Gene Ontology graph embeddings with Dynamic Thresholding based Deep Neural Networks for Multi-label protein...

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Protein subcellular localization is crucial for cellular function.
• Traditional methods for localization are time-consuming.
• Machine learning and deep learning offer enhanced prediction accuracy.

Purpose of the Study:

• To develop a more efficient and accurate method for protein subcellular localization prediction.
• To leverage Gene Ontology (GO) embeddings and deep neural networks for improved predictions.
• To enable accurate multi-label classification of protein locations.

Main Methods:

• Feature extraction using node embeddings from Gene Ontology graphs.
• Capturing hierarchical relationships within GO terms (molecular functions, biological processes, cellular components).
• Multi-label classification with a dynamic thresholding Deep Neural Network.

Main Results:

• Achieved 44.69% Overall Actual Accuracy and 91.21% Relaxed Accuracy on the DeepLoc 2.0 dataset.
• Achieved 82.43% Overall Actual Accuracy and 97.25% Relaxed Accuracy on the Plant-mSubP dataset.
• Demonstrated significant improvements over state-of-the-art models, with minimum increases of 5.69% (Overall Actual Accuracy) and 12.67% (Relaxed Accuracy).

Conclusions:

• The proposed two-step approach effectively combines GO embeddings and dynamic thresholding for superior multi-label protein localization.
• The model significantly enhances prediction accuracy compared to existing methods.
• This advancement offers a more efficient tool for understanding protein functions within cellular environments.