SwarmTCR: a computational approach to predict the specificity of T Cell Receptors

Abstract

Motivation: Computationally predicting the specificity of T cell receptors can be a powerful tool to shed light on the immune response against infectious diseases and cancers, autoimmunity, cancer immunotherapy, and immunopathology. With more T cell receptor sequence data becoming available, the need for bioinformatics approaches to tackle this problem is even more pressing. Here we present SwarmTCR, a method that uses labeled sequence data to predict the specificity of T cell receptors using a nearest-neighbor approach. SwarmTCR works by optimizing the weights of the individual CDR regions to maximize classification performance. Results: We compared the performance of SwarmTCR against a state-of-the-art method (TCRdist) and showed that SwarmTCR performed significantly better on epitopes EBV-BRLF1, NS4B with single cell data and epitopes EBV-BRLF1, IAV-M1 with bulk sequencing data (alpha and beta chains). In addition, we show that the weights returned by SwarmTCR are biologically interpretable. Availability: SwarmTCR is distributed freely under the terms of the GPL-3 license. The source code and all sequencing data are available at GitHub (https://github.com/thecodingdoc/SwarmTCR).

Stitchr: stitching coding TCR nucleotide sequences from V/J/CDR3 information

“…TCRs proposed to enter pre-clinical testing require empirical validation of their specificities, due to effects like bystander activation (55), non-specific MHC multimer reagent staining (56,57), and cross-reactivity (58,59). Moreover several bioinformatic strategies to predict antigen specificity from sequence make use of information encoded at regions outside those typically sequenced in CDR3-centric protocols (60)(61)(62).…”

Challenges in neoantigen-directed therapeutics

“…This shows that modelling pMHC-TCR interactions should be possible. Standalone tools implementing such pMHC-TCR prediction models using known pMHC-TCR complexes include NetTCR(2), TCRex, ATMTCR, TCRdist, SwarmTCR, SETE and ERGO(-II) [214][215][216][217][218][219][220][221] . Other tools and pipelines model these interactions through machine learning by using confirmed immunogenic and non-immunogenic peptides as positive and negative training sets [214][215][216]222 .…”