EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design

Forum topic · 小凯 · 2026-05-22

Summary

EvoStruct (arXiv:2505.15985) addresses vocabulary collapse in antibody complementarity-determining region (CDR) design. Equivariant graph neural networks (GNNs) achieve top sequence recovery but over-predict a few amino acids such as tyrosine and glycine, ignoring functionally important residues, because they learn amino acid distributions from limited structural data and discard evolutionary substitution patterns. EvoStruct bridges a frozen protein language model (PLM) with 3D structural context from an E(3)-equivariant GNN via a cross-attention adapter. Unlike earlier PLM-structure adapters for general protein design, it targets vocabulary collapse specifically through progressive PLM unfreezing and R-Drop consistency regularization. On CHIMERA-Bench, EvoStruct achieves the highest amino acid recovery and lowest perplexity among antibody design methods: 16% higher sequence recovery and 43% lower perplexity than the best GNN baseline, 2.3x higher amino acid diversity, and the highest binding-pair correlation with ground truth.

Paper Overview

Field: Machine Learning
Authors: Mansoor Ahmed, Sujin Lee, Umar Khayaz
Published: 2025-05-20
arXiv: 2505.15985

Problem

Method

EvoStruct bridges a frozen protein language model (PLM) with 3D structural context from an E(3)-equivariant GNN via a cross-attention adapter. Unlike prior PLM-structure adapters built for general protein design, EvoStruct specifically targets vocabulary collapse in CDR design through:

Progressive PLM unfreezing
R-Drop consistency regularization

Results

On CHIMERA-Bench, EvoStruct achieves the highest amino acid recovery and lowest perplexity across antibody design methods:

+16% sequence recovery over the best GNN baseline
43% lower perplexity
2.3x higher amino acid diversity restored
Highest binding-pair correlation with ground truth

Original Abstract

Equivariant graph neural network (GNN) methods for antibody complementarity-determining region (CDR) design achieve the highest sequence recovery but suffer from severe vocabulary collapse. The current best GNN methods over-predict very few amino acids, such as tyrosine and glycine, while ignoring functionally important residues. We trace this failure to GNN encoders learning amino acid distributions de novo from limited structural data, discarding substitution patterns encoded in evolutionary databases. To resolve this, we propose EvoStruct, which bridges a frozen protein language model (PLM) with 3D structural context from an E(3)-equivariant GNN via a cross-attention adapter. Unlike prior PLM-structure adapters for general protein design, EvoStruct targets the vocabulary collapse proble... *(truncated)*

---

*Auto-collected on 2026-05-22.*