Explainable Deep Learning Framework for SERS Bioquantification.

Area of Science:

• Spectroscopy and Analytical Chemistry
• Biomarker Discovery
• Computational Biology and Machine Learning

Background:

• Surface-enhanced Raman spectroscopy (SERS) is a promising technique for rapid and cost-effective biomarker quantification.
• Existing SERS analysis methods lag behind state-of-the-art machine learning, necessitating advanced computational frameworks.
• Lack of model explainability in SERS hinders understanding of confounding factors in biomarker-disease relationships.

Purpose of the Study:

• To present a robust computational framework for SERS bioquantification integrating spectral processing, quantification, and explainability.
• To develop an explainability method tailored for SERS mixture analysis.
• To demonstrate the framework's utility using serotonin quantification in urine.

Main Methods:

• A three-step framework: spectral processing, quantification, and explainability.
• Denoising autoencoder for spectral enhancement; Convolutional Neural Networks (CNNs) and Vision Transformers for quantification.
• Development of a Context Representative Interpretable Model Explanation (CRIME) method for SERS analysis.

Main Results:

• Optimized serotonin quantification in urine using denoised spectra and a CNN achieved a mean absolute error of 0.15 μM and mean percentage error of 4.67%.
• The CRIME method identified six unique prediction contexts for the CNN model, with three directly associated with serotonin.
• The proposed framework demonstrates effective SERS bioquantification and provides crucial model explainability.

Conclusions:

• The developed framework significantly enhances SERS bioquantification accuracy and addresses the need for model explainability.
• The CRIME method offers insights into prediction contexts, aiding in the assessment of confounding factors.
• This approach facilitates novel, untargeted biomarker discovery by leveraging SERS's speed and cost-effectiveness.