Deep learning-assisted metabolic fingerprint profiling based on V-groove and wrinkle-shaped 3D surface-enhanced Raman scattering substrate for early colorectal cancer diagnosis.

Colorectal cancer (CRC) remains one of the leading causes of cancer-related mortality worldwide, yet more than 90 % of deaths are preventable through early detection. Here, we presented a high-performance surface-enhanced Raman scattering (SERS) substrate featuring a three-dimensional superstructure composed of micro V-groove arrays, wrinkle patterns, and silver nanoparticles (VGA/WS/AgNPs 3D-SERS), capable of detection with a sensitivity as low as 10-14 M. Using this platform, we acquired distinctive SERS fingerprints of serum metabolites associated with early CRC (eCRC), including corticosterone, homogentisic acid, and 5,6-dimethyl-4-oxo-4H-pyran-2-carboxylic acid. By integrating these metabolic signatures with a Convolutional Neural Network (CNN), we established a SERS-based eCRC Metabolites Model (SCMM), which robustly discriminated eCRC patients from healthy controls with an area under the curve (AUC) of 0.98 in the test set. The model achieved an accuracy of 0.944 and a specificity of 1.0. Notably, SCMM also demonstrated a 92.73 % positive detection rate among patients who were carcinoembryonic antigen (CEA) negative. Collectively, this work introduced a powerful platform for precise identification of eCRC patients, providing a valuable complement to conventional CEA-based diagnostics.