Research map / Label-efficient and robust learning / Recognition and long-tail
Learning to engineer protein flexibility
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.
Builds sequence- and structure-based protein flexibility predictors (Flexpert-Seq/3D) on ProtTrans with LoRA and fine-tunes ProteinMPNN with a flexibility loss (Flexpert-Design) to steer inverse folding toward increased flexibility.
Problem
Generative inverse folding models ignore protein flexibility, and flexibility data (experimental B-factors, MD RMSF, ENM estimates) is scarce, heterogeneous, and costly, so there is no fast way to predict or engineer per-residue flexibility in designed proteins.
Approach
Train a fast protein-language-model regressor (ProtTrans with LoRA, optionally with ANM-based CNN adaptor) to predict MD-derived RMSF, then use it as a differentiable flexibility loss (via straight-through Gumbel-Softmax) to fine-tune a flexibility-conditioned ProteinMPNN so it can be steered to increase flexibility in chosen segments.
Key results
- Flexpert-3D reaches PCC 0.83 with MD RMSF on ATLAS test (139 proteins), vs 0.76 for ANM and GNM, with an estimated upper bound of 0.88.
- Flexpert-Design achieves median enrichment ratio 1.52 and 83% flexibility-increasing mutations, vs 1.07 and 61% for ProteinMPNN on CATH4.3 with |S|=50, tau=5.
- Flexpert-Design sequence recovery drops only about one point (0.49 to 0.48) relative to ProteinMPNN at theta=0.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Flexpert-Seq instead of AlphaFold2 for sequence-only flexibility prediction | ATLAS test set · ATLAS test set | PCC to MD RMSF: 0.72 -> 0.78 | Helped |
| Add backbone structure and ANM features to the sequence-based predictor | Flexpert-Seq · ATLAS test set | PCC to MD RMSF: 0.78 -> 0.83 | Helped |
| Use Flexpert-3D instead of ANM for structure-based flexibility prediction | ATLAS test set · ATLAS test set | PCC to MD RMSF: 0.76 -> 0.83 | Helped |
| Provide flexibility instructions and train Flexpert-Design with native flexibility pseudolabels | ProteinMPNN · CATH4.3 test set, |S|=50, τ=5 | Median enrichment ratio: 1.07 -> 1.43 | Helped |
| Add the L_Flex loss (ablation) | Flexpert-Design · CATH4.3 test set, |S|=50, τ=5 | Median enrichment ratio: 1.43 -> 1.52 | Helped |
| Train Flexpert-Seq on raw B-factors instead of MD RMSF (ablation) | Flexpert-Seq · B-factor training evaluation | PCC to MD: 0.76 -> 0.47 | Hurt |
| Train Flexpert-Seq on per-protein normalized B-factors instead of MD RMSF (ablation) | Flexpert-Seq · B-factor training evaluation | PCC to MD: 0.76 -> 0.56 | Hurt |
| Use PDB structures rather than MD-relaxed structures as ANM input (ablation) | ANM · ATLAS proteins | PCC to MD RMSF: 0.77 -> 0.73 | Hurt |
| Use ESMFold-predicted structures rather than MD-relaxed structures as ANM input (ablation) | ANM · ATLAS proteins | PCC to MD RMSF: 0.77 -> 0.73 | Hurt |
| Apply ProteinMPNN redesign before ESMFold structure prediction for ANM input (ablation) | ANM · ATLAS proteins | PCC to MD RMSF: 0.73 -> 0.71 | Hurt |
| Increase the flexibility-loss weight θ from 0.8 to 0.9 (ablation) | Flexpert-Design · CATH4.3 test set | Sequence recovery: 0.48 -> 0.47 | Hurt |
| Increase the flexibility-loss weight θ from 0.8 to 0.95 (ablation) | Flexpert-Design · CATH4.3 test set | Sequence recovery: 0.48 -> 0.47 | Hurt |
Surprising findings
- Flexpert-Design struggles to decrease flexibility (tau=-10): the best variant achieves median ratio 0.99 with 51% decreasing mutations, which the authors treat as a limitation.
- B-factors from PDB are hard to learn from (PCC 0.47-0.56 vs 0.76 for MD-trained), attributed to crystal packing artifacts.
Limitations
- Flexpert-Design cannot reliably decrease flexibility; biased toward positive instructions because training flexibilities are non-negative.
- Engineered sequences show lower AF2 recovery of ground-truth structure (RMSD 3.36 vs 2.00 A) and lower pLDDT in engineered regions.
- Predictor performance drops for temperatures far from the 300 K training temperature in mdCATH (0.34 at 450 K).
Experimental setup
Datasets: ATLAS CATH4.3 PDB B-factors mdCATH
Benchmarks: ATLAS test set mdCATH CATH4.3 test set
Metrics: Pearson correlation coefficient median enrichment ratio proportion of flexibility-increasing mutations sequence recovery RMSD pLDDT
Baselines: AlphaFold2 pLDDT ESMFold pLDDT ANM GNM B-factors ProteinMPNN
Builds on: ProteinMPNN ProtTrans LoRA Gumbel-Softmax ProDy
Closest papers by meaning
- Flexibility-conditioned protein structure design with flow matching (ICML 2025 · similarity 0.91)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.89)
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.89)
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.89)
- Predicting directional flexibility in proteins (NeurIPS 2026 · 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.