Research map / Graph learning / Knowledge and temporal graphs
Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RAAD, a relation-aware equivariant graph network that co-designs antigen-specific CDR sequences and structures in one forward pass without pre-given epitopes, plus a specificity metric (SP-score) and contrastive constraint for antibody optimization.
Problem
Antibody CDR design struggles with long CDR sequences (insufficient context), relies on pre-given static epitopes, and optimization yields non-specific "universal" antibodies that bind off-target antigens.
Approach
An attributed heterogeneous graph with 8 edge relations and relation-aware equivariant message passing that dynamically updates antigen-antibody edge relations (epitope-unknown), generates sequences and backbone coordinates in one shot, and adds a mutual-information-based contrastive specificity constraint and an SP-score metric during iterative target augmentation.
Key results
- On SAbDab 10-fold CV, RAAD reaches 58.87% AAR on CDR-H2 (vs 55.41% dyMEAN) and 0.72 RMSD on CDR-H2 (12.20% lower than dyMEAN).
- On RAbD CDR-H3 design, RAAD achieves 41.26 AAR, 0.9874 TM-score and 1.46 RMSD, the best among all compared methods.
- In SKEMPI-based optimization, RAAD with L_SP reaches SP-score 11.549 with ΔΔG_neg of -0.238, versus -1.605 for MEAN and -1.543 for dyMEAN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add node features beyond residue type (ablation) | RAAD · SAbDab modeling | AAR: 66.12 -> 67.64 | Helped |
| add edge features beyond edge type (ablation) | RAAD · SAbDab modeling | AAR: 65.62 -> 67.64 | Helped |
| use relation-aware message passing (ablation) | RAAD · SAbDab modeling | AAR: 64.85 -> 67.64 | Helped |
| dynamically update edge relations (ablation) | RAAD · SAbDab modeling | AAR: 66.73 -> 67.64 | Helped |
| add global nodes and their message passing (ablation) | RAAD · SAbDab modeling | AAR: 65.63 -> 67.64 | Helped |
| add the contrastive specificity constraint (ablation) | RAAD antibody optimization · SKEMPI V2.0 | SP-score: 10.415 -> 11.549 | Helped |
| add the contrastive specificity constraint (ablation) | RAAD antibody optimization · SKEMPI V2.0 | ΔΔG_neg: -1.748 -> -0.238 | Helped |
| use full antigen input instead of epitope input (ablation) | RAAD with heavy and light chains · Antibody modeling | AAR: 66.54 -> 67.64 | Helped |
| use full antigen input instead of epitope input (ablation) | RAAD with heavy and light chains · CDR-H3 generation | AAR: 39.42 -> 41.26 | Helped |
| add light-chain context (ablation) | RAAD with heavy-chain and full-antigen context · CDR-H3 generation | AAR: 40.17 -> 41.26 | Mixed |
| pre-train the encoder on UniRef with ESM-1b (ablation) | RAAD · SAbDab modeling | AAR: 67.64 -> 68.10 | Helped |
| pre-train the encoder on OAS with AbBERT (ablation) | RAAD · SAbDab modeling | AAR: 67.64 -> 68.32 | Helped |
Surprising findings
- Optimization with ITA alone tends to produce universal antibodies that bind non-target antigens with high affinity, motivating the specificity constraint.
- Light chain context contributes little to design performance, while heavy-chain framework context matters most.
Limitations
- Optimization depends on an iterative target augmentation loop that relies on the quality of the external ΔΔG predictor.
- Pre-training effects are only preliminarily explored in the appendix.
Experimental setup
Datasets: SAbDab RAbD SKEMPI V2.0 UniRef OAS AlphaFoldDB
Benchmarks: SAbDab RAbD SKEMPI V2.0
Metrics: AAR RMSD TM-score SP-score ΔΔG ΔL
Baselines: LSTM C-LSTM RefineGNN C-RefineGNN HERN C-HERN DiffAb AbDiffuser ADesigner MEAN dyMEAN RosettaAD
Builds on: MEAN dyMEAN RefineGNN Iterative Target Augmentation
Closest papers by meaning
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.89)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.88)
- Synergy of GFlowNet and Protein Language Model Makes a Diverse Antibody Designer (AAAI 2025 · similarity 0.88)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.87)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.87)
- AsEP: Benchmarking Deep Learning Methods for Antibody-specific Epitope Prediction (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.