Protein solubility is an important physicochemical property in food science. Accurate prediction tools for solubility are in demand, as they can be used for screening food protein candidates. This study aims to build a state-of-the-art (SOTA) protein solubility prediction model with molecular descriptors extracted from predicted protein structures. Based on a solubility dataset of 3001 E. coli proteins, AlphaFold 2 was employed to generate 3D structures of the corresponding proteins. From the acquired structures, 22 structure-level features, three residue-level descriptors, and protein contact maps were extracted. A multilayer perceptron (MLP) model was then created using the 22 structure features as input. Moreover, using the residual features and contact maps, a graph convolutional network (GCN) model was constructed. Finally, the two models were treated as meta-predictors, whose outputs were input into a support vector machine (SVM) for the stacking model. Model performances were evaluated with an independent E. coli protein test dataset and an external S. cerevisiae dataset. In both datasets, the stacking model achieved great performances: R² of 0.494 and 0.410 with RMSE values of 0.228 and 0.225, respectively. To assess the applicability of the model in the food industry, a wet-chemistry validation is also conducted using food proteins. This work shows the potential of applying deep-learning techniques in food science as well as providing a SOTA regression model for exact protein solubility prediction.

Workflow of the modeling process used in this study

Yonghui Li

Appearances