Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

PharmacoMatch: Efficient 3D Pharmacophore Screening via Neural Subgraph Matching

ICLR 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Zero-shot robustness and transfer · 0 citations (OpenAlex)

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 PharmacoMatch, a self-supervised GNN that embeds 3D pharmacophores into a non-negative order embedding space so that pharmacophore matching becomes fast vector comparison for zero-shot virtual screening pre-screening.

Problem

3D pharmacophore screening is computationally expensive for very large (billion-compound) libraries because each query must be aligned to every database conformer.

Approach

Reformulates pharmacophore matching as approximate neural subgraph matching: a GNN encoder trained contrastively with an order-embedding max-margin loss on augmented query-target pairs (point deletion and tolerance-sphere displacement), so database embeddings are precomputed and queries are matched by a cheap penalty comparison.

Key results

• On DUD-E 10 targets, PharmacoMatch AUROC is generally below CDPKit (e.g. ADA 83 vs 94, ANDR 76 vs 71 in favor of PharmacoMatch on some targets), with relative AUROC of 79.5-98.8% between the two methods' outputs.
• On DEKOIS2.0 and LIT-PCBA pre-screening, PharmacoMatch reaches BEDROC 15.1 vs PharmacoNet 12.3 on DEKOIS2.0 and AUROC 62.5 vs 60.9 favoring PharmacoNet, with better early enrichment (EF0.5% 5.5 vs 4.4).
• Vector matching takes 0.3 µs per pharmacophore versus 66 µs for alignment, roughly two orders of magnitude faster; PharmacoMatch runtime per ligand is 3.3e-6 s versus 5.2e-3 s for PharmacoNet.

Measured effects reported in the paper

Surprising findings

• PharmacoMatch's AUROC is generally lower than the CDPKit alignment algorithm on most DUD-E targets, though it reaches higher BEDROC on some targets such as ANDR and ADA relative to its own AUROC.
• The embedding process itself (67 µs per pharmacophore) takes about as long as a single alignment, so speed-up depends on embedding once and reusing vectors.

Limitations

• Embedding is not lossless and may reduce 3D geometric precision compared with alignment.
• The E(3)-invariant encoder cannot distinguish a pharmacophore from its mirror image, which may increase false positives.
• Alignment-prediction AUROC is generally lower than the CDPKit alignment baseline on most DUD-E targets.

Experimental setup

Datasets: ChEMBL (~1.2M drug-like small molecules for training) DUD-E (10 targets for alignment comparison) DEKOIS2.0 LIT-PCBA

Benchmarks: DUD-E (10 targets) DEKOIS2.0 LIT-PCBA

Metrics: AUROC BEDROC enrichment factor runtime per ligand

Baselines: CDPKit alignment algorithm PharmacoNet

Builds on: order embeddings (Li et al.) NNConv edge-conditioned convolution SchNet distance encoding DenseNet skip connections

Closest papers by meaning

• Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.86)
• ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design (ICLR 2025 · similarity 0.86)
• Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.86)
• M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.86)
• DrugHash: Hashing Based Contrastive Learning for Virtual Screening (AAAI 2025 · similarity 0.86)
• Drug-TTA: Test-Time Adaptation for Drug Virtual Screening via Multi-task Meta-Auxiliary Learning (ICML 2025 · similarity 0.85)