Rapid Identification of Trace Pharmacodynamic Substances in Traditional Chinese Medicine via SERS and Deep Learning

In the modernization of traditional Chinese medicine (TCM), trace detection of pharmacodynamic substances faces a critical challenge: insufficient sensitivity, which significantly hinders accurate quality assessment and standardization. Conventional techniques often fail to measure trace components in complex sample matrices. Therefore, the development of a rapid, effective, sensitive, and reliable analytical method, along with a corresponding quality evaluation system, is of great importance. This study used moth wing (MW) scales as a template to fabricate an Ag/MW SERS substrate via magnetron sputtering. The optimal Ag₃₀/MW SERS substrate (30 min sputtering) achieved an enhancement factor of 6.47 × 10⁶ and good reproducibility (minimum RSD: 7.03%). Principal component analysis (PCA) was integrated with four deep learning algorithms (MLP, Transformer, ResNet, DNN) to detect three typical TCM pharmacodynamic substances in pure standard solutions: atractylon, cimifugin, and timosaponin A-III. The models enabled rapid identification, with the MLP model reaching 95.00% accuracy. This research provides a novel, highly accurate, and efficient detection method with potential for TCM pharmacodynamic substances, demonstrating feasibility for bioactive compound identification in model systems, and shows promising potential for future application in TCM composition analysis and quality control.