Hybrid Ensemble Learning and XAI for Cholinesterase Inhibitor Prediction in Alzheimer’s Therapy: An Interpretable Ligand-Based Framework

Abstract

Cholinesterase inhibition remains a central strategy in symptomatic therapy for neurodegenerative disorders, yet robust ligand-based prioritization under realistic chemical generalization settings remains challenging. In this study, we developed a scaffold-aware ensemble framework to predict the inhibitory activity of acetylcholinesterase (AChE) and butyrylcholinesterase (BuChE). Molecules were standardized using an RDKit pipeline comprising largest-fragment selection, uncharging, tautomer canonicalization, sanitization, and InChIKey-based deduplication. Each compound was initially represented by an 8,372-dimensional candidate feature space comprising a 4,096-bit ECFP6 fingerprint, 2,048 hashed Morgan-count features, a 2,048-bit RDKit fingerprint, 167 MACCS keys, and 13 physicochemical descriptors. Within each training fold, constant features were removed, and the top 2,500 surviving features were retained by mutual-information selection, so that the 8,372-dimensional pool is an overcomplete candidate set whose effective per-model dimensionality is roughly 1,450–1,724 features. The SVD-logistic-regression branch further projected this selected representation to 384 latent components. A heterogeneous first layer comprising LightGBM, Random Forest, ExtraTrees, and SVD-Logistic Regression was integrated via a sigmoid-calibrated logistic regression meta-model trained on out-of-fold base probabilities. Performance was assessed using nested 5-fold scaffold cross-validation for both targets, along with an external BindingDB evaluation for AChE and a broader, cholinesterase-based provisional external stress test for BuChE after overlap filtering. On internal scaffold cross-validation, the proposed ensemble achieved ROC-AUC <inline-formula> <tex-math notation="LaTeX">$= 0.9149\pm 0.0165$ </tex-math></inline-formula>, PR-AUC <inline-formula> <tex-math notation="LaTeX">$= 0.9234\pm 0.0195$ </tex-math></inline-formula>, BACC <inline-formula> <tex-math notation="LaTeX">$= 0.8316\pm 0.0243$ </tex-math></inline-formula>, and MCC <inline-formula> <tex-math notation="LaTeX">$= 0.6624\pm 0.0447$ </tex-math></inline-formula> for AChE, and ROC-AUC <inline-formula> <tex-math notation="LaTeX">$= 0.9198\pm 0.0313$ </tex-math></inline-formula>, PR-AUC <inline-formula> <tex-math notation="LaTeX">$= 0.9300\pm 0.0172$ </tex-math></inline-formula>, BACC <inline-formula> <tex-math notation="LaTeX">$= 0.8455\pm 0.0373$ </tex-math></inline-formula>, and MCC <inline-formula> <tex-math notation="LaTeX">$= 0.6938\pm 0.0757$ </tex-math></inline-formula> for BuChE. On the target-specific external AChE BindingDB benchmark (<inline-formula> <tex-math notation="LaTeX">$n=119$ </tex-math></inline-formula>), the ensemble achieved ROC-AUC = 0.8611, BACC = 0.8311, and MCC = 0.5369. These results demonstrate that the proposed calibrated ensemble maintains strong scaffold-aware predictive capability while highlighting challenges caused by external distribution shifts. The framework is positioned as a robust ligand-prioritization tool rather than a definitive predictor of prospective biochemical activity.