Source-linked AI summary
Iterative Refinement Graph Neural Network for Antibody Sequence-Structure Co-design
Wengong Jin, Jeremy Wohlwend, Regina Barzilay, Tommi Jaakkola
TL;DR
Computational antibody design must generate CDRs with desired properties even though their optimal structures are generally unknown beforehand. The paper introduces a graph-based model that autoregressively generates sequences while iteratively refining their global structures and efficiently modeling framework context. It reports improved test likelihood and stronger performance than prior baselines across antibody-generation tasks, including SARS-CoV-2 neutralization design.
Problem
CDR design has a vast constrained sequence space, while prior methods either omit structure or require a desired 3D structure known in advance.
Method
RefineGNN co-designs CDR sequences and structures as graphs by interleaving autoregressive residue generation with iterative global-structure refinement and coarse-grained context modeling.
Results
The method outperforms sequence-based and graph-based baselines across three antibody-generation tasks, including lower test perplexity and stronger property-guided design performance.
Takeaways & Limitations
The results support graph-based sequence-structure co-design for generating antibodies when suitable CDR structures are not known beforehand.
Takeaways & Limitations
Efficient whole-antibody refinement remains constrained because recomputing context representations is expensive and framework coordinates must adjust as CDR coordinates change.
Abstract
from arXiv · showhide
Antibodies are versatile proteins that bind to pathogens like viruses and stimulate the adaptive immune system. The specificity of antibody binding is determined by complementarity-determining regions (CDRs) at the tips of these Y-shaped proteins. In this paper, we propose a generative model to automatically design the CDRs of antibodies with enhanced binding specificity or neutralization capabilities. Previous generative approaches formulate protein design as a structure-conditioned sequence generation task, assuming the desired 3D structure is given a priori. In contrast, we propose to co-design the sequence and 3D structure of CDRs as graphs. Our model unravels a sequence autoregressively while iteratively refining its predicted global structure. The inferred structure in turn guides subsequent residue choices. For efficiency, we model the conditional dependence between residues inside and outside of a CDR in a coarse-grained manner. Our method achieves superior log-likelihood on the test set and outperforms previous baselines in designing antibodies capable of neutralizing the SARS-CoV-2 virus.
1 INTRODUCTION
The paper frames antibody CDR design as a difficult sequence-generation problem because desired binding properties are constrained and CDR structures are rarely known in advance. It proposes graph-based co-design of CDR sequences and structures, with iterative structural refinement and coarse-grained context modeling, and reports improved generation performance across three tasks.
- Over 20^60 possible CDR sequences must be searched for candidates satisfying binding affinity, stability, and synthesizability constraints.
- Predefined-structure generation is unsuitable because desired antibody structures are rarely known a priori, motivating sequence-structure co-design.
- The model represents sequence-structure pairs as graphs to model CDR-context dependence at both sequence and structure levels.
- An interleaved architecture autoregressively generates amino-acid nodes while iteratively refining a global graph structure rather than expanding a teacher-forced partial graph.
- Across language modeling, SARS-CoV-2 neutralization optimization, and antigen-binding design, the method achieves lower test perplexity and outperforms prior baselines in property-guided design.
2 RELATED WORK
Prior antibody-design methods use either computationally expensive energy optimization or predominantly sequence-based generative modeling, while protein methods often condition on known structures. This paper instead co-designs sequences and structures because suitable CDR structures are often unknown for new pathogens.
- Physics-based antibody-design methods iteratively modify sequences and structures toward local energy minima but are computationally expensive.
- Most generative antibody-design methods are sequence-based, whereas related protein models condition generation on a backbone structure or protein fold.
- Existing autoregressive graph models generate edges sequentially and cannot modify previously generated subgraphs when new nodes arrive.
- State-of-the-art protein-folding models are not directly applicable because they require complete sequences and prespecified multiple-sequence alignments.
- Unlike methods refining complete molecules or point clouds, this approach refines 3D structures for incomplete graphs while interleaving refinement with graph generation.
3 ANTIBODY SEQUENCE AND STRUCTURE CO-DESIGN
The paper represents antibody sequence and 3D structure jointly as a graph, then generates CDR residues while iteratively refining the global structure. It conditions CDR design on framework context and uses coarse-grained context modeling for efficiency.
- Background: Antibodies bind antigens through variable domains containing CDRs, which are the main determinants of binding and neutralization.The paper focuses on heavy-chain CDR generation conditioned on the framework region.
- Graph representation: Each antibody graph node represents a residue with sequence and dihedral-angle features, while edges encode sequence distance, spatial distance, direction, and orientation.The graph uses a K-nearest-neighbors structure with K = 8.
- Iterative refinement: RefineGNN autoregressively predicts the next amino acid, updates its node label, and then predicts coordinates for all residues to accommodate the new residue.Separate sequence and structure message-passing networks perform the two predictions.
- Training: During training, teacher forcing is applied only to discrete amino-acid prediction, while structure prediction continuously refines the full structure without teacher forcing.The graph-generation objective combines sequence cross-entropy with structure losses based on distances and angles.
