Research map / Image generation / Autoregressive and latent generation
Antigen-Specific Antibody Design via Direct Energy-based Preference 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 AbDPO, a residue-level direct energy-based preference optimization that fine-tunes a pre-trained conditional antibody diffusion model on self-generated, energy-labelled samples, with gradient surgery to resolve attraction/repulsion conflicts.
Problem
Antigen-specific antibody sequence-structure co-design is typically evaluated with AAR and RMSD, which do not capture rationality or binding affinity; existing co-design methods often produce high-energy, clash-laden CDR-H3 structures due to scarce high-quality training data.
Approach
Adapt DPO to diffusion models with residue-level reward assignment (energy decomposed per CDR residue), fine-tune on the model's own synthetic samples ranked by Rosetta energies, decompose energy into total, non-repulsive and repulsive terms, and resolve gradient conflicts via PCGrad-style gradient surgery.
Key results
- AbDPO achieves mean CDR E_total of 629.44 and CDR-Ag ΔG of 307.56 across 55 complexes, versus 1729.51 and 1297.25 for DiffAb.
- AbDPO yields 9 successful complexes out of 55 (AbDPO+ 5), while all baselines have 0.
- AbDPO reports AAR of 31.25%, lower than DiffAb's 34.92% and dyMEAN's 40.04%, with RMSD 1.98 vs 1.92 for DiffAb.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune with residue-level energy preferences using AbDPO | DiffAb · RAbD test set (55 complexes) | CDR E_total: 1729.51 -> 629.44 | Helped |
| Fine-tune with residue-level energy preferences using AbDPO | DiffAb · RAbD test set (55 complexes) | CDR-Ag ΔG: 1041.43 -> 307.56 | Helped |
| Fine-tune with residue-level energy preferences using AbDPO | DiffAb · RAbD test set (55 complexes) | N_success: 0 -> 9 | Helped |
| Add pLL and PHR preferences | AbDPO · RAbD test set (55 complexes) | AAR: 31.25% -> 36.27% | Helped |
| Add pLL and PHR preferences | AbDPO · RAbD test set (55 complexes) | CDR E_total: 629.44 -> 1106.48 | Hurt |
| Use CDR-level preference instead of residue-level preference (ablation) | AbDPO · 1a14 | optimization efficiency | Hurt |
| Optimize ΔG directly without energy decomposition (ablation) | AbDPO · 1a14 | E_nonRep | Hurt |
| Remove gradient surgery (ablation) | AbDPO · 1a14 | E_Rep | Hurt |
| Replace preference optimization with supervised fine-tuning on top 10% high-quality samples (ablation) | AbDPO · 1a14 | CDR E_total (Avg.): 336.02 -> 1053.82 | Hurt |
| Replace preference optimization with supervised fine-tuning on top 10% high-quality samples (ablation) | AbDPO · 1a14 | CDR-Ag ΔG (Avg.): 88.64 -> 374.27 | Hurt |
Surprising findings
- A simple baseline that outputs the most frequent amino acid per CDR position reaches 38.77% AAR on RAbD, suggesting AAR can be gamed.
- Optimizing the full binding energy ΔG directly causes the model to push antibodies away from the antigen, eliminating repulsion and attraction together.
Limitations
- Preferences are defined via Rosetta energies, whose correlation with wet-lab binding is unresolved.
- Clashes remain in generated antibodies, so most designs still have high energy; only 9 of 55 complexes succeed.
- AAR and RMSD are slightly worse than some baselines (e.g., dyMEAN AAR, DiffAb RMSD).
Experimental setup
Datasets: Structural Antibody Database (SAbDab) RAbD benchmark
Benchmarks: RAbD
Metrics: AAR RMSD CDR E_total CDR-Ag ΔG pLL PHR N_success
Baselines: HERN MEAN dyMEAN DiffAb DiffAb-SFT
Builds on: direct preference optimization denoising diffusion probabilistic model PCGrad gradient surgery Diffusion-DPO DiffAb
Closest papers by meaning
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.93)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.91)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.90)
- Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization (AAAI 2025 · similarity 0.88)
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.88)
- Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences (ICLR 2025 · similarity 0.87)
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.