Classifying Implantable Biomaterial Neurotoxicity via Mechanistic Targets and Machine Learning

Sánchez-Gómez J.S.; Sobrino I.C.A. · Proceedings - 2025 IEEE Latin American Conference on Nanotechnology, LANANO 2025 (2025) · DOI: 10.1109/lanano68466.2025.11512626

Resumen

Implantable devices can release ions, monomers, additives, and nanoparticulate by-products that interact with neural systems, yet neurotoxicity is seldom evaluated with mechanism-aware, data-driven methods. We introduce a concise framework that classifies implantable-biomaterial compounds against four mechanistic targets of neural injury: mitochondrial membrane-potential disruption (SR-MMP), oxidative-stress response (SR-ARE), p53-mediated apoptosis (SR-p53), and aryl hydrocarbon receptor activation (NR-AhR). Chemistries are curated and encoded with Morgan fingerprints plus physicochemical and metal descriptors, then learned by supervised models under scaffold-stratified cross-validation with AUCPR-based early stopping and Platt calibration. On held-out scaffolds, multilayer perceptrons outperform linear and tree baselines, with the strongest gains for SR-ARE and NR-AhR, while calibration supports risk-tier decisions. Post-hoc explainability using TreeSHAP and DeepSHAP highlights substructures consistent with mechanistic priors, cationic amphiphiles (MMP), electrophilic motifs (ARE), and halogenated aromatics (AhR), and flags out-of-domain chemotypes. The tool enables rapid, cost-effective triage of coatings, additives, and degradation products for scenarios involving chronic, low-dose exposure and blood-brain barrier considerations. Planned extensions include multitask learning, implant-specific ion-release features, and BBB/ADME descriptors, plus prospective in vitro confirmation, yielding a practical, mechanism-guided pipeline for safer material selection in nanotechnology-enabled implants.

Instituciones: Universidad Peruana de Ciencias Aplicadas