pan-MHC and cross-Species Prediction of T Cell Receptor-Antigen Binding

Yi Han, Yuqiu Yang, Yanhua Tian, Farjana Fattah, Mitchell S. von Itzstein, Yifei Hu, Minying Zhang, Xiongbin Kang, Donghan M. Yang, Jialiang Liu, Yaming Xue, Chaoying Liang, Indu Raman, Chengsong Zhu, Olivia Xiao, Jonathan E. Dowell, Jade Homsi, Sawsan Rashdan, Shengjie Yang, Mary Gwin, David Hsiehchen, Yvonne Gloria-McCutchen, Ke Pan, Fangjiang Wu, Don L. Gibbons, Xinlei Wang, Cassian Yee, Junzhou Huang, Alexandre Reuben, Chao Cheng, Jianjun Zhang, David E. Gerber, Tao WangbioRxiv 2023 · DOI

TL;DR

This paper introduces pMTnet-omni, a novel AI system that accurately predicts T cell receptor (TCR) binding to peptide-MHC (pMHC) complexes, a crucial but challenging problem in immunology. By integrating sequence and structural information, pMTnet-omni achieves state-of-the-art performance and shows promise for clinical applications like predicting adverse drug reactions and designing immunotherapies.

What Problem Does This Paper Address?

Understanding how T cell receptors (TCRs) recognize specific peptide-MHC (pMHC) complexes is fundamental to immunology. This interaction is the basis for adaptive immunity, where T cells identify and eliminate threats like pathogens and cancer cells. However, experimentally determining these TCR-pMHC pairings is incredibly time-consuming, expensive, and low-throughput. This bottleneck limits our ability to understand immune responses, develop effective immunotherapies (like TCR-based cancer treatments), and predict adverse drug reactions. Existing computational methods often struggle with accuracy, limited scope (e.g., only human TCRs or specific MHC classes), and don't fully leverage structural information.

Methods

The core innovation is pMTnet-omni, a hybrid AI system that combines protein sequence and structural information. It employs a "conquer-by-division" strategy:

  • pMHC Encoder: A structure-aware model (using ESM2 embeddings) for MHCs and a sequence-based model for peptides. This part is trained to predict peptide-MHC binding itself, achieving performance comparable to existing tools like NetMHCpan. Crucially, it leverages large-scale information sharing across MHC class I and II, and human and mouse data.
  • TCR Encoders: Four separate variational autoencoders (VAEs) are used to encode TCRs: one for the V gene and one for the CDR3 region, for both the alpha (α) and beta (β) chains. This is motivated by the distinct roles of V genes (binding MHC) and CDR3 regions (binding peptide) and their differing sequence diversity.
  • Head Module: These encoded representations are then fed into a final module that predicts the TCR-pMHC binding affinity. This module is trained using contrastive learning, where true binding pairs are distinguished from randomly sampled non-binding pairs. The output is a percentile ranking (rank%) score, where lower scores indicate stronger predicted binding.

Datasets

The study utilized a comprehensive collection of TCR-pMHC pairing data:

  • Training Data: 168,484 human/mouse class I/II TCR-pMHC pairs curated from 69 public datasets.
  • Validation Data: A separate set of 50 datasets, ensuring no overlap with the training data.
  • Additional Datasets:
    • Mouse tumor progression scRNA/TCR-seq data (344SQ/129Sv mouse model).
  • ICI irAE patient cohort data.
  • COVID-19 vaccination data.
  • Visium-TCR-seq data from a melanoma patient.
  • 10X scRNA-seq/scTCR-seq datasets with antigen affinity information.

The authors also used existing datasets for training individual components, such as the IEDB database for pMHC binding and IMGT for TCR V gene sequences.

Key Insights and Ideas

  • Hybrid Approach is Key: Combining sequence and structural information for pMHCs, while using sequence for TCRs (given their inherent flexibility and somatic hypermutation), is a powerful strategy that outperforms purely sequence- or structure-based methods.
  • Information Sharing is Crucial: Pooling data across MHC classes (I and II) and species (human and mouse) for the pMHC encoder significantly improves performance, especially for underrepresented classes like MHC class II and mouse pairs.
  • CDR3 Beta is Dominant, but Alpha Matters: Masking out TCR alpha chains reduced performance more significantly than masking beta chains, highlighting the critical role of the beta chain but also the contribution of the alpha chain to TCR-pMHC binding.
  • Distinguishing Binding Affinity: pMTnet-omni can not only predict binding but also differentiate between strong and weak binders, even for TCRs with very similar sequences, suggesting it captures subtle differences in binding affinity.
  • Clinical Relevance Demonstrated: The model's ability to predict immune-related adverse events (irAEs) in cancer patients and track TCR repertoire changes during vaccination and tumor progression underscores its potential for real-world clinical applications.

