Research map / Probabilistic and scientific modeling / Molecules and proteins
Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies
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 AlignAb, a three-stage (BERT pre-training, diffusion transfer, energy-based alignment) framework for antibody CDR-H3 sequence-structure co-design, aligning a diffusion model with Pareto-optimal multi-objective DPO on Rosetta energies with iterative online sampling and temperature scaling.
Problem
Antibody sequence-structure co-design models generate CDRs that are structurally irrational or weakly binding because training relies on AAR/RMSD and ignores energy functions, and structured antibody-antigen data is scarce.
Approach
Extends AbDPO to diffusion models with a ground-truth reward margin and multi-objective (weighted-sum of attraction and repulsion energies) preference pairs, trained iteratively on online samples with decaying/scaled sampling temperature, avoiding extra data.
Key results
- AlignAb achieves the smallest average energy gap (39.00) versus baselines such as AbX (63.43) and DiffAb (263.44) on RAbD-based Rosetta energies.
- In ablation, Iterative POEA reaches a gap of 32.39 versus 51.59 for single-round POEA and 158.74 for DPO.
- AlignAb gets RMSD 1.51 (best among compared methods) but AAR 35.34, below dyMEAN's 40.95.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add energy-based alignment (ablation) | AlignAb · CDR-H3 design, first 10 antigens | Energy gap: 205.82 -> 32.39 | Helped |
| add energy-based alignment (ablation) | AlignAb · CDR-H3 design, first 10 antigens | CDR E_total: 128.10 -> 34.13 | Helped |
| add antibody sequence pre-training (ablation) | AlignAb · CDR-H3 design, first 10 antigens | Energy gap: 56.33 -> 32.39 | Helped |
| use POEA instead of original DPO (ablation) | AlignAb · CDR-H3 design, first 10 antigens | Energy gap: 158.74 -> 51.59 | Helped |
| use iterative online POEA instead of single-iteration POEA (ablation) | AlignAb · CDR-H3 design, first 10 antigens | Energy gap: 51.59 -> 32.39 | Helped |
| apply temperature scaling at T=1.5 | DiffAb · CDR-H3 sequence diversity | CDR-H3 entropy: 3.29 -> 3.84 | Helped |
| use AlignAb rather than AbX | antigen-binding CDR-H3 design · CDR-H3 design | Energy gap (Top): 37.75 -> 17.91 | Helped |
| use AlignAb rather than AbX | antigen-binding CDR-H3 design · CDR-H3 design | Energy gap (Avg.): 63.43 -> 39.00 | Helped |
| add alignment stage (ablation) | AlignAb · CDR-H3 design | RMSD: 2.25 -> 1.51 | Helped |
| add alignment stage (ablation) | AlignAb · CDR-H3 design, first 10 test antigens | Boltz-2 IPTM: 0.669 ± 0.118 -> 0.698 ± 0.104 | Helped |
Surprising findings
- Naive DPO alignment performs worse than the no-alignment-free POEA variants on total and repulsion energies, with DPO's gap (158.74) far above the non-aligned AlignAb* variant in some metrics.
- Several baselines (HERN, dyMEAN, ABGNN) produce extremely high CDR E_total values, indicating strong steric clashes despite competitive AAR/RMSD.
Limitations
- Multi-objective weights are fixed at a 1:3 ratio across all test cases due to limited compute, so per-antigen Pareto tuning is not explored.
- Evaluation relies on Rosetta energies as rewards and evaluation proxies, which may not reflect wet-lab binding.
- Only CDR-H3 design on 42 RAbD complexes is evaluated; AAR is lower than some baselines.
Experimental setup
Datasets: Observed Antibody Space (OAS) SAbDab RAbD
Benchmarks: RAbD (42 complexes)
Metrics: CDR E_total CDR-Ag dG CDR-Ag E_att CDR-Ag E_rep energy gap (MAE vs reference) AAR RMSD Boltz-2 IPTM TAP developability scores
Baselines: MEAN HERN dyMEAN ABGNN DiffAb AbX AbDPO
Builds on: DPO AbDPO Diffusion-DPO DiffAb AbGNN
Closest papers by meaning
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.93)
- Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences (ICLR 2025 · similarity 0.88)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.88)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.88)
- Synergy of GFlowNet and Protein Language Model Makes a Diverse Antibody Designer (AAAI 2025 · similarity 0.87)
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · 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.