Research map / Probabilistic and scientific modeling / Molecules and proteins
Enhancing Protein Mutation Effect Prediction through a Retrieval-Augmented Framework
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 a retrieval-augmented model (MSM-Mut with MSM-IPA) that retrieves local structure motifs from a PDB-wide ESM-IF embedding database to predict protein mutation effects on stability and binding affinity.
Problem
Predicting the effects of protein mutations (ddG for stability and binding/PPI surfaces) is hard because existing models fail to extract mutation-related local structure motifs from protein databases.
Approach
Build a vector database (SMEDB) of per-residue ESM-IF encoder embeddings for all PDB residues, retrieve similar local structure motifs via GPU kNN (CUHNSW), align them by central residue frame, and aggregate them with a multi-structure motif invariant point attention module to predict mutation ddG.
Key results
- On SKEMPI2.0, MSM-Mut reaches per-structure Pearson 0.4736 and Spearman 0.4354, above RDE-Net (0.4448/0.4010) and DiffAffinity (0.4220/0.3970).
- On S669, MSM-Mut achieves Pearson 0.54 and RMSE 1.51, outperforming Stability Oracle (0.52 / 1.43 RMSE is lower for Stability Oracle, so only Pearson is higher).
- On the Novozymes dataset, MSM-Mut reaches Spearman 0.484 versus 0.438 for Rosetta and 0.323 for MSM-Mut without retrieval.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add retrieval-augmented local structure motifs (ablation) | MSM-Mut · SKEMPI2.0 | Per-structure Pearson: 0.4325 -> 0.4736 | Helped |
| add retrieval-augmented local structure motifs (ablation) | MSM-Mut · S669 | Pearson: 0.45 -> 0.54 | Helped |
| add retrieval-augmented local structure motifs (ablation) | MSM-Mut · Novozymes | Spearman: 0.323 -> 0.484 | Helped |
| add pretraining (ablation) | MSM-Mut with 16 neighbors · S669 | Pearson: 0.43 -> 0.54 | Helped |
| increase retrieved neighbors from 1 to 16 (ablation) | MSM-Mut · S669 | Pearson: 0.49 -> 0.54 | Helped |
| increase retrieved neighbors from 16 to 1024 (ablation) | MSM-Mut · S669 | Pearson: 0.54 -> 0.51 | Hurt |
| select 16 motifs randomly from top 10000 rather than use top 16 neighbors (ablation) | MSM-Mut · S669 | Pearson: 0.54 -> 0.43 | Hurt |
| replace ESM-IF retrieval embeddings with CBAE (ablation) | MSM-Mut · SKEMPI2.0 | Per-structure Pearson: 0.4619 -> 0.4736 | Hurt |
| replace ESM-IF retrieval embeddings with CBAE (ablation) | MSM-Mut · S669 | Pearson: 0.52 -> 0.54 | Hurt |
| combine MSA-Profile with MSM-Profile (ablation) | mutation impact profiles · S669 | Pearson: 0.19 -> 0.23 | Helped |
| add retrieval-augmented local structure motifs (ablation) | MSM-Mut · S669 | RMSE: 1.43 -> 1.51 | Hurt |
Surprising findings
- Retrieved motifs can come from structurally unrelated proteins (low TM-score) yet still improve mutation prediction.
- A simple MSM-profile built from top-100 retrieved motifs outperforms MSA-profile and many unsupervised language models on SKEMPI2.0.
Limitations
- Performance is bounded by third-party submodules such as ESM-IF and CUHNSW.
- On S669 the model does not beat Stability Oracle on RMSE and the authors note it does not match Stability Oracle overall.
Experimental setup
Datasets: PDB PDB-REDO cDNA120K SKEMPI2.0 S669 Novozymes thermostability dataset
Benchmarks: SKEMPI2.0 S669 Novozymes dataset SARS-CoV-2 antibody optimization
Metrics: Pearson Spearman per-structure Pearson per-structure Spearman RMSE MAE
Baselines: Rosetta FoldX PSSM ESM-1v MSA Transformer Tranception ESM-IF ProteinMPNN MIF-Net RDE-Net DiffAffinity DDGPred
Builds on: ESM-IF AlphaFold2 invariant point attention CUHNSW RDE-Net Stability Oracle
Closest papers by meaning
- Learning to Predict Mutational Effects of Protein-Protein Interactions by Microenvironment-aware Hierarchical Prompt Learning (ICML 2024 · similarity 0.86)
- MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering (NeurIPS 2024 · similarity 0.86)
- Understanding protein function with a multimodal retrieval-augmented foundation model (NeurIPS 2025 · similarity 0.86)
- Predicting mutational effects on protein binding from folding energy (ICML 2025 · similarity 0.84)
- Evolution-Inspired Loss Functions for Protein Representation Learning (ICML 2024 · similarity 0.83)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.83)
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.