A Deep Learning-Enabled SERS Platform for Decoding Tumor Heterogeneity and Dynamic Therapeutic Monitoring at the Cellular Level

Abstract

Accurate and convenient identification of tumor heterogeneity remains a significant challenge in precise tumor diagnosis and treatment. To address this, we developed an integrated platform combining label-free surface-enhanced Raman spectroscopy (SERS) with a residual network (ResNet) deep learning algorithm for decoding tumor heterogeneity and predicting therapeutic vulnerability at the cellular level. First, a general pan-cancer SERS database of eleven cell lines (nine cancer cell lines representing various organs and two normal cell lines) was constructed for broader tumor discrimination. An Isolation Forest (Isoforest) algorithm was applied to cleanse spectral anomalies, and the optimized ResNet18 model achieved an average classification accuracy of 98.18% across 11 cell lines, 98.67% across five distinct cancer types and normal controls, and 98.33% in discriminating clinical tissue samples. Second, a breast cancer subtype-specific SERS dataset (comprising ten breast cancer cell lines and one normal cell line) was established to enable precise classification of four molecular subtypes (Luminal A, Luminal B, HER2-enriched (HER2-E), and triple-negative breast cancer (TNBC)). The platform demonstrated accuracies of 95.73% for cell lines, 92.60% for subtype categories, and 97.33% for clinical tissue specimens. Notably, the platform successfully decoded cellular proportions in heterogeneous mixed samples and enabled real-time tracking of dynamic phenotypic transitions in HER2-E cells during targeted therapy and drug withdrawal. Combined with SHapley Additive exPlanations (SHAP) spectral feature interpretation and Western blot assays, a comprehensive “phenotype–mechanism–protein” cross-verification framework was established. Overall, this SERS-deep learning platform offers a rapid, precise, and label-free strategy for clinical cancer screening and dynamic therapy monitoring.