A Fused Deep Learning Approach to Transform Novel Drug Repositioning

Dongsheng Cao, Kun Li, Jiacai Yi, Qing Ye, Xixi Yang, Yu Long, Youchao Deng, Chengkun Wu, Tingjun Hou, Dejun JiangResearch Square 2024 · DOI

Abstract

Drug repositioning holds significant promise for discovering new therapeutic applications of existing drugs, thereby accelerating drug development, reducing associated costs, and improving overall efficiency. However, current methodologies encounter difficulties in effectively managing diverse network representations, tackling cold start issues, and handling intrinsic attribute representations. In this study, we introduce UKEDR, a fused deep learning framework that seamlessly integrates knowledge graph embedding, sophisticated pre-training strategies, and recommendation systems to address these challenges in drug repositioning. In UKEDR, a straightforward yet effective semantic similarity-driven embedding approach that leverages both pre-trained embeddings and knowledge graph structure was proposed to overcome the intractable cold start issue. Our comprehensive evaluations reveal that UKEDR outperforms various state-of-the-art baselines, including classical machine learning, network-based and deep learning approaches. In a cold start scenario simulating real-world drug discovery, UKEDR achieves 24.2% higher AUPR compared to the latest state-of-the-art, highlighting its superior capability in handling unseen nodes and generalizing to novel compounds. Furthermore, we demonstrate the effectiveness of UKEDR through repurposing case studies for various diseases such as falciparum malaria, prostate cancer and so on. Finally, model interpretability is enhanced through knowledge graph visualization, providing valuable insights into the drug repositioning process.

Introduction

Traditional drug discovery is often described as a "valley of death" venture, typically requiring 15 years and up to $2 billion in investment, yet suffering from a 90% failure rate. Drug repositioning—the process of identifying new therapeutic uses for existing, FDA-approved drugs—offers a critical shortcut to this process. However, computational drug repositioning faces three primary hurdles: the Cold Start Problem, where models struggle with compounds lacking prior interaction data; Attribute Neglect, where models focus on network topology while ignoring intrinsic chemical or pathological properties; and Complex Feature Interaction, where simple mathematical models fail to capture the non-linear relationships between drugs and diseases.

The UKEDR Framework

To address these challenges, we developed UKEDR (Unified Knowledge-graph Embedding for Drug Repositioning), a multi-modal pipeline that treats repositioning as a recommendation system problem. The framework consists of three core components:

  1. Knowledge Graph Embedding (KGE): We utilize PairRE, a model that represents relations as paired vectors, to map a massive biomedical network into a low-dimensional latent space. This captures the relational "neighborhood" of entities.
  2. Dual-Stream Pre-training: To capture intrinsic "identity," we employ CReSS for drugs (using contrastive learning on molecular SMILES strings and 13C NMR spectral data) and DisBERT for diseases (a domain-specific language model fine-tuned on 400,000 disease descriptions).
  3. Attentional Factorization Machines (AFM): Instead of standard dot products, UKEDR uses AFM to fuse these features. This allows the model to use an attention mechanism to weight specific feature interactions—such as chemical structure versus pathway position—for each drug-disease pair.
Performance heatmaps comparing various Knowledge Graph Embedding (KGE) methods across different Recommendation System (RS) backbones.

A critical innovation in UKEDR is the Semantic Similarity Cold-Start Strategy. When an unseen drug is introduced, the model identifies its neighbors in the pre-trained semantic space and calculates its initial graph position as the centroid of its k-nearest neighbors. This effectively "anchors" new compounds into the existing Knowledge Graph.

Experimental Results

The study utilized a custom-built "Mega-KG" integrating 11 data sources, encompassing 2.3 million entities and over 157 million relationships.

Performance Benchmarks

  • Cold Start Superiority: In simulations of real-world discovery (predicting outcomes for drugs in clinical trials not yet in the training set), UKEDR achieved an AUPR of 0.961, outperforming the previous state-of-the-art (DREAMwalk) by 24.2%.
  • Warm Start Performance: On the RepoAPP dataset, the model achieved an AUC of 0.958 and an AUPR of 0.968, consistently ranking at the top across all standard benchmarks.

Case Studies and Interpretability

The model's practical utility was demonstrated through several case studies. UKEDR successfully "rediscovered" the potential of Losartan for treating hypercholesterolemia and identified Fluticasone propionate as a candidate for falciparum malaria. These predictions are supported by recent independent literature, validating the framework's predictive power. Furthermore, the use of knowledge graph visualization enhances model interpretability, allowing researchers to trace the relational paths that lead to a specific recommendation.

Conclusion

UKEDR represents a significant advancement in drug repositioning by effectively bridging the gap between relational network data and intrinsic semantic attributes. By solving the cold start problem with a 24.2% improvement over existing methods, UKEDR provides a robust tool for identifying novel therapeutic applications for both existing and entirely new compounds. Future work will focus on incorporating real-world evidence (RWE) and clinical records to further refine these predictions.

Summary updated

Abstract

<title>Abstract</title> Drug repositioning holds significant promise for discovering new therapeutic applications of existing drugs, thereby accelerating drug development, reducing associated costs, and improving overall efficiency. However, current methodologies encounter difficulties in effectively managing diverse network representations, tackling cold start issues, and handling intrinsic attribute representations. In this study, we introduce UKEDR, a fused deep learning framework that seamlessly integrates knowledge graph embedding, sophisticated pre-training strategies, and recommendation systems to address these challenges in drug repositioning. In UKEDR, a straightforward yet effective semantic similarity-driven embedding approach that leverages both pre-trained embeddings and knowledge graph structure was proposed to overcome the intractable cold start issue. Our comprehensive evaluations reveal that UKEDR outperforms various state-of-the-art baselines, including classical machine learning, network-based and deep learning approaches. In a cold start scenario simulating real-world drug discovery, UKEDR achieves 24.2% higher AUPR compared to the latest state-of-the-art, highlighting its superior capability in handling unseen nodes and generalizing to novel compounds. Furthermore, we demonstrate the effectiveness of UKEDR through repurposing case studies for various diseases such as falciparum malaria, prostate cancer and so on. Finally, model interpretability is enhanced through knowledge graph visualization, providing valuable insights into the drug repositioning process.
BibTeX
@inproceedings{cao2024fused,
  author = {Dongsheng Cao and Kun Li and Jiacai Yi and Qing Ye and Xixi Yang and Yu Long and Youchao Deng and Chengkun Wu and Tingjun Hou and Dejun Jiang},
  title = {{A Fused Deep Learning Approach to Transform Novel Drug Repositioning}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-5416722/v1},
}
Figure