- Conditional generation: For conditional CDR generation, the framework context is encoded and clustered into residue blocks whose mean coordinates and embeddings support iterative global graph refinement.This coarse-grained representation reduces context length while retaining structural interactions with the CDR.
4 EXPERIMENTS
The experiments evaluate RefineGNN on language modeling, structure prediction, antigen-binding design, and SARS-CoV-2 neutralization optimization against sequence and graph-generation baselines.
- Evaluation setup: The evaluation covers perplexity and RMSD on held-out antibody structures, amino acid recovery on antigen-binding design, and neutralization optimization against SARS-CoV-2.The experiments also use a 60-complex antigen-binding benchmark and predictor-based neutralization scoring.
- Language modeling and structure prediction: RefineGNN outperforms sequence-only and AR-GNN baselines on CDR-H3 language modeling, reducing PPL by 13% versus the sequence-only model and 10% versus AR-GNN.The reported PPL values are 8.38 versus 9.70 and 9.44, respectively.
- Language modeling and structure prediction: RefineGNN predicts CDR-H3 structures more accurately than AR-GNN, achieving a 30% relative RMSD reduction.Figure 3 illustrates the comparison with predicted RMSDs of 0.57 for RefineGNN and 2.16 for AR-GNN on PDB 4bkl.
- Ablation studies: Removing attention or context coarsening raises PPL from 8.38 to 8.86 or 9.01, while removing both produces a substantially worse unconditional model.These ablations support the reported multi-resolution conditioning strategy.
- Antigen-binding antibody design: On antigen-binding design, RefineGNN achieves the highest AAR score, improving by approximately 7% absolute over the best baseline.The benchmark uses fixed antibody frameworks, and the methods generate CDR-H3 sequences without conditioning on antigen structure.
- SARS-CoV-2 neutralization optimization: On SARS-CoV-2 optimization, RefineGNN improves average neutralization score by 3% over the best baseline and achieves CoVAbDab PPL of 7.86 versus 8.67.Models are pretrained on SAbDab and finetuned on CoVAbDab using ITA.
5 CONCLUSION
RefineGNN performs antibody sequence–structure co-design by revising generated subgraphs as new residues are added, and outperforms sequence- and graph-based approaches across three antibody generation tasks.
- RefineGNN revises a generated subgraph to accommodate the addition of new residues.This distinguishes it from previous graph generation methods.
- The model significantly outperforms sequence-based and graph-based approaches on three antibody generation tasks.
A.1 REFINEGNN
The RefineGNN representation encodes residue geometry through dihedral-angle node features and local-coordinate edge features, with attention used in the model architecture.
- Each residue node feature encodes three dihedral angles using their sine and cosine values.The feature vector contains cos φ_i, cos ψ_i, cos ω_i, sin φ_i, sin ψ_i, and sin ω_i.
- Each residue’s orientation matrix defines a local coordinate system from backbone direction vectors and their derived binormal and normal vectors.
- The model uses a standard bilinear attention mechanism in its attention layer.
A.2 AR-GNN
AR-GNN autoregressively encodes the current antibody subgraph, predicts new-residue structure and sequence attributes, and relies on teacher forcing with data augmentation during training.
- AR-GNN encodes the current subgraph induced by generated residues into representation vectors at each generation step.
- AR-GNN predicts the next residue’s dihedral-angle feature, pairwise distances to previous residues, and amino acid type.
- Teacher forcing requires ground-truth structure and sequence during training, creating a training–testing discrepancy addressed with corrupted-graph data augmentation.Gaussian coordinate noise is added to antibody graphs before message passing.
- AR-GNN uses attention to combine residue representations with context vectors for conditional generation.
- Table 3 presents examples of generated CDR-H3 sequences and their predicted neutralization improvements over original antibodies S1D7 and C694.
B EXPERIMENTAL DETAILS
The experiments specify model hyperparameters, SAbDab dataset composition, bound/unbound structure handling, baseline design settings, and SARS-CoV-2 neutralization optimization procedures.
- RefineGNN used hidden dimension 256 and four message-passing layers, while AR-GNN used hidden dimension 256 and three layers.All models used Adam, dropout 0.1, and learning rate 0.001.
- After deduplication, SAbDab contained separate train, validation, and test splits for CDR-H1, CDR-H2, and CDR-H3.The reported split sizes were 4050/359/326, 3876/483/376, and 3896/403/437, respectively.
- The study retained bound and unbound antibody structures because their mismatch was considered relatively small, with 65% of training data from bound structures.A cited study found average VH-domain RMSD below 0.7 across eight antibodies.
- RAbD selected a known CDR structure matching the original CDR-H3 length but not its exact sequence before grafting and energy minimization.
- For SARS-CoV-2 neutralization, models were pretrained on SAbDab and fine-tuned with ITA for 3000 steps using M = 100.
- The neutralization predictor was trained on CoVAbDab to predict SARS-CoV-1 and SARS-CoV-2 labels for spike-protein receptor-binding-domain targets.