Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization
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 BounDr.E, a one-class drug-likeness predictor that aligns molecular fingerprints with knowledge-graph drug embeddings and tightens a hyperspherical boundary around approved drugs via an EM-like loop.
Problem
Drug-likeness prediction lacks reliable negatives, and existing methods rely on ambiguous negative sets or purely structural features, yielding poorly separated or overly broad decision boundaries.
Approach
Model drug-likeness as a compact hypersphere around approved drugs in a space where structural fingerprints are aligned with biomedical knowledge-graph embeddings (softened CLIP with ATC similarity plus geodesic mixup), and iteratively refine the boundary with an EM-like procedure that pushes pseudo-negative non-drugs outside.
Key results
- On the time-based split, BounDr.E reaches F1 0.846 vs 0.740 for DeepDL (the previous best), with ICR 0.009 vs 0.054.
- Zero-shot false-positive rate on withdrawn drugs is 0.51 vs 0.91 for DeepDL and 0.96 for FP-XGB.
- In a targeted pipeline on 10,543 BCR-ABL generated compounds, BounDr.E filtering followed by other filters yields 38 candidates (0.36%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BOUNDR.E instead of DeepDL | Drug-likeness prediction · DrugBank vs ZINC20, time-based split | F1: 0.740 (0.0584) -> 0.846 (0.0165) | Helped |
| use BOUNDR.E instead of FP-OCSVM | Zero-shot toxic compound filtering · Withdrawn toxic compounds | false-positive rate: 0.69 (0.002) -> 0.51 (0.014) | Helped |
| use BOUNDR.E instead of FP-OCSVM | Zero-shot toxic compound filtering · Hepatotoxic compounds | false-positive rate: 0.53 (0.003) -> 0.54 (0.009) | Hurt |
| use BOUNDR.E instead of FP-OCSVM | Zero-shot toxic compound filtering · Cardiotoxic compounds | false-positive rate: 0.25 (0.006) -> 0.20 (0.009) | Helped |
| use BOUNDR.E instead of DrugMetric | Zero-shot toxic compound filtering · Carcinogenic compounds | false-positive rate: 0.82 (0.087) -> 0.19 (0.014) | Helped |
| add knowledge-integrated alignment rather than using fingerprints alone (ablation) | BOUNDR.E · Drug-like compound identification | F1: 0.54 (0.032) -> 0.85 (0.017) | Helped |
| replace manifold alignment with the proposed alignment (ablation) | BOUNDR.E · Drug-like compound identification | F1: 0.40 (0.045) -> 0.85 (0.017) | Helped |
| replace CLIP with softened CLIP and geodesic mixup (ablation) | BOUNDR.E · Drug-like compound identification | F1: 0.59 (0.022) -> 0.85 (0.017) | Helped |
| add softened CLIP and geodesic mixup rather than using Geodesic Mixup alone (ablation) | BOUNDR.E · Drug-like compound identification | F1: 0.69 (0.045) -> 0.85 (0.017) | Helped |
| add geodesic mixup to softened CLIP alignment (ablation) | BOUNDR.E · Drug-like compound identification | F1: 0.73 (0.037) -> 0.85 (0.017) | Helped |
| use EM-like boundary optimization instead of MLP, SVM, XGB, naive PU, or DeepSVDD (ablation) | BOUNDR.E on aligned space · Drug-like compound identification | F1: +0.08 vs MLP; -0.01 vs SVM; +0.10 vs XGB; +0.03 vs naive PU; +0.53 vs DeepSVDD | Mixed |
| add EM-like optimization (ablation) | BOUNDR.E · Drug-like compound identification | F1: 0.44 (0.162) -> 0.85 (0.017) | Helped |
Surprising findings
- Supervised and PU baselines degrade heavily under cross-dataset shift while the one-class approach is largely unaffected.
- Replacing ZINC non-drugs with property-matched PubChem compounds lowers performance for all models, showing that property-matched backgrounds are harder.
Limitations
- Generalization to novel modalities such as PROTACs or rare-disease compounds is acknowledged as limited.
- EM-like procedure lacks theoretical guarantees for global optima and remains sensitive to initialization.
Experimental setup
Datasets: DrugBank v5.1.12 ZINC20 PubChem ChEMBL ATC classification Wu et al. 2023 toxicity sets
Benchmarks: DrugBank vs ZINC20 time-based split scaffold-based split cross-dataset PubChem/ChEMBL to ZINC toxic compound sets (withdrawn, hepatotoxic, cardiotoxic, carcinogenic)
Metrics: F1 IDR ICR AUROC Average Precision MCC false-positive rate
Baselines: FP-Similarity-cutoff FP-KNN classifier FP-SVM FP-XGB FP-OCSVM FP-DeepSVDD FP-nnPU FP-PU with NN DrugMetric D-GCAN DeepDL
Builds on: CLIP Geodesic Mixup Deep SVDD DeepDL D-GCAN
Closest papers by meaning
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.87)
- Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.86)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.84)
- Monroe: A Molecular Foundation Model for In-context Probabilistic Inference (NeurIPS 2026 · similarity 0.84)
- PharmacoMatch: Efficient 3D Pharmacophore Screening via Neural Subgraph Matching (ICLR 2025 · similarity 0.84)
- MF-LAL: Drug Compound Generation Using Multi-Fidelity Latent Space Active Learning (ICML 2025 · 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.