complete set of molecular ligands for proteins in cells and organisms. Preferably, the positive set of input-output pairs is constructed from ligandome data from training cells.

[0063] Preferably, the deep learning model according to the present invention is at least one of a deep semantic similarity model, a convolutional deep semantic similarity model, an iterative deep semantic similarity model, a deep relevance matching model, a deep and wide model, a deep language model, a transformer network, a long short-term memory network, a trained deep learning text embedding, a trained named entity recognition, a Siamese neural network, an interacting Siamese network, or a lexical and semantic matching network, or any combination thereof.

[0064] Preferably, training the deep learning model includes determining a score function. More preferably, the score function is one or more of a squared error score function, an average score function, or a maximum score function. Preferably, the score function is constructed as a sum of squared errors between the probability output by the model and the HLA-neoantigen determinant relationship information associated with the training data set. Furthermore, this can be implemented by using scores 0 and 1, which represent the values ​​ascribed to the ground truth in the training data set: "not presented" (=0) and "presented" (=1).

[0065] In a further embodiment of the present invention, the model's coefficients are adjusted at every training step to minimize a score function. A neural network is composed of interconnected neurons. At the same time, each connection in the neural network is associated with a weight that, when multiplied by the input value, indicates the importance of the relationship in the neuron. In order for a neural network to learn, the weights associated with neuron connections must be updated after a forward pass of data through the network. These weights are adjusted, often via a process called backpropagation, to help reconcile discrepancies between the actual and predicted results of subsequent passes.

[0066] Preferably, the deep learning model according to the present invention is a sequence-to-sequence model. The term "sequence-to-sequence model (seq2seq)" as used herein refers to a term known in the state of the art and is also called an encoder-decoder model. This should preferably be understood as a model in which an encoder reads an input sequence and outputs a single vector, and a decoder reads the vector and produces an output sequence. Thus, such a model aims to map fixed and / or non-fixed length inputs, where the input and output lengths may differ, to fixed and / or non-fixed length outputs. The use of the seq2seq approach, in which HLA alleles are modeled by the amino acid sequences of specific, functionally relevant parts of the overall structure, has the advantage that the model can extrapolate and predict the likelihood of presentation of neoantigenic determinants for untrained HLA alleles. Most preferably, the seq2seq model is a transformer network.

[0067] According to a preferred embodiment, the present invention provides for processing the input of a plurality of input-output pairs into an embedded input numeric vector by converting corresponding items of antigenic determinant sequences using a neo-antigenic determinant embedder and a positional encoder. The embedded input numeric vector contains information about a set of amino acids constituting the antigenic determinant sequence of the corresponding item and the positions of the amino acids within the antigenic determinant sequence. According to a further preferred embodiment, the present invention provides for processing the output of a pair into an embedded output numeric vector by converting corresponding items of α-chain peptide sequences using an allele embedder and a positional encoder. The embedded output numeric vector contains information about a set of amino acids constituting the peptide sequence of the corresponding item and the positions of the amino acids within the peptide sequence. The above-mentioned embedder and encoder enable the conversion of input and output of a deep learning model into an appropriate format before or after processing, during training, validation, or use.

[0068] Most preferably, the deep learning model is a Transformer network or Transformer. Transformer networks were developed to solve problems of sequence transformation or neural machine translation, i.e., any task of transforming or matching an input sequence to an output sequence. For a model to perform sequence transformation, it is necessary to have some kind of memory. This requires understanding the connectivity, including dependencies and long-range connections, between inputs. These Transformer neural networks utilize the concept of self-attention. Transformer neural networks can replace previous approaches, such as long-term short-term memory (LSTM) or convolutional neural networks (CNN), which used attention between the model's encoder and decoder. The self-attention mechanism allows the model's inputs to interact with each other and find elements or parts that should receive more attention. The output is a collection of these interactions and attention scores.

[0069] More specifically, the attention function can be described as mapping a query, i.e., a sequence and a set of key-value pairs, to an output. Here, the query (q), key (k), value (v), and output are all vectors. The keys and values ​​can be considered as the model's memory, meaning all previously processed queries. A score is calculated to determine the self-attention of a token, i.e., an amino acid, in the sequence. Each token in the sequence needs to be scored for the token for which the self-attention calculation is desired. The score determines how much focus should be placed on other parts of the sequence once the token is encoded at a particular position. This score is calculated by taking the dot product of the query vector and the key vector of each scored token. By employing scaled dot-product attention, the output is calculated as a weighted sum of values. The weight assigned to each value is determined by the dot product of the query with all the keys.

[0070] There are various motivations for using self-attention methods. A primary advantage of using transformer neural networks is the ability to parallelize the encoder's self-attention, thus reducing overall model training time. Another is the path length between long-range dependencies in the network. Learning long-range dependencies is a key challenge in many sequence transformation tasks. One important factor affecting the ability to learn such dependencies is the length of the paths that the forward and reverse signals must traverse within the network. The shorter these paths are between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies.

