AlphaFold 2-based stacking model for protein solubility prediction and its transferability on seed storage proteins

Area of Science:

• Biochemistry
• Computational Biology
• Protein Science

Background:

• Accurate protein solubility prediction is vital for selecting food-grade proteins.
• Current models often neglect crucial protein structural information, limiting their effectiveness.

Purpose of the Study:

• To develop an advanced regression model for predicting protein solubility.
• To enhance prediction accuracy by incorporating both protein sequence and structural data.

Main Methods:

• Developed a multilayer perceptron (MLP) model using sequence and predicted structural properties.
• Constructed a graph convolutional network (GCN) utilizing residue-level features and contact maps.
• Integrated MLP and GCN predictions into a stacking model with support vector regressor (SVR).

Main Results:

• The stacking model achieved R² values of 0.502 on test and 0.468 on external validation datasets.
• Demonstrated superior performance compared to existing protein solubility regression models.
• Validated the model's transferability on seed storage proteins and through molecular dynamic simulations.

Conclusions:

• The developed stacking model effectively leverages sequence and structural features for accurate protein solubility prediction.
• The model shows significant potential for applications in food science, agriculture, and beyond microbial proteins.