Predicting circRNA-drug sensitivity associations via graph attention auto-encoder

Lei Deng, Lei Deng 0002, Zixuan Liu, Yurong Qian, Jingpu ZhangBMC Bioinformatics 2022 · DOI

TL;DR

This paper introduces GATECDA, a novel computational framework that uses a graph attention auto-encoder to predict associations between circular RNAs (circRNAs) and drug sensitivity. By integrating multiple similarity networks, GATECDA effectively learns representations of circRNAs and drugs, achieving high accuracy in predicting these crucial relationships for drug discovery.

What Problem Does This Paper Address?

Circular RNAs (circRNAs) are increasingly recognized for their roles in cancer development and therapy resistance. Their expression levels can significantly influence how effectively cells respond to drugs. Identifying these circRNA-drug sensitivity associations is vital for advancing cancer therapy and discovering new drugs. However, traditional experimental methods for validating these associations are time-consuming and expensive. This paper addresses the need for an efficient and accurate computational approach to predict these unknown associations, thereby accelerating drug discovery and development.

Methods

The core of this work is the GATECDA framework, which leverages a Graph Attention Auto-Encoder (GATE). Here's a breakdown of the key components:

  • Data Integration: GATECDA integrates data from multiple sources, including circRNA host gene sequences, drug structures, and known circRNA-drug sensitivity associations.
  • Similarity Networks: To capture relationships, the framework constructs various similarity networks:
  • Sequence similarity of host genes: Measures similarity between circRNA host genes using Levenshtein distance.
  • GIP kernel similarity: Calculates similarity based on the Gaussian interaction profile, assuming circRNAs associated with similar drugs are themselves similar.
  • Structural similarity of drugs: Uses RDKit and the Tanimoto method to assess similarity based on drug structures.
  • Similarity Fusion: To create a comprehensive view, similarities from different aspects are fused. The paper explores both SNF (Similarity Network Fusion) and an arithmetic average strategy, ultimately favoring the latter for its computational efficiency.
  • Graph Attention Auto-Encoder (GATE): This unsupervised learning model is central to GATECDA. It learns low-dimensional representations (embeddings) for circRNAs and drugs. GATE uses a self-attention mechanism to weigh the importance of neighbors, allowing it to capture complex relationships within the graph structure. The auto-encoder architecture reconstructs node attributes and graph structures, ensuring that the learned representations retain critical information.
  • Prediction Model: The learned representations from GATE are then fed into a fully connected neural network (DNN) to predict the association score between each circRNA-drug pair.

Datasets

The study utilizes several publicly available datasets:

  • circRic database: Provides circRNA-drug sensitivity associations, derived from the GDSC database. This dataset contains information on 80076 associations involving 404 circRNAs and 250 drugs.
  • National Center for Biotechnology Information (NCBI) Gene database: Used to obtain sequences of host genes for circRNAs.
  • PubChem database: Used to retrieve structural information for drugs.

The training set was curated from the circRic database, focusing on significant associations (FDR < 0.05), resulting in 4134 associations involving 271 circRNAs and 218 drugs.

Key Insights and Ideas

  • The Power of Graph Attention: The use of Graph Attention Auto-Encoders (GATE) is a key innovation, allowing the model to learn rich, context-aware representations of circRNAs and drugs by considering their network neighborhoods. This is more sophisticated than traditional methods that might only consider pairwise interactions.
  • Multi-Modal Similarity Fusion: Combining different types of similarity (sequence, GIP, structural) provides a more holistic understanding of circRNA-drug relationships. The paper demonstrates that fusing these networks leads to better predictive performance than using single similarity measures.
  • Addressing the "Cold Start" Problem: The paper acknowledges the challenge of predicting associations for new drugs with no prior known interactions. While their current approach of random assignment is a limitation, it highlights an important area for future research.

Key Results

  • High Predictive Accuracy: GATECDA achieved an impressive average AUC of 89.18% under 10-fold cross-validation and 88.46% under 5-fold cross-validation.
  • Superior Performance: When compared to other methods (KATZ, VGAE, VGAMF, GCNMDA), GATECDA consistently outperformed them across various metrics, including AUC and AUPR, demonstrating its effectiveness in predicting circRNA-drug sensitivity associations.
  • Effectiveness of Fusion: The study showed that GATECDA, which fuses multiple similarity networks, significantly outperformed configurations using only sequence similarity (GATECDA-S) or GIP kernel similarity (GATECDA-G).
  • Case Studies: Case studies on drugs like PAC-1 and Foretinib, as well as de novo predictions for new drugs, provided qualitative evidence of GATECDA's ability to identify relevant circRNAs.

Limitations and Future Work

The authors acknowledge a key limitation: the "cold start" problem when predicting associations for new drugs that have no known interactions in the training dataset. Their current solution of random assignment for initial drug features is noted as not being particularly effective.

For future work, the authors plan to:

  • Collect More Data: Acquire a larger dataset of circRNA-drug sensitivity associations.
  • Integrate More Biological Information: Incorporate additional biological data, such as circRNA-circRNA similarities and associations between circRNAs, drugs, and diseases, to reduce reliance on known associations.
  • Enhance New Drug Prediction: Develop more sophisticated methods to address the cold start problem for new drugs, moving beyond simple random assignment.

Summary updated

BibTeX
@article{deng2022predicting,
  author = {Lei Deng and Lei Deng and Zixuan Liu and Yurong Qian and Jingpu Zhang},
  title = {{Predicting circRNA-drug sensitivity associations via graph attention auto-encoder}},
  journal = {BMC Bioinformatics},
  year = {2022},
  doi = {10.1186/s12859-022-04694-y},
}
Figure