[0071] According to a preferred embodiment, the transformer network comprises an encoder and a decoder. Neoantigenic determinant embedder o position encoder o One or more sequence encoders, each containing two sublayers: i. Multi-head self-attention sublayer ii. Forward propagation sublayer The decoder includes: One or more sequence decoders, each containing the following three sublayers: i. Multi-head self-attention sublayer ii. Multi-head Encoder-Decoder Attention Sublayer iii. Forward propagation sublayer oHLA sequence embedder o A probability generator, comprising: i. Linear projector ii. Softmax layer

[0072] An "embedder" converts each input into a vector or tensor using an embedding algorithm. This conversion is necessary because many machine learning algorithms, including deep neural networks, require inputs to be vectors of continuous values ​​because they do not operate on plain strings of text. Using an embedder provides the benefits of dimensionality reduction and contextual similarity. Reducing the dimensionality of a feature or dataset improves model accuracy, allows algorithms to train faster, requires less storage space, and removes redundant features and noise. The similarity between pairs of inputs can be calculated by some similarity or distance measure applied to corresponding pairs of vectors, providing a more expressive representation of the data.

[0073] In the Transformer, self-attention ignores the position of tokens within a sequence. However, the position and order of tokens, i.e., amino acids, are an important part of the sequence. To overcome this limitation, the Transformer explicitly adds a "positional encoding." This is a piece of information added to each token about its position within the sequence. Both input and output embedded sequences are position-encoded, allowing the self-attention process to accurately infer position-related interdependencies. These are added to the input or output embeddings before the sum enters the first attention layer.

[0074] The "sequence coding" consists of a stack of several identical layers. Each layer has two sublayers. The first is a "multihead self-attention" mechanism, and the second is a simple "forward propagation network." Instead of computing attention only once, the multihead mechanism performs scaled dot-product attention multiple times in parallel. The independent attention outputs are simply concatenated and linearly transformed to the expected dimension. This expands the model's ability to focus on different locations. The output of the self-attention layer is fed into a simple forward propagation neural network, where information further moves in only one direction. Remaining connections, or shortcuts, are used around each of the two sublayers. This allows the model to use fewer layers in the initial training phase, thereby simplifying the network. Each layer ends with a normalization of its output and the sum of the residual connections. The "sequence decoder" is very similar to the encoder, but has an additional "multihead encoder-decoder attention sublayer." The encoder-decoder sublayer is different from the encoder or decoder attention sublayer. Unlike multi-head self-attention, the encoder-decoder attention sublayer creates a query matrix from the layer below it, which is the decoder self-attention, and obtains a matrix of keys and values ​​from the encoder layer output, which helps the decoder focus on the right location in the input array.

[0075] The decoder output is converted into a predicted next token probability by using a "linear projection" or transformation and a "softmax function" or "softmax layer." The linear projection layer reduces the dimensionality of the data as well as the number of network parameters. The softmax layer is a multi-class operation and is used in determining the probabilities of multiple classes at once. Because the outputs of the softmax function can be interpreted as probabilities (i.e., they must sum to 1), the softmax layer is typically the final layer used in neural network functions.

[0076] According to a preferred embodiment, training of a deep learning model comprises multiple training steps, each of which involves processing a pair of a plurality of input-output pairs according to the following steps: o Processing the inputs of the pair into an embedded input numeric vector by converting corresponding items of the antigenic determinant sequence using a neo-antigenic determinant embedder and a positional encoder, where the embedded input numeric vector contains information about the set of amino acids that make up the antigenic determinant sequence of the corresponding item and the positions of the amino acids within the antigenic determinant sequence. o Processing the output of the pair into an embedded output numeric vector by transforming the corresponding entry of the α-chain peptide sequence using an allele embedder and a positional encoder, the embedded output numeric vector containing information about the set of amino acids that make up the peptide sequence of the corresponding entry and their positions in the peptide sequence. o Processing the embedded input numeric vector into an encoded input numeric vector using at least one sequence encoder including a multi-head self-attention sublayer and a forward propagation sublayer, the encoded input numeric vector including information about the antigenic determinant sequence features of the corresponding item in the antigenic determinant sequence. o processing the embedded output numeric vector using a multi-head self-attention sublayer into an output attention numeric vector, the output attention numeric vector containing information about the interdependencies of multiple amino acids that make up the peptide sequence of a corresponding item of the alpha chain peptide sequence; o Processing the coded input numeric vectors and the corresponding output attention vectors into correlated numeric vectors using a multi-head encoder-decoder attention sublayer and a forward propagation sublayer, the correlated numeric vectors containing correlation information between the coded input numeric vectors and the corresponding output attention vectors. o Processing the correlation numeric vectors into corresponding probabilities between the embedded input numeric vectors and the embedded output numeric vectors using a probability generator.