Key Results

  • High Accuracy: pMTnet-omni achieved an overall AUROC of 0.888 and AUPRC of 0.786 on an independent validation cohort, surpassing existing benchmark software.
  • Cross-Species and Cross-MHC Class Performance: The model performed comparably for human (AUROC=0.888) and mouse (AUROC=0.897) TCR-pMHC pairs, and for Class I (AUROC=0.895) and Class II (AUROC=0.876) MHCs.
  • Distinguishing Similar TCRs: The model showed high accuracy in distinguishing stronger vs. weaker binders among TCR pairs with only 1-2 amino acid differences.
  • Biomarker for irAEs: A pMTnet-omni-based risk score was found to be predictive of immune-related adverse events in a cohort of cancer patients treated with ICIs.
  • Monitoring Immune Responses: The model successfully characterized TCR repertoire evolution in response to SARS-CoV-2 infection and vaccination, and spatial heterogeneity of TCR-antigen interactions within tumors.

Limitations and Future Work

  • Limitations:
    • The model's performance is dependent on the quality and quantity of available training data.
  • While it integrates structural information for pMHCs, TCR structures are not directly modeled due to their flexibility and the scarcity of solved structures.
  • The model does not account for all non-TCR factors that can influence binding, such as TCR expression levels, clustering, or co-stimulatory molecules, which can vary with biological conditions.
  • Future Work:
    • Exploring alternative model designs, potentially incorporating structural information for TCRs or peptides more directly.
  • Developing ensemble methods ("wisdom of the crowds") by combining various sequence- and structure-based models for even better prediction.
  • Fine-tuning the general pMTnet-omni model with smaller, condition-specific datasets to adapt it to particular biological contexts.
  • Further development of TCR-immunotherapeutic design and diagnostic tools based on pMTnet-omni predictions.

Summary updated

Abstract

SUMMARY Profiling the binding of T cell receptors (TCRs) of T cells to antigenic peptides presented by MHC proteins is one of the most important unsolved problems in modern immunology. Experimental methods to probe TCR-antigen interactions are slow, labor-intensive, costly, and yield moderate throughput. To address this problem, we developed pMTnet-omni, an Artificial Intelligence (AI) system based on hybrid protein sequence and structure information, to predict the pairing of TCRs of αβ T cells with peptide-MHC complexes (pMHCs). pMTnet-omni is capable of handling peptides presented by both class I and II pMHCs, and capable of handling both human and mouse TCR-pMHC pairs, through information sharing enabled this hybrid design. pMTnet-omni achieves a high overall Area Under the Curve of Receiver Operator Characteristics (AUROC) of 0.888, which surpasses competing tools by a large margin. We showed that pMTnet-omni can distinguish binding affinity of TCRs with similar sequences. Across a range of datasets from various biological contexts, pMTnet-omni characterized the longitudinal evolution and spatial heterogeneity of TCR-pMHC interactions and their functional impact. We successfully developed a biomarker based on pMTnet-omni for predicting immune-related adverse events of immune checkpoint inhibitor (ICI) treatment in a cohort of 57 ICI-treated patients. pMTnet-omni represents a major advance towards developing a clinically usable AI system for TCR-pMHC pairing prediction that can aid the design and implementation of TCR-based immunotherapeutics.
BibTeX
@misc{han2023pan,
  author = {Yi Han and Yuqiu Yang and Yanhua Tian and Farjana Fattah and Mitchell S. von Itzstein and Yifei Hu and Minying Zhang and Xiongbin Kang and Donghan M. Yang and Jialiang Liu and Yaming Xue and Chaoying Liang and Indu Raman and Chengsong Zhu and Olivia Xiao and Jonathan E. Dowell and Jade Homsi and Sawsan Rashdan and Shengjie Yang and Mary Gwin and David Hsiehchen and Yvonne Gloria-McCutchen and Ke Pan and Fangjiang Wu and Don L. Gibbons and Xinlei Wang and Cassian Yee and Junzhou Huang and Alexandre Reuben and Chao Cheng and Jianjun Zhang and David E. Gerber and Tao Wang},
  title = {{pan-MHC and cross-Species Prediction of T Cell Receptor-Antigen Binding}},
  year = {2023},
  doi = {10.1101/2023.12.01.569599},
  howpublished = {bioRxiv},
}
Figure