Research map / Probabilistic and scientific modeling / Molecules and proteins
Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions
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 Boltzmann Alignment, which uses ProteinMPNN log-likelihoods within a thermodynamic cycle (bound vs. unbound chains) for unsupervised (BA-Cycle) and fine-tuned supervised (BA-DDG) ΔΔG prediction on SKEMPI v2.
Problem
Predicting binding free energy changes (ΔΔG) from mutations is limited by scarce experimental labels, and existing inverse-folding pre-training approaches use plain supervised fine-tuning without alignment or explicit treatment of the unbound state.
Approach
Derive ΔΔG from the Boltzmann distribution and Bayes' theorem as a difference of inverse-folding log-likelihoods for the bound complex versus the separately evaluated unbound chains (a thermodynamic cycle), then fine-tune ProteinMPNN with a Boltzmann alignment loss that adds a KL penalty to the pre-trained distribution.
Key results
- BA-Cycle (training-free) reaches Spearman 0.3201 unsupervised on SKEMPI v2, versus prior SoTA 0.2632 (RDE-Linear).
- BA-DDG reaches Spearman 0.5134 supervised on SKEMPI v2, versus prior SoTA 0.4324 (Surface-VQMAE).
- On SAbDab binding energy, BA-DDG achieves Spearman 0.385, outperforming other methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| include the unbound state in the thermodynamic cycle (ablation) | ProteinMPNN inverse-folding scoring (BA-Cycle) · SKEMPI v2 | overall Spearman: 0.3112 -> 0.4097 | Helped |
| replace bound-state-only scoring with BA-Cycle (ablation) | ProteinMPNN inverse-folding scoring · SKEMPI v2 | overall Spearman: 0.2806 -> 0.4097 | Helped |
| use Boltzmann supervision instead of SFT (ablation) | BA-DDG · SKEMPI v2 | overall Spearman: 0.6240 -> 0.6346 | Helped |
| use Boltzmann supervision instead of DPO (ablation) | BA-DDG · SKEMPI v2 | overall Spearman: 0.4870 -> 0.6346 | Helped |
| use AlphaFold3 predicted structures instead of crystal structures (ablation) | BA-DDG · SKEMPI common subset | overall Spearman: 0.7946 -> 0.7821 | Hurt |
| fine-tune ProteinMPNN with BA-DDG rather than use the original ProteinMPNN | ProteinMPNN · SARS-CoV-2 antibody mutations | average normalized per-amino acid perplexity: 2.403 -> 1.325 | Helped |
| fine-tune ProteinMPNN with BA-DDG rather than use the original ProteinMPNN | ProteinMPNN · SARS-CoV-2 antibody mutations | average preference: 42.47% -> 47.03% | Helped |
Surprising findings
- Boltzmann-style supervision outperforms SFT and weighted DPO on the same backbone, and DPO underperforms SFT here.
Limitations
- Requires crystal or predicted complex structures as input; mutant backbones are assumed equal to wild-type.
- Does not model side-chain conformations.
- Multi-point mutation performance is weaker than single-point due to autoregressive decoding coupling.
Experimental setup
Datasets: SKEMPI v2 SAbDab
Benchmarks: SKEMPI v2 SAbDab binding energy subset DIPS test set (docking)
Metrics: Pearson Spearman RMSE MAE AUROC per-structure Pearson/Spearman C-RMSD success fraction
Baselines: Rosetta FoldX ESM-IF ESM-1v MIF-Network Prompt-DDG ProMIM Surface-VQMAE
Builds on: ProteinMPNN direct preference optimization
Closest papers by meaning
- Predicting mutational effects on protein binding from folding energy (ICML 2025 · similarity 0.90)
- Zero-shot protein stability prediction by inverse folding models: a free energy interpretation (NeurIPS 2025 · similarity 0.89)
- Aligning Protein Conformation Ensemble Generation with Physical Feedback (ICML 2025 · similarity 0.87)
- Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks (NeurIPS 2026 · similarity 0.86)
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.86)
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · 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.