[0077] In other embodiments, the embedding of both the input of the pair, the antigenic determinant sequence, and the output of the pair, the HLA peptide sequence, can follow one of different modalities.

[0078] According to a first possible format, each amino acid position is one-hot coded, which means that since there are 20 standard amino acids, it is converted into a 1x20 vector. Every position in the vector is 0 (zero), except for one position, which is 1 (one). This latter position represents the actual amino acid present. In this way, for example, a 9-mer is converted into a 9x20 matrix, with only 9 positions being 1, while all other positions are 0.

[0079] According to the second possible format, each amino acid is tokenized individually: a dictionary of amino acid logarithms is constructed, and each amino acid is represented by a numerical value. For example, proline is represented as 1, while valine is represented as 2, etc. In this way, a 9-mer is converted into a vector with a length of 9 numbers.

[0080] According to a third possible format, each amino acid is replaced by an embedding vector of n numbers. These n numbers relate to specific characteristics of the amino acid, which may be defined physically, chemically, or in other ways. In a preferred example, the amino acid is embedded by the values ​​of its n main components, which are derived from a set of physicochemical properties / characteristics. Thus, the 9-mer is transformed into a matrix of 9xn numbers in this example.

[0081] Three possible embedding modes can be performed directly at individual amino acid positions: one amino acid is embedded in one embedding vector; in another mode, the sequence can be split into strings with a length greater than one to embed both the antigen determinant sequence (input) and the HLA sequence (output). In this way, instead of considering individual amino acids, k-mers are considered.

[0082] According to a further preferred embodiment, the processing of pairs of the plurality of input-output pairs further comprises the following steps: Obtaining data points for the score function for training by comparing the probability of correspondence between the embedded input numeric vectors and the embedded output numeric vectors with corresponding related information associated with the training data set. o Tuning parameters associated with the deep learning model to optimize said score function. Preferably, the score function is one or more of a squared error score function, a mean score function, or a maximum score function.

[0083] In one embodiment, the score function may be a binary cross-entropy loss function.

[0084] In one embodiment of the present invention, as previously described, positive input-output pairs can be assigned different weights, preferably depending on their frequency of occurrence in the mass spectrometry data used to construct the positive training set. The weights modify the influence the pairs have on the training of the deep learning model. A larger weight will result in a larger adjustment of the parameters associated with the deep learning model when training the model using such input-output pairs.

[0085] According to another preferred embodiment, the Transformer network includes an encoder but no decoder. In this network, both the input antigenic determinant sequence and the embedded vector of the input HLA sequence are processed as a single vector. A type of masking is performed to indicate whether the values ​​of the input embedded vector are associated with either a neoantigenic determinant or an HLA. This means, for example, that the sign of the numerical value associated with the antigenic determinant input is changed, while the sign associated with the HLA input remains unchanged. Additionally, in this network model, custom separation values ​​are inserted at various positions in the input embedded vector, particularly at the beginning and / or end of the vector, and between the antigenic determinant-associated value and the HLA-associated value. In this way, further distinctions can be made between both input sequences while processing them as a single vector.

[0086] According to another preferred embodiment, after training the model, one or more of the following is obtained: - A set of coefficients that can be used to reproduce a function given the correct structure - a set of parameters that describe all aspects of the model's training - A structural scheme that can be used to regenerate the model for inference / testing purposes -HLA dictionary seen during model training

[0087] According to one embodiment, the present invention provides a method in which other semi-independent models can be trained in conjunction with the primary architecture to take into account other relevant biological parameters, including RNA expression of the gene from which the neoantigenic determinant is derived, RNA expression of all other genes in the specimen, expression of non-coding RNAs, post-translational modification status, RNA editing events, immune rates of all immune cell types, clonality of the specimen, confidence scores of all genome modification events, peptide-MHC binding affinity predicted by other means, peptide-MHC complex stability, peptide stability and turnover, neighboring amino acids in the neoantigenic determinant's original protein, proteasome activity, and peptide processing activity. The model structure is configured in such a way that any missing data in this list does not prevent the model from outputting a representation probability.

