Analyzing the structure-activity relationship of raspberry polysaccharides using interpretable artificial neural network model.

The structure-activity relationship has been a hot topic in the field of polysaccharide research. Six polysaccharides and three polysaccharide fragments were obtained from raspberry pulp. Based on their structural information and immune-enhancing activity data, an artificial neural network (ANN) model was used for prediction, and Gradient-weighted class activation mapping (Grad-CAM) algorithm was exploited for explanation structure-activity relationship of these raspberry polysaccharides in the present study. The structural information and immune activity data of raspberry polysaccharides were respectively used as input and output in the ANN model. The training and testing losses of ANN model was no longer decreased after trained for 200 epochs. The mean-square error (MSE) of training set and test set stabilized around 0.003 and 0.013, and the mean absolute percentage error (MAPE) of training set and test set were 0.21 % and 0.98 %, indicating the trained ANN model converged well and exhibited strong robustness. The interpretability analysis showed that molecular weight, content of arabinose, galactose or galacturonic acid, and glycosyl linkage patterns of →3)-Arap-(1→, Araf-(1→, →4)-Galp-(1 â†’ were the main structural factors greatly affecting the immune-enhancing activity of raspberry polysaccharides. This work may provide a new perspective for the study of structure-activity relationship of polysaccharides.