Structure-aware protein solubility prediction from sequence through graph convolutional network and predicted contact...

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Protein solubility is critical for the development of biocatalysts and therapeutic agents, directly impacting production costs.
• Accurate prediction of protein solubility from amino acid sequences is a significant challenge.
• Existing computational models often rely on one-dimensional amino acid embeddings, neglecting crucial spatial structural information.

Purpose of the Study:

• To develop a novel computational method for accurately predicting protein solubility using only the amino acid sequence.
• To address the limitations of existing sequence-based methods by incorporating structural information.

Main Methods:

• Developed GraphSol, a structure-aware method utilizing an attentive graph convolutional network (GCN).
• Constructed a protein topology attribute graph from predicted contact maps derived solely from the amino acid sequence.
• Evaluated model performance on the eSOL dataset through cross-validation and independent testing.

Main Results:

• GraphSol demonstrated substantially superior performance compared to existing sequence-based protein solubility prediction methods.
• The model achieved a consistent R-squared value of 0.48 in both cross-validation and independent testing, indicating robust stability.
• This represents the first application of GCNs for sequence-based protein solubility prediction.

Conclusions:

• GraphSol offers a significant advancement in predicting protein solubility from amino acid sequences.
• The GCN-based architecture provides a more accurate and structure-aware approach compared to traditional methods.
• The GraphSol framework is adaptable and can be extended to various other protein prediction tasks requiring only sequence data.