Integrating ensemble machine-learning and fibril docking to discover potent, novel triazole-naphthalene tau-aggregation inhibitors.

Abstract

Tau-protein aggregation is a central pathological feature of Alzheimer's disease, so blocking fibril growth is an attractive therapeutic goal. We curated a high-quality set of 289 literature IC₅₀ measurements for human-tau aggregation and trained a stacked-ensemble QSAR model (SVR + RF + XGB) that achieves fivefold CV Q² = 0.63, external R² = 0.57 and RMSE = 0.73 log-units. Applicability-domain analysis revealed no high-influence outliers in the calibration set, and a 5-nearest-neighbour density test confirmed that each of sixteen previously unreported 1,2,4-triazole-naphthalene derivatives (TND, TND-1…TND-15) lies in locally populated chemical space, albeit at the edge of the global domain. The model predicts pIC₅₀ = 6.75-7.53 (IC₅₀ ≈ 30-177 nM), nominating TND-9, TND-15 and TND-5 as top-ranked candidates based on predicted potency. Nearly all TNDs fall within the BBB window (MW ≈ 350-450 Da, TPSA < 90 Ų); most obey cLogP ≤ 5, and the few slightly above still map to the BOILED-Egg CNS-positive zone. Retrospective docking against phosphorylated-tau fibrils (PDB ID 6HRF) highlighted TND, TND-5 and TND-14 with sub-micromolar predicted affinity, forming key contacts in the microtubule-binding cleft. These docking results support binding plausibility rather than quantitative aggregation inhibition. TND-8, although highly ranked by docking, was deprioritised owing to low predicted GI absorption. Physicochemical and CNS-oriented ADMET filters further support developability of the top leads. The integrated workflow-combining rigorously validated QSAR, structure-based docking on the 6HRF polymorph and developability profiling-provides an open-source blueprint for tau-aggregation inhibitor discovery. Consensus ranking prioritises TND-5 for immediate in-silico follow-up, with TND, TND-14, TND-9 and TND-15 as secondary leads.