Research map / Probabilistic and scientific modeling / Molecules and proteins
AffinityFlow: Guided Flows for Antibody Affinity Maturation
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 AffinityFlow, an AlphaFlow-based alternating optimization framework that guides structure generation with a structure-based affinity predictor, then mutates sequences via ProteinMPNN and selects with a sequence-based predictor, trained with a co-teaching module on noisy Rosetta energies.
Problem
Antibody affinity maturation is costly; the paper targets sequence-only in silico affinity maturation of single-domain antibodies (sdAbs) given only antibody and antigen sequences, with scarce labeled binding data.
Approach
Combine AlphaFlow sequence-conditioned structure generation with predictor guidance toward low binding energy, followed by ProteinMPNN inverse folding and sequence-predictor selection, and train both predictors via co-teaching where each filters pairwise Rosetta-derived labels for the other.
Key results
- AffinityFlow reaches 91.2% IMP over all CDR positions, versus 90.0% for the next best (AbDPO), with lowest Sim 0.528.
- On CDR-H3, IMP falls from 93.3% to 73.3% in the one-iteration ablation and to 63.3% without AlphaFlow.
- Predictor Spearman R improves from about 0.10 (sequence) and -0.004 (structure) without energy data to 0.51 and 0.52 with co-teaching sample selection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AffinityFlow instead of the strongest baseline, AbDPO | Affinity maturation methods · CDR-H1 | IMP: 88.3% -> 88.9% | Helped |
| Use AffinityFlow instead of the strongest baseline, AbDPO | Affinity maturation methods · CDR-H2 | IMP: 91.1% -> 93.3% | Helped |
| Use AffinityFlow instead of the strongest baseline, AbLang | Affinity maturation methods · CDR-H3 | IMP: 88.5% -> 89.7% | Helped |
| Use AffinityFlow instead of the strongest baseline, AbDPO | Affinity maturation methods · All CDRs | IMP: 90.0% -> 91.2% | Helped |
| Use three alternating iterations rather than one iteration with multiple simultaneous mutations (ablation) | AffinityFlow · CDR-H3 | IMP: 73.3% -> 93.3% | Helped |
| Add Amber relaxation predictor-corrector step (ablation) | AffinityFlow · CDR-H3 | IMP: 83.3% -> 93.3% | Helped |
| Use AlphaFlow guided structure generation instead of direct gradient optimization (ablation) | AffinityFlow · CDR-H3 | IMP: 63.3% -> 93.3% | Helped |
| Use AlphaFlow guided structure generation instead of direct gradient optimization (ablation) | AffinityFlow · CDR-H3 | Nat: 0.314 -> 0.330 | Helped |
| Add biophysical energy data to predictor training (ablation) | AffinityFlow · CDR-H3 | IMP: 66.7% -> 93.3% | Helped |
| Use sample selection when training on pairwise energy data (ablation) | AffinityFlow · CDR-H3 | IMP: 76.7% -> 93.3% | Helped |
| Use sample selection when training on pairwise energy data (ablation) | AffinityFlow · CDR-H3 | Sim: 0.523 -> 0.514 | Helped |
| Add biophysical energy data for predictor fine-tuning (ablation) | Sequence-based predictor · Test set | Spearman's R: 0.0956 -> 0.40 | Helped |
Surprising findings
- Under a standard supervised-only predictor, AffinityFlow's IMP on CDR-H3 drops from 89.7% to 68.3%, and baseline IMP drops sharply too, indicating that much of the reported gain depends on the energy-trained sequence predictor shared across methods.
- Structure-based ESM2-GVP predictor was not expected to outperform ESM-2 alone, as the authors note the lack of reliable antibody structures.
Limitations
- Predictors trained on only 120 labeled sdAb-antigen pairs; the structure-based predictor is weak in isolation (R about -0.004 without energy data).
- Many baselines' high IMP is attributed to the shared energy-trained sequence predictor, making comparisons dependent on that choice.
- Evaluation relies on Rosetta binding energies and IgFold/GeoDock predictions rather than wet-lab measurements.
Experimental setup
Datasets: SAbDab sdAb subset GeoDock docked complexes Rosetta binding energies
Benchmarks: SAbDab sdAb subset (60 maturation targets)
Metrics: IMP (improvement percentage) Sim (inter-antigen sequence similarity) Nat (AntiBERTy inverse perplexity) Spearman R
Baselines: ESM AbLang nanoBERT dWJS DiffAb AbDPO GearBind
Builds on: AlphaFlow AlphaFold2 ProteinMPNN Co-teaching DiffAb
Closest papers by meaning
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.89)
- dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen (AAAI 2025 · similarity 0.89)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.89)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.89)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.88)
- Synergy of GFlowNet and Protein Language Model Makes a Diverse Antibody Designer (AAAI 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.