TCRMatch: Predicting T-cell receptor specificity based on sequence similarity to previously characterized receptors

Abstract

The adaptive immune system in vertebrates has evolved to recognize non-self-antigens, such as proteins expressed by infectious agents and mutated cancer cells. T cells play an important role in antigen recognition by expressing a diverse repertoire of antigen-specific receptors, which bind epitopes to mount targeted immune responses. Recent advances in high-throughput sequencing have enabled the routine generation of T-cell receptor (TCR) repertoire data. Identifying the specific epitopes targeted by different TCRs in these data would be valuable. To accomplish that, we took advantage of the ever-increasing number of TCRs with known epitope specificity curated in the Immune Epitope Database (IEDB) since 2004. We compared six metrics of sequence similarity to determine their power to predict if two TCRs have the same epitope specificity. We found that a comprehensive k-mer matching approach produced the best results, which we have implemented into TCRMatch, an openly accessible tool (http://tools.iedb.org/tcrmatch/) that takes TCR β-chain CDR3 sequences as an input, identifies TCRs with a match in the IEDB, and reports the specificity of each match. We anticipate that this tool will provide new insights into T cell responses captured in receptor repertoire and single cell sequencing experiments and will facilitate the development of new strategies for monitoring and treatment of infectious, allergic, and autoimmune diseases, as well as cancer.

NetTCR-2.0 enables accurate prediction of TCR-peptide binding by using paired TCRα and β sequence data

“…This model is a variant of the model proposed earlier by us for pan-specific prediction of kinase-specific phosphorylation 25 . This model significantly outperformed simpler sequence-based models implemented using the TCRMatch 22 and TCRdist 7 frameworks.…”

“…Given the relatively limited length variation of the CDR3 sequences included in the current work (90% of the paired CDR3α and CDR3β sequences from the paired data set have a length in the range of 9-13 amino acids), this shortcoming does not have large impacts for the current work. However, it will be essential to consider alternative and less length-biased approaches, such as, for instance, the kernel similarity method underlying TCRmatch 22 , if the work is extended to cover full-length TCRs and/or include the complete set of CDR sequences.…”

“…A baseline model was designed to establish the predictive power of simple similarity-based methods. The similarity-scoring approach used in the baseline model was the kernel-scoring method introduced by Shen et al 36 with default parameters, as described earlier in the MAIT Match 19 and TCRMatch 22 methods. In the model, the prediction score for a given TCR is calculated as the highest score obtained when scoring the CDR3β against a database of positive CDR3βs.…”

“…Most are constructed based on data from the IEDB, VDJdb, and/or McPAS-TCR and, in addition to the epitope information, make use of either CDR3β sequences alone [13][14][15] , a mixture of CDR3α and CDR3β sequences 16 , or smaller data sets entailing all 6 CDR3 sequences and potentially additional cellular information 17,18 . Methodologically, the different studies range from simple CDR3β alignmentbased methods 19,22 , over CDR similarity-weighted distances such as TCRdist 7 , k-mer feature spaces in combination with PCA and decision trees (SETE 13 ), random forests 20,21 such as TCRex 23 , CNN-based (ImRex) 16 , and Gaussian process classification methods (TCRGP 17 ), to more complex approaches integrating natural language processing (NLP) methods (ERGO 14 ). The overall conclusion from these earlier works is that while the prediction of TCR specificity is feasible, the volume and accuracy of current data limit the performance of the developed models.…”

TCR-BERT: learning the grammar of T-cell receptors for flexible antigen-xbinding analyses

“…Recently, the growing popularity and accessibility of sequencing technologies has enabled highthroughput profiling of TCR sequences 24 25 , which has in turn empowered numerous computational methods studying and predicting TCR-antigen binding. Conventional approaches like GLIPH 15 26 , TCRMatch 27 , and TCRdist 28 typically rely on sequence motif comparisons and manually engineered heuristics to predict which TCRs are likely to share common antigen binding partners. More recently, researchers have applied various machine learning methods to predicting TCR-antigen binding.…”