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

Abstract

Protein clustering and classification are critical for understanding protein functions and interactions, particularly within structure-based predictions. Traditional sequence-based clustering often overlooks the pivotal role of tertiary structure in determining protein function. Structural clustering remains limited and challenging, with existing methods struggling to achieve high accuracy and manage complex data. This study focuses on the tertiary structures of Verticillium dahliae proteins, employing deep learning techniques for effective clustering and classification. Using AlphaFold2, we predicted protein structures and generated Cα atom distance matrices. We introduced a novel Unique Nuclear Sequence Element (UNSE) neural network to enhance feature extraction, constructing weighted distance matrices by integrating Cα distances with Pfam annotations. This method effectively captures complex structural relationships. Additionally, Basic Local Alignment Search Tool (BLAST) sequence alignments validated the sequence similarity within protein families, ensuring the biological relevance of clustering results. We applied clustering algorithms to both raw and weighted matrices, comparing their performance against traditional sequence-based and other structure-based methods, including DeepGO and DeepFRI. Evaluation metrics such as Silhouette Score, ${F}_{max}$, and AUPR demonstrated that our weighted matrix approach significantly outperforms conventional methods in accuracy and robustness. These findings confirm that integrating deep learning with weighted distance matrices effectively captures structural and functional protein characteristics, providing a robust tool for structural biology.