EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation

2026-05-20 · Source: Takara TLDR - Daily AI Papers · Field: Science & Research — Life Sciences & Biology, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

EvoStruct is a novel method designed for antibody Complementarity-Determining Region (CDR) design, addressing the "vocabulary collapse" issue prevalent in existing equivariant graph neural network (GNN) approaches. Current GNNs over-predict a limited set of amino acids, such as tyrosine and glycine, because they learn distributions de novo from restricted structural data, overlooking crucial evolutionary substitution patterns. EvoStruct resolves this by integrating a frozen protein language model (PLM) with 3D structural context from an E(3)-equivariant GNN via a cross-attention adapter. It employs progressive PLM unfreezing and R-Drop consistency regularization to specifically combat vocabulary collapse. Evaluated on the CHIMERA-Bench dataset, EvoStruct achieved the highest amino acid recovery and lowest perplexity, improving sequence recovery by 16% and reducing perplexity by 43% relative to GNN baselines. It also recovered 2.3x greater amino acid diversity and demonstrated the highest binding-pair correlation with ground truth.

Key takeaway

For research scientists developing antibody Complementarity-Determining Region (CDR) design models, EvoStruct demonstrates a critical advancement in overcoming vocabulary collapse. Your current GNN-based methods may be over-predicting common amino acids, limiting functional diversity. Consider adopting hybrid architectures that bridge protein language models with structural GNNs, incorporating progressive unfreezing and consistency regularization to achieve significantly higher amino acid diversity and improved binding-pair correlation in your designs.

Key insights

EvoStruct integrates PLMs and GNNs to enhance antibody CDR design diversity and accuracy by leveraging evolutionary and structural priors.

Principles

• GNNs learning de novo from limited structural data causes vocabulary collapse.
• Evolutionary databases offer crucial amino acid substitution patterns.
• Bridging PLMs with structural context improves amino acid diversity.

Method

EvoStruct bridges a frozen protein language model (PLM) with 3D structural context from an E(3)-equivariant GNN via a cross-attention adapter, using progressive PLM unfreezing and R-Drop consistency regularization.

In practice

• Combine PLM and GNN features using cross-attention adapters.
• Implement progressive PLM unfreezing for targeted fine-tuning.
• Apply R-Drop consistency regularization to boost design diversity.

Topics

• Antibody CDR Design
• Protein Language Models
• Graph Neural Networks
• Evolutionary Priors
• Structural Priors
• Vocabulary Collapse
• CHIMERA-Bench

Best for: AI Scientist, Research Scientist

Related on AIssential

• AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design — AgentPLM enhances protein language models by integrating external biophysical feedback and learning to effectively utilize it during sequence generation.
• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• Mutation Without Variation: Convergence Dynamics in LLM-Driven Program Evolution — LLM-driven program mutation inherently converges to limited structural forms, hindering open-ended exploration.
• HD-Prot: A Protein Language Model for Joint Sequence-Structure Modeling with Continuous Structure Tokens — A hybrid diffusion pLM effectively integrates continuous protein structure tokens to overcome discretization's information loss.
• New Benchmarking Shows Limited Generalization Power of TCR Antigenic Epitope Prediction Models — Existing TCR antigen prediction models lack generalization due to insufficient, unbiased benchmark datasets, limiting their utility.
• 🔬ESM: The Bitter Lesson is Coming for Proteins - Alex Rives, BioHub — Scaling protein language models with vast, diverse data reveals emergent biological capabilities and enables programmable design.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.