A fragment-aware model for novel psychoactive substances analysis with uncertainty quantification

The rapid emergence of novel psychoactive substances (NPS) poses significant challenges to forensic toxicology, as traditional detection methods struggle to keep pace with their increasing complexity and diversity. To address this, we propose the NPS Fragment-Aware Chemical Language Model (NPS-FACL), a large language model (LLM) framework that enhances NPS detection by leveraging their inherent chemical substructures. Our approach utilizes fragment-aware tokenization, achieving a 20.33% reduction in token representation complexity, which contributes to a 2.01% increase in F1-score. Furthermore, the model integrates explainable uncertainty quantification, enhancing prediction reliability and providing insights into substructure-driven biases for improved forensic analysis of NPS. This work marks a shift towards explainable AI-assisted decision-making, providing proactive alerts and uncertainty-driven insights for early warning of emerging synthetic drug threats, in contrast to conventional Liquid Chromatography-Mass Spectrometry (LC-MS) detection methods, which focus on molecule detection through extensive sample preparation and lack forward-looking predictive capabilities. Beyond forensic toxicology, the approach supports broader applications in public health surveillance, drug regulation, and harm reduction strategies by enabling rapid and scalable identification of novel substances.