ConTact: Contact-First Antibody CDR Design via Explicit Interface Reasoning

2026-05-23 · Source: cs.LG updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Engineering & Applied Sciences · Depth: Expert, extended

Summary

ConTact is a novel computational architecture for antibody Complementarity-Determining Region (CDR) design that explicitly addresses a limitation in existing methods by separating contact identification from sequence prediction. This three-stage cascade involves learning surface complementarity fingerprints, predicting CDR-antigen contacts, and injecting contact-gated antigen features into the sequence head. The model incorporates a distance-biased cross-attention module for geometric priors and a contact-weighted cross-entropy loss to focus learning on binding-critical positions. Evaluated on Chimera-Bench, ConTact achieved superior performance, demonstrating the lowest RMSD (1.63 Å, 7% better than the next-best baseline), highest epitope F1 score (0.79, 10% over GNN baselines), and competitive amino acid recovery (0.38 AAR) across eleven baselines. It has 9.68M parameters and trains in 1.6 hours on an NVIDIA H100 80GB GPU.

Key takeaway

For machine learning engineers developing antibody design models, you should consider explicitly decomposing contact prediction from sequence generation. ConTact's "contact-first" approach significantly enhances structural quality and epitope awareness, outperforming methods that conflate these tasks. Incorporating a supervised contact prediction stage and contact-gated antigen feature injection can improve your model's ability to leverage antigen information effectively, leading to more accurate and specific CDR designs.

Key insights

Explicitly predicting antigen contacts before sequence design significantly improves antibody CDR design performance.

Principles

• Existing CDR design conflates contact identification and sequence prediction.
• Antigen information flow should be gated by predicted contact confidence.
• Contact-weighted loss concentrates gradient signal on binding-critical positions.

Method

ConTact's three-stage decoder performs complementarity fingerprinting, supervised contact prediction, and contact-gated antigen feature injection into the sequence head.

In practice

• Use distance-biased cross-attention for geometric priors.
• Apply focal binary cross-entropy for imbalanced contact prediction.
• Implement double gating for antigen feature injection.

Topics

• Antibody Design
• CDR Design
• Contact Prediction
• Graph Neural Networks
• Chimera-Bench
• Protein-Antigen Interaction

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• AgForce Enables Antigen-conditioned Generative Antibody Design — Existing antibody design models fail to condition on antigen, suffering from antigen blindness and vocabulary collapse due to cross-entropy loss.
• Deciphering Fingerprints of 3D Molecular Surfaces for Accurate Epitope Prediction — Direct analysis of 3D molecular surfaces significantly enhances epitope prediction accuracy and generalization.
• Reusability report: Meta-learning for antigen-specific T cell receptor binder identification — PanPep generalizes well to unseen antigens with few known binders under specific negative sampling.
• Scaling antibody language models improves structure aware representation for antibody engineering — Scaling antibody language models with advanced tokenization improves structure-aware representation for engineering.
• Enhancing success rates in therapeutic antibody design through generative models — DualGPT-AB is a dual-stage GPT framework optimizing therapeutic antibody design for specificity and developability.
• AWS and Hopkins Engineering announce groundbreaking database for AI/ML antibody design — A new diverse, experimentally validated dataset enables rigorous benchmarking of AI models for antibody developability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.