Explainable Deep Learning Framework for SERS Bioquantification

Surface-enhanced Raman spectroscopy (SERS) is rapidly gaining attention as a fast and inexpensive method of biomarker quantification, which can be combined with deep learning to elucidate complex biomarker-disease relationships. Current standard practices in SERS analysis are behind the state-of-the-art machine learning approaches; however, the present challenges of SERS analysis could be effectively addressed with a robust computational framework. Furthermore, there is a need for improved model explainability for SERS analysis, which at present is insufficient in assessing the contexts in which confounding factors affect prediction outcomes. This study presents a framework for SERS bioquantification rooted in a three-step process, including spectral processing, quantification, and explainability. A serotonin quantification task in urine was assessed as a model task, with 682 SERS spectra measured in a micromolar range using cucurbit[8]uril chemical spacers. A denoising autoencoder was utilized for spectral enhancement, while convolutional neural networks (CNNs) and vision transformers were utilized for biomarker quantification. In addition, a context representative interpretable model explanation (CRIME) method was developed to suit the current needs of SERS mixture analysis explainability. Serotonin quantification was most efficient in denoised spectra analyzed using a CNN with a three-parameter logistic output layer (mean absolute error = 0.15 μM, mean percentage error = 4.67%). Subsequently, the CRIME method revealed the CNN model to present six unique prediction contexts, of which three were associated with serotonin. The proposed framework could unlock a novel, untargeted hypothesis-generating method of biomarker discovery, considering the rapid and inexpensive nature of SERS measurements and the potential to identify biomarkers from CRIME contexts.