A novel structure-based encoding for machine-learning applied to the inference of SH3 domain specificity

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Understanding protein-protein and protein-ligand interactions is vital for cell machinery.
• Peptide Recognition Modules (PRMs) are key to protein-protein interactions, binding short protein sequences.
• High-throughput methods for scanning PRM-peptide interactions have high false-positive rates, necessitating in silico validation tools.

Purpose of the Study:

• To develop a computational tool for inferring domain-peptide interactions.
• To improve the accuracy and reliability of predicting interactions for PRMs, specifically the SH3 domain family.
• To create a novel encoding technique for representing binding information efficiently.

Main Methods:

• Developed a machine-learning approach using interaction data from high-throughput techniques.
• Integrated sequence and structure data of interacting protein partners.
• Proposed a novel encoding technique based on domain-peptide contact residues, minimizing dimensionality.

Main Results:

• Achieved >90% accuracy in detecting new binders for known SH3 domains.
• Outperformed existing methods including neural models, profile methods, and statistical predictors.
• Demonstrated generalization capability by inferring specificity for previously unknown SH3 domains.

Conclusions:

• The developed machine-learning method accurately predicts SH3 domain-peptide interactions.
• The novel encoding technique effectively represents binding information, avoiding the curse of dimensionality.
• This approach offers a reliable in silico tool for validating experimental findings and discovering new interactions.