AI Developments for T and B Cell Receptor Modeling and Therapeutic Design

TL;DR

The chapter surveys how AI advances—especially protein language models, multimodal learning, and generative design—are transforming TCR and BCR modeling for immunoepidemiology and therapeutic design. It highlights data resources (AIRR, OAS, VDJdb, ImmuneCODE) and structure-aware approaches that combine sequence with 3D information to improve prediction of binding, developability, and immunogenicity, while detailing generative frameworks (diffusion, autoregressive, flow-based) and multi-objective optimization for designing immune receptors. It also emphasizes integrative, multi-omics and single-cell contexts to capture receptor function within cellular states, and discusses transfer learning as a path to robust generalization amid data scarcity. Despite progress, the text notes data sparsity, label noise, interpretability, and translational bottlenecks, arguing for immune-focused foundation models, physics-informed AI, standardized benchmarks, and regulatory-ready pipelines to realize clinically actionable receptor engineering. Overall, the work outlines a roadmap where predictive and design-oriented AI co-evolve with experimental validation to accelerate immune therapeutics and vaccines in a safe, scalable manner.

Abstract

Artificial intelligence (AI) is accelerating progress in modeling T and B cell receptors by enabling predictive and generative frameworks grounded in sequence data and immune context. This chapter surveys recent advances in the use of protein language models, machine learning, and multimodal integration for immune receptor modeling. We highlight emerging strategies to leverage single-cell and repertoire-scale datasets, and optimize immune receptor candidates for therapeutic design. These developments point toward a new generation of data-efficient, generalizable, and clinically relevant models that better capture the diversity and complexity of adaptive immunity.

Figures (7)

• Figure 1: Masked language modeling architecture for immune receptor representation. Antibody and TCR sequences are tokenized, with selected positions masked, and then processed through transformer encoder layers from a PLM that learn contextualized embeddings via self-attention. Through the neural network's feedforward modules, the embeddings are further processed for downstream analyses, such as visualization in a lower-dimensional space. Pre-training on large repertoire databases (OAS for antibodies; VDJdb, ImmuneCODE for TCRs) enables the model to predict masked residues and capture sequence patterns relevant to antigen recognition. The resulting embeddings support downstream tasks including binding prediction, specificity modeling, and repertoire analysis.
• Figure 2: Conversion of protein structures to graph representations for geometric deep learning using Graphein jamasb2022graphein. TCR-p-MHC complex structures (1AO7) from the Protein Data Bank (PDB) berman2000protein can be transformed into molecular graphs where nodes represent atoms or residues and edges encode spatial proximity or chemical bonds, enabling graph neural networks to learn interaction patterns.
• Figure 3: Structural diversity of antibodies targeting the SARS-CoV-2 spike protein (light blue). (Left) unbound spike protein; (Right) multiple antibodies with distinct sequences (shown in rainbow colors) recognize different epitopes across the spike, including sites in the receptor-binding domain (RBD), N-terminal domain (NTD), and S2 stalk region, thereby contributing to the breadth of neutralizing immune responses.
• Figure 4: AI framework for TCR-epitope specificity prediction using bimodal attention networks. TCRs and epitopes are encoded separately and then matched through attention mechanisms to predict binding probability, enabling repertoire-scale specificity screening.
• Figure 5: Transition from predictive to generative language models in immune receptor design. (Predictive): A protein language model (PLM) encodes an input sequence into embeddings, which are then used for downstream tasks such as classification (e.g., binding vs. non-binding) or regression (e.g., affinity prediction). (Generative): The model is inverted, given desired properties or partial sequence constraints (indicated by "?"), a PLM with a language modeling head (LM-Head) generates novel sequences predicted to exhibit target characteristics, enabling de novo design of antibodies and TCR mimic binders for applications including viral neutralization, protein engineering, and therapeutic development.