[0088] According to a preferred embodiment, the present invention further comprises the following steps: training a semi-independent neural network with semi-independent training data sets including at least a positive data set of the deep learning model or a variant thereof and an associated prediction-improvement parameter training data set, wherein the associated prediction-improvement parameter training data set is related to one or more biological parameters of: RNA expression of a gene from which the neoantigenic determinant is derived; RNA expression of a plurality of genes in the cancerous tissue specimen; expression of non-coding RNA sequences; post-translational modification information; RNA editing events; immunoprevalence of a plurality of immune cell types; clonality of the cancerous tissue specimen; confidence scores of a plurality of genomic modification events; peptide-MHC binding affinity; peptide-MHC complex stability; peptide stability and / or turnover; adjacent amino acids in the neoantigenic determinant sequence; proteasomal activity; and peptide processing activity. Preferably, the associated prediction-improvement parameter training data set is related to at least adjacent amino acids in the neoantigenic determinant sequence. - determining the semi-independent presentation likelihood of each of the set of neoantigens for the HLA peptide sequence by a trained semi-independent neural network. - for each of the set of neoantigens, combining the determined semi-independent presentation likelihood with the presentation likelihood obtained by the trained model to obtain an overall presentation likelihood. Preferably, the combination is performed by a trained single layer neural network. Preferably, the semi-independent neural network is a single layer neural network. Preferably, at least one of exome or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data from tumor cells and normal cells associated with the tumor of the subject is obtained from cancerous tissue specimens and healthy tissue specimens of the subject, respectively.

[0089] According to one embodiment, training of all sublayers is performed by using an Adam-type optimization algorithm. Optimization is an algorithm or method used to modify attributes of a neural network, such as weights and learning rates, to help reduce loss or error and obtain results faster. The algorithm utilizes the power of adaptive learning rate methods to find individual learning rates for each parameter. Adam uses estimates of the first and second moments of the gradient to adapt the learning rate for each weight of the neural network.

[0090] According to one embodiment, a deep learning model, preferably a Transformer network, is trained for five epochs of five-fold cross-validation. k-fold cross-validation is easily understood, easily implemented, and generally produces skill estimates for the model on new data that have lower bias than other methods. There is a bias-variance tradeoff associated with the choice of k in k-fold cross-validation. Performing k-fold cross-validation using k=5 produces test error rate estimates that are neither overly biased nor very variance.

[0091] As used herein, "epoch" refers to a term known in the state of the art and should preferably be understood as an indication of the number of passes through the entire training data set that a machine learning algorithm completes. One epoch is one cycle through the complete training data set.

[0092] As used herein, "K-fold cross-validation" refers to a term known in the state of the art and should preferably be understood as a statistical method for estimating the skill of a machine learning model. This technique involves repeatedly randomly dividing a set of observations into k groups or folds of approximately equal size. The first fold is treated as a validation set. The method is applied to the remaining k-1 folds. Results of k-fold cross-validation runs are often summarized with a mean value for the model skill score. It is also good practice to include a measure of the dispersion of the skill score, such as the standard deviation or standard error.

[0093] The present invention is further described by the following non-limiting examples, which further illustrate the invention and are not intended, or should be construed, to limit the scope of the invention.

[0094] Example Example 1: This example relates to training a sequence-to-sequence transformer model according to the present invention.

[0095] The sequence-to-sequence transformer model has the following architecture: -encoder Neoantigenic determinant embedder o position encoder o One or more sequence encoders, each containing two sublayers: i. Multi-head self-attention sublayer ii. Forward propagation sublayer -Decoder One or more sequence decoders, each containing the following three sublayers: i. Multi-head self-attention sublayer ii. Multi-head Encoder-Decoder Attention Sublayer iii. Forward propagation sublayer oHLA sequence embedder o A probability generator, comprising: i. Linear Projector ii. Softmax layer

[0096] The sequence-to-sequence transformer model described above is trained by processing a set of positive and negative input-output pairs through the model.

[0097] A positive set of input-output pairs is constructed from ligandome data from monoallelic human cell lines or biallelic human tissues (healthy or cancerous). Each positive input consists of the sequence of an antigenic determinant shown to be present on the cell surface in a given data set (8-15 amino acids). Each associated positive output is made from the concatenated amino acid sequence of the alpha chains of HLA alleles expressed by cells in the same data set (71 amino acids).

[0098] A negative set of input-output pairs is constructed from the human proteome. Each input is a random 8- to 15-mer sequence from the human proteome that is not present in either ligandome dataset. Each associated output is the concatenation of the alpha chain sequences of a random set of HLA alleles present in the positive dataset.

