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

Modeling recognition between T-cell receptors (TCRs) and peptide-MHC (pMHC) complexes is a fundamental challenge in computational immunology, constrained by sparse paired interaction data relative to abundant unpaired sequences. We introduce DecoderTCR, a masked language model framework that addresses this through two contributions: (1) a compositional continual pre-training curriculum that learns component representations from marginal data before refining cross-chain dependencies, and (2) Iterative Entropy-Guided Refinement (IEGR), a non-autoregressive decoding algorithm that resolves high-confidence positions first to provide context for uncertain regions. On held-out benchmarks, DecoderTCR achieves 0.96 AUROC for zero-shot pMHC binding prediction and 0.76 AUROC for epitope-specific TCR recognition, approaching supervised baselines without epitope-specific training. Learned representations recover structural contacts without coordinate supervision, and generated sequences exhibit realistic recombination statistics. Experimental validation across two rounds of wet-lab screening reveals a prediction-generation gap that can be narrowed via a lab-in-the-loop paradigm for TCR design.

Our immune system relies on molecules called T-cell receptors to identify and fight viruses, bacteria, and cancer. Predicting which receptors recognize a specific disease target is critical for developing new immunotherapies. However, researchers face a major data bottleneck: we have cataloged millions of individual molecules but possess very few confirmed matches between them. To address this, we built DecoderTCR, an AI model that first learns general biological rules from vast amounts of isolated data before focusing on the rare confirmed pairs. When designing a new receptor, our model anchors the most certain parts of the molecule first and uses them to guide the highly variable regions. This allows DecoderTCR to predict immune recognition far more accurately than general-purpose protein language models, even pinpointing exactly where molecules physically touch without ever seeing a 3D structure. Furthermore, testing AI-generated designs in the laboratory and feeding the results back into the model quadrupled our success rate from 3% to 12%. This establishes a practical cycle where each physical experiment teaches our AI model to design better, significantly accelerating the discovery of targeted treatments.