Abstract
RNA-binding proteins (RBPs) are essential regulators of gene expression, and identifying their binding sites is critical for understanding cellular processes and disease mechanisms. Traditional prediction models often treat RNA as a linear sequence, neglecting the vital role of its secondary structure. We propose RNASSR-Net, a novel deep learning framework that combines Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs) to model RNA as a topological graph. By integrating sequence motifs with structural features through a base importance guidance mechanism, RNASSR-Net achieves state-of-the-art performance on the RBP-24 benchmark, particularly in data-scarce scenarios.
Introduction
The interaction between RNA and proteins is a fundamental biological process, influencing everything from splicing and transport to translation and viral replication. RNA-binding proteins (RBPs) recognize specific targets based on a combination of primary sequence motifs and complex secondary structures, such as hairpins, loops, and stems.
Existing computational methods for RBP binding prediction often struggle to capture the non-Euclidean nature of RNA folding. While 1D CNNs excel at detecting local sequence patterns, they are inherently limited in representing long-range interactions between nucleotides that are spatially close in a folded structure but distant in the linear sequence. To address this, we developed RNASSR-Net, which treats RNA secondary structure as a graph, allowing for a more faithful representation of the molecule's physical topology.
Methods
RNASSR-Net utilizes a dual-track architecture to process RNA data simultaneously as a sequence and a graph.
Dual-Track Architecture
- Sequence Track (CNN): This branch processes one-hot encoded RNA sequences using standard convolutional layers. It is designed to identify short, conserved nucleotide motifs that serve as primary binding signals.
- Structure Track (GCN): RNA secondary structures are predicted using
RNAfoldand converted into graphs where nucleotides are nodes and base-pairing probabilities define the edges. A Graph Convolutional Network then extracts features from this topological representation, capturing the spatial context of the binding site.
Base Importance Guidance
A key innovation of RNASSR-Net is the "base importance" mechanism. The model uses spatial importance weights derived from the CNN track to guide the GCN. This cross-track communication ensures that the graph-based learning focuses on structural regions most likely to contain relevant sequence motifs, effectively fusing sequence and structure information during the feature extraction process.
Results and Evaluation
The model was evaluated using the RBP-24 benchmark dataset, which includes 24 sub-datasets from various CLIP-seq experiments.
Performance Metrics
RNASSR-Net achieved a mean Area Under the Curve (AUC) of 94.4%, representing a significant improvement (1.3% to 5.7%) over existing baselines such as GraphProt and iDeepE. Notably, the model demonstrated robust performance on smaller datasets, such as the C17ORF85 protein, where it showed a 5.9% improvement. This suggests that incorporating structural priors is particularly beneficial when sequence data is limited.
Biological Validation and Motif Discovery
Beyond predictive accuracy, RNASSR-Net provides interpretable results by identifying binding motifs. The model successfully recovered well-known biological motifs, such as the poly-U stretches recognized by HNRNPC.
Discussion and Conclusion
The success of RNASSR-Net highlights the importance of topological representation in RNA biology. By treating RNA as a graph, the model captures long-range structural dependencies that are invisible to 1D sequence models. The integration of sequence and structure tracks via the base importance mechanism provides a comprehensive view of the RBP binding landscape.
While RNASSR-Net sets a new standard for binding prediction, several avenues for improvement remain. The model currently relies on predicted secondary structures; incorporating experimental structural data could further enhance accuracy. Additionally, extending the framework to account for 3D tertiary structures and optimizing the computational efficiency of the graph convolutions for long transcripts are key areas for future research.
Summary updated