Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment
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.
Builds a 5.5M pseudo-ligand-pocket dataset by fragmenting PDB protein chains and pretrains a pocket encoder via contrastive alignment to a frozen pretrained Uni-Mol molecule encoder (ProFSA).
Problem
Pocket representations usually ignore ligand interactions because only ~100k non-redundant pocket-ligand complexes exist in the PDB, limiting large-scale pretraining for interaction modeling.
Approach
Treat protein fragments as pseudo-ligands and their surrounding residues as pockets, creating millions of pseudo-complexes from protein-only data, then align the pocket encoder contrastively to fixed pretrained small-molecule encoder embeddings of the fragments.
Key results
- Zero-shot druggability RMSE improves over Uni-Mol (e.g., Total SASA 30.50 vs 49.00; Fpocket 0.1228 vs 0.1419).
- Pocket matching finetuned: Kahraman AUC 0.85 vs Uni-Mol 0.71; TOUGH-M1 AUC 0.94 vs DeeplyTough 0.91.
- LBA at 30% identity: RMSE 1.377 vs best prior 1.403 (EGNN-PLM), Pearson 0.628 vs 0.577 (GeoSSL).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pretrain pocket representations with fragment-surroundings alignment | ProFSA vs Uni-Mol · Druggability prediction, zero-shot | Fpocket RMSE: 0.1419 -> 0.1228 | Helped |
| pretrain pocket representations with fragment-surroundings alignment | ProFSA vs Uni-Mol · Pocket matching, Kahraman (w/o PO4), zero-shot | AUC: 0.66 -> 0.80 | Helped |
| pretrain pocket representations with fragment-surroundings alignment | ProFSA vs Uni-Mol · PDBBind v2019, sequence identity 30% | Pearson: 0.558 -> 0.628 | Helped |
| pretrain pocket representations with fragment-surroundings alignment | ProFSA vs Uni-Mol · Ligand efficacy prediction | AUPRC: 0.695±0.07 -> 0.806±0.04 | Helped |
| use FRAD instead of Uni-Mol as molecular encoder (ablation) | ProFSA · Pocket matching, Kahraman (w/o PO4) | AUC: 0.7870 -> 0.7401 | Hurt |
| use PTVD instead of Uni-Mol as molecular encoder (ablation) | ProFSA · Pocket matching, TOUGH-M1 | AUC: 0.8178 -> 0.7548 | Hurt |
| remove distributional alignment (ablation) | ProFSA · PDBBind v2019, sequence identity 30% | RMSE: 1.353 -> 1.441 | Hurt |
| remove distributional alignment (ablation) | ProFSA · Pocket matching, TOUGH-M1 | AUC: 0.8178 -> 0.7589 | Hurt |
| remove fixed molecular encoder (ablation) | ProFSA · Pocket matching, Kahraman (w/o PO4) | AUC: 0.7870 -> 0.6905 | Hurt |
| remove fixed molecular encoder (ablation) | ProFSA · Druggability prediction (zero-shot), Total Sasa | RMSE: 31.17 -> 32.17 | Hurt |
| reduce pocket distance threshold from 6Å to 4Å (ablation) | ProFSA · Pocket matching, Kahraman (w/o PO4) | AUC: 0.7870 -> 0.7062 | Hurt |
| increase pocket distance threshold from 6Å to 8Å (ablation) | ProFSA · Pocket matching, Kahraman (w/o PO4) | AUC: 0.7870 -> 0.8322 | Helped |
Surprising findings
- Zero-shot ProFSA beats the finetuned IsoMIF-level performance on Kahraman, indicating interaction knowledge transfer from protein-only data.
- Pretraining with pseudo-ligands from protein structures needs no pocket-ligand complex labels.
Limitations
- Pseudo-ligands (peptide fragments) differ from real ligands in properties such as logP and rotatable bonds.
- Pocket-ligand interaction types such as rare non-canonical ones are only partly represented in synthetic data.
Experimental setup
Datasets: ProFSA pseudo-complex dataset (from PDB) PDBBind v2019 Kahraman dataset TOUGH-M1 dataset Uni-Mol druggability dataset BioLip2
Benchmarks: Uni-Mol druggability dataset Kahraman pocket matching TOUGH-M1 PDBBind LBA (Atom3D 30%/60% split) ATOM3D ligand efficacy
Metrics: RMSE AUC AUROC Pearson Spearman AUPRC
Baselines: Uni-Mol CoSP DeeplyTough IsoMIF GeoSSL EGNN-PLM ProNet TM-align
Builds on: Uni-Mol DeeplyTough ATOM3D
Closest papers by meaning
- Generalized Protein Pocket Generation with Prior-Informed Flow Matching (NeurIPS 2024 · similarity 0.88)
- Protein-ligand binding representation learning from fine-grained interactions (ICLR 2024 · similarity 0.88)
- A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning (NeurIPS 2026 · similarity 0.87)
- Pre-Training Protein Bi-level Representation Through Span Mask Strategy On 3D Protein Chains (ICML 2024 · similarity 0.86)
- DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction (NeurIPS 2025 · similarity 0.86)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.85)
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.