SERS‐VGG27: A Deep Learning Framework With Two‐Stage Feature Selection for Precise SERS‐Based Drug Classification

Surface‐enhanced Raman spectroscopy (SERS) has become a powerful tool for drug screening; however, its potential is often constrained by complex spectral profiles and the lack of robust analytical tools. Here, we introduce SERS‐VGG27, a deep learning framework that synergizes a two‐stage feature selection with a spectrally optimized neural network architecture to overcome these limitations. This method integrates biomolecule‐guided prescreening with data‐driven refinement, effectively eliminating redundant features while preserving essential biochemical signatures. The refined spectra are then analyzed using the SERS‐VGG27 network—an enhanced architecture derived from VGG19—engineered to reduce the parameter complexity of the feature extraction layers by 21.50% and to sensitively capture subtle, drug‐induced molecular variations. Experimental validation shows that 85.77% of redundant information in the initial SERS dataset is removed during coarse‐grained selection, further reduced to 6.38% after fine‐grained refinement. Leveraging these optimizations, SERS‐VGG27 achieves an accuracy of 96.73% in classifying 21 drug types across eight distinct mechanisms of action, significantly outperforming conventional models. Furthermore, extensive benchmarking across three heterogeneous external datasets—encompassing clinical oncology, toxicology, and microbiology—validates the cross‐domain generalization and reliability of our SERS‐VGG27 framework.