<scp>NGCN</scp>: <scp>Drug‐target</scp> interaction prediction by integrating information and feature learning from heterogeneous network

Junyue Cao, Qingfeng Chen, Junlai Qiu, Yiming Wang, Wei Lan, Xiaojing Du, Kai TanJournal of Cellular and Molecular Medicine 2024 · DOI

TL;DR

This paper introduces NGCN, a novel method for predicting drug-target interactions (DTIs) by integrating diverse biological data into a heterogeneous network. NGCN effectively learns low-dimensional topological representations of drugs and targets using a graph convolutional neural network, leading to significant improvements in prediction accuracy over existing state-of-the-art methods.

What Problem Does This Paper Address?

Developing new drugs is a complex, costly, and time-consuming process. A critical step is understanding how drugs interact with their biological targets (proteins). Traditional experimental methods for identifying these interactions are laborious and expensive. Computational approaches, particularly those leveraging machine learning, offer a more efficient alternative. However, existing methods often struggle with the inherent complexity, high dimensionality, and noise present in the vast and diverse biological data available (e.g., drug properties, protein functions, disease associations). This paper aims to address these challenges by proposing a method that can effectively integrate and learn from these heterogeneous data sources to improve DTI prediction accuracy.

Methods

The core of NGCN lies in its ability to learn meaningful representations from a heterogeneous network. The process involves several key steps:

  • Network Construction: NGCN integrates multiple types of biological networks, including drug-drug interactions, protein-protein interactions, drug-disease associations, and protein-disease associations, into a unified heterogeneous network.
  • Diffusion State Calculation (RWR): For each individual network, a Random Walk with Restart (RWR) algorithm is applied to compute the "diffusion state" of each drug and protein node. This diffusion state captures the topological relationships of a node with all other nodes in that network.
  • Feature Reduction (clusDCA): To handle the high dimensionality and potential noise in the diffusion states, an improved diffusion component analysis method (clusDCA) is employed. This step reduces the diffusion state features into a lower-dimensional representation, preserving essential topological information.
  • Feature Updating (Spectral GCN): The low-dimensional features are then further refined using a spectral graph convolutional neural network (GCN). This GCN layer updates the node features by aggregating neighborhood information, effectively learning topology-preserving representations.
  • DTI Prediction: Finally, the updated drug and target features are used to reconstruct a drug-target interaction matrix, predicting potential DTIs.

The novelty lies in the combination of RWR for capturing diffusion states, clusDCA for dimensionality reduction, and spectral GCN for feature refinement within a heterogeneous network framework.

Datasets

The paper utilizes a dataset that includes four types of nodes: drug nodes, protein nodes, disease nodes, and side-effect nodes. The dataset also comprises six types of association networks: drug-protein association, protein-protein association, drug-drug interaction, drug-disease, protein-disease association, and drug-side effect networks. For validation, the authors also used four drug-target interaction datasets created by Yamanishi et al. (Enzyme, Ion Channels, GPCR, and Nuclear Receptors).

Key Insights and Ideas

  • Heterogeneous Networks are Key: The paper strongly emphasizes that integrating information from multiple, diverse biological networks is crucial for accurate DTI prediction. No single network type is sufficient.
  • Topology Matters: NGCN's success hinges on its ability to learn the "topology" of the network – how nodes are connected and their relative positions. This is achieved through RWR and GCN.
  • Dimensionality Reduction is Essential: The high dimensionality and noise in biological data can be a major bottleneck. NGCN's use of clusDCA effectively addresses this by creating more manageable and informative feature representations.
  • GCN Enhances Feature Representation: The spectral GCN layer plays a vital role in refining the learned features, leading to superior performance compared to methods that rely solely on raw diffusion states.
  • Extensible Framework: The proposed NGCN framework is designed to be flexible, allowing for the incorporation of additional network types and information, suggesting its potential for future advancements.

Key Results

NGCN demonstrated substantial performance improvements over several state-of-the-art methods.

  • AUPR Improvement: The paper reports a nearly 1.0% increase in AUPR (Area Under the Precision-Recall Curve) compared to other leading methods.
  • Robustness: NGCN showed robust performance even when faced with sparse positive samples (a 1:10 ratio of positive to negative examples), outperforming other algorithms whose performance degraded significantly.
  • Validation on Multiple Datasets: NGCN outperformed existing methods on four different drug-target interaction datasets (Enzyme, Ion Channels, GPCR, and Nuclear Receptors), indicating its generalizability.
  • Feature Update Benefit: Experiments showed that the feature updating operation (using GCN) significantly improved prediction performance, as highlighted in Table 1.

Limitations and Future Work

  • Handling Low-Degree Nodes: The authors acknowledge that significant differences in node degrees within the graph network can pose a challenge for effective information extraction, particularly from low-degree nodes.
  • Future Directions:
    • Integrating More Diverse Data: The authors plan to enhance the framework by integrating even more diverse network data to achieve a more comprehensive understanding of drug-target interactions.
  • Addressing Node Degree Differences: Future work will focus on developing strategies to effectively extract information from low-degree nodes, improving the model's ability to handle imbalanced network structures.

Summary updated

Abstract

Drug-target interaction (DTI) prediction is essential for new drug design and development. Constructing heterogeneous network based on diverse information about drugs, proteins and diseases provides new opportunities for DTI prediction. However, the inherent complexity, high dimensionality and noise of such a network prevent us from taking full advantage of these network characteristics. This article proposes a novel method, NGCN, to predict drug-target interactions from an integrated heterogeneous network, from which to extract relevant biological properties and association information while maintaining the topology information. It focuses on learning the topology representation of drugs and targets to improve the performance of DTI prediction. Unlike traditional methods, it focuses on learning the low-dimensional topology representation of drugs and targets via graph-based convolutional neural network. NGCN achieves substantial performance improvements over other state-of-the-art methods, such as a nearly 1.0% increase in AUPR value. Moreover, we verify the robustness of NGCN through benchmark tests, and the experimental results demonstrate it is an extensible framework capable of combining heterogeneous information for DTI prediction.
BibTeX
@article{cao2024scp,
  author = {Junyue Cao and Qingfeng Chen and Junlai Qiu and Yiming Wang and Wei Lan and Xiaojing Du and Kai Tan},
  title = {{<scp>NGCN</scp>: <scp>Drug‐target</scp> interaction prediction by integrating information and feature learning from heterogeneous network}},
  journal = {Journal of Cellular and Molecular Medicine},
  year = {2024},
  doi = {10.1111/jcmm.18224},
}
Figure