[0099] Each training input-output pair is processed through the following model: -The input peptides are padded with "." tokens as needed to a length of 15. The resulting sequences are then embedded into a 21*15 one-hot tensor by the neo-antigenic determinant embedder. - The sequence-based model embeds all HLA into a 21*71 one-hot tensor by an allele embedder according to the sequence of the two peptide-interacting α-helices. The embedded sequences of both input and output are then position-encoded, allowing self-attention processing to accurately infer position-related interdependencies. The embedded input sequences are processed sequentially by the sequence encoder. The self-attention sublayer learns the intra-peptide interdependencies. The forward propagation sublayer processes the input embeddings accordingly. The result of this encoding process is an encoding that represents a fixed-dimensional feature of the input nascent peptide. The embedded HLA sequence input is then processed and combined with the encoded neoantigenic determinant input in each decoder sequentially to progressively form the embedded output sequence. The self-attention sublayer learns intra-allelic interdependencies. The peptide attention sublayer correlates the encoded peptide representation with the embedded output. The forward propagation sublayer applies modifications to the embedded output accordingly. In this step, a correspondence between input and output is established. Note that the attention sublayer, which allows for the detection of intra-sequence interdependencies, significantly improves the overall predictive power of the model. -Finally, the embedded output is processed through a generator to output the probability of the correspondence between the embedded input and the embedded output, representing the probability of presentation (0 to 1, 1 being the highest probability). The score function is constructed as the sum of squared errors between the probability output by the model and the actual HLA-peptide relationship (0: the peptide was not presented on the surface of cells expressing this allele, i.e., the peptide was part of the negative data set mentioned above - 1: the peptide was presented on the surface of cells expressing this allele, i.e., it was part of the positive data set mentioned above). Other ways of aggregating the data are possible, such as considering a mean score function or a maximum score function. At every training step, i.e., at each processing of a new input-output pair, the model's coefficients are adjusted to minimize the score function thus defined.

[0100] The model is trained as follows: -The model is trained for 5 epochs of 5-fold cross-validation. -Training this model follows these steps: First, the model is trained on all samples and simply predicts, amino acid by amino acid, the complete sequence of amino acids represented by a particular set of alleles (self-supervised learning). Next, in order for the model to learn specific peptide-HLA relationships, only samples derived from mono-allelic HLA datasets (e.g., from mono-allelic cell lines) are used for training (a "burn-in" period). Finally, to generalize model learning to real patient data, HLA multi-allelic instances are used for training. -All layers of the model are trained using an ADAM-type optimizer.

[0101] At the end of training, the model outputs a set of coefficients that can be used to reproduce the function given the correct structure, a set of parameters that describe all aspects of the model's training, a structural scheme that can be used to recreate the model for inference / testing, and a dictionary of HLAs seen during model training.

[0102] Example 2: This example relates to the use of a model trained according to Example 1 in a workflow according to the invention.

[0103] The present embodiment provides a workflow for predicting the likelihood of presentation on the surface of cancer cells of variable length neoantigenic determinants given the set of HLA alleles expressed by the cells.

[0104] This workflow uses a sequence-to-sequence transformer model, which allows for the extrapolation and prediction of the likelihood of presentation of a neoantigenic determinant for any HLA allele, even if it has not been trained.

[0105] The workflow for this work is as follows: First, new peptides are discovered using next-generation sequencing data on cancer biopsies. Both DNA and RNA sequencing data are used to extract a set of aberrant genomic events that potentially carry neoantigenic determinants. These events are given a confidence score, which is based on the number of sequencing reads that support them and their penetrance in the genome. The antigenic determinants from the highest confidence events are selected for follow-up steps. -The HLA composition of the biopsy will also be assessed using the same genomic data. The sequences of the selected peptides are provided to a trained model along with the sequences of known HLA. -The model calculates the probability of peptide presentation for each HLA in the provided set and outputs an overall peptide probability based on these individual values.

[0106] Additionally, this workflow may or may not include steps to refine the probability predictions by providing other biological parameters to the model, such as RNA expression levels, MHC binding likelihood, or neo-antigenic determinant protein context.

[0107] Example 3: This example relates to an alternative implementation of the Transformer model according to Example 1.

[0108] As described above in Example 1, the input neoantigenic determinant sequence is padded to a length of 15 using "." tokens as needed. The resulting sequence is then embedded into a 21 x 15 one-hot tensor by the neoantigenic determinant embedder. Therefore, the model in Example 1 requires that the sequence be within the correct length range. However, the model can also be implemented to allow for antigenic determinants and HLAs of any length. Similarly, the model can be implemented to allow for variable length embedding. Furthermore, the model can be implemented to allow for embedding into matrices of different sizes up to 300 x 15.

[0109] As described above in Example 1, the model is sequence-based, and the allele embedder embeds all HLAs into a 21 × 71 one-hot tensor according to the sequence of two peptide-interacting α-helices. Alternatively, the model can process the associated HLAs as a categorical encoding. Categorical encoding refers to converting categorical features into one or more numerical features. All HLAs are encoded according to a central repository that reorganizes all HLA sequences known at the time the model is built. Alternatively, the model can be non-sequence-based, whereby HLAs are one-hot encoded based on their previous central repository encoding. Associated HLA sequences are processed one by one. Thus, a particular neoantigenic determinant would be processed once for each HLA sequence found to be associated in a particular specimen. The model would not be able to output a prediction if the HLA allele amino acid sequence was unknown. This is highly unlikely, but possible, for some rare HLA alleles.

[0110] Example 4: This example relates to the use of the workflow according to Example 2 to determine a subject's treatment.

