A Deep Learning Framework for Identifying Essential Proteins by Integrating Multiple Types of Biological Information

Area of Science:

• Computational biology
• Systems biology
• Bioinformatics

Background:

• Identifying essential proteins is crucial for understanding cellular survival and evolution.
• Traditional centrality and machine learning methods for essential protein identification have limitations in capturing biological complexity and automatically selecting features.
• Prior knowledge-based scoring functions in centrality methods are insufficient, and selected features in machine learning methods may not represent complete biological properties.

Purpose of the Study:

• To propose a novel deep learning framework for automated feature learning to identify essential proteins without prior knowledge.
• To improve the accuracy and comprehensiveness of essential protein identification by integrating multiple biological data types.
• To evaluate the performance of the proposed framework against existing computational methods.

Main Methods:

• Utilized node2vec for automatic learning of protein-protein interaction (PPI) network topologies.
• Employed bidirectional long short-term memory (BiLSTM) cells to capture non-local relationships in gene expression data.
• Incorporated subcellular localization information using a high-dimensional indicator vector.

Main Results:

• The proposed deep learning framework demonstrated superior performance compared to traditional centrality methods and existing machine learning-based methods in identifying essential proteins in S. cerevisiae.
• Ablation studies revealed that protein-protein interaction network embedding was the most significant contributor to performance improvement.
• Gene expression profiles and subcellular localization information also positively contributed to enhancing the identification of essential proteins.

Conclusions:

• The developed deep learning framework effectively identifies essential proteins by automatically learning complex biological features.
• Integrating PPI network topology, gene expression, and subcellular localization data is vital for accurate essential protein prediction.
• This approach offers a more robust and automated alternative to existing methods for essential protein identification.