GraphLncLoc

GraphLncLoc predicts the subcellular localization of long non-coding RNAs (lncRNAs) to support interpretation of their biological functions.

Key Features:

• de Bruijn graph transformation: Transforms lncRNA sequences into de Bruijn graphs to represent sequence structure and k-mer connectivity.
• Graph convolutional networks (GCNs): Applies GCNs to de Bruijn graphs to extract high-level graph features and learn latent sequence representations.
• Graph classification framing: Redefines sequence classification as a graph classification problem to leverage graph-structured learning.
• Fully connected classifier: Feeds extracted feature vectors into a fully connected layer for final subcellular localization prediction.
• Contrast to k-mer frequency methods: Preserves sequence order information and captures variable-length patterns that k-mer frequency features do not.
• Motif identification: Identifies significant motifs associated with nuclear subcellular localization in case studies.
• Improved performance: Demonstrates improved accuracy and more distinguishable, robust features compared to traditional machine learning predictors based on k-mer frequencies.

Scientific Applications:

• lncRNA localization prediction: Predicts subcellular localization of lncRNAs to inform hypotheses about their cellular roles.
• Functional inference: Supports interpretation of lncRNA biological function based on predicted localization.
• Motif discovery for nuclear localization: Identifies motifs associated with nuclear localization to guide further experimental validation.
• Benchmarking of predictors: Provides a graph-based approach for comparative evaluation against k-mer frequency-based localization predictors.

Methodology:

Convert lncRNA sequences into de Bruijn graphs, apply graph convolutional networks to extract high-level features and latent representations, and input the resulting feature vectors into a fully connected layer for localization prediction.