[0111] Treatment decisions are as follows: - selecting a subset of the set of identified neoantigens based on the determined presentation likelihood to obtain a selected subset of neoantigens. The subset is obtained by comparing the presentation likelihood of each of the set of neoantigens with a threshold. If the associated presentation likelihood exceeds said threshold, the neoantigen is added to said subset. - identifying one or more T cells that are antigen-specific for at least one of the neoantigens within said subset;

[0112] Example 5: This example relates to an improved model that includes a sequence-to-sequence Transformer model according to Example 1 and one or more semi-independent models for the aforementioned Transformer model. The improved model can be used in a workflow according to Example 2 to determine a treatment for a subject.

[0113] According to this example, multiple semi-independent single-layer neural network models are trained in association with a primary Transformer architecture to consider other relevant biological parameters. Thus, each of the multiple semi-independent models is trained by training a single-layer neural network with a semi-independent training data set that includes a training data set for a sequence-to-sequence Transformer model and an associated prediction-improvement parameter training data set. By considering parameters from the prediction-improvement parameter training data set, the overall prediction accuracy is improved.

[0114] The parameter training data set for each of the multiple semi-independent single-layer neural network models relates to one or more biological parameters including RNA expression of the gene from which the neoantigenic determinant is derived, RNA expression of all genes in the cancerous tissue specimen except for the gene from which the neoantigenic determinant is derived, expression of non-coding RNA sequences, post-translational modification status, RNA editing events, immune rates of all immune cell types, clonality of the cancerous tissue specimen, confidence scores of all genomic modification events, peptide-MHC binding affinity predicted by other means, peptide-MHC complex stability, peptide stability and turnover, adjacent amino acids in the original protein of the neoantigenic determinant, proteasome activity, and peptide processing activity.

[0115] After training each of the semi-independent models, the trained semi-independent neural network determines a semi-independent presentation likelihood for each of the set of neoantigens for the HLA peptide sequence. The determined semi-independent presentation likelihoods are then combined with the presentation likelihoods obtained by the trained model for each of the set of neoantigens to obtain an overall presentation likelihood. In this example, the combination is performed by a trained single-layer neural network.

[0116] Example 6: This example concerns a comparison between a model according to the present invention and the prior art algorithms, the EDGE algorithm and the MHCflurry algorithm.

[0117] A sequence-to-sequence transformer model according to the present invention was developed and trained on: The positive dataset contains 326,297 publicly available input-output pairs. Each pair contains, as input, an entry for an antigenic determinant sequence. The antigenic determinant sequence is identified or predicted from surface-bound or secreted HLA / peptide complexes encoded by the corresponding HLA allele expressed by the training cells. Each pair further contains, as output, an entry for the alpha chain peptide sequence encoded by the corresponding HLA allele. Publicly available from Abelin et al., 2017; Bulik-Sullivan et al., 2019; di Marco et al., 2017; Sarkizova et al., 2019; and Trolle et al., 2016. The negative dataset includes 652,594 input-output pairs. Each input includes, as input, an entry for a peptide sequence. The peptide sequence is a random sequence of the human proteome. Each pair further includes, as output, a peptide sequence encoded from a random HLA allele.

[0118] The model was then tested on a test data set that included: -729 positive pairs, selected from the published test dataset of the EDGE algorithm (Bulik-Sullivan et al. 2019). -1.822.500 negative pairs. Each pair contains as input an entry of a peptide sequence. The peptide sequence is a random sequence of the human proteome. Each pair further contains as output a peptide sequence encoded from a random HLA allele.

[0119] Care was taken to avoid including pairs in the test dataset that had already been included in the training phase of the model.

[0120] A fit-recall curve (PR curve) was generated for the test dataset. Fit is measured as the proportion of truly presented positive antigenic determinants called, while recall measures the proportion of truly positive antigenic determinants that were correctly called positive. Thus, the fit-recall curve is a good measure of the model's ability to accurately call the desired positive result without making errors. The better the model, the more the fit-recall curve slopes toward the upper right corner.

[0121] The results are shown in Figure 1A. The results of the Transformer model according to the present invention are shown in blue (most sloping towards the upper right corner), while the results of the EDGE algorithm are shown in black. Additionally, the (substantially flat) green line represents the best accuracy achieved by the affinity-based model MHCflurry.

[0122] The results clearly show that the model according to the invention outperforms the close prior art algorithm EDGE and current state-of-the-art industry methods such as MHCflurry on the same test data set.

[0123] Example 7: This example concerns the capabilities of the model according to the invention for extrapolation and prediction.

[0124] As a sequence-to-sequence algorithm, the model derives its predictive power not from taxonomic data, but from comparing and noting correlations between two sequences. This means that predictions can be made for HLA alleles for which no training data was available, provided the protein sequence is known.

