Research map / 3D vision / Face and shape
Learning Compressed Shape-Aware Molecular Representations for Virtual Screening
Robin Winter, Julian Cremer, Djork-Arné Clevert
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train end-to-end with quantization-aware IVF-PQ instead of post-hoc FAISS k-means (ablation) | SAND · Enamine 10M subset, compression sweep | compression at equal correlation: M=16 matches post-hoc M=64 (about 4x better compression) | Improved |
| quantization-aware training changes embedding geometry (ablation) | SAND · 1M test molecules | principal components for >90% variance: 317 -> 83 | Improved |
| retrain FAISS codebooks post-hoc on quantization-aware model embeddings (ablation) | SAND · compression sweep | quantization performance | No clear effect |
| replace Spearman loss with triplet loss (ablation) | SAND (GINE) · Enamine, Recall@100 | recall: 0.50 -> 0.42 | Hurt |
| replace Spearman loss with MSE loss (ablation) | SAND (GINE) · Enamine, Recall@100 | recall: 0.50 -> 0.46 | Hurt |
| replace GINE (28M) with GPS (87M) encoder (ablation) | SAND · Enamine and ChEMBL | recall: 0.50 -> 0.51 (Enamine, 100) | No clear effect |
| replace GINE with GATv2 encoder (ablation) | SAND · Enamine | recall: 0.50 -> 0.50 | No clear effect |
| replace IVF-PQ with residual vector quantization (RVQ) (ablation) | SAND · Enamine / ChEMBL | Spearman correlation: Enamine 0.78 -> 0.84; ChEMBL 0.73 -> 0.81 | Mixed |
| apply 64x compression (M=32 quantization) (ablation) | SAND · Enamine / ChEMBL | Spearman correlation: 0.86 -> 0.78 (Enamine); 0.81 -> 0.73 (ChEMBL) | Hurt |
| train on shape overlay scores with rank-preserving contrastive loss vs ECFP/pretrained 3D baselines | SAND · Enamine, Recall@1000 | recall: 0.30 (ECFP) -> 0.89 | Improved |
| use pretrained 3D-aware representations (GraphMVP, DrugCLIP) instead of ECFP | shape retrieval baselines · Enamine | top-100 recall vs correlation: R 0.20 -> 0.37 but recall 0.13 -> 0.05 | Mixed |
| retrieve analogs with SAND instead of ECFP or random for Flowr-generated ligands | hit expansion pipeline · JNK3 and p38 MAPK docking | Chemgauss4 score and post-optimization RMSD | Improved |
Closest papers by meaning
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.87)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.87)
- 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.86)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.86)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.85)
- Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks (ICML 2024 · 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.
