Leveraging Network Target Theory for Efficient Prediction of Drug-Disease Interactions: A Transfer Learning Approach

Abstract

Efficient virtual screening methods can expedite drug discovery and facilitate the development of innovative therapeutics. This study presents a novel transfer learning model based on network target theory, integrating deep learning techniques with diverse biological molecular networks to predict drug-disease interactions. By incorporating network techniques that leverage vast existing knowledge, the approach enables the extraction of more precise and informative drug features, resulting in the identification of 88,161 drug-disease interactions involving 7,940 drugs and 2,986 diseases. Furthermore, this model effectively addresses the challenge of balancing large-scale positive and negative samples, leading to improved performance across various evaluation metrics such as an Area under curve (AUC) of 0.9298 and an F1 score of 0.6316. Moreover, the algorithm accurately predicts drug combinations and achieves an F1 score of 0.7746 after fine-tuning. Additionally, it identifies two previously unexplored synergistic drug combinations for distinct cancer types in disease-specific biological network environments. These findings are further validated through in vitro cytotoxicity assays, demonstrating the potential of the model to enhance drug development and identify effective treatment regimens for specific diseases.

Keywords: cancer; drug combination; drug‐disease interaction (DDIs); few‐shot learning; network target.

Figure 1

The workflow diagram depicting the methodology employed in this study. The study employs the InfoGraph method to obtain the chemical embeddings, while the node2vec method is utilized to acquire gene and disease embeddings. These embeddings are input into a fully connected neural network to predict chemical‐gene interactions. Subsequently, genes undergo random walk on the signed PPI network to obtain drug embeddings. Finally, another fully connected neural network is used to predict interactions between the drugs and diseases based on the drug and disease embeddings.

Figure 2

The model's performance on the prediction of single drug‐disease interactions. A) Receiver Operating Characteristic curve (ROC) of the model. B) Precision‐Recall Curve (PRC) of the model.

Figure 3

Comparison of F1 scores A), and AUC B) between the original model and the model without the drug embedding module across fivefold cross‐validation, respectively.

Figure 4

The performance of the model in predicting interactions between drug combinations and diseases. A) For each drug combination, the boxplot shows the comparison between the average outputs of individual drugs and the combined output of the drug combination. B) For each drug combination, the boxplot compares the direct output of the drug combination with the output of the adjusted drug combination, where a random drug is substituted randomly. C–E) The performance of the efficacy of various methods on three distinct metrics: 1) the average outputs of individual drugs in a drug combination, 2) the direct output of the drug combination as a whole, and 3) the output of modified drug combination in which a randomly selected drug is replaced.

Figure 5

A comparison between the zero‐shot and few‐shot models across various metrics within the context of drug combination datasets.

Figure 6

Cell growth inhibition by specific drugs, and the synergistic inhibitory effects observed in drug combinations. A) The cytotoxic effects of cytarabine, oxaliplatin, and their combination were analyzed in HCT‐15 cells. B) Heatmaps displaying synergy scores of the response to a multi‐dose drug combination in HCT‐15 cells. C) Cytotoxic analysis of cytarabine, docetaxel, and their combination was conducted in HGC‐27 cells. D) Heatmaps showing synergy scores of the response to a multi‐dose drug combination in HGC‐27 cells. A synergy score exceeding 10 suggests potential synergistic effects between the two drugs.