CFRPI: Predicting ncRNA-Protein Interactions via Counterfactual Learning-Integrated Heterogeneous Graph Neural Networks

CFRPI is an innovative project developed by zParaselene, open-sourced on GitHub (released on May 31, 2026). This project combines a counterfactual learning framework with heterogeneous graph neural networks to predict interactions between non-coding RNAs (ncRNAs) and proteins. It aims to provide an efficient and interpretable new tool for bioinformatics research, addressing the issues of high cost, long cycle time of traditional methods, and lack of interpretability in black-box models.

Non-coding RNAs (ncRNAs) play critical roles in biological processes such as gene regulation, cell differentiation, and disease occurrence. Understanding the interactions between ncRNAs and proteins (ncRPI) is of great significance for revealing gene expression regulatory mechanisms and disease pathology. However, experimental methods for identifying ncRPI are costly and time-consuming, making it difficult to meet large-scale screening needs. Thus, developing efficient and accurate computational methods has become an important direction in the field of bioinformatics.

ncRPI prediction faces two core challenges: First, data heterogeneity—ncRNAs and proteins, as different biological molecules, have large differences in sequence features, structural properties, and functional patterns, so effectively fusing heterogeneous information is key to modeling. Second, interpretability of prediction results—biologists not only need to know the interaction relationships but also want to understand the underlying mechanisms, which traditional black-box models are difficult to satisfy.

The core innovation of CFRPI lies in the introduction of a counterfactual learning framework. Counterfactual learning helps the model identify key factors affecting predictions by constructing hypothetical scenarios (e.g., changes in results when a certain feature is different). In ncRPI prediction, this method can distinguish between causal effects and statistical correlations of sequence features, significantly improving the interpretability of predictions and providing clear feature importance indicators for biologists.

CFRPI uses a heterogeneous graph neural network as its basic architecture: ncRNAs and proteins are represented as different types of nodes, and known interactions form edges. Through a message-passing mechanism, ncRNA nodes aggregate features of interacting proteins, and protein nodes aggregate features of related ncRNAs. Bidirectional information flow allows the model to learn cross-molecular type representations and capture complex patterns of ncRPI.

CFRPI was evaluated on multiple public datasets. Compared with matrix factorization, random forests, and homogeneous graph neural networks, it showed significant improvements in metrics such as AUC and AUPR. More importantly, the key sequence segments identified through counterfactual analysis are highly consistent with known biological knowledge, verifying its advantages in prediction accuracy and biological interpretability.

CFRPI provides an efficient and interpretable computational tool for ncRPI research: Wet-lab biologists can use it to screen candidate interaction pairs and narrow down the experimental scope; computational biologists can draw on its counterfactual learning framework for predicting interactions between other biological molecules; the open-source implementation facilitates community reproduction of results, expansion of functions, and promotion of method improvements.

CFRPI successfully combines counterfactual learning with heterogeneous graph neural networks, providing a new solution for ncRPI prediction. It enhances interpretability on the basis of high prediction accuracy, bridging the gap between computational prediction and biological understanding. In the future, with the development of single-cell sequencing technology, integrating cell context information will be a further direction for this method and similar research.