Machine Learning-Driven Ensemble Screening of Multitarget Kinase Inhibitors for Tauopathy-Associated Neurodegeneration Using All-Atom and Steered MD Simulations

Tauopathies arise when normal functions of the tau protein in axonal transport and neuronal maintenance are disrupted by an imbalance between kinases and phosphatases. Dysregulation of key kinases such as dual-specificity Tyrosine-Regulated Kinase 1A (DYRK1A), Tau Tubulin Kinase 1 (TTBK1), and ABL Proto-Oncogene 1, and Non-Receptor Tyrosine Kinase (ABL1) drives excessive tau phosphorylation and neurofibrillary tangle accumulation. DYRK1A regulates MAPT exon 10 splicing and phosphorylates tau at multiple Ser/Thr residues, priming it for further phosphorylation by other kinases. TTBK1 phosphorylates tau at disease-associated epitopes within the microtubule-binding domain, promoting detachment from microtubules and aggregation. ABL1 phosphorylates tau at tyrosine residues, linking tau modification with Aβ-induced synaptic dysfunction. These events collectively drive tau hyperphosphorylation, misfolding, and neurofibrillary pathology characteristic of tauopathies. To identify natural product-derived multitarget inhibitors for these kinases, we developed a comprehensive machine learning (ML) workflow trained on bioactivity data from ChEMBL and BindingDB. We implemented five distinct classifiers: CatBoost, Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Naive Bayes, and XGBoost. Stratified sampling and SMOTE were employed to address class imbalance for DYRK1A and ABL1, while Bemis-Murcko scaffold splitting was used to ensure rigorous evaluation of the data-scarce TTBK1 data set. A soft-voting ensemble model, integrating optimized CatBoost, XGBoost, and SVM, demonstrated superior performance. This robust ensemble was deployed to screen ∼695,000 natural compounds from the COCONUT 2.0 database. The resulting hits were refined through consensus molecular docking and deep learning-based rescoring (GNINA), leading to the identification of two high-potential lead molecules, CNP0591834.1 and CNP0484145.0. Validation using 1 μs molecular dynamics simulations confirmed their conformational stability and strong binding affinities. Steered MD further demonstrated their superior mechanical resistance to unbinding, particularly in DYRK1A and ABL1 complexes. Overall, this integrative computational framework highlights these two natural compounds as potent multitarget leads with strong potential to mitigate tau-hyperphosphorylation-driven neurodegeneration.