GCNfold

GCNfold predicts RNA secondary structures by integrating graph convolutional networks, Transformer Encoders, and a UNet-based long-distance dependency extractor to infer structural elements such as stems, hairpins, and internal loops.

Key Features:

• Graph Convolutional Network (GCN) Integration: A three-layer GCN Structure Extractor mines structural information from RNA sequences, identifying elements including stems, hairpins, and internal loops.
• Structure and Sequence Fusion: Structural information is embedded into sequences using Transformer Encoders to jointly represent sequence and structural features.
• Long-distance Dependency Extraction: A UNet-based Long-distance Dependency Extractor captures long-range pairwise relationships within RNA sequences.
• Efficiency and Performance: The model uses a small number of parameters and fast inference, achieving overall accuracy exceeding 80%, with GCNfold-Small inferring structures in ~90 milliseconds and attaining close to 90% accuracy on average.

Scientific Applications:

• Functional Genomics: Understanding how RNA secondary structures influence gene expression and regulation.
• Drug Discovery: Identifying potential therapeutic targets by analyzing RNA interactions and conformations.
• Molecular Biology Research: Investigating the structural basis of RNA function in cellular processes.

Methodology:

A three-layer GCN extracts structural elements, Transformer Encoders embed structural information into sequences, and a UNet-based module captures long-range pairwise relationships.