Pan-Screening the Structural Predictability of Drug Toxicity: Validation of Molecular Fingerprints

Abstract

Objective To systematically evaluate the predictability of drug chemical structures against the complete set of MedDRA Preferred Terms (PTs), construct a comprehensive structure– adverse reaction landscape, validate the utility of molecular fingerprint representations in pan-screening tasks, and establish quantitative tools for drug safety pre-screening. Methods Drug–adverse event association data were compiled from SIDER 4.1 and its derivative compilation Nsider, encompassing 2,005 approved drugs, of which 1,120 small-molecule drugs with successfully retrieved chemical structures were retained for analysis. For each Preferred Term (PT) meeting the minimum sample-size threshold (≥25 positive and ≥30 negative drugs; n = 768), XGBoost classifiers were trained using MACCS 166-bit fingerprints, with model performance evaluated via 5-fold stratified cross-validation. A PT was designated as high-confidence if it achieved a validation AUC ≥ 0.75, a cross-validation AUC ≥ 0.70, and passed Benjamini–Hochberg false discovery rate (FDR) correction (q < 0.05). The entire analytical pipeline was replicated using Morgan ECFP4 2048-bit fingerprints to assess the robustness of fingerprint selection. The robustness of significance screening was independently validated by permutation testing (500 random label permutations), and model generalizability was evaluated through Bemis–Murcko scaffold-based stratified split validation. SHAP (SHapley Additive exPlanations) analysis was employed to identify critical structural features, and mechanism-based case studies further demonstrated concordance between model-identified features and established pharmacophores. Results Among the 768 modelable PTs, 71 (9.2%) met the high-confidence significance criteria (FDR q < 0.05), predominantly comprising nervous system disorders (10 PTs) and psychiatric disorders (9 PTs). The toxicities exhibiting the strongest structural predictability included hyperprolactinaemia (AUC = 0.968), dermatitis acneiform (AUC = 0.966), tardive dyskinesia (AUC = 0.952), and breast tenderness (AUC = 0.979). Under a unified XGBoost hyperparameter configuration, the low-dimensional MACCS fingerprint (166-bit) demonstrated near-equivalent performance to the high-dimensional Morgan fingerprint (2048-bit) in pan-screening tasks (ΔAUC = −0.012, Cohen’s d = 0.16), and achieved comparable performance at a substantially lower dimensionality across 18 of 22 organ systems. SHAP analysis identified key structural fragments corresponding to established pharmacophores (six-membered rings, tertiary amine cores, heteroatomic bonds, and bridged heteroatomic chains), and mechanism-based case studies further demonstrated that the model captured structural patterns concordant with known pharmacophores of D2 antagonists, EGFR inhibitors, and related agents. Bemis–Murcko scaffold-based stratified split validation revealed a median AUC decrease of merely 0.042, with 66% of PTs maintaining an AUC ≥ 0.70 under scaffold-split conditions. Immune-mediated and multifactorial endpoints remained structurally opaque under the current data conditions. Conclusion This study constructed a comprehensive structure–adverse reaction predictability landscape encompassing 768 MedDRA PTs; seventy-one PTs achieved high-confidence criteria. The MACCS 166-bit fingerprint achieved near-equivalent performance to the Morgan 2048-bit fingerprint at a markedly lower dimensionality, and its chemically explicit features enable direct translation of model outputs into structural alerts, providing a computational tool for prospective screening of potential adverse reaction profiles for any novel compound. A tiered assessment strategy is proposed: machine learning ensemble models for high-confidence screening, substructure alert rules for high-specificity exclusion, and pharmacovigilance signal detection for complementary coverage of structurally unpredictable endpoints. Bemis–Murcko scaffold-based stratified split validation further demonstrated that the median generalization performance decrease was merely 0.042 AUC, with 66% of PTs maintaining an AUC ≥ 0.70 under scaffold-split conditions, corroborating the existence of a genuine chemical basis underlying structure–adverse reaction mappings.