TAPIR: a T-cell receptor language model for predicting rare and novel targets.

T-cell receptors (TCRs) are involved in most human diseases, but linking their sequences with their targets remains an unsolved grand challenge in the field. In this study, we present TAPIR (T-cell receptor and Peptide Interaction Recognizer), a T-cell receptor (TCR) language model that predicts TCR-target interactions, with a focus on novel and rare targets. TAPIR employs deep convolutional neural network (CNN) encoders to process TCR and target sequences across flexible representations (e.g., beta-chain only, unknown MHC allele, etc.) and learns patterns of interactivity via several training tasks. This flexibility allows TAPIR to train on more than 50k either paired (alpha and beta chain) or unpaired TCRs (just alpha or beta chain) from public and proprietary databases against 1933 unique targets. TAPIR demonstrates state-of-the-art performance when predicting TCR interactivity against common benchmark targets and is the first method to demonstrate strong performance when predicting TCR interactivity against novel targets, where no examples are provided in training. TAPIR is also capable of predicting TCR interaction against MHC alleles in the absence of target information. Leveraging these capabilities, we apply TAPIR to cancer patient TCR repertoires and identify and validate a novel and potent anti-cancer T-cell receptor against a shared cancer neoantigen target (PIK3CA H1047L). We further show how TAPIR, when extended with a generative neural network, is capable of directly designing T-cell receptor sequences that interact with a target of interest.

Predicting HLA class II antigen presentation through integrated deep learning.

Accurate prediction of antigen presentation by human leukocyte antigen (HLA) class II molecules would be valuable for vaccine development and cancer immunotherapies. Current computational methods trained on in vitro binding data are limited by insufficient training data and algorithmic constraints. Here we describe MARIA (major histocompatibility complex analysis with recurrent integrated architecture; https://maria.stanford.edu/ ), a multimodal recurrent neural network for predicting the likelihood of antigen presentation from a gene of interest in the context of specific HLA class II alleles. In addition to in vitro binding measurements, MARIA is trained on peptide HLA ligand sequences identified by mass spectrometry, expression levels of antigen genes and protease cleavage signatures. Because it leverages these diverse training data and our improved machine learning framework, MARIA (area under the curve = 0.89-0.92) outperformed existing methods in validation datasets. Across independent cancer neoantigen studies, peptides with high MARIA scores are more likely to elicit strong CD4+ T cell responses. MARIA allows identification of immunogenic epitopes in diverse cancers and autoimmune disease.