DSResSol: A Sequence-Based Solubility Predictor Created with Dilated Squeeze Excitation Residual Networks.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Protein solubility is crucial for protein function and production yield in research and industry.
• Experimental solubility prediction is costly, time-consuming, and often inaccurate.
• Existing computational tools lack accuracy and applicability for diverse protein classes.

Purpose of the Study:

• To develop a highly accurate computational tool for predicting protein solubility using only protein sequence.
• To overcome the limitations of existing in silico solubility prediction models.

Main Methods:

• Developed a novel deep learning model, DSResSol, integrating squeeze excitation residual networks with dilated convolutional neural networks.
• The model analyzes amino acid k-mers and their local/global interactions, including long-range dependencies.
• Input is solely the protein sequence.

Main Results:

• DSResSol outperforms all existing sequence-based solubility predictors by at least 5% in accuracy.
• The model shows significantly reduced bias towards insoluble proteins.
• Prediction accuracy for soluble proteins is at least 13% higher than existing models.
• Identified glutamic acid and serine as critical amino acids for solubility prediction.

Conclusions:

• DSResSol provides a fast, reliable, and inexpensive method for predicting protein solubility.
• The model aids in guiding experimental design for protein development.
• Highlights the importance of sequence-based analysis and long-range interactions for accurate solubility prediction.