DecoderTCR: Compositional Pretraining and Entropy-Guided Decoding for TCR-pMHC Interactions.

Area of Science:

• Computational immunology
• Bioinformatics
• Machine learning for immunology

Background:

• Modeling T-cell receptor (TCR) and peptide-MHC (pMHC) interactions is crucial but challenging due to limited paired data.
• Unpaired TCR and pMHC sequence data are abundant, presenting an opportunity for novel modeling approaches.

Purpose of the Study:

• To introduce DecoderTCR, a masked language model framework for TCR-pMHC recognition modeling.
• To address data sparsity by leveraging both paired and unpaired sequence data.
• To improve zero-shot prediction capabilities in TCR-pMHC binding and epitope-specific recognition.

Main Methods:

• Implemented a compositional continual pre-training curriculum using marginal data before refining cross-chain dependencies.
• Developed Iterative Entropy-Guided Refinement (IEGR), a non-autoregressive decoding algorithm for efficient context resolution.
• Utilized masked language modeling for learning representations from sequence data.

Main Results:

• Achieved 0.96 AUROC for zero-shot pMHC binding prediction.
• Attained 0.76 AUROC for epitope-specific TCR recognition, nearing supervised baseline performance without epitope-specific training.
• Learned representations that recover structural contacts without coordinate supervision and generated sequences with realistic recombination statistics.

Conclusions:

• DecoderTCR effectively models TCR-pMHC interactions and achieves high predictive performance with sparse data.
• A prediction-generation gap exists, indicating that while discrimination is strong, reliable sequence generation remains an open challenge.
• The framework demonstrates the potential of masked language models in computational immunology for understanding immune recognition.