DeepciRGO

DeepciRGO predicts gene ontology (GO) biological process functions of circular RNAs (circRNAs) by integrating heterogeneous network representation learning with a deep multi-label hierarchical classification model.

Key Features:

• Heterogeneous network integration: Constructs a global heterogeneous network by integrating circRNA co-expression data, protein-protein interactions (PPI), and protein-circRNA associations.
• Representation learning (HIN2Vec): Uses HIN2Vec to extract topology features for proteins and circRNAs from the heterogeneous network.
• Deep multi-label hierarchical classification: Trains a deep hierarchical multi-label classifier to predict GO biological process terms for each circRNA.
• Sequence and structure features: Incorporates RNA sequence and RNA structure information alongside network-derived features.
• Benchmark evaluation: Evaluated on the manually curated circRNA2GO-62 dataset (62 circRNAs, 185 GO annotations) with reported maximum F-measure 0.412, recall 0.400, and accuracy 0.425.
• Network integration advantage: Demonstrates improved predictive performance through integration of circRNA co-expression, PPI, and protein-circRNA associations.

Scientific Applications:

• Functional annotation of circRNAs: Prediction of GO biological process functions for circRNAs.
• Identification of disease-associated circRNAs: Prioritization of circRNAs potentially involved in diseases and cancers based on predicted functions and network context.
• CircRNA–protein network analysis: Analysis of topology-derived features of proteins and circRNAs within integrated interaction networks.
• Feature integration studies: Assessment of the impact of combining co-expression, PPI, protein-circRNA associations, and sequence/structure information on prediction performance.

Methodology:

Construct a global heterogeneous network from circRNA co-expression, PPI, and protein-circRNA associations; apply HIN2Vec to extract topology features for proteins and circRNAs; combine network-derived features with RNA sequence and structure information to train a deep multi-label hierarchical classifier; evaluate on the circRNA2GO-62 benchmark dataset.