Research map / Probabilistic and scientific modeling / Molecules and proteins
PScalpel: A Machine Learning-based Guider for Protein Phase-Separating Behaviour Alteration
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 PScalpel, a graph-based contrastive learning and genetic algorithm framework that predicts how missense mutations alter protein LLPS propensity and recommends mutations with desired phase-separation behaviour.
Problem
Missense mutations can alter protein liquid-liquid phase separation (LLPS), linked to diseases like Alzheimer's and cancer, but existing LLPS predictors fail to distinguish LLPS propensity among highly similar mutant sequences and labeled mutant data is scarce.
Approach
Combines a distilled AlphaFold2-like contact-map predictor (BetaFold) with Twin Towers Graph Contrastive Learning that jointly trains two auxiliary tasks (ClinVar benign/pathogenic mutants and LLPSDB/PDB phase-separating vs. non-separating proteins), then uses a genetic algorithm with the trained model as fitness function to recommend mutations.
Key results
- PSDM reached AUROC 0.7226 and F1 0.6474 on mutant proteins, versus 0.5403-0.6445 AUROC for PSPredictor, PSAP, DeePhase and FNN.
- TLPSDM achieved AUROC 0.776 on cGAS and 0.667 on TDP-43, the highest among compared methods.
- BetaFold reached an AUROC of about 0.87 on PDB subsets and was reported as outperforming ResPRE, DeepHomo and DeepHomo2 on CASP14.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PSDM instead of the strongest listed baseline | PSDM · Mutant proteins | AUROC: 0.6445 -> 0.7226 | Helped |
| Use PSDM instead of the strongest listed baseline | PSDM · Mutant proteins | F1-score: 0.5310 -> 0.6474 | Helped |
| Add transfer learning by freezing PSDM layers and fine-tuning the final layer | PSDM · cGAS | AUROC: 0.756 -> 0.776 | Helped |
| Add transfer learning by freezing PSDM layers and fine-tuning the final layer | PSDM · TDP-43 | AUROC: 0.612 -> 0.667 | Helped |
| Add the BetaFold-predicted protein graph structure with GAT (ablation) | PSDM · Mutant protein phase-separation dataset | AUROC and F1 Score | Helped |
| Add T3GCL for feature extraction (ablation) | PSDM · Mutant protein phase-separation dataset | AUROC and F1 Score | Helped |
Surprising findings
- Existing natural-protein LLPS predictors (AUROC 0.91-0.97 on natural proteins) drop to near-random (0.54-0.57 AUROC) on mutant proteins.
Limitations
- Mutant evaluation set (283 sequences) was used only for testing and is small; TLPSDM validated on only two single-protein datasets.
- Mutation search limited to at most k=2 sites per sample.
- Wet-lab validation of recommended mutants is not clearly reported in the text.
Experimental setup
Datasets: AlphaFold Protein Structure Database RCSB PDB NCBI ClinVar LLPSDB CASP14 cGAS mutant dataset TDP-43 mutant dataset
Benchmarks: CASP14 (34 proteins) RCSB PDB test (731 proteins) LLPSDB mutant proteins (283) cGAS dataset TDP-43 dataset
Metrics: AUROC F1-score precision
Baselines: PSPredictor PSAP DeePhase FNN PSPPI ResPRE DeepHomo DeepHomo2
Builds on: AlphaFold2 Contrastive learning (SimCLR-style) Graph Attention Network ProtVec Performer kernel attention
Closest papers by meaning
- Towards precision protein-ligand affinity prediction benchmark: A Complete and Modification-Aware DAVIS Dataset (NeurIPS 2025 · similarity 0.83)
- PSBench: a large-scale benchmark for estimating the accuracy of protein complex structural models (NeurIPS 2025 · similarity 0.83)
- Enhancing Protein Mutation Effect Prediction through a Retrieval-Augmented Framework (NeurIPS 2024 · similarity 0.83)
- Predicting mutational effects on protein binding from folding energy (ICML 2025 · similarity 0.82)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.82)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.82)
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.