Integration of Neighbor Topologies Based on Meta-Paths and Node Attributes for Predicting Drug-Related Diseases

Ping Xuan, Zixuan Lu, Tiangang Zhang, Yong Liu, Toshiya NakaguchiInternational Journal of Molecular Sciences 2022 · DOI

TL;DR

This paper introduces NAPred, a novel method for predicting new disease indications for existing drugs by deeply integrating drug and disease network topology with their attributes. NAPred leverages multi-scale meta-paths and attention mechanisms to capture complex relationships, outperforming existing methods in drug-disease association prediction and demonstrating its utility in identifying potential drug repositioning candidates.

What Problem Does This Paper Address?

Developing new drugs is a notoriously long, expensive, and high-risk endeavor. Drug repositioning, which involves finding new uses for existing, approved drugs, offers a promising avenue to accelerate this process and reduce costs, as these drugs already have established safety profiles. The challenge lies in accurately predicting which diseases an existing drug might effectively treat. While previous computational methods have explored drug and disease similarities and associations, they often fail to deeply integrate the intricate topological relationships within complex biological networks or effectively combine node attributes. This paper aims to bridge this gap by proposing a method that comprehensively captures and leverages both the structural (topological) and feature (attribute) information of drugs and diseases within heterogeneous networks.

Methods

NAPred employs a two-branch architecture:

  1. Neighbor Topology Encoding Branch: This branch focuses on understanding the structural relationships.
  2. Heterogeneous Networks: Three drug-disease heterogeneous networks are constructed based on different drug similarity metrics (chemical substructure, protein structural domain, and gene ontology).
  3. Multi-Scale Meta-Paths: The method utilizes meta-paths of varying lengths and types (e.g., drug-drug-drug, drug-disease-drug) to define "neighbor" sets for drug and disease nodes at different scales.
  4. Attention Mechanisms: * Neighbor-Scale-Level Attention: This mechanism adaptively combines feature representations from neighbors at different scales, assigning importance based on their contribution.
  5. Neighbor-Topology-Level Attention: This further refines the representation by learning the relative importance of different types of neighbor topologies (e.g., drug neighbors vs. disease neighbors).
  6. CNN-Based Encoding: A Convolutional Neural Network (CNN) with attention is used to learn the pairwise neighbor topology representations for drug-disease pairs.
  7. Attribute Encoding Branch: This branch focuses on the intrinsic properties of drugs and diseases.
  8. Attribute Embedding: Node attributes (chemical substructures, protein structural domains, gene ontologies) are embedded for drug-disease pairs.
  9. CAE-Based Encoding: A Convolutional Autoencoder (CAE) is employed to learn a compressed, representative encoding of these pairwise attributes, aiming to capture essential information while reducing dimensionality and sparsity.

Finally, the outputs from both branches are integrated and weighted to produce a final association score, indicating the likelihood of a drug-disease relationship.

Datasets

The study utilizes the following datasets:

  • UMLS: For drug-disease association data, containing information on 763 drugs, 681 diseases, and 3051 known associations.
  • PubChem: For drug chemical substructure data.
  • InterPro: For drug target protein structural domain data.
  • UniProt: For gene ontology information of target proteins.

The paper also mentions using the CC dataset to test the robustness of their method (NAPredDD), replacing the original data with its chemistry, targets, and network information.

Key Insights and Ideas

  • The Power of Heterogeneity and Multi-Scale: Simply looking at direct drug-disease links isn't enough. NAPred's success stems from constructing multiple heterogeneous networks that capture diverse drug similarities and then exploring relationships through meta-paths of varying lengths. This allows for the discovery of indirect but significant connections.
  • Attention is All You Need (for Neighbors): The paper highlights the critical role of attention mechanisms. By learning to weigh the importance of neighbors at different scales and across different topological structures, NAPred can focus on the most informative signals, leading to more accurate predictions.
  • Synergy of Topology and Attributes: The dual-branch approach, combining structural network information with intrinsic node attributes, is key. Neither topology nor attributes alone are sufficient; their integration provides a more holistic understanding of drug-disease relationships.
  • Deep Learning for Complex Relationships: The application of CNNs and CAEs allows for the extraction of complex, non-linear patterns from the rich, high-dimensional data, which simpler models might miss.

Key Results

  • Superior Performance: NAPred significantly outperformed six state-of-the-art drug-disease association prediction methods, achieving the best mean AUC (0.978) and AUPR values. For example, it outperformed GFPred by 3.3% in AUC and by 14.8% in AUPR.
  • Improved Recall: The method demonstrated a higher recall rate, meaning it was more successful at identifying actual drug-disease associations among the top-ranked candidates. For instance, NAPred achieved recall rates of 86.14%, 89.19%, 93.24%, 95.54%, and 97.33% for top-30, -60, -90, -120, and -150 candidates, respectively, generally outperforming other methods.
  • Robustness: NAPred showed good robustness when tested on a different dataset (CC dataset).
  • Case Studies: Case studies on five drugs (ampicillin, ceftriaxone, doxorubicin, erythromycin, and itraconazole) further illustrated NAPred's ability to identify plausible, literature-supported potential drug-disease associations.

Limitations and Future Work

  • Parameter Tuning: The paper mentions that optimal parameters were used for each method during cross-validation, implying that parameter tuning is crucial for achieving peak performance.
  • Data Sparsity: While CAE helps, the inherent sparsity of biological data remains a challenge.
  • Future Work: The authors suggest that their predictive model can serve as a valuable tool for biologists to guide wet laboratory experiments, implying a future direction towards experimental validation and refinement of the predictions. The paper also implicitly suggests that further exploration of different meta-path strategies and network construction methods could yield even better results.

Summary updated

Abstract

Identifying new disease indications for existing drugs can help facilitate drug development and reduce development cost. The previous drug-disease association prediction methods focused on data about drugs and diseases from multiple sources. However, they did not deeply integrate the neighbor topological information of drug and disease nodes from various meta-path perspectives. We propose a prediction method called NAPred to encode and integrate meta-path-level neighbor topologies, multiple kinds of drug attributes, and drug-related and disease-related similarities and associations. The multiple kinds of similarities between drugs reflect the degrees of similarity between two drugs from different perspectives. Therefore, we constructed three drug-disease heterogeneous networks according to these drug similarities, respectively. A learning framework based on fully connected neural networks and a convolutional neural network with an attention mechanism is proposed to learn information of the neighbor nodes of a pair of drug and disease nodes. The multiple neighbor sets composed of different kinds of nodes were formed respectively based on meta-paths with different semantics and different scales. We established the attention mechanisms at the neighbor-scale level and at the neighbor topology level to learn enhanced neighbor feature representations and enhanced neighbor topological representations. A convolutional-autoencoder-based module is proposed to encode the attributes of the drug-disease pair in three heterogeneous networks. Extensive experimental results indicated that NAPred outperformed several state-of-the-art methods for drug-disease association prediction, and the improved recall rates demonstrated that NAPred was able to retrieve more actual drug-disease associations from the top-ranked candidates. Case studies on five drugs further demonstrated the ability of NAPred to identify potential drug-related disease candidates.
BibTeX
@article{xuan2022integration,
  author = {Ping Xuan and Zixuan Lu and Tiangang Zhang and Yong Liu and Toshiya Nakaguchi},
  title = {{Integration of Neighbor Topologies Based on Meta-Paths and Node Attributes for Predicting Drug-Related Diseases}},
  journal = {International Journal of Molecular Sciences},
  year = {2022},
  doi = {10.3390/ijms23073870},
}
Figure