Research map / Probabilistic and scientific modeling / Molecules and proteins

Predicting mutational effects on protein binding from folding energy

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein interface and docking prediction · 2 citations (OpenAlex)

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 StaB-ddG, a transfer-learning predictor for protein-protein binding ddG that parameterizes binding energy as folding energy of the complex minus its partners, initialized from ProteinMPNN and fine-tuned on Megascale stability and SKEMPIv2 binding data.

Problem

Predicting mutational effects on protein-protein binding free energy (ddG_bind) is hard because binding affinity data are scarce, and deep learning predictors underperform force-field methods like FoldX and Flex ddG.

Approach

Use the thermodynamic state-function identity to write binding energy as complex folding energy minus the folding energies of the binding partners, with folding energy given by ProteinMPNN log-likelihoods; fine-tune first on folding stability data and then on binding data, with antithetic variates and Monte Carlo ensembling to reduce variance.

Key results

• StaB-ddG achieves per interface Spearman 0.45 on the SKEMPIv2.0 test split, outperforming prior deep learning predictors, though not significantly different from FoldX (0.48) or Flex ddG (0.42).
• Averaging StaB-ddG and FoldX gives per interface Spearman 0.53, the best among all methods reported.
• StaB-ddG runs about 0.2 seconds per mutation versus about 210 seconds per mutation for FoldX, reported as a roughly 1000x speedup.

Measured effects reported in the paper

Surprising findings

• Zero-shot StaB-ddG (using ProteinMPNN weights with variance reduction) is competitive with fine-tuned variants and outperforms prior DL methods.
• On the TCR mimic case study the predictor achieves only Spearman 0.13 ± 0.39, so the authors cannot conclude it is predictive there.

Limitations

• Does not model backbone changes upon mutation, which matters for flexible interfaces.
• Performance is better on rigid, smaller complexes, reflecting bias in the folding stability data toward small domains under 80 residues.
• Yeast display case study shows the binding proxy is confounded by expression, and TCR mimic case study is inconclusive due to small sample size.

Experimental setup

Datasets: Megascale stability dataset SKEMPIv2.0 TCR3d database yeast surface display site saturation mutagenesis dataset

Benchmarks: SKEMPIv2.0 (re-split, 80-81 test complexes) yeast surface display mini-binder dataset TCR mimic case study

Metrics: per interface Spearman per interface Pearson RMSE overall Pearson overall Spearman AUROC

Baselines: FoldX Flex ddG Surface-VQMAE RDE-Net Prompt-DDG DiffAffinity PPIformer ProMIM ESM-IF1 ProteinMPNN

Builds on: ProteinMPNN ThermoMPNN FoldX Flex ddG Megascale stability dataset

Closest papers by meaning

• Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions (ICLR 2025 · similarity 0.90)
• Zero-shot protein stability prediction by inverse folding models: a free energy interpretation (NeurIPS 2025 · similarity 0.89)
• Learning to engineer protein flexibility (ICLR 2025 · similarity 0.87)
• Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.87)
• Aligning Protein Conformation Ensemble Generation with Physical Feedback (ICML 2025 · similarity 0.87)
• A Physics-Augmented Deep Learning Framework for Classifying Single Molecule Force Spectroscopy Data (ICML 2025 · similarity 0.86)