Graph neural networks and transformers for antimalarial drug discovery under distribution shift
Abstract
Description: Complete reproducibility records for “Graph neural networks and transformers for antimalarial drug discovery: honest-negative results underdistribution shift” submitted to the Journal of Computer-Aided Molecular Design (JCAMD). This deposit provides: - Core dataset (19,836 molecules from African natural product-inspired library) - Benchmark results (5 GNN/Transformer architectures vsECFP4-RF baseline) - Random and scaffold split evaluations (5-fold cross-validation, 5 seeds) - Extended robustness campaign (25 configurations, 625 fold-seedrecords) - External validation (22,267 molecule ChEMBL-derived disjoint panel) - Structural complexity analysis (Fsp3 ≥0.45, rings ≥4 cohorts) - Topologicalfusion results (GIN-TFP, GIN-TNE integrating P3 representations) - ChemBERTa sequence model benchmarks (125/125 configurations complete) - Distanceaware conformal triage filter (NN-Tanimoto + ECE calibration) - Analysis scripts (Python 3.11+ with malaria_md environment, PyTorch 2.13.0, PyG 2.8.0) -Complete documentation (DAR, methods supplements, theoretical framework) Key findings: - ECFP4-RF baseline dominates under scaffold split (0.8300 vs GIN 0.8047, GIN-TFP 0.8138) - Honest-negative result: GNNs/Transformers do notoutperform classical fingerprints at this scale - 1-WL expressivity bottleneck on complex African NP topologies (Fsp3 ≥0.45, rings ≥4) - Topological fusion (TFP)provides modest +0.0091 AUC gain under scaffold split - Distance-aware triage filter boosts OOD screening precision by +14.2% - Performance orderinginvariant across 5 split families and 3 capacity settingsAll files are sha256-verified. See README.md for complete usage instructions. Keywords (comma-separated): antimalarial, graph neural networks, GNN, transformers, ChemBERTa, drug discovery, distributionshift, scaffold split, African natural products, Plasmodium falciparum, ECFP4, topological data analysis, PyTorchGeometric, molecular machine learning, honest-negative results, conformal prediction
