Explainable Analysis for New Psychoactive Substance Identification with Chemical Insights

The emergence of new psychoactive substances (NPS) poses a significant challenge to global public health due to their structural diversity and rapid evolution. Traditional detection methods, such as gas chromatography-mass spectrometry and liquid chromatography-mass spectrometry, which focus on component analysis and comparison with known molecules, lack adaptability for novel variants. In contrast, deep learning offers promise for proactive identification. We introduce NPS-EDR, an Explainable Deep Reasoning model for NPS identification. It features a two-stage prediction-explanation framework that enhances analysis and builds confidence. The architecture utilizes the strengths of cooperative training-mode-specific experts and reinforcement learning to ensure consistency between explanations and answers. Built on a chain-of-thought data set of over 2,900 NPS and drug molecules, NPS-EDR integrates chemical prior knowledge for precise structural-functional interpretation. Unlike detection, our model leverages biochemical insights to provide interpretable structural and functional analyses of potential NPS molecules, achieving superior accuracy, precision, and transparency compared to mainstream large language models and biomolecular-specific chemical language models. Its transparent reasoning enhances public health strategies, pharmacological research, and forensic science, addressing the NPS crisis.