Rprot-Vec: a deep learning approach for fast protein structure similarity calculation.

Area of Science:

• Computational biology
• Bioinformatics
• Structural bioinformatics

Background:

• Predicting protein structural similarity and detecting homologous sequences are crucial for inferring protein functions.
• Current methods often rely on 3D structural data, which is not universally available.
• There is a need for efficient, sequence-based methods for large-scale structural similarity prediction.

Purpose of the Study:

• To develop a deep learning model for predicting protein structural similarity and homology detection using only primary sequence data.
• To create a fast and lightweight solution that overcomes the limitations of traditional 3D structure-dependent methods.

Main Methods:

• The Rprot-Vec (Rapid Protein Vector) model integrates bidirectional GRU and multi-scale CNN layers.
• ProtT5-based encoding is utilized for efficient sequence representation.
• The model was trained and evaluated using curated datasets.

Main Results:

• Rprot-Vec achieved a 65.3% accurate similarity prediction rate for homologous proteins (TM-score > 0.8).
• The model demonstrated an average prediction error of 0.0561 across all TM-score intervals.
• Rprot-Vec outperformed existing TM-vec baselines in all tested scenarios, despite having fewer parameters.

Conclusions:

• Rprot-Vec provides a rapid and efficient sequence-based approach for structural similarity prediction.
• The model has broad applications in protein homology detection, structure-function inference, and drug repurposing.
• Open-source availability and released datasets will promote community adoption and further research.