Research map / Image generation / Autoregressive and latent generation
Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and 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 RADAb, a retrieval-augmented diffusion model that conditions antibody CDR sequence design on structurally similar CDR-like motifs retrieved from the PDB via MASTER, with a dual-branch (global IPA and local tied-row axial attention) denoiser.
Problem
Generative antibody design typically creates CDR sequences de novo without template constraints, which leads to optimization difficulties, limited data diversity (SAbDab, under ten thousand complexes), overfitting, and unnatural sequences.
Approach
Retrieve sets of structurally homologous CDR-like fragments from a PDB-derived motif database using backbone RMSD and feed their sequences as evolutionary conditions into a dual-branch denoising diffusion model (global IPA context branch plus local CDR-focused axial attention over a pseudo-MSA), agnostic to the base generator.
Key results
- On SAbDab CDR-H3 design, RADAb reaches 57.02% AAR versus 52.99% for the best baseline AbMPNN* and 49.17% for Diffab-fix.
- On long CDR-H3 (length >14), RADAb achieves 51.35% AAR versus 43.27% for AbMPNN* and 42.26% for Diffab-fix.
- In binding-energy optimization, RADAb yields ΔΔG-seq of 7.06 versus 14.05 for Diffab-fix, with 37.30% of designs beating the original sequence (IMP-seq).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add retrieval augmentation (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Helped |
| Add evolutionary embedding (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Helped |
| Use retrieval augmentation and evolutionary embedding together (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Helped |
| Use ground-truth CDR sequence as retrieval input (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Helped |
| Increase the number of retrieved CDR-like fragments from low values (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Helped |
| Use a low number of retrieved CDR-like fragments (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Hurt |
| Increase retrieved fragments beyond k=15 (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Hurt |
| Use k=15 retrieved CDR-like fragments (ablation) | RADAb · CDR-H3 inverse folding | AAR (%) | Helped |
| Use RADAb instead of AbMPNN | CDR-H3 inverse folding · SAbDab CDR-H3 | AAR (%): 52.99 -> 57.02 | Helped |
| Use RADAb instead of ProteinMPNN | Long CDR-H3 inverse folding · Long CDR-H3 (length >14) | AAR (%): 46.63 -> 51.35 | Helped |
| Use RADAb instead of Diffab-fix | Antibody functionality optimization · Antibody functionality optimization | ΔΔG-seq: 14.05 -> 7.06 | Helped |
Surprising findings
- Performance drops when the number of retrieved fragments k is low, attributed to overfitting, and degrades again beyond k=15 due to noise.
- Retrieved CDR-like fragments are filtered to exclude identical CDR sequences during training, and the authors note retrieval based on CDR structure risks data leakage for co-design tasks.
Limitations
- Not validated in wet-lab experiments.
- Retrieved linear motifs from MASTER can have structural issues such as length mismatch or discontinuity.
- Current retrieval scheme only supports inverse folding and risks data leakage for structure-sequence co-design.
Experimental setup
Datasets: SAbDab non-redundant PDB (CDR-like fragments) OAS
Benchmarks: SAbDab test set (63 complexes) long CDR-H3 subset (length >14)
Metrics: AAR scRMSD plausibility (pseudo-log-likelihood) ΔΔG ΔΔG-seq IMP-seq
Baselines: Grafting ProteinMPNN ESM-IF1 Diffab-fix AbMPNN
Builds on: DiffAb MASTER ESM-2 Invariant Point Attention MSA Transformer
Closest papers by meaning
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.93)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.90)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.89)
- Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences (ICLR 2025 · similarity 0.89)
- dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen (AAAI 2025 · similarity 0.88)
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.88)
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.