Research map / Label-efficient and robust learning / Semi-supervised learning
S-MolSearch: 3D Semi-supervised Contrastive Learning for Bioactive Molecule Search
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 S-MolSearch, a semi-supervised contrastive learning framework for ligand-based virtual screening that combines ChEMBL affinity-labeled pairs with ~19M unlabeled molecules, using smooth optimal transport to generate soft labels from a teacher encoder.
Problem
Ligand-based virtual screening is limited by scarce, noisy binding affinity labels covering a small chemical space, and standard data augmentation does not preserve molecular activity for 3D molecules.
Approach
Two Uni-Mol-initialized encoders: a supervised encoder trained with InfoNCE on affinity-defined clusters from ChEMBL, and a full-data encoder trained on soft labels obtained by smooth optimal transport over the supervised encoder's similarity matrix, motivated by an inverse optimal transport view of contrastive learning.
Key results
- On DUD-E (zero-shot), S-MolSearch(0.9) reaches 92.56% AUROC, 75.37% BEDROC, and 47.94% EF 1%, versus DrugCLIP's 80.93%, 50.52%, and 31.89%.
- On LIT-PCBA (zero-shot), S-MolSearch(0.9) achieves 61.78% AUROC and 8.48% BEDROC, outperforming Gnina (60.93%) and DrugCLIP (6.23% BEDROC).
- Ablation on DUD-E shows removing soft labels drops EF 1% from 34.60 to 33.32 (regularizer removed) or 30.73 (soft labels removed, using hard labels).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a 0.9 rather than 0.4 similarity threshold for ChEMBL filtering | S-MolSearch · DUD-E zero-shot | BEDROC (%): 54.22 -> 75.37 | Helped |
| Use a 0.9 rather than 0.4 similarity threshold for ChEMBL filtering | S-MolSearch · LIT-PCBA zero-shot | AUROC (%): 57.34 -> 61.78 | Helped |
| Add IOT soft labels instead of using hard similarity labels (ablation) | S-MolSearch · DUD-E | EF 1%: 30.73 -> 34.60 | Helped |
| Add IOT soft labels instead of using hard similarity labels (ablation) | S-MolSearch · LIT-PCBA | EF 1%: 6.19 -> 6.28 | Helped |
| Add KoLeo regularizer (ablation) | S-MolSearch · DUD-E | EF 1%: 33.32 -> 34.60 | Helped |
| Add KoLeo regularizer (ablation) | S-MolSearch · LIT-PCBA | EF 1%: 5.24 -> 6.28 | Helped |
| Use Uni-Mol pretrained checkpoint instead of random initialization (ablation) | S-MolSearch · DUD-E | EF 1%: 31.86 -> 34.60 | Helped |
| Use Uni-Mol pretrained checkpoint instead of random initialization (ablation) | S-MolSearch · LIT-PCBA | EF 1%: 5.24 -> 6.28 | Helped |
| Use semi-supervised S-MolSearch instead of scaffold-cluster self-supervised learning (ablation) | S-MolSearch · DUD-E | EF 1%: 20.13 -> 34.60 | Helped |
| Use semi-supervised S-MolSearch instead of scaffold-cluster self-supervised learning (ablation) | S-MolSearch · LIT-PCBA | EF 1%: 3.26 -> 6.28 | Helped |
| Use semi-supervised S-MolSearch instead of self-supervised pretraining followed by supervised finetuning (ablation) | S-MolSearch · DUD-E | EF 1%: 29.20 -> 34.60 | Helped |
| Use semi-supervised S-MolSearch instead of self-supervised pretraining followed by supervised finetuning (ablation) | S-MolSearch · LIT-PCBA | EF 1%: 5.71 -> 6.28 | Helped |
Surprising findings
- Performance on LIT-PCBA is much lower than on DUD-E for all methods, which the authors attribute to cell-based assays introducing noise into actives.
- The 0.4 similarity-threshold variant has lower AUROC than the 0.9 variant on LIT-PCBA, while the 0.4 variant is preferred for avoiding overlap with benchmark targets.
Limitations
- Limited interpretability compared with traditional 3D shape-based search methods.
- Two encoders increase memory consumption relative to single-encoder methods.
- Focuses only on molecular affinity data, omitting broader biochemical interactions.
Experimental setup
Datasets: ChEMBL v33 Uni-Mol unlabeled molecule corpus (~19M) DUD-E LIT-PCBA
Benchmarks: DUD-E LIT-PCBA
Metrics: AUROC BEDROC EF 0.5% EF 1% EF 5%
Baselines: ROCS Phase Shape LIGSIFT SHAFTS Glide-SP Vina Pafnucy OnionNet Planet DrugCLIP Gnina DeepDTA
Builds on: Uni-Mol InfoNCE inverse optimal transport KoLeo regularizer DrugCLIP
Closest papers by meaning
- 3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery (ICLR 2025 · similarity 0.86)
- AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation (NeurIPS 2025 · similarity 0.85)
- UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery (ICLR 2025 · similarity 0.85)
- PharmacoMatch: Efficient 3D Pharmacophore Screening via Neural Subgraph Matching (ICLR 2025 · similarity 0.85)
- Offline Model-based Optimization for Real-World Molecular Discovery (ICML 2025 · similarity 0.85)
- TurboHopp: Accelerated Molecule Scaffold Hopping with Consistency Models (NeurIPS 2024 · similarity 0.84)
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.