Progress in structure prediction and design of adaptive immune receptors.
TL;DR
Recent progress in structure-based deep learning approaches for AIRs are reviewed, including AIR-specific language models, structure prediction, and epitope-conditioned design, and emerging generative design strategies such as inverse folding, diffusion-based backbone generation, sequence-space diffusion, and sequence-structure co-design are reviewed.
Abstract
Adaptive immune receptors (AIRs), including antibodies and T-cell receptors (TCRs), mediate antigen recognition and represent a major class of therapeutic biomolecules. Their unique architecture, combining conserved framework with highly diverse complementarity-determining regions (CDRs), poses challenges for structure prediction and design. Recent advances in deep learning have transformed these fields, yet AIR-antigen interactions remain difficult targets due to limited structural data, weak co-evolutionary signals, and conformational heterogeneity. Herein, we review recent progress in structure-based deep learning approaches for AIRs, including AIR-specific language models, structure prediction, and epitope-conditioned design. We discuss commonly used datasets, evaluation metrics, and sources of bias that complicate cross-study comparisons and highlight the need for improved benchmarks. We also review emerging generative design strategies such as inverse folding, diffusion-based backbone generation, sequence-space diffusion, and sequence-structure co-design, and outline key challenges, including accurate modeling of flexible CDR loops, reliable ranking of AIR-antigen complexes, and scalable epitope-specific AIR design.