Background: Surface-enhanced Raman spectroscopy (SERS) analysis of urine is a promising liquid biopsy technique for cancer detection. However, its clinical translation is hindered by two major challenges that impact classification efficacy. First, the SERS signal of urine is confounded by fluctuations induced by physiological differences in urine composition such as pH and dilution. Second, the molecular origin of the SERS signal of urine is incompletely understood, limiting the interpretability of machine learning classifiers in terms of specific biochemical markers. Methods: In this pilot study, we analyzed urine samples from breast cancer patients (n = 18) and control subjects (n = 10) at three pH levels (5, 7, and 9). Additionally, we analyzed simulated urine mixtures consisting of uric acid, hypoxanthine, xanthine, and creatinine in physiological concentrations to explain the variation in the SERS spectra at different pH values. Results: Urine at pH 9 yielded the most detailed spectral features. The SERS spectral pattern under alkaline pH reflected greater contributions from hypoxanthine, uric acid, and creatinine, while xanthine contributions diminished due to competitive interactions at the SERS substrate surface. Normalizing SERS signals to the creatinine band at 1420 cm⁻¹ effectively mitigated the confounding effects of urine dilution. Conclusions: Optimizing the pH to 9 and normalizing to creatinine significantly enhances the interpretability and accuracy of SERS-based urine analysis for cancer detection. These findings offer important theoretical and practical advancements for the development of SERS-based liquid biopsy tools for cancer detection.

Background: Surface-enhanced Raman spectroscopy (SERS) analysis of urine is a promising liquid biopsy technique for cancer detection. However, its clinical translation is hindered by two major challenges that impact classification efficacy. First, the SERS signal of urine is confounded by fluctuations induced by physiological differences in urine composition such as pH and dilution. Second, the molecular origin of the SERS signal of urine is incompletely understood, limiting the interpretability of machine learning classifiers in terms of specific biochemical markers. Methods: In this pilot study, we analyzed urine samples from breast cancer patients (n=18) and control subjects (n=10) at three pH levels (5, 7, and 9). Additionally, we analyzed simulated urine mixtures consisting of uric acid, hypoxanthine, xanthine, and creatinine in physiological concentrations to explain the variation in the SERS spectra at different pH values. Results: Urine at pH 9 yielded the most detailed spectral features. The SERS spectral pattern under alkaline pH reflected greater contributions from hypoxanthine, uric acid, and creatinine, while xanthine contributions diminished due to competitive interactions at the SERS substrate surface. Normalizing SERS signals to the creatinine band at 1420 cm⁻¹ effectively mitigated the confounding effects of urine dilution. These adjustments improved the classification accuracy of a random forest model. Conclusions: Optimizing the pH to 9 and normalizing to creatinine significantly enhance the interpretability and accuracy of SERS-based urine analysis for cancer detection. These findings offer important theoretical and practical advancements in the development of SERS-based liquid biopsy tools for cancer detection.