Research map / Probabilistic and scientific modeling / Molecules and proteins
Synergy of GFlowNet and Protein Language Model Makes a Diverse Antibody Designer
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 PG-AbD, which combines a GFlowNet generator with a product-of-experts reward (frozen ProGen2 PLM times a trainable Potts model) to generate diverse, developable, novel antibody CDR sequences.
Problem
Computational antibody CDR design with pretrained protein language models tends to produce low-diversity, low-novelty candidates because low-temperature sampling collapses to a narrow set of plausible sequences.
Approach
Uses GFlowNet to sample antibody CDRs proportionally to a reward given by a Product of Experts of a frozen pretrained PLM (ProGen2) and a trainable Potts model, jointly training the PoE by contrastive divergence on GFlowNet-generated negatives and the GFlowNet by trajectory balance; a FiLM module injects a template CDR.
Key results
- On RabDab, PG-AbD reaches DIV 62.73 (about 13.5% over the best prior method) with AAR 58.40 and PPL 4.26.
- On SabDab, PG-AbD improves over reprogBERT, with reported average gains of 18.3% in AAR and 31.1% in DIV.
- AlphaFold-Multimer complexes of generated CDR-H3 antibodies reach 85.68 average and 91.59 maximum pLDDT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PG-AbD with FiLM rather than without FiLM (ablation) | PG-AbD · RabDab CDR-H3 | AAR: 35.17 -> 58.40 | Helped |
| Use PG-AbD with FiLM rather than without FiLM (ablation) | PG-AbD · RabDab CDR-H3 | DIV: 76.59 -> 62.73 | Hurt |
| Add the Potts Model to the PoE (ablation) | PG-AbD · SabDab | AAR: 46.38 -> 55.67 | Helped |
| Add the Potts Model to the PoE (ablation) | PG-AbD · SabDab | DIV: 39.26 -> 54.80 | Helped |
| Add the PLM to the PoE (ablation) | PG-AbD · SabDab | AAR: 33.29 -> 55.67 | Helped |
| Add the PLM to the PoE (ablation) | PG-AbD · SabDab | DIV: 45.70 -> 54.80 | Helped |
| Use ProGen2 as the reward PLM instead of the tested alternative PLMs (ablation) | GFlowNet self-exploring · RabDab CDR-H3 | DIV: 31.26; 42.31; 52.11; 54.42; 62.73 | Helped |
| Use PG-AbD rather than reprogBERT | Antibody optimization · SabDab CDR-H3 | AAR: 32.6 -> 39.5 | Helped |
| Use PG-AbD rather than reprogBERT | Antibody optimization · SabDab CDR-H3 | DIV: 67.4 -> 67.8 | No clear effect |
| Use PG-AbD rather than the compared methods | Antigen-specific antibody design · RabDab CDR-H3 | DIV: 13.5% | Helped |
| Use PG-AbD rather than reprogBERT | Antibody optimization · SabDab | DIV: 31.1% | Helped |
Surprising findings
- Fine-tuned ProtBert and ESM-2 give high AAR but low diversity or high PPL, and zero-shot AbLang and ESM-2 give high diversity at poor developability.
Limitations
- Evaluation relies on proxy metrics (PPL, AAR, diversity, novelty) rather than experimental binding validation.
- Structural evaluation uses AlphaFold-Multimer pLDDT confidence, which is not a direct measure of binding affinity.
Experimental setup
Datasets: SabDab RabDab
Benchmarks: RabDab SabDab
Metrics: AAR PPL NOV DIV pLDDT
Baselines: LSTM AR-GNN refine-GNN CEM AntibodyDiff ProtBert AbLang-H AbLang-L ESM-2 reprogBERT RosettaAntibodyDesign
Builds on: GFlowNet ProGen2 Potts model / direct coupling analysis Products of Experts refine-GNN
Closest papers by meaning
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.88)
- Relation-Aware Equivariant Graph Networks for Epitope-Unknown Antibody Design and Specificity Optimization (AAAI 2025 · similarity 0.88)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.87)
- How to make the most of your masked language model for protein engineering (NeurIPS 2026 · similarity 0.87)
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.87)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (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.