IGCNSDA: unraveling disease-associated snoRNAs with an interpretable graph convolutional network

Lei Deng, Xiaowen Hu, Pan Zhang, Dayun Liu, Jiaxuan Zhang, Yuanpeng Zhang, Yihan Dong, Yanhao Fan, Lei Deng 0002Briefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces IGCNSDA, a novel graph convolutional network (GCN) approach that accurately predicts associations between small nucleolar RNAs (snoRNAs) and diseases. Unlike traditional "black box" methods, IGCNSDA offers interpretability, providing insights into the underlying mechanisms of these associations and aiding in the development of new diagnostic and therapeutic strategies.

What Problem Does This Paper Address?

Small nucleolar RNAs (snoRNAs) are non-coding RNAs involved in various cellular processes, and their dysregulation is increasingly linked to human diseases. Identifying these snoRNA-disease associations is crucial for advancing disease detection and treatment. While experimental methods exist, they are time-consuming and costly. Computational methods, particularly deep learning, have emerged as promising alternatives. However, many existing deep learning models for predicting snoRNA-disease associations function as "black boxes," lacking interpretability. This means they can predict associations but cannot explain why they make those predictions, hindering a deeper understanding of the biological mechanisms involved. This paper aims to bridge this gap by developing an interpretable GCN model for snoRNA-disease association prediction.

Methods

The core of IGCNSDA is a Graph Convolutional Network (GCN) that leverages the relationships between snoRNAs and diseases. The process involves several key steps:

  • SnoRNA-Disease Bipartite Graph Construction: Known associations between snoRNAs and diseases are represented as a bipartite graph, where snoRNAs and diseases are nodes, and an edge signifies a known association.
  • Subgraph Generation: This is a novel contribution. IGCNSDA employs a subgraph generation algorithm to group similar snoRNAs and their associated diseases into distinct subgraphs. This is framed as a multi-classification task where each snoRNA is assigned to a subgraph. This step is crucial for isolating potentially similar biological contexts and reducing noise.
  • High-Order Embedding Propagation: Within each subgraph, a GCN (specifically, a LightGCN variant) is used to propagate information and update node embeddings. This allows the model to capture higher-order relationships between snoRNAs and diseases within these localized contexts.
  • Layer Combination: Embeddings from different GCN layers are combined to form the final representations for snoRNAs and diseases.
  • Association Score Prediction: The final association score between a snoRNA-disease pair is calculated using the inner product of their learned embeddings.
  • Optimization: The model is trained using the Bayesian Personalized Ranking (BPR) loss function, which aims to rank known associations higher than unknown ones.

Datasets

  • RNADisease v4.0: This is the primary dataset used for training and evaluation. It contains experimentally validated and predicted associations between non-coding RNAs and diseases. For this study, it was filtered to include 471 snoRNAs, 84 diseases, and 1095 known associations. This dataset was split into training (80%) and testing (20%) sets.
  • ncRPheo: This dataset was used for independent testing to further validate the model's performance. It provided 439 associations between 82 diseases and 13 snoRNAs, excluding those already present in RNADisease.

Key Insights and Ideas

  • Subgraph Generation for Noise Reduction: The novel subgraph generation mechanism is a key innovation. By grouping similar snoRNAs and their associated diseases, IGCNSDA effectively isolates relevant biological contexts, mitigating the impact of noisy data that can arise from aggregating information across the entire, potentially heterogeneous, graph.
  • Interpretability through Subgraphs: The subgraph approach not only improves performance but also enhances interpretability. It allows researchers to focus on specific clusters of related snoRNAs and diseases, providing a more targeted understanding of their interactions.
  • Capturing Higher-Order Relationships: By propagating information within these subgraphs, IGCNSDA can capture more nuanced, higher-order associations that might be missed by methods that consider the entire graph at once.
  • Superior Performance with Interpretability: The study demonstrates that achieving interpretability does not come at the cost of predictive accuracy. IGCNSDA outperforms several state-of-the-art methods while offering valuable insights.

Key Results

  • Performance: IGCNSDA achieved superior performance on the RNADisease dataset, with an average AUC of 0.8438, AUPR of 0.8744, and accuracy of 0.7831. These metrics showed significant improvements over other methods.
  • Independent Testing: On the ncRPheo dataset, IGCNSDA also demonstrated the best predictive performance, with an AUC of 0.7142, AUPR of 0.8746, and accuracy of 0.7290.
  • Efficiency: IGCNSDA was significantly faster and more memory-efficient than comparative methods, with a runtime approximately 70 times faster than the suboptimal method and consuming less memory.
  • Robustness to Noise: Sensitivity analysis showed that IGCNSDA's subgraph generation algorithm effectively mitigates the impact of noisy data, especially in the training set.
  • Interpretability: t-SNE visualizations confirmed that IGCNSDA effectively clusters embeddings of related snoRNAs and diseases. Case studies on Prostate Cancer and Lung Cancer identified known and potential snoRNA-disease associations, validating the model's practical utility.

Limitations and Future Work

  • Overfitting with More Layers: The authors acknowledge that, like many GCN-based models, increasing the number of graph convolutional layers can lead to overfitting.
  • Future Work: The authors plan to explore pre-training extensive models on large-scale datasets using transfer learning. This will be followed by fine-tuning on downstream datasets and customizing interpretable tasks to further enhance generalization capability, mitigate over-smoothing, and improve interpretability.

Summary updated

BibTeX
@article{deng2024igcnsda,
  author = {Lei Deng and Xiaowen Hu and Pan Zhang and Dayun Liu and Jiaxuan Zhang and Yuanpeng Zhang and Yihan Dong and Yanhao Fan and Lei Deng},
  title = {{IGCNSDA: unraveling disease-associated snoRNAs with an interpretable graph convolutional network}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae179},
}
Figure