SDN2GO

SDN2GO predicts protein functions from protein sequences, domains, and protein-protein interaction (PPI) networks using deep learning to assign Gene Ontology (GO) terms for large-scale annotation of proteins in databases such as UniProtKB.

Key Features:

• Deep Learning Architecture: Employs convolutional neural networks (CNNs) to extract features from protein sequences, domains, and PPI networks.
• Integration of Multiple Data Sources: Integrates protein sequences, domain structures, and interaction networks to improve prediction accuracy.
• Gene Ontology (GO) Term Prediction: Predicts GO terms covering the three GO sub-ontologies: molecular function, biological process, and cellular component.
• Innovative Use of Domain Information: Processes protein domain information with an NLP-inspired pre-trained deep learning sub-model to derive domain features.
• Feature Integration and Weight Classifier: Combines extracted sequence, domain, and network features using a weight classifier for GO term classification.
• Performance and Validation: Validated against two competitive methods and BLAST using CAFA time-delayed datasets, demonstrating superior performance across GO sub-ontologies.

Scientific Applications:

• Large-scale protein annotation: Annotating proteins in databases such as UniProtKB when experimental evidence is limited.
• Functional inference from sequence and networks: Inferring molecular functions, biological processes, and cellular components from sequence, domain, and PPI data.
• Genomics: Providing functional annotations for gene products to support genomics studies.
• Proteomics: Assigning GO-based functions to proteins for proteomics analyses.
• Systems biology: Enabling integrated functional annotations across interaction networks for systems-level analyses.
• Drug discovery: Supporting target identification and functional interpretation in drug discovery research.

Methodology:

Preprocesses protein sequences, domain information, and PPI networks; applies CNNs for feature extraction; uses an NLP-inspired pre-trained deep learning sub-model to extract domain features; integrates extracted features via a weight classifier to predict GO terms.