Predicting antigen-specificity of single T-cells based on TCR CDR3 regions

Abstract

It has recently become possible to assay T-cell specificity with respect to large sets of antigens as well as T-cell receptor sequence in high-throughput single-cell experiments. We propose multiple sequence-data specific deep learning approaches to impute TCR to epitope specificity to reduce the complexity of new experiments. We found that models that treat antigens as categorical variables outperform those which model the TCR and epitope sequence jointly. Moreover, we show that variability in single-cell immune repertoire screens can be mitigated by modeling cell-specific covariates.Antigen recognition is one of the key factors of T-cell-mediated immunity. The ability to accurately predict T-cell activation upon epitope recognition would have transformative effects on many research areas from in infectious disease, autoimmunity, vaccine design, and cancer immunology, but has been thwarted by lack of training data and adequate models. Although tremendous effort has been spent on elucidating the common rules that govern the TCR-pMHC interaction, it still remains elusive. The T-cell receptor (TCR) interacts with peptides immobilized on MHC multimers (pMHC) through its three complementarity determining region (CDR) loops of the ɑ-and β-chain. The hypervariable loops CDR3ɑ and CDR3β are most commonly aligned with the presented epitope 1 and are hypothesized to be the main driver of T-cell specificity 2 . Due to lack of sufficient data, previous models for T-cell specificity were only based on the CDR3β loop 3,4 , 5 .In this study, we exploit a newly developed single-cell technology that enables the simultaneous sequencing of the paired TCR ɑ-and β-chain while determining the T-cell specificity to train multiple deep learning architectures modeling the TCR-pMHC interaction including both chains. The models include single-cell specific covariates accounting for the variability found in such data, thereby fully exploit the multiplicity of observations that can be easily sampled in single-cell screens. We show that models that include both ɑ-and β-chain

Current challenges for epitope-agnostic TCR interaction prediction and a new perspective derived from image classification

“…The ROC and PR plots for both CNNs are shown in Figure 6. Both models perform poorly under this epitope-agnostic setting (mean AUC 0.58 ± 0.05SD, 0.45 ± 0.06SD; average precision 0.57 ± 0.04SD, 0.44 ± 0.04SD), which is also in line with previous studies (15,20). However, we do find some indication that the interaction map approach is able to capture at least a small additional part of the signal in the data, in contrast to the dual input CNN.…”

“…The overall performance is also comparable to previously pro- posed general (i.e. not epitope-specific) TCR-epitope recognition models (15,20). This suggests that the feature encoding used to create the interaction maps can provide the same information to a network as the standard approach, and even make it more tractable for the network to learn from it.…”

“…Despite the advent of single cell TCR paired chain sequencing, currently available TCR-epitope binding data mostly consists of single-chain, often beta-chain, data, while in reality both the alpha-and beta-chains are thought to contribute to binding specificity. Indeed, recent studies have concluded that utilising both chains could increase the accuracy of predictive TCR-epitope models (14,15).…”

“…However, recent studies have shown that while generic TCR-epitope prediction models can achieve a reasonable performance on interactions that involve epitopes already encountered training, they cannot yet reliable extrapolate to novel epitopes (i.e. epitope-agnostic models) (15,20). This is likely caused by the large heterogeneity in potential TCR and epitope sequences, and the limited amount of readily available training data that covers this sequence space.…”

“…Using this type of interaction map as the model input differs from the prevailing approach seen in sequence-based prediction models, not only in the field of TCR-epitope prediction, but also for the modeling of other molecular interactions. The interacting molecules are usually supplied to a model as separate or concatenated inputs (1, 3,5,6,15,18,20,25,26). Those types of models need to learn an internal representation for each molecule separately, before being combined again in deeper layers.…”