Research map / Architectures and efficiency / RNNs and normalization
DrugHash: Hashing Based Contrastive Learning for Virtual Screening
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 DrugHash, a hashing-based contrastive learning method that learns end-to-end binary hash codes for protein pockets and molecules to make zero-shot virtual screening faster and far less memory-hungry.
Problem
Virtual screening over billion-scale molecular libraries is too slow with docking, and real-valued embedding representations from contrastive methods like DrugCLIP incur large memory and retrieval costs.
Approach
Treats virtual screening as cross-modal retrieval and learns binary hash codes end-to-end for both protein and molecule encoders by adding a regularizing loss that pulls real-valued embeddings toward sign-based binary codes, on top of an InfoNCE contrastive objective.
Key results
- On DUD-E, DrugHash reaches AUROC 83.73 and BEDROC 57.16 versus DrugCLIP's 79.45 and 47.82, with EF0.5% of 43.03 vs 37.86.
- On LIT-PCBA, DrugHash achieves the best BEDROC (7.14) and EF0.5% (9.65), but its AUROC (54.58) trails Banana (62.78).
- Storing the Enamine REAL database (~6.5B molecules) needs 96GB with DrugHash versus 3080GB for DrugCLIP and Banana (32x saving); retrieval over REAL takes 301 s vs 1064 s for DrugCLIP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DrugCLIP real-valued retrieval representations with the DrugHash binary hashing strategy | DrugCLIP · DUD-E | AUROC: 79.45 -> 83.73 | Helped |
| replace DrugCLIP real-valued retrieval representations with the DrugHash binary hashing strategy | DrugCLIP · LIT-PCBA | EF^0.5%: 7.77 -> 9.65 | Helped |
| replace DrugCLIP real-valued retrieval representations with the DrugHash binary hashing strategy | DrugCLIP · LIT-PCBA | AUROC: 56.36 -> 54.58 | Hurt |
| replace DrugCLIP real-valued retrieval representations with the DrugHash binary hashing strategy | DrugCLIP · REAL (~6.5B) | memory cost: 3080GB -> 96GB | Helped |
| replace DrugCLIP real-valued retrieval representations with the DrugHash binary hashing strategy | DrugCLIP · REAL (~6.5B) | time cost: 1064 seconds -> 301 seconds | Helped |
| increase code length from 128 to 256 (ablation) | DrugHash · DUD-E | AUROC: 83.73 -> 84.96 | Helped |
| increase code length from 128 to 512 (ablation) | DrugHash · DUD-E | BEDROC: 57.16 -> 60.50 | Helped |
| set λ to 0, removing the hashing training objective (ablation) | DrugHash · DUD-E | accuracy | Hurt |
| vary λ from 0 to 1.0 (ablation) | DrugHash · DUD-E | accuracy | Mixed |
| add the hashing strategy during training (ablation) | DrugHash · validation set | BEDROC | Helped |
Surprising findings
- Binary hash codes improve accuracy over real-valued embeddings, which the authors attribute to regularization against overfitting.
- Without the hashing loss (lambda=0) accuracy drops sharply even though binary codes are still used at inference.
Limitations
- Only one simple sign-based hashing strategy explored; other binarization methods not studied.
- Trained only on PDBBind-scale data; effect of larger complex datasets untested.
- Encoder choice not explored beyond the Uni-Mol-based encoder.
Experimental setup
Datasets: PDBBind HomoAug DUD-E LIT-PCBA ZINC Enamine REAL CASF-2016
Benchmarks: DUD-E LIT-PCBA
Metrics: AUROC BEDROC EF0.5% EF1% EF5% memory cost time cost
Baselines: Glide-SP AutoDock Vina NN-score RF-Score Pafnucy Planet Banana DrugCLIP Surflex Gnina DeepDTA
Builds on: DrugCLIP Uni-Mol CLIP InfoNCE
Closest papers by meaning
- AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation (NeurIPS 2025 · similarity 0.86)
- PharmacoMatch: Efficient 3D Pharmacophore Screening via Neural Subgraph Matching (ICLR 2025 · similarity 0.86)
- Drug-TTA: Test-Time Adaptation for Drug Virtual Screening via Multi-task Meta-Auxiliary Learning (ICML 2025 · similarity 0.85)
- CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening (ICLR 2025 · similarity 0.85)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.84)
- S-MolSearch: 3D Semi-supervised Contrastive Learning for Bioactive Molecule Search (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.