Machine learning-based prediction of proteins' architecture using sequences of amino acids and structural alphabets

Area of Science:

• Structural Biology
• Bioinformatics
• Machine Learning

Background:

• The Protein Data Bank (PDB) and AlphaFold predictions have vastly increased protein structure data.
• Classifying these structures using the CATH domain database is crucial but challenging due to manual annotation limitations.
• Next-generation sequencing generates numerous protein sequences lacking structural annotations, necessitating efficient pre-annotation methods.

Purpose of the Study:

• To develop a fully automated machine-learning model for classifying protein domains into CATH Architectures.
• To address the challenge of annotating rapidly expanding protein structure repositories.
• To enable precise pre-annotation of protein sequences with structural features.

Main Methods:

• Developed a novel machine-learning model for protein domain classification.
• Utilized amino acid sequences as input for the model.
• Incorporated structural alphabets alongside amino acid sequences for enhanced classification.

Main Results:

• Achieved an F1 Score of 0.92 using only amino acid sequences.
• Attained an F1 Score of 0.94 when using both amino acid sequences and structural alphabets.
• Demonstrated the model's capability to classify both known and unknown protein structures.

Conclusions:

• The developed machine-learning model offers a highly accurate and automated solution for CATH Architecture classification.
• This approach can significantly aid in annotating large protein structure datasets and pre-annotating novel sequences.
• The model provides a valuable tool for structural and functional annotation in bioinformatics.