[0125] This ability for extrapolation / prediction is a real advantage considering that acquiring new training data is a lengthy and expensive process.

[0126] To test this capability, a model was trained as in Example 6. A new test data set was constructed from 2,039 positive pairs uniquely associated with the HLA-A*74:02 allele, along with 5,097,500 negative pairs for which no data existed in the training set. Each pair includes, as input, an entry for a peptide sequence. These peptide sequences are random sequences in the human proteome. Each pair further includes, as output, a peptide sequence encoded from a random HLA allele.

[0127] The results are shown in Figure 1 B. The fit-recall curves clearly show that the model according to the invention has very good predictive power, even for previously unseen alleles.

Claims

1. 1. A computer-implemented method for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor in a subject, the method comprising: obtaining sequencing data from tumor cells associated with the tumor and normal cells of the subject, wherein the sequencing data is at least one of exome or whole genome nucleotide sequencing data and transcriptome nucleotide sequencing data; obtaining a set of aberrant genomic events associated with the tumor by comparing the sequencing data from the tumor cells to the sequencing data from the normal cells; obtaining data representing a peptide sequence for each of a set of neoantigens identified based at least in part on the set of aberrant genomic events, wherein the peptide sequence of each neoantigen includes at least one modification that distinguishes it from a corresponding wild-type peptide sequence identified from the normal cells of the subject; obtaining data representing HLA peptide sequences from the tumor cells based on the tumor sequencing data; training a deep learning model with a training dataset comprising a positive dataset, the positive dataset comprising a plurality of input-output pairs, each input-output pair comprising, as an input, an item of an antigenic determinant sequence, the antigenic determinant sequence being identified or inferred from a surface-bound or secreted HLA / peptide complex encoded by a corresponding HLA allele expressed by a training cell, each input-output pair further comprising, as an output, an item of an alpha chain peptide sequence encoded by the corresponding HLA allele; the training of the deep learning model comprising a plurality of training steps, each training step comprising: processing the input of the input-output pair into an embedded input numeric vector by converting the corresponding item of an antigenic determinant sequence using a neo-antigenic determinant embedder and a positional encoder, the embedded input numeric vector containing information about a set of amino acids that make up the antigenic determinant sequence of the corresponding item and the positions of the amino acids within the antigenic determinant sequence; processing the output of the input-output pair into an embedded output numeric vector by transforming the corresponding entry of the peptide sequence of the α chain using an allele embedder and a positional encoder, the embedded output numeric vector containing information about the set of amino acids that make up the peptide sequence of the corresponding entry and the positions of the amino acids within the peptide sequence; processing the embedded input numeric vector into an encoded input numeric vector using at least one sequence encoder including a multi-head self-attention sublayer and a forward propagation sublayer, wherein the encoded input numeric vector contains information about the features of the antigenic determinant sequence of the corresponding item in the antigenic determinant sequence; processing the embedded output numeric vector into an output attention numeric vector using a multi-head self-attention sublayer, the output attention numeric vector containing information about interdependencies of the amino acids constituting the peptide sequence of the corresponding item of the peptide sequence of the alpha chain; processing the coded input numeric vectors and corresponding output attention vectors into correlated numeric vectors using a multi-head encoder-decoder attention sublayer and a forward propagation sublayer, the correlated numeric vectors including correlation information between the coded input numeric vectors and the corresponding output attention vectors; processing said correlation numeric vectors into probabilities of correspondence between said embedded input numeric vectors and said embedded output numeric vectors using a probability generator; and determining the likelihood of presentation of each of the set of neoantigens for the peptide sequence of the HLA by the trained deep learning model; A method comprising:

2. 10. The method of claim 1, associating a confidence score with each of the set of aberrant genomic events based at least in part on the number of sequencing reads in the sequencing data that support each associated aberrant genomic event; obtaining a subset of aberrant genomic events by comparing the confidence score of each aberrant genomic event in the set of aberrant genomic events to a threshold, where an event is added to the subset if the associated confidence score exceeds the threshold; Including, the set of neoantigens is identified based at least in part on the subset of aberrant genomic events. method.

3. 3. The method of claim 1, wherein the positive dataset comprises a mono-allelic dataset and a multi-allelic dataset, the mono-allelic dataset comprising input-output pairs obtained from training cells from a mono-allelic cell line, and the multi-allelic dataset comprising input-output pairs obtained from training cells from a multi-allelic tissue.

4. 4. The method of claim 3, wherein the training of the deep learning model comprises two or more training cycles, each training cycle comprising a plurality of training steps, each training step comprising processing a pair of the plurality of input-output pairs, one of the two or more training cycles comprising training the deep learning model on the single-allele dataset, and one of the two or more training cycles comprising training the deep learning model on both the single-allele dataset and the multi-allelic dataset.

