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

Protein subcellular localization is key to understanding cellular function. Traditional methods are slow, prompting the use of machine learning and deep learning to enhance prediction accuracy. This study aims to leverage these approaches for more efficient and accurate localization prediction. This study presents a novel two-step approach for protein localization. First, feature extraction is performed using node embeddings derived from Gene Ontology graphs that capture hierarchical relationships between Gene Ontology terms representing molecular functions, biological processes, and cellular components. Embedding these Gene Ontology terms into a continuous vector space enhances protein representations within complex cellular environments. The second step involves multi-label classification using a dynamic thresholding based Deep Neural Network to classify proteins into multiple subcellular locations. Unlike fixed thresholds, which typically yield single-label predictions, dynamic thresholding adapts to model outputs, enabling simultaneous prediction across multiple locations. The proposed model is evaluated on two benchmark datasets, namely DeepLoc 2.0 and a curated version of the Plant-mSubP dataset. The model achieves 44.69% Overall Actual Accuracy and 91.21% Relaxed Accuracy on the DeepLoc 2.0 dataset, and 82.43% Overall Actual Accuracy and 97.25% Relaxed Accuracy on the Plant-mSubP dataset. These results show significant improvements, with a minimum increase of 5.69% in Overall Actual Accuracy on DeepLoc 2.0 and 12.67% in Relaxed Accuracy on Plant-mSubP compared to state-of-the-art models. The proposed model significantly improves multi-label protein localization by combining Gene Ontology embeddings with dynamic thresholding, outperforming existing methods.