Deep learning-enhanced clustering and classification of protein molecule tertiary structures using weighted distance...

Area of Science:

• Structural biology
• Bioinformatics
• Computational biology

Background:

• Protein clustering and classification are vital for understanding protein function and interactions.
• Traditional sequence-based methods often neglect the importance of tertiary structure.
• Existing structural clustering methods face challenges in accuracy and data complexity.

Purpose of the Study:

• To develop an effective deep learning-based method for protein clustering and classification using tertiary structures.
• To improve the accuracy and robustness of protein structure-based analysis.
• To provide a robust tool for structural biology research.

Main Methods:

• Protein structure prediction using AlphaFold2 and generation of Cα atom distance matrices.
• Development of a Unique Nuclear Sequence Element (UNSE) neural network for feature extraction.
• Construction of weighted distance matrices integrating Cα distances with Pfam annotations.
• Application of clustering algorithms and comparison with traditional and other structure-based methods (DeepGO, DeepFRI).
• Validation using Basic Local Alignment Search Tool (BLAST) for sequence similarity.

Main Results:

• The weighted distance matrix approach significantly outperformed conventional sequence-based and other structure-based methods.
• Evaluation metrics (Silhouette Score, Fmax, AUPR) demonstrated superior accuracy and robustness.
• The method effectively captures complex structural relationships and functional characteristics.

Conclusions:

• Integrating deep learning with weighted distance matrices provides a powerful approach for protein clustering and classification.
• This method enhances the understanding of protein structure-function relationships.
• The developed approach offers a robust tool for structural biology and bioinformatics.