Current sequencing methods allow for detailed samples of T cell receptors (TCR) repertoires. To determine from a repertoire whether its host had been exposed to a target, computational tools that predict TCR-epitope binding are required. Currents tools are based on conserved motifs and are applied to peptides with many known binding TCRs. We employ new Natural Language Processing (NLP) based methods to predict whether any TCR and peptide bind. We combined large-scale TCR-peptide dictionaries with deep learning methods to produce ERGO (pEptide tcR matchinG predictiOn), a highly specific and generic TCR-peptide binding predictor. A set of standard tests are defined for the performance of peptide-TCR binding, including the detection of TCRs binding to a given peptide/antigen, choosing among a set of candidate peptides for a given TCR and determining whether any pair of TCR-peptide bind. ERGO reaches similar results to state of the art methods in these tests even when not trained specifically for each test. The software implementation and data sets are available at https://github.com/louzounlab/ERGO . ERGO is also available through a webserver at: http://tcr.cs.biu.ac.il/ .

One Sentence Summary: The combination of advanced tools from natural language processing and large-scale dictionaries of T cell receptors and their target peptide precisely predicts whether a T cell would bind a specific target. AbstractThe T cell repertoire is composed of T cell receptors (TCR) selected by their cognate MHCpeptides and naive TCR that do not bind known peptides. While the task of distinguishing a peptide-binding TCR from a naive TCR unlikely to bind any peptide can be performed using sequence motifs, distinguishing between TCRs binding different peptides requires more advanced methods. Such a prediction is the key for using TCR repertoires as disease-specific biomarkers. We here used large scale TCR-peptide dictionaries with state-of-the-art natural language processing (NLP) methods to produce ERGO (pEptide tcR matchinG predictiOn), a highly specific classifier to predict which TCR binds to which peptide. We successfully employed ERGO for two related tasks: discrimination between peptide binding and naive TCRs and the more complicated task of distinguishing between TCRs that bind different peptides. We show that ERGO significantly outperforms all current methods for classification of TCRs that bind peptides, but more importantly can distinguish the specific target of a TCR among a large set of peptides. The software implementation and data sets are available at: https://github.com/IdoSpringer/ERGO

Background Targeting epitopes derived from neo-antigens (or “neo-epitopes”) represents a promising immunotherapy approach with limited off-target effects. However, most peptides predicted using MHC binding prediction algorithms do not induce a CD8 + T cell response, and there is a crucial need to refine the predictions to readily identify the best antigens that could mediate T-cell responses. Such a response requires a high enough number of epitopes bound to the target MHC. This number is correlated with both the peptide-MHC binding affinity and the number of peptides reaching the ER. Beyond this, the response may be affected by the properties of the neo-epitope mutated residues. Methods Herein, we analyzed several experimental datasets from cancer patients to elaborate better predictive algorithms for T-cell reactivity to neo-epitopes. Results Indeed, potent classifiers for epitopes derived from neo-antigens in melanoma and other tumors can be developed based on biochemical properties of the mutated residue, the antigen expression level and the peptide processing stage. Among MHC binding peptides, the present classifiers can remove half of the peptides falsely predicted to activate T cells while maintaining the absolute majority of reactive peptides. Conclusions The classifier properties further highlight the contribution of the quantity of peptides reaching the ER and the mutation type to CD8 + T cell responses. These classifiers were then validated on neo-antigens obtained from other datasets, confirming the validity of our prediction. Algorithm Availability: http://peptibase.cs.biu.ac.il/Tcell_predictor/ or by request from the authors as a standalone code.

Epitopes presented on MHC class I molecules pass multiple processing stages before their presentation on MHC molecules, the main ones being proteasomal cleavage and TAP binding. Transporter associated with antigen processing (TAP) binding is a necessary stage for most, but not all, MHC-I-binding peptides. The molecular determinants of TAP-binding peptides can be experimentally estimated from binding experiments and from the properties of peptides inducing a CD8 T cell response. We here propose novel optimization formalisms to combine binding and activation experimental results to produce a classifier for TAP binding using dual-output kernel and deep learning approaches. The application of these algorithms to the human and murine TAP binding leads to predictors that are much more precise than current state of the art methods. Moreover, the computed score is highly correlated with the observed binding energy. The new predictors show that TAP binding may be much more selective than previously assumed in humans and mice and sensitive to the properties of most positions of the peptides. Beyond the improved precision for TAP binding, we propose that the same approach holds in most molecular binding problems, where functional and binding measures are simultaneously available, and can be used to significantly improve the precision of binding prediction algorithms in general and immune system molecules specifically.