5. 5. The method of claim 1, wherein the training dataset for training the deep learning model further comprises a negative dataset comprising a plurality of input-output pairs, each input-output pair comprising an item of a peptide sequence as an input, the peptide sequence being a random sequence of the human proteome, and each input-output pair comprising the negative dataset further comprises, as an output, a peptide sequence encoded from a random HLA allele.

6. 6. The method of claim 1, wherein the deep learning model is at least one of a deep semantic similarity model, a convolutional deep semantic similarity model, an iterative deep semantic similarity model, a deep relevance matching model, a deep and wide model, a deep language model, a transformer network, a long short-term memory network, trained deep learning text embeddings, trained named entity recognition, a Siamese neural network, an interacting Siamese network, or a lexical and semantic matching network, or a combination thereof.

7. 7. The method of claim 1, wherein the deep learning model is a Transformer network.

8. 8. The method of claim 1, wherein the processing of the input-output pairs of the plurality of input-output pairs comprises: obtaining score function data points for training by comparing the probability of correspondence between the embedded input numeric vectors and the embedded output numeric vectors with corresponding relationship information associated with the training data set; adjusting parameters associated with the deep learning model to optimize the score function; further comprising: method.

9. The method of claim 8, wherein the score function is one or more of a sum of squared errors score function, a mean score function, or a maximum score function.

10. 8. The method of claim 7, wherein the transformer network includes an encoder and a decoder, the encoder comprising: a neoantigenic determinant embedder; a position encoder; one or more constellation encoders, each including two sublayers: a multi-head self-attention sublayer and a forward propagation sublayer; and the decoder comprises: one or more constellation decoders, each of which includes three sublayers: a multi-head self-attention sublayer, a multi-head encoder-decoder attention sublayer, and a forward propagation sublayer; an HLA sequence embedder; a probability generator, the probability generator including a linear projector and a softmax layer; A method comprising:

11. 11. The method according to any one of claims 1 to 10, training a semi-independent neural network with semi-independent training datasets comprising at least the positive dataset of the deep learning model or a variant thereof and an associated prediction-improvement parameter training dataset; the associated prediction-improvement parameter training data set is related to one or more biological parameters of: RNA expression of a gene from which a neoantigenic determinant is derived; RNA expression of a plurality of genes in a cancerous tissue specimen; expression of a non-coding RNA sequence; post-translational modification information; RNA editing event information; immune rates of a plurality of immune cell types; clonality of the cancerous tissue specimen; confidence scores of a plurality of genomic modification events; peptide-MHC binding affinity; peptide-MHC complex stability; peptide stability and / or turnover; adjacent amino acids in a neoantigenic determinant sequence; proteasome activity; and peptide processing activity. determining, by the trained semi-independent neural network, a semi-independent presentation likelihood for each of the set of neoantigens for the peptide sequence of the HLA; for each of the set of neoantigens, combining the determined semi-independent presentation likelihoods with the presentation likelihoods obtained by the trained deep learning model to obtain an overall presentation likelihood; Including, method.

12. The method of claim 11, wherein the associated prediction-improvement parameter training data set relates to at least adjacent amino acids within a neoantigenic determinant sequence.

13. The method of claim 11, wherein combining the determined semi-independent presentation likelihood and the presentation likelihood for each of the set of neoantigens is performed by a trained single-layer neural network.

14. The method of claim 11, wherein the semi-independent neural network is a single-layer neural network.

15. 15. A computer system for determining the likelihood of presentation of a set of neoantigens by tumor cells of a tumor in a subject, the computer system being configured to perform the computer-implemented method of any one of claims 1 to 14.

16. 15. A computer program for causing a computer system to determine the likelihood of presentation of a set of neoantigens by tumor cells of a tumor in a subject, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the computer-implemented method of any one of claims 1 to 14.

17. The method according to any one of claims 1 to 14, selecting a subset of the identified set of neoantigens based on the determined presentation likelihood to obtain a selected subset of neoantigens; identifying one or more T cells that are antigen-specific for at least one of the neoantigens within the subset; Including, method.

18. A method of using the computer system of claim 15, comprising: selecting a subset of the identified set of neoantigens based on the determined presentation likelihood to obtain a selected subset of neoantigens; identifying one or more T cells that are antigen-specific for at least one of the neoantigens within the subset; Including, method.

19. A method of using the computer program of claim 16, comprising: selecting a subset of the identified set of neoantigens based on the determined presentation likelihood to obtain a selected subset of neoantigens; identifying one or more T cells that are antigen-specific for at least one of the neoantigens within the subset; Including, method.

20. A method according to any one of claims 17 to 19, wherein the subset of the set of identified neoantigens is obtained by comparing the presentation likelihood of each of the set of neoantigens with a threshold, and if the associated presentation likelihood exceeds the threshold, the neoantigen is added to the subset.

Images