Research map / Probabilistic and scientific modeling / Molecules and proteins
Protein Inverse Folding From Structure Feedback
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.
Fine-tunes inverse-folding models (ProteinMPNN, InstructPLM) with DPO plus an SFT regularizer, using TM-Score preference labels from ESMFold predictions of self-generated sequences, improving CATH 4.2 TM-Score from 0.77 to 0.81.
Problem
Inverse folding models are trained to maximize sequence likelihood, which does not reliably ensure that designed sequences fold into the target structure; existing preference-optimization efforts depend on costly wet-lab labels.
Approach
Use a folding model (ESMFold) as an in-silico judge: sample sequences from the inverse-folding model, fold them, rank by TM-Score to the target structure, and treat top-half vs bottom-half as chosen/rejected pairs for DPO with an added SFT loss on chosen sequences, optionally iterated over rounds.
Key results
- InstructPLM-DPO raises CATH 4.2 TM-Score from 0.77 to 0.81 and recovery from 53.58% to 55.21%.
- Multi-round DPO on 10 hard structures raises mean TM-Score by 79.5%, from 0.39 to 0.70.
- ProteinMPNN-DPO improves CATH 4.2 recovery from 40.27% to 41.29%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DPO fine-tuning with wild-type sequences as chosen samples | ProteinMPNN · CATH 4.2 | recovery: 40.27 -> 41.29 | Helped |
| Apply DPO fine-tuning | InstructPLM · CATH 4.2 | recovery: 53.58% to 55.21% | Helped |
| Apply DPO fine-tuning | InstructPLM · TS50 | recovery: 60.20 % to 62.01 % | Helped |
| Apply DPO fine-tuning | InstructPLM · TS500 | recovery: 66.13 % to 66.37 % | Helped |
| Apply DPO fine-tuning | InstructPLM · CATH 4.2 test set, All | TM-Score: 0.77 to 0.81 | Helped |
| Train only on chosen examples (SFT) instead of jointly using chosen and rejected examples (ablation) | InstructPLM · CATH 4.2 test set, All | TM-Score: from 0.77 to 0.78 | Helped |
| Add rejected examples to chosen-example training, using λ = 10 (ablation) | InstructPLM-DPO · CATH 4.2 test set, All | TM-Score: 0.78 to 0.80 | Helped |
| Add rejected examples to chosen-example training, using λ = 10 (ablation) | InstructPLM-DPO · CATH 4.2 test set, TM-Score < 0.5 | TM-Score: 0.37 to 0.43 | Helped |
| Increase rejected-example weight from λ = 10 to λ = 1 (ablation) | InstructPLM-DPO · CATH 4.2 test set, All | TM-Score: 0.80 to 0.81 | Helped |
| Set λ = 0 so rejected samples outweigh chosen samples (ablation) | InstructPLM-DPO · CATH 4.2 test set, All | TM-Score: 0.81 to 0.81 | Hurt |
| Increase training steps up to 4,000 (ablation) | InstructPLM-DPO · CATH 4.2 | TM-Score | Helped |
| Increase contrastive samples per structure (ablation) | InstructPLM-DPO · CATH 4.2 | TM-Score | Hurt |
Surprising findings
- Scaling the number of contrastive samples per structure increases recovery but degrades TM-Score and increases RMSD, so data quality matters more than quantity.
- ProteinMPNN failed to converge when trained only on its own predictions, requiring wild-type sequences as added chosen samples.
Limitations
- Increasing contrastive samples per structure can reduce structural similarity despite higher recovery.
- ProteinMPNN fails to converge on self-generated data and needs wild-type sequences; the model cannot solve the hardest cases (e.g., 1mkn.A), suggesting a stronger baseline is needed.
- Weak correlation between recovery change and TM-Score change (R^2=0.19); multi-round DPO only briefly explored due to compute limits.
Experimental setup
Datasets: CATH 4.2 TS50 TS500
Benchmarks: CATH 4.2 TS50 TS500
Metrics: sequence recovery perplexity TM-Score RMSD
Baselines: ProteinMPNN InstructPLM PiFold LM-Design KW-Design ESM-IF SPIN
Builds on: DPO RLHF LoRA ProteinMPNN InstructPLM ESMFold TM-Align
Closest papers by meaning
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.90)
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.88)
- Zero-shot protein stability prediction by inverse folding models: a free energy interpretation (NeurIPS 2025 · similarity 0.88)
- Fast Uncovering of Protein Sequence Diversity from Structure (ICLR 2025 · similarity 0.88)
- All-atom inverse protein folding through discrete flow matching (ICML 2025 · similarity 0.88)
- Predicting mutational effects on protein binding from folding energy (ICML 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.