Biomolecule classification by multiscale one-dimensional convolutional neural network.

Abstract

Raman spectroscopy has been used as a powerful tool to identify biomolecules in various fundamental and clinical studies. In recent years, machine learning has been used in the spectroscopy field to extract the spectrum feature for biomolecule classification and recognition. However, few methods have been developed for the classification of non-preprocessed spectra of biomolecules. The main challenge to deal with the spectra is the weak signals, strong backgrounds, and high bio-complexity. In this study, we developed a method to classify the membrane protein behavior in a nanometer-thin bilayer after an activator was applied. First, we used data augmentation to solve the problem of the sparsity of our spectrum data. Second, we constructed a multiscale 1D-CNN model for the spectrum classification. The ability to capture spectral features at different scales with the multiscale 1D-CNN model can deal with the difficulty to classify spectra of the complex cell membrane system. Last, score-CAM was used for model visualization to explain which spectrum features were used for the classification. The results show that our proposed multiscale 1D-CNN can improve the model performance metrics of other conventional machine algorithms by 4% to 13%. In addition, the obtained heatmaps from score-CAM also provided reasonable interpretations for the classification results in the amide I and amide III spectrum regions. These results show that the method we developed can give a persuasive classification for target biomolecules and has the potential to be applied in the study of disease treatment and biochemical reactions.