Understanding how proteins interact with RNA is crucial for uncovering how genes are regulated and how diseases develop. RNA-binding proteins (RBPs) control many essential biological processes, including how RNA is processed, transported, and translated into proteins. However, existing computational tools often focus on specific proteins and rely heavily on the experimental setup used to collect data, which limits their ability to work across different studies.

Researchers from Soochow University have developed a new computational tool called PaRPI, designed to overcome these limitations. PaRPI combines data from a wide range of experiments and cell types, allowing it to learn common and unique patterns of how proteins bind to RNA. By integrating information from different datasets, PaRPI creates a unified model that can predict where RNA and proteins interact, even when analyzing previously unseen molecules.

The architecture of PaRPI

a The overall framework of PaRPI. After tokenization, RNA sequences are input into a multi-head self-attention layer to generate dynamic contextual embeddings. The icSHAPE pipeline is used to generate RNA secondary structure profiles, while RNAplfold provides predicted RNA secondary structure information. RNA features are then structured into a graph network. The input protein sequences are processed by the pre-trained ESM-2 model to extract protein features. b Learning interactions module. RNA representations are learned through GraphConv and Transformer to capture both local and global information, with the DPRBP network module performing stepwise dimensionality reduction on sequence length. Interaction module facilitates the interaction between RNA sequence representations and protein representations. c Graphconv module. Node spatial features consist of sequence adjacency edges and secondary structure features from RNAplfold. GraphSAGE is used to aggregate information from neighboring nodes. d Interaction module. Learning the interactions between RNA representations and protein representations through element-wise multiplication. e DPRBP module. A multi-layer feature selector module is developed to capture context-sensitive information by gradually reducing the dimensionality of the sequence.

The team tested PaRPI on 261 datasets from eCLIP and CLIP-seq experiments, and the model outperformed all existing methods in accurately identifying RNA-protein binding sites. More impressively, PaRPI demonstrated strong predictive power for new RNA and protein combinations, highlighting its versatility.

Beyond improving predictions, PaRPI can also help scientists explore how genetic variants linked to diseases might alter RNA-protein interactions. This makes it a powerful tool for advancing our understanding of gene regulation and identifying new molecular mechanisms behind complex diseases.

By integrating large-scale data and applying machine learning, PaRPI paves the way for deeper insights into RNA-protein communication and its role in human health.

Availability – The code is available on the GitHub repository https://github.com/ljquanlab/PaRPI or the Zenodo repository https://doi.org/10.5281/zenodo.16794270