Deep learning system for predicting the T cell receptor binding specificity of neoantigens
View Patent ↗Neoantigens play a key role in the recognition of tumor cells by T cells. However, only a small proportion of neoantigens truly elicit T cell responses, and fewer clues exist as to which neoantigens are recognized by which T cell receptors (TCRs). To help determine the TCRs that interact with particular neoantigens, prediction models that predict TCR-binding specificities of neoantigens presented by different classes of major histocompatibility complex (MHCs) were developed. To confirm the applicability of the model to clinical settings, the prediction models were comprehensively validated by a series of analyses. The validated prediction models used a flexible transfer learning approach and differential learning schema to achieve highly accurate prediction of TCR binding specificity only using TCR sequence data, antigen sequence data, and MHC alleles.
1 . A method of predicting T cell receptor (TCR) binding specificities comprising:
determining a set of MHC embeddings that encode neoantigen and major histocompatibility complex (MHC) data for a plurality of MHC proteins (pMHC);
determining a set of TCR embeddings that encode TCR data for a plurality of TCR sequences;
pre-training a prediction model on the set of MHC embeddings and the set of TCR embeddings;
training the prediction model using a differential learning schema that feeds a binding TCR-pMHC pair and a non-binding TCR-pMHC pair into the prediction model during each training cycle; and
determining a prediction for binding specificity of an input TCR-pMHC pair based on the prediction model.
2 . The method of claim 1 , further comprising:
obtaining a set of TCR-pMHC pairs that are experimentally validated as immunogenic, the set of TCR-pMHC pairs including the input TCR-pMHC pair; and
validating the prediction model by comparing the binding specificity prediction for the input TCR-pMHC pair to a known binding specificity for the input TCR-pMHC pair.
3 . The method of claim 1 , further comprising:
determining a clonal expansion of a plurality of T cells, the clonal expansion including multiple TCR clones having known binding interactions with a set of pMHCs and a clone size for each of the multiple TCR clones;
determining a prediction for binding specificity between each of the multiple TCR clones and each of the pMHCs included in the set of pMHCs based on the prediction model; and
validating the prediction model by comparing the clone size for each of the TCR clones to the predicted binding specificity.
4 . The method of claim 1 , wherein each of the MHC embeddings includes a numeric representation of one or more pMHCs, and the method further comprising:
training a MHC numeric embedding layer on a MHC training dataset including textual representations of pMHCs; and
determining the numeric representation of the one or more pMHCs for each of the MHC embeddings based on the MHC numeric embedding layer.
5 . The method of claim 1 , wherein the MHC embeddings are determined using a multi-layer neural network that determines a probability that a particular pMHC molecule binds to one or more neo-antigen protein sequences.
6 . The method of claim 1 , wherein each of the TCR embeddings include a numeric representation of one or more TCR protein sequences, and the method further comprising:
training a TCR numeric embedding layer on a TCR training dataset including multiple training TCR protein sequences, the TCR training dataset including a structured data representation of one or more biochemical properties of multiple amino acids included in the training TCR protein sequences; and
determining the numeric representation of the one or more TCR protein sequences based on the TCR numeric embedding layer.
7 . The method of claim 6 , wherein the multiple amino acids are included in a complementary determining region (CDR) of the training TCR protein sequences, and the method further comprising:
manipulating the structured data representation to enable amino acids from multiple CDRs of the training TCR protein sequences to be added to the TCR training dataset.
8 . The method of claim 1 , wherein the TCR embeddings are determined using an auto-encoder that includes multiple encoder layers and multiple decoder layers.
9 . The method of claim 1 , further comprising normalizing the MHC embeddings and the TCR embeddings to enable the prediction model to be pre-trained on multiple classes of pMHCs.
10 . The method of claim 1 , wherein the prediction for binding specificity includes a variable that describes a percentile rank of a predicted binding strength between the input TCR-pMHC pair, with respect to a background distribution including the predicted binding strength between each TCR included in a pool of 10,000 randomly sampled TCRs and a pMHC included in the input TCR-pMHC pair.
11 . A system for predicting T cell receptor (TCR) binding specificities comprising:
a memory including executable instructions; and
a processor configured to execute the executable instructions and cause the system to:
determine a set of MHC embeddings that encode neoantigen and major histocompatibility complex (MHC) data for a plurality of MHC proteins (pMHC);
determine a set of TCR embeddings that encode TCR data for a plurality of TCR sequences;
pre-train a prediction model on the set of MHC embeddings and the set of TCR embeddings;
train the prediction model using a differential learning schema that feeds a binding TCR-pMHC pair and a non-binding TCR-pMHC pair into the prediction model during each training cycle; and
determine a prediction for binding specificity of an input TCR-pMHC pair based on the prediction model.
12 . The system of claim 11 , wherein the processor is further configured to:
obtain a set of TCR-pMHC pairs that are experimentally validated as immunogenic, the set of TCR-pMHC pairs including the input TCR-pMHC pair; and
validate the prediction model by comparing the binding specificity prediction for the input TCR-pMHC pair to a known binding specificity for the input TCR-pMHC pair.
13 . The system of claim 11 , wherein the processor is further configured to:
determine a clonal expansion of a plurality of T cells, the clonal expansion including multiple TCR clones having known binding interactions with a set of pMHCs and a clone size for each of the multiple TCR clones;
determine a prediction for binding specificity between each of the multiple TCR clones and each of the pMHCs included in the set of pMHCs based on the prediction model; and
validate the prediction model by comparing the clone size for each of the TCR clones to the predicted binding specificity.
14 . The system of claim 11 , wherein each of the MHC embeddings includes a numeric representation of one or more pMHCs, and the processor is further configured to:
train a MHC numeric embedding layer on a MHC training dataset including textual representations of pMHCs; and
determine the numeric representation of the one or more pMHCs for each of the MHC embeddings based on the MHC numeric embedding layer.
15 . The system of claim 11 , wherein the MHC embeddings are determined using a multi-layer neural network that determines a probability that a particular pMHC molecule binds to one or more neo-antigen protein sequences.
16 . The system of claim 11 , wherein each of the TCR embeddings include a numeric representation of one or more TCR protein sequences, and the processor is further configured to:
train a TCR numeric embedding layer on a TCR training dataset including multiple training TCR protein sequences, the TCR training dataset including a structured data representation of one or more biochemical properties of multiple amino acids included in the training TCR protein sequences; and
determine the numeric representation of the one or more TCR protein sequences based on the TCR numeric embedding layer.
17 . The system of claim 16 , wherein the multiple amino acids are included in a complementary determining region (CDR) of the training TCR protein sequences, and the processor is further configured to:
manipulate the structured data representation to enable amino acids from multiple CDRs of the training TCR protein sequences to be added to the TCR training dataset.
18 . The system of claim 11 , wherein the TCR embeddings are determined using an auto-encoder that includes multiple encoder layers and multiple decoder layers.
19 . The system of claim 11 , wherein the processor is further configured to normalize the MHC embeddings and the TCR embeddings to enable the prediction model to be pre-trained on multiple classes of pMHC.
20 . The system of claim 11 , wherein the prediction for binding specificity includes a variable that describes a percentile rank of a predicted binding strength between the input TCR-pMHC pair, with respect to a background distribution including the predicted binding strength between each TCR included in a pool of 10,000 randomly sampled TCRs and a pMHC included in the input TCR-pMHC pair.