Study on the quantitative analysis of Tilianin based on Raman spectroscopy combined with deep learning

Tilianin is a commonly used pharmaceutical ingredient with various biological activities such as antioxidant, anti-inflammatory, and anticancer, which is able to exert antitumor effects by inhibiting tumor cell proliferation, inducing apoptosis and inhibiting angiogenesis. Studies have demonstrated to be particularly useful in a variety of cancers such as liver, lung and gastric cancers. Quantitative analysis of Tilianin can improve the quality control of related drugs and assist in guiding clinical application and disease treatment. However, there are limited studies on the quantitative analysis of Tilianin. High performance liquid chromatography (HPLC) and mass spectrometry (MS) are commonly used methods for the quantitative analysis of the components, but they often require complex pretreatment steps and specialized analytical capabilities, and are sample-destructive. The method based on Raman spectroscopy and deep learning is a widely used non-destructive analysis method. For this reason, this paper proposes a residual self-attention mechanism model based on Raman spectroscopy and deep learning for quantitative analysis of 6 concentrations of Tilianin. Six different concentrations of Tilianin-methanol solutions were prepared, and a total of 120 spectral samples were collected, which were pre-processed and inputted into our Raman Spectrum with Self-Attention Quantification Net (RSAQN) for analyzing and predicting. The structure of this model not only focuses on the deep and shallow features of the spectrum, but also the information between different channels, and the self-attention mechanism further extracts the features and outputs the predicted values of Tilianin concentration through the fully connected layer. In this paper, five sets of comparison models are set up, including two machine learning models (Random Forest, K-Nearest Neighbors, Artificial Neural Network) and two deep learning models (Convolutional Neural Network and Variational Autoencoder), and the results show that the model in this paper fits the best, obtaining an R2 of 0.9144, as well as a small error.