Predicting recognition between T cell receptors and epitopes with TCRGP

Emmi Jokinen, Jani Huuhtanen, Satu Mustjoki, Markus Heinonen, Harri LähdesmäkiPLOS Computational Biology 2021 · OpenReview

Abstract

Humans have dedicated cells called T cells that recognize potentially harmful invaders using T cell receptors (TCRs) and protect us from infections. Each person has billions of T cells and millions of different kinds of TCRs that enable us to recognize a large variety of invaders. When for example a virus enters the body, T cells with TCRs recognizing that virus multiply and fight it. After such attacks, T cells that recognized the virus, remain in the body, forming an immunological memory. T cells thus characterize a person’s infection history, but in order to reveal it, we need to determine which epitopes, molecular signatures of the invaders, different TCRs can recognize. We can characterize which TCRs recognize certain epitopes experimentally, but this is a time and sample consuming task and is not often possible with scarce patient samples such as biopsies. However, previously produced experimental data has enabled us to develop a computational method, TCRGP, that can predict which epitopes a TCR recognizes with a higher accuracy than previous methods. As a computational method, TCRGP spares resources from time and sample consuming experimental validations and offers an interesting way to analyze publicly available TCR data.

Introduction

The adaptive immune system relies on T cell receptors (TCRs) to identify molecular signatures of pathogens, known as epitopes. Because the variety of possible TCRs is astronomical (estimated at $10^{18}$), determining which TCR recognizes which epitope is a monumental task. While experimental methods like pMHC-multimer assays can identify these pairings, they are time-consuming, expensive, and often impossible to perform on small clinical samples like biopsies.

TCRGP is a novel Gaussian process-based framework designed to address this gap. By utilizing multiple structural regions of the TCR and a robust probabilistic approach, it predicts whether a TCR will recognize a specific target antigen. TCRGP outperforms existing state-of-the-art methods and enables "digital sorting" of antigen-specific T cells from complex single-cell sequencing data.

Methods

The authors developed TCRGP (T Cell Receptor Gaussian Process), which stands out through several key technical choices:

  • Gaussian Process (GP) Classification: Unlike standard neural networks that provide a "point estimate," GPs are non-parametric Bayesian models. They learn a distribution over possible functions, allowing the model to be robust against overfitting—a critical feature when dealing with the small datasets typical of rare epitopes.
  • Multiple Kernel Learning: TCRGP creates separate "kernels" (similarity measures) for different Complementarity Determining Regions (CDRs): CDR1, CDR2, CDR2.5, and CDR3 for both the $\alpha$ and $\beta$ chains. The model automatically learns which of these regions are most important for recognizing a specific epitope.
  • Alignment-Based Encoding: To handle sequences of different lengths, the authors used IMGT definitions to align sequences and then encoded amino acids using principal components of a modified BLOSUM62 substitution matrix.
  • Scalability: To ensure the model could handle millions of sequences, they implemented Sparse Variational Inference, allowing the GP to scale efficiently.
Overview of the TCRGP pipeline, from sequence alignment to probabilistic prediction.

Datasets and Evaluation

The model was rigorously evaluated using several high-quality datasets:

  • Dash Data: A benchmark set of TCR$\alpha\beta$ sequences for 10 epitopes.
  • VDJdb: A large, curated database of TCR sequences with known antigen specificity (22 epitopes used).
  • Zheng Data: Single-cell RNA + TCR$\alpha\beta$ sequencing from liver cancer patients.
  • Cheng Data: HBV-reactive T cell populations used to train specific HBV classifiers.

Key Insights and Results

TCRGP demonstrated superior accuracy, outperforming the previous state-of-the-art method, TCRdist, in 57 out of 60 comparisons on the Dash dataset. It achieved a mean AUROC of 0.913 on the Dash data and 0.863 on the VDJdb data.

The Importance of the "Whole" Receptor

While the CDR3 region is the primary driver of specificity, TCRGP showed that including CDR1, CDR2, and the "loop" region (CDR2.5) significantly boosts prediction accuracy. For some epitopes, these "germline" regions were actually more informative than the hypervariable CDR3.

Probabilistic Uncertainty and In Silico Sorting

Because TCRGP is Bayesian, it outputs a probability score (0 to 1). This allows researchers to set high-confidence thresholds. The authors demonstrated that TCRGP can "digitally sort" T cells from single-cell RNA-seq data. In a liver cancer case study, TCRGP identified HBV-specific T cells and revealed they were predominantly in an "exhausted" state, characterized by high expression of inhibitory markers like TIGIT and TIM-3.

Limitations and Future Work

The model can struggle to distinguish between highly similar epitopes due to natural TCR cross-reactivity. Additionally, while TCRGP handles small data well, its performance still scales with data volume; accuracy for very rare epitopes (fewer than 40-50 sequences) remains a challenge. Future work may involve incorporating MHC information to improve predictions across different patient HLA types.

Summary updated

Abstract

Author summary Humans have dedicated cells called T cells that recognize potentially harmful invaders using T cell receptors (TCRs) and protect us from infections. Each person has billions of T cells and millions of different kinds of TCRs that enable us to recognize a large variety of invaders. When for example a virus enters the body, T cells with TCRs recognizing that virus multiply and fight it. After such attacks, T cells that recognized the virus, remain in the body, forming an immunological memory. T cells thus characterize a person’s infection history, but in order to reveal it, we need to determine which epitopes, molecular signatures of the invaders, different TCRs can recognize. We can characterize which TCRs recognize certain epitopes experimentally, but this is a time and sample consuming task and is not often possible with scarce patient samples such as biopsies. However, previously produced experimental data has enabled us to develop a computational method, TCRGP, that can predict which epitopes a TCR recognizes with a higher accuracy than previous methods. As a computational method, TCRGP spares resources from time and sample consuming experimental validations and offers an interesting way to analyze publicly available TCR data.
BibTeX
@article{jokinen2021predicting,
  author = {Emmi Jokinen and Jani Huuhtanen and Satu Mustjoki and Markus Heinonen and Harri Lähdesmäki},
  title = {{Predicting recognition between T cell receptors and epitopes with TCRGP}},
  journal = {PLOS Computational Biology},
  year = {2021},
  url = {https://openreview.net/forum?id=pddrs3654b},
}
Figure