PanGIA: A universal framework for identifying association between ncRNAs and diseases

Xiao-Yuan Liu, Xi Lu, Qiuhao Chen, Jimeng Sun, Tianyi Zhao, Yan ZhuGigaScience 2025 · DOI

TL;DR

This paper introduces PanGIA, a novel computational framework that unifies the prediction of associations between four major types of non-coding RNAs (ncRNAs) – miRNAs, lncRNAs, circRNAs, and piRNAs – and human diseases. By leveraging a heterogeneous graph attention network and a mixture-of-experts approach, PanGIA significantly outperforms existing type-specific methods, offering a more comprehensive and accurate understanding of ncRNA-disease relationships.

What Problem Does This Paper Address?

The intricate roles of non-coding RNAs (ncRNAs) in human health and disease are increasingly recognized. However, experimentally identifying these associations is a laborious and costly process. While computational methods have emerged to predict these links, they typically focus on a single type of ncRNA (e.g., miRNAs, lncRNAs). This approach overlooks the crucial fact that different ncRNA types interact and cooperate, forming complex regulatory networks that collectively influence disease pathogenesis. Existing methods fail to capture these cross-type interactions, limiting their ability to provide a holistic view of ncRNA-disease associations. This paper addresses this gap by proposing a unified framework, PanGIA, capable of simultaneously predicting associations across multiple ncRNA types and diseases.

Methods

PanGIA employs a multi-stage approach:

  1. Pretraining ncRNA Embeddings: The model utilizes DNABERT6, a transformer-based model adapted for DNA sequences, to generate rich, context-aware embeddings for each ncRNA sequence. This captures intricate sequence-level features that are crucial for downstream analysis.
  2. Heterogeneous Graph Construction: A comprehensive heterogeneous graph is built, integrating various data sources:
  3. ncRNA-disease associations: Known experimentally validated links.
  4. ncRNA-ncRNA similarity: Calculated based on both sequence embeddings (using cosine similarity) and functional similarity (using Gaussian Interaction Profile - GIP kernel).
  5. Disease-disease similarity: Derived from both semantic similarity (using Disease Ontology) and functional similarity (using GIP kernel).
  6. Multitask Association Prediction: A novel framework combines a Heterogeneous Graph Attention Network (HAN) with a Mixture-of-Experts (MoE) architecture.
  7. HAN: Processes the heterogeneous graph to learn node representations by attending to relevant neighbors and meta-paths.
  8. MoE: Employs multiple "expert" networks, allowing for specialized processing of features. A gating mechanism selects the most relevant experts for each task (predicting associations for each ncRNA type).
  9. Cross-Task Attention: A key innovation that allows the model to learn interdependencies between different ncRNA-disease prediction tasks, further enhancing overall performance and generalization. This mechanism enables collaborative modeling across tasks.

Datasets

The study utilizes several publicly available databases for constructing the heterogeneous graph:

  • ncRNA-disease associations: HMDD v4.0 (miRNA), LncRNADisease v3.0 (lncRNA/circRNA), piRDisease v1.0 (piRNA).
  • ncRNA sequences: miRBase (miRNA), circBase (circRNA), GENCODE/NONCODE (lncRNA), piRBase/piRNAdb (piRNA).
  • Disease information: Disease Ontology database for disease similarity.

The paper does not introduce a new benchmark dataset but rather integrates existing data to create a unified prediction framework.

Key Insights and Ideas

  • The Power of Integration: The most significant insight is that treating ncRNAs as a collective ("pan-ncRNA") and modeling their interrelationships, rather than studying them in isolation, dramatically improves prediction accuracy.
  • Heterogeneous Graph as a Unified View: Building a graph that explicitly encodes relationships between different ncRNA types and diseases provides a powerful representation for capturing complex biological interactions.
  • Mixture-of-Experts for Task Specialization: The MoE architecture allows the model to learn specialized representations for each ncRNA type while enabling knowledge sharing through cross-task attention.
  • DNABERT's Utility for ncRNA Sequences: Adapting DNA sequence models like DNABERT for ncRNA sequences proves effective in capturing nuanced sequence features that are critical for predicting functional roles and associations.

Key Results

  • Superior Performance: PanGIA consistently outperformed existing state-of-the-art methods across all evaluated ncRNA types (miRNA, lncRNA, circRNA, piRNA) and when integrating all types. Metrics like AUC, AUPR, and Rank Index showed significant improvements. For the pan-ncRNA task, PanGIA achieved an AUC of 0.988 ± 0.002 and AUPR of 0.985 ± 0.004.
  • Robustness: The model demonstrated robustness when nodes or ncRNA types were removed, indicating its resilience to data sparsity and the value of its integrated approach.
  • Biological Interpretability: Case studies confirmed that high-confidence associations predicted by PanGIA were supported by existing literature, highlighting the model's biological relevance and potential for guiding experimental research.
  • Ablation Studies: Removing core components like HAN, MoE, or the cross-task attention mechanism led to noticeable performance degradation, underscoring the importance of each module in PanGIA's success.

Limitations and Future Work

  • Data Dependency: The study acknowledges that PanGIA's performance is sensitive to the quantity of training data, with a decline in predictive capability under data-sparse conditions.
  • Future Directions: The authors suggest incorporating additional omics data (e.g., gene expression, epigenomic data) and optimizing the network architecture to further enhance predictive power and expand its applications in disease mechanism exploration and biomarker discovery.

Summary updated

Abstract

PanGIA demonstrates significant advantages in predicting disease associations for different types of ncRNAs, including miRNAs, lncRNAs, circRNAs, and piRNAs. Case studies further confirm the accuracy of the model's predictions, as all high-confidence associations were supported by literature evidence. This demonstrates the model's strong biological interpretability and promising potential for practical applications. The successful application of PanGIA provides a new paradigm for exploring disease-associated ncRNAs, highlighting their immense potential in the field of biomedical research.
BibTeX
@inproceedings{liu2025pangia,
  author = {Xiao-Yuan Liu and Xi Lu and Qiuhao Chen and Jimeng Sun and Tianyi Zhao and Yan Zhu},
  title = {{PanGIA: A universal framework for identifying association between ncRNAs and diseases}},
  booktitle = {GigaScience},
  year = {2025},
  doi = {10.1093/gigascience/giaf123},
}
Figure