Research map / Probabilistic and scientific modeling / Molecules and proteins
ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search
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 ProtInvTree, a training-free reward-guided MCTS framework over ESM-3 for protein inverse folding that uses focus-and-grounding actions and jumpy denoising for fast rewards, improving scTM-score and diversity/novelty on CATH and TS45.
Problem
Protein inverse folding often optimizes sequence recovery and overlooks the one-to-many nature of the task, producing low diversity; existing methods also rarely use test-time computation to improve structural consistency and diversity.
Approach
Reformulates iterative denoising of protein sequences as a tree-based MDP searched with UCT-based MCTS, where each step is a two-stage focus (choose positions) and grounding (choose residues) action, and node values come from a single-step DDIM-style jumpy denoising followed by folding and TM-score reward, all on a frozen ESM-3 policy.
Key results
- On CATH 4.2 (All), ProtInvTree (ESM-3) reaches scTM-score 0.881 and RMSD 1.513 versus 0.858 / 1.566 for the best baseline KW-Design.
- On CATH 4.3, ProtInvTree achieves scTM-score 0.835 vs 0.818 for KW-Design (ESM-2).
- On TS45 de novo designs, ProtInvTree reaches scTM-score 0.724 and RMSD 2.513, versus 0.711 and 2.643 for KWDesign.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply reward-guided tree search to frozen ESM-3 | ESM-3 · CATH v4.2 Short | scTM-score: 0.668 -> 0.768 | Helped |
| Apply reward-guided tree search to frozen ESM-3 | ESM-3 · CATH v4.2 Single-chain | scTM-score: 0.692 -> 0.800 | Helped |
| Apply reward-guided tree search to frozen ESM-3 | ESM-3 · CATH v4.2 All | scTM-score: 0.816 -> 0.881 | Helped |
| Use reward-guided tree search | ESM-3 · CATH v4.3 | scTM-score: 0.775 -> 0.835 | Helped |
| Use reward-guided tree search | ESM-3 · TS45 | scTM-score: 0.690 -> 0.724 | Helped |
| Use autoregressive position selection instead of random position selection (ablation) | ProtInvTree · Structure consistency (selection-strategy comparison) | scTM-score: 0.881 -> 0.877 | Hurt |
| Use entropy-based position selection instead of random position selection (ablation) | ProtInvTree · Structure consistency (selection-strategy comparison) | scTM-score: 0.881 -> 0.870 | Hurt |
| Increase the number of candidate expansions | ProtInvTree · Test-time scaling analysis | sc-TMscore | Helped |
| Increase planning depth | ProtInvTree · Test-time scaling analysis | sc-TMscore | Helped |
| Increase the number of candidate expansions | ProtInvTree · Test-time scaling analysis | diversity | Hurt |
| Increase planning depth | ProtInvTree · Test-time scaling analysis | diversity | Hurt |
Surprising findings
- Baselines optimized for sequence recovery achieve limited novelty at matched sc-TMscore levels.
- In the ablation, random position selection performed slightly better than autoregressive or entropy-based selection for the focus step.
Limitations
- No experimental validation in real-world biological settings.
- Diversity increases with search but structural consistency and diversity trade off; gains saturate with deeper planning.
- Reward relies on ESMFold predictions, so the reward inherits folding-model inductive bias.
Experimental setup
Datasets: CATH v4.2 CATH v4.3 TS45 (CASP15 TS-domains)
Benchmarks: CATH v4.2 CATH v4.3 TS45
Metrics: scTM-score RMSD diversity novelty
Baselines: StructGNN GraphTrans GCA GVP ProteinMPNN AlphaDesign PiFold UniIF LM-Design KW-Design ESM-3
Builds on: Monte Carlo Tree Search UCT ESM-3 DDIM ESMFold
Closest papers by meaning
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.90)
- Learning to engineer protein flexibility (ICLR 2025 · similarity 0.89)
- Fast Uncovering of Protein Sequence Diversity from Structure (ICLR 2025 · similarity 0.89)
- Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.88)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.88)
- FoldToken: Learning Protein Language via Vector Quantization and Beyond (AAAI 2025 · similarity 0.88)
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.