Explainable Deep Learning Framework for SERS Bio-quantification

TL;DR

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.

Abstract

Surface-enhanced Raman spectroscopy (SERS) is a potential fast and inexpensive method of analyte quantification, which can be combined with deep learning to discover biomarker-disease relationships. This study aims to address present challenges of SERS through a novel SERS bio-quantification framework, including spectral processing, analyte quantification, and model explainability. To this end,serotonin quantification in urine media 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, and convolutional neural networks (CNN) and vision transformers were utilized for biomarker quantification. Lastly, a novel context representative interpretable model explanations (CRIME) method was developed to suit the current needs of SERS mixture analysis explainability. Serotonin quantification was most efficient in denoised spectra analysed using a convolutional neural network 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 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.

Figures (13)

• Figure 1: SERS deep learning framework development pipeline. Illustrated are the SERS measurement process applied (A), and the computational framework pipeline. Benchmark comparisons of alternative methodology are presented on the right. Preprocessing methods (B) are marked in orange and light red, quantification methods (C) in blue, and explainability methods (D) in dark red. Asymmetric least squares (ALS) baselining is applied to all spectra prior to assessing the framework or the benchmarks. SERS = surface-enhanced raman spectroscopy, AuNP = gold nanoparticle, CB[8] = cucurbit[8]uril, CNN = convolutional neural network, XGBoost = extreme gradient boosting trees, PLSR = partial least squares regression, SVM = support vector machines, CRIME = context representative interpretable model explanations, LEN = logic explained networks, SHAP = Shapley additive explanations.
• Figure 1: Examples of denoised spectra in the artificial training data (A-C), and in lyophilized urine spectra (D-F).Y-axis (relative intensity) is omitted for clarity.
• Figure 2: Results of the final models in the validation and test sets for the four model types in both raw (A) and denoised datasets (B). Validation set results are shown in grey, and test set results are shown in color: the linear CNN model is shown in yellow (diamond), the vision transformer model in blue (circle), the scale-adjusting CNN in green (triangle), and the three-parameter logistic output layer CNN model in red (square). The shown values were obtained from the final test set. Validation set results are presented in Table \ref{['valresults']} in the appendix. MAE = mean absolute error, MPE = mean percentage error.
• Figure 2: Predictions of the trained ensembles on the validation set for all types of neural networks on both raw spectra and denoised spectra.
• Figure 3: Results for Context Representative Interpretable Model Explanations (CRIME) analysis. Six distinct contexts were identified, which are visualized across mean spectra in subfigures A - F. Positive prediction weights are presented in green, negative prediction weights in yellow, and perturbation limits have been shaded in teal. Red regions in the mean spectra correspond to average perturbation limits at either the top or bottom of the feature weight range for the simplicity of the plot. Latent spaces are visualized by context and concentrations, and compound similarity matching was done using cosine similarity. The highest similarity score is presented alongside the matched compound.