AI-SNPS2: A Multi-Layered LC-MS/MS Platform Integrating Molecular Networking and Retention Time Prediction for Identifying Controlled and New Psychoactive Substances

The emergence of unknown controlled substances poses a significant challenge in forensic and analytical sciences. While liquid chromatography–tandem mass spectrometry (LC–MS/MS) enables identification of compounds through spectral database matching, it remains limited for synthesized analogues not present in existing libraries. To address this gap, we developed AI-SNPS2 (Artificial Intelligence Screener for Narcotic Drugs and New Psychoactive Substances 2), an enhanced version of our previously reported screening software. AI-SNPS2 is structured into five integrated layers: LC–MS Viewer, AI Classifier, Identifier, NetBuilder (a GNPS-inspired molecular networking module), and RT Predictor (a machine learning-based retention time prediction module). These layers allow structural analogue detection via spectral similarity and chromatographic plausibility filtering, thereby extending identification capabilities beyond conventional spectral search. The RT Predictor layer incorporates four regression models─artificial neural network (ANN), support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost)─trained on 42 molecular descriptors from 164 controlled substances. All models exhibited strong performance, with XGBoost achieving the highest accuracy (R2 = 0.964, MAE = 0.585). When applied with the RT calibration function, deviations were typically within a few minutes on a 110 min gradient, demonstrating the RT predictor's utility for candidate filtering. The software's utility was further evaluated by spiking JWH-019, JWH-015, and JWH-302 into two complex matrices; both were successfully identified through integration of molecular networking (MN) and hybrid similarity search (HSS) algorithm. Furthermore, evaluation using five additional compounds demonstrated that AI-SNPS2 is a highly promising tool for detecting compounds absent from existing databases.