DeepGOA

DeepGOA predicts Gene Ontology (GO) annotations for proteins by integrating graph convolutional networks (GCNs) on the GO directed acyclic graph (DAG) of over 40,000 terms with convolutional neural networks (CNNs) on amino acid sequences to combine GO hierarchy information with sequence-derived features.

Key Features:

• Graph Convolutional Network (GCN) Integration: Uses GCNs to capture semantic relationships and latent inter-relations between GO terms and to update edge weights within the GO DAG.
• Correlation Quantification: Quantifies correlations or edges between GO terms and uses these measures to modify the DAG edge weights.
• Semantic Representation Learning: Learns semantic representations of GO terms by applying GCNs on the updated DAG.
• Convolutional Neural Network (CNN) for Amino Acid Sequences: Extracts feature representations from amino acid sequences via CNNs.
• Dot Product-based End-to-End Training: Computes the dot product between GO term semantic representations and sequence feature representations to enable coherent end-to-end training.
• Large GO Repository Support: Operates on the GO hierarchy comprising over 40,000 entries organized as a DAG.

Scientific Applications:

• Protein Function Annotation: Predicts Gene Ontology annotations to improve the accuracy of protein function assignment.
• Cross-species Annotation: Has been applied to datasets including Maize (Zea mays ssp. mays L.) and human protein sequences.
• Gene Function and Biotechnology Research: Supports studies of gene function in model organisms and applications relevant to crop improvement and biotechnology.

Methodology:

Quantify correlations between GO terms and update DAG edge weights; apply GCNs to the updated DAG to learn GO term semantic representations; extract sequence features with CNNs; compute dot products between term and sequence representations to enable end-to-end training.