EpicPred: Predicting phenotypes driven by epitope binding TCRs using attention-based multiple instance learning.

Abstract

Motivation: Correctly identifying epitope binding TCRs is important to both understand their underlying biological mechanism in association to some phenotype and accordingly develop T-cell mediated immunotherapy treatments. Although the importance of the CDR3 region in TCRs for epitope recognition is well recognized, methods for profiling their interactions in association to a certain disease or phenotype remains less studied. We developed EpicPred to identify phenotype specific TCR-epitope interactions. EpicPred first predicts and removes unlikely TCR-epitope interactions to reduce false positives using the Open-set Recognition. Subsequently, multiple instance learning was used to identify TCR-epitope interactions specific to a cancer type or severity levels of COVID-19 patients.

Results: From six public TCR databases, 244,552 TCR sequences and 105 unique epitopes were used to predict epitope binding TCRs and to filter out non-epitope binding TCRs using the open-set recognition method. The predicted interactions were used to further predict the phenotype groups in two cancer and four COVID-19 TCR-seq datasets of both bulk and single-cell resolution. EpicPred outperformed the competing methods in predicting the phenotypes, achieving an average AUROC of 0.80 ± 0.07.

Availability and implementation: The EpicPred Software is available at https://github.com/jaeminjj/EpicPred.

Supplementary information: Supplementary data are available at Bioinformatics online.