利用機器學習及二級質譜資訊對新興影響精神物質進行分類與結構解析

Abstract

近年來，新興影響精神物質（NPS）的迅速發展對公共衛生與法律監管制度造成了極大的危害。由於NPS的化學結構的多樣性與快速變化，傳統法規難以有效監管其濫用。此外，質譜分析技術依賴標準品與數據庫，對於未知或新興NPS的檢測能力受限，難以應對其結構變異所帶來的挑戰。 本研究旨在應用機器學習對NPS二級質譜圖進行分類，並探討其在質譜數據分析中的應用潛力。採用多種監督式機器學習模型，包括支援向量機（SVM）、K-近鄰（KNN）、提升樹（BT）與人工神經網路（ANN），以提升NPS分類的準確性與泛用性。 研究中使用超效能液相層析串聯質譜（UPLC-MS/MS）技術，於質荷比範圍m/z 50 - m/z 500內記錄質譜數據，並在碰撞能量10、20、40 eV條件下獲取產物離子圖譜。共收集394種包含NPS及相關藥物的質譜數據，並依其化學結構分為五大類：合成卡西酮（synthetic cathinones）、合成類大麻（synthetic cannabinoids）、苯二氮平類（benzodiazepines）、苯乙胺類（phenethylamines）及其他類（others）。 為了評估分類模型效能，本研究採用10折交叉驗證（10-fold cross-validation），並以準確率（accuracy）與F1 score作為評估指標。於驗證集資料，模型的F1 score介於0.58-1.00之間。在獨立測試集中，透過模型投票機制獲得整體準確率為87.5%（範圍62.5% - 100%）。 此外，本研究透過變數重要性分析(variable importance analysis)與質譜數據解釋，進一步解析合成卡西酮的結構特徵。在碰撞誘導解離（CID）之下，顯示N-吡咯烷（N-pyrrolidinyl）、N-烷基化（N-alkylated）、3,4-亞甲二氧基（3,4-methylenedioxy）修飾的合成卡西酮具有特定的中性丟失模式，可用於推測各類別卡西酮的碎裂途徑。 預期本研究可有效應用於NPS的分類與鑑定，並可進一步用於未知NPS的結構解析與法醫毒理學的檢測與鑑定。

The rapid emergence of new psychoactive substances (NPS) has posed significant challenges to public health and regulatory frameworks. Due to the structural diversity and continuous evolution of NPS, conventional legal and analytical approaches struggle to effectively control their abuse. Furthermore, mass spectrometry techniques, which rely on reference standards and established databases, often struggle to identify novel or structurally modified NPS. This study aims to explore the potential of machine learning algorithms in NPS classification and their application in mass spectrometric data analysis. We employed supervised machine learning models, including support vector machine (SVM), k-nearest neighbors (KNN), boosted trees (BT), and artificial neural networks (ANN), to improve the accuracy and generalization of NPS classification. Ultra-performance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS) was utilized to acquire mass spectra within the range of m/z 50 - 500, using collision energies of 10, 20, and 40 eV to generate product ion spectra. A total of 394 substances, including NPS and related drugs, were analyzed and categorized into five classes: synthetic cathinones, synthetic cannabinoids, benzodiazepines, phenethylamines, and others. Model performance was assessed using 10-fold cross-validation, with accuracy and F1 score serving as evaluation metrics. The models achieved F1 scores ranging from 0.58 to 1.00 on the validation dataset. In an independent testing dataset, ensemble model voting resulted in an overall accuracy of 87.5% (ranging from 62.5% to 100%). Furthermore, variable importance analysis and mass spectral interpretation were conducted to elucidate the structural characteristics of synthetic cathinones. Collision-induced dissociation revealed distinct neutral loss patterns for N-pyrrolidinyl, N-alkylated, and 3,4-methylenedioxy derivative cathinones, providing insights into potential fragmentation pathways. We expect that this research will demonstrate the efficacy of machine learning in the classification and identification of NPS, offering a robust approach for the detection and structural elucidation of unknown NPS in forensic toxicology.