Cross-Species Multitask Learning with Molecular and ADME Descriptors for Liver Microsomal Metabolic Stability

Liver microsomal metabolic stability is a key determinant of in vivo exposure and an essential filter in lead optimization, yet cross-species prediction remains difficult because of heterogeneous metabolic pathways and limited model interpretability. We propose a cross-species multitask learning framework that integrates complementary molecular modalities-SMILES-derived fingerprints (Morgan and MACCS/RDKit), molecular graphs, and in silico absorption, distribution, metabolism, and excretion (ADME)/physicochemical descriptors-to predict binary microsomal stability (unstable: t _1/2 ≤ 30 min; stable: t _1/2 > 30 min) in human (HLM), rat (RLM), and mouse (MLM) liver microsomes. We curated 18,921 PubChem BioAssay measurements (6,685 HLM; 5,753 RLM; 6,483 MLM). Under stratified 10-fold Bemis-Murcko scaffold cross-validation with ensemble prediction and species-specific thresholds, the model achieved AUROC values of 0.811, 0.806, and 0.794 and AUPR values of 0.854, 0.860, and 0.862 for HLM, RLM, and MLM, respectively, consistently outperforming conventional machine-learning and single-task deep-learning baselines. SHapley Additive exPlanations (SHAP) identified transport/permeability indicators, CYP interaction flags, and the lipophilicity-polarity axis as the features most strongly associated with predicted stability. EdgeSHAPer, a graph neural network explanation method based on SHAP, highlighted stabilizing and destabilizing substructures. Recurrent destabilizing attributions were observed in alkene and allylic/benzylic contexts, whereas amide/carbamate motifs exhibited stabilizing attributions, with nitriles and halogens showing context-dependent effects. Fragment-ADME enrichment analysis characterized associations between local structural motifs and whole-molecule properties including lipophilicity, solubility, and blood-brain barrier permeability. This multi-modal, cross-species framework demonstrates that integrating structural encodings with ADME descriptors enhances both predictive performance and interpretability, yielding hypothesis-generating attributions for structural optimization that warrant prospective experimental validation.