Research map / Probabilistic and scientific modeling / Molecules and proteins
Bridge-IF: Learning Inverse Protein Folding with Markov Bridges
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 Bridge-IF, a Markov diffusion bridge that refines a structure-derived PiFold prior sequence toward the native sequence, using an ESM-based PLM with timestep and structure conditioning for inverse protein folding.
Problem
Inverse protein folding (designing sequences for a backbone structure) suffers from error accumulation in discriminative autoregressive/one-shot models and fails to capture the one-to-many mapping from structure to plausible sequences.
Approach
Replace the uninformative noise prior of discrete diffusion with a deterministic structure-derived sequence prior (from PiFold) and learn a Markov bridge from that prior to the native sequence; use a reparameterized, simplified cross-entropy loss and adapt a frozen pretrained PLM with AdaLN-Bias and a structural cross-attention adapter.
Key results
- On CATH v4.2, Bridge-IF (ESM-2 650M) reaches 58.59% all-chain recovery vs. 57.38% for KW-Design, and 3.83 all-chain perplexity.
- On CATH v4.3, Bridge-IF with ESM-2 3B reaches a reported 61.27% recovery, and the 650M-scale variant reaches 58.93% with ESM-1b.
- In de novo design with Chroma backbones, Bridge-IF beats ProteinMPNN on scTM (0.73 vs 0.69) and designability (0.85 vs 0.80).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Bridge-IF instead of the strongest listed baseline for CATH v4.2 All | Bridge-IF (ESM-2 650M) · CATH v4.2 All | recovery rate: 57.38 -> 58.59 | Helped |
| Use Bridge-IF instead of LM-Design for foldability | Foldability test set · Foldability test set | TM-score: 0.73 -> 0.81 | Helped |
| Use Bridge-IF instead of ProteinMPNN for de novo design | De novo protein design · De novo protein design | scTM: 0.69 vs. 0.73 | Helped |
| Pre-train and freeze the structure encoder before training the PLM, rather than training them jointly (ablation) | Bridge-IF · CATH v4.2 | recovery rate: 56.53 -> 58.59 | Helped |
| Add AdaLN-Bias instead of using vanilla AdaLN (ablation) | Bridge-IF · CATH v4.2 All | recovery rate: 57.92% -> 58.59% | Helped |
| Use simplified cross-entropy loss instead of the variational lower bound loss (ablation) | Bridge-IF · CATH v4.2 All | recovery rate: 56.43 -> 58.59 | Helped |
Surprising findings
- Discriminative models with high recovery (PiFold, LM-Design) score lower on TM-score foldability than probabilistic models like Bridge-IF and ProteinMPNN.
Limitations
- No wet-lab validation of designed sequences.
- Relies on a pre-trained PiFold structure encoder as the prior, so performance depends on the quality of that encoder.
Experimental setup
Datasets: CATH v4.2 CATH v4.3 PDB multi-chain complex dataset (30% identity clusters)
Benchmarks: CATH v4.2 CATH v4.3 PDB multi-chain complexes 82-sample foldability test set Chroma de novo backbones
Metrics: perplexity median recovery rate TM-score scTM designability
Baselines: ProteinMPNN PiFold GraDe-IF LM-Design KW-Design ESM-IF GVP GraphTrans
Builds on: Markov bridge models discrete diffusion models PiFold LM-Design DiT adaLN
Closest papers by meaning
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.90)
- Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.90)
- Learning to engineer protein flexibility (ICLR 2025 · similarity 0.89)
- All-atom inverse protein folding through discrete flow matching (ICML 2025 · similarity 0.88)
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.88)
- Zero-shot protein stability prediction by inverse folding models: a free energy interpretation (NeurIPS 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.