Research map / Probabilistic and scientific modeling / Molecules and proteins
Fast Uncovering of Protein Sequence Diversity from Structure
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 InvMSAFold, an inverse folding model that maps a backbone structure to parameters of a low-rank Potts/pairwise (or autoregressive) sequence distribution trained on MSAs, enabling fast, diverse CPU sampling of sequences.
Problem
Inverse folding models trained only on native sequences give peaked, low-diversity outputs and sampling from autoregressive transformers like ESM-IF1 and ProteinMPNN is expensive for large-scale virtual screening.
Approach
Train a GVP-based encoder plus transformer decoder to output low-rank pairwise (Potts) couplings and fields for a structure, supervised on MSAs of homologs rather than the single native sequence, so the generated lightweight model can be sampled quickly on CPU; two variants: pseudo-likelihood Potts (PW) and autoregressive (AR).
Key results
- InvMSAFold-AR gives the best covariance reconstruction, with average KL on PCA projections of 0.49 (inter-cluster) and 0.67 (intra-cluster) versus 15.8 and 11.9 for ESM-IF1.
- InvMSAFold-AR sampling on a single CPU core is orders of magnitude faster than ESM-IF1 on GPU, with near-constant time across lengths when batched.
- At high normalized Hamming distance, ESM-IF1 refolding quality drops below both InvMSAFold variants, which remain robust.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InvMSAFold-AR instead of ESM-IF1 to generate sequences | Protein sequence generation · Inter-Cluster | KL divergence: 15.8 -> 0.49 | Helped |
| Use InvMSAFold-AR instead of ESM-IF1 to generate sequences | Protein sequence generation · Intra-Cluster | KL divergence: 11.9 -> 0.67 | Helped |
| Use InvMSAFold-PW instead of ESM-IF1 to generate sequences | Protein sequence generation · Inter-Cluster | KL divergence: 15.8 -> 4.6 | Helped |
| Use InvMSAFold-PW instead of ESM-IF1 to generate sequences | Protein sequence generation · Intra-Cluster | KL divergence: 11.9 -> 11.2 | No clear effect |
| Use InvMSAFold-AR rather than InvMSAFold-PW | InvMSAFold · Inter-Cluster | KL divergence: 4.6 -> 0.49 | Helped |
| Use InvMSAFold-AR rather than InvMSAFold-PW | InvMSAFold · Intra-Cluster | KL divergence: 11.2 -> 0.67 | Helped |
| Generate sequences using InvMSAFold-AR rather than ESM-IF1 | Protein sequence generation · Covariance reconstruction, inter-cluster and intra-cluster test sets | Pearson correlation of amino-acid covariances | Helped |
| Generate sequences using InvMSAFold-AR rather than InvMSAFold-PW | InvMSAFold · Covariance reconstruction | Pearson correlation of amino-acid covariances | Helped |
| Generate sequences using InvMSAFold-AR rather than ESM-IF1 | Protein sequence generation · Sequence patterns in natural-homolog PCA space | coverage of sequence space | Helped |
| Generate sequences using InvMSAFold-PW rather than ESM-IF1 | Protein sequence generation · Sequence patterns in natural-homolog PCA space | coverage of sequence space | Helped |
| Generate sequences using InvMSAFold-AR or InvMSAFold-PW rather than ESM-IF1 at larger native-sequence distances | Protein sequence generation · Intra-cluster predicted structures | RMSD to original structure | Helped |
| Use InvMSAFold-AR rather than ESM-IF1 for sampling | Protein sequence generation · CPU sampling speed | sampling speed | Helped |
Surprising findings
- ESM-IF1 samples concentrate tightly around the native sequence, and its covariance reconstruction is worse than both InvMSAFold variants.
- InvMSAFold-PW, trained with pseudo-likelihood and sampled via MCMC, underperforms the autoregressive variant on covariance reconstruction.
Limitations
- Pseudo-likelihood training of the PW variant does not capture covariances well and requires MCMC sampling.
- Evaluation relies on predicted properties (Thermoprot, Protein-Sol) and AlphaFold refolding rather than experimental validation.
- Biosecurity concerns: generated sequences could evade screening, though the authors argue homology detection mitigates this.
Experimental setup
Datasets: CATH UniProt50 (MSAs via MMseqs2)
Benchmarks: CATH inter-cluster test set CATH intra-cluster test set MSA test set
Metrics: Pearson correlation of covariances KL divergence on PCA projections normalized Hamming distance RMSD after AlphaFold 2 refolding sampling speed predicted thermostability and solubility
Baselines: ESM-IF1 ProteinMPNN
Builds on: ESM-IF1 GVP-GNN arDCA bmDCA
Closest papers by meaning
- ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.89)
- Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.88)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.87)
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.87)
- AlphaFold Database Debiasing for Robust Inverse Folding (NeurIPS 2025 · similarity 0.87)
- All-atom inverse protein folding through discrete flow matching (ICML 2025 · similarity 0.86)
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.