Applications of machine learning for the general unknown screening of HRMS data within forensic toxicology

This review is intended for forensic toxicologists and cheminformaticians seeking an understanding of the past implementations and future directions of artificial intelligence (AI) and machine learning (ML) for high-resolution mass spectrometry (HRMS) data interrogation in forensic toxicology. It provides a comprehensive overview of the data processing steps required to generate valid ML inputs, including molecular representation, augmentation, tokenization, embedding, and spectral deconvolution. We examine the advantages and disadvantages of different modeling strategies and summarize existing models from forensic toxicology and related domains. Applications are grouped into spectra-to-compound, compound-to-spectra, and classification models, with attention to recent advances and the practical challenges of limited data, polysubstance use, and validation. By leveraging advances from related fields, ML can enhance forensic HRMS workflows, enabling more efficient unknown screening, structural elucidation, and classification of emerging substances. This review aims to bridge disciplinary perspectives and support the practical integration of ML into routine forensic toxicology.