Research map / Image generation / Autoregressive and latent generation
Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints
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 AbX, a continuous-time score-based diffusion model for CDR sequence-structure co-design that uses ESM-2 as an evolutionary prior and adds physical (van der Waals, violation) and geometric (FAPE, distogram, lDDT) training losses.
Problem
Antibody CDR design is limited by scarce labeled antibody-antigen complex data, causing overfitting in diffusion-based generative models.
Approach
Jointly model discrete CDR sequences (CTMC) and SE(3) backbone structures (score-based SDE) with a conditional score network, using a pre-trained ESM-2 as an evolutionary prior and auxiliary physical and geometric losses applied at low noise levels (t<1/4).
Key results
- On the RAbD test set, AbX reaches IMP 18.64% vs 12.07% for DiffAb and 0.00% for dyMEAN, and loop AAR 30.80% vs 21.25% for DiffAb.
- On the DiffAb test set, AbX reaches IMP 42.93% vs 37.13% for DiffAb, with loop RMSD 3.41 vs 3.60.
- Ablation removing the evolutionary constraint lowers plausibility from -0.67 to -0.70 and H3 AAR from 49.17 to 53.21 (ablation numbers in Table 4 are mixed; IMP drops to 19.36 without the evolutionary constraint).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AbX instead of DiffAb | antibody design on RAbD test dataset · RAbD test dataset | IMP (%): 12.07 -> 18.64 | Helped |
| use AbX instead of dyMEAN | antibody design on RAbD test dataset · RAbD test dataset | Plausibility: -1.21 -> -1.01 | Helped |
| use AbX instead of dyMEAN | antibody design on RAbD test dataset · RAbD test dataset | Loop AAR: 22.25 -> 30.80 | Helped |
| use AbX instead of DiffAb | antibody design on RAbD test dataset · RAbD test dataset | Loop RMSD: 3.45 -> 3.24 | Helped |
| use AbX instead of DiffAb | antibody design on DiffAb test dataset · DiffAb test dataset | IMP (%): 37.13 -> 42.93 | Helped |
| use AbX instead of DiffAb | antibody design on DiffAb test dataset · DiffAb test dataset | Plausibility: -0.70 -> -0.67 | Helped |
| use AbX instead of dyMEAN | antibody design on DiffAb test dataset · DiffAb test dataset | Loop AAR: 29.31 -> 30.25 | Helped |
| use AbX instead of DiffAb | antibody design on DiffAb test dataset · DiffAb test dataset | Loop RMSD: 3.60 -> 3.41 | Helped |
| add geometric and physical constraints (ablation) | AbX · DiffAb test dataset | IMP (%): 19.36 -> 54.82 | Helped |
| add physical constraint (ablation) | AbX · DiffAb test dataset | IMP (%): 52.02 -> 54.82 | Helped |
| add evolutionary constraint (ablation) | AbX · DiffAb test dataset | Plausibility: -0.77 -> -0.67 | Helped |
| add evolutionary constraint (ablation) | AbX · DiffAb test dataset | H3 AAR (%): 45.32 -> 49.17 | Hurt |
Surprising findings
- Removing the evolutionary constraint (ESM-2) sharply lowered IMP to 19.36% from 54.82%, a larger drop than removing physical constraints.
Limitations
- No wet-lab experimental validation; all evaluation is in silico.
- Binding affinity is approximated by Rosetta binding energy, not measured affinity.
- Physical constraints are limited to backbone van der Waals and violation losses; full-atom force fields not yet included.
Experimental setup
Datasets: SAbDab RAbD test set DiffAb test set
Benchmarks: RAbD test set (60 complexes) DiffAb test set (19 antigens)
Metrics: amino acid recovery CDR RMSD plausibility (AntiBERTy likelihood) improvement percentage (Rosetta binding energy) loop AAR loop RMSD
Baselines: DiffAb dyMEAN RAbD
Builds on: FrameDiff tauLDR DiffAb AlphaFold2 losses
Closest papers by meaning
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.93)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.91)
- 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 0.89)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.89)
- Synergy of GFlowNet and Protein Language Model Makes a Diverse Antibody Designer (AAAI 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.