Research map / Probabilistic and scientific modeling / Molecules and proteins
dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen
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 dyAb, an end-to-end antibody design framework that starts from AlphaFold2-predicted pre-binding antigen structures and combines coarse-grained interface alignment with fine-grained sequence-structure flow matching to model antigen conformational change upon binding.
Problem
Existing antibody design methods use post-binding antigen structures as input and ignore antigen conformational changes upon antibody binding, reducing reliability of designed antibodies.
Approach
Initialize from AlphaFold2-predicted pre-binding antigen structure, coarsely place the antibody by aligning a predicted binding interface (Kabsch alignment), then refine antibody CDR sequence and complex structure with flow matching that directly predicts the end state.
Key results
- On RAbD epitope-binding CDR-H3 generation, dyAb reaches 37.89% AAR, 0.9264 TMscore and 9.86 RMSD, versus 35.97% AAR and 11.67 RMSD for dyMEAN.
- On SKEMPI affinity optimization, dyAb achieves overall ΔΔG of -14.44 with ΔL 3.44, versus -11.13 and 6.84 for dyMEAN.
- On complex structure prediction, dyAb reports TMscore 0.9224 and DockQ 0.446, outperforming dyMEAN (0.2450 TMscore, 0.391 DockQ) and matching multi-stage pipelines on TMscore.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dyAb instead of dyMEAN for CDR-H3 design | Antibody design with dynamic antigen structures · RAbD benchmark | TMscore: 0.4584 -> 0.9264 | Helped |
| Use dyAb instead of dyMEAN for affinity optimization | Antibody design · SKEMPI, Overall | ΔΔG: -11.13 -> -14.44 | Helped |
| Use dyAb instead of dyMEAN for complex structure prediction | Antibody-antigen complex prediction · IgFold test set | DockQ: 0.391 -> 0.446 | Helped |
| Add coarse-grained alignment (ablation) | dyAb · RAbD benchmark | TMscore: 0.1403 -> 0.9264 | Helped |
| Add fine-grained flow matching (ablation) | dyAb · RAbD benchmark | TMscore: 0.5072 -> 0.9264 | Helped |
| Set flow-matching sampling step to 1 instead of the default 10 (ablation) | dyAb · RAbD benchmark | AAR: 37.89% -> 31.53% | Hurt |
| Set flow-matching sampling step to 50 instead of the default 10 (ablation) | dyAb · RAbD benchmark | AAR: 37.89% -> 32.80% | Hurt |
Surprising findings
- Removing the coarse-grained interface alignment collapses generation performance (AAR 13.44%, TMscore 0.1403), showing the alignment stage is critical.
- Multi-stage pipelines with ground-truth post-binding antigen structures do not clearly outperform dyAb on docking metrics, with IgFold→HDock w/ GT having lower DockQ than w/ AF2.
Limitations
- Wet-lab experiments are needed to validate the designed antibodies' capability.
- Lower lDDT for complex structure prediction suggests a focus on global rather than local structural accuracy.
Experimental setup
Datasets: SAbDab SKEMPI
Benchmarks: RAbD SKEMPI IgFold test set
Metrics: AAR CAAR TMscore lDDT RMSD DockQ ΔΔG ΔL
Baselines: RosettaAb HERN DiffAb MEAN dyMEAN IgFold+HDock IgFold+HERN
Builds on: flow matching dyMEAN MEAN AlphaFold2 Kabsch algorithm
Closest papers by meaning
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.89)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.89)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.88)
- Cross-Gate MLP with Protein Complex Invariant Embedding Is a One-Shot Antibody Designer (AAAI 2024 · similarity 0.87)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.87)
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · 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.