NEXGB: A Network Embedding Framework for Anticancer Drug Combination Prediction

Abstract

Compared to single-drug therapy, drug combinations have shown great potential in cancer treatment. Most of the current methods employ genomic data and chemical information to construct drug-cancer cell line features, but there is still a need to explore methods to combine topological information in the protein interaction network (PPI). Therefore, we propose a network-embedding-based prediction model, NEXGB, which integrates the corresponding protein modules of drug-cancer cell lines with PPI network information. NEXGB extracts the topological features of each protein node in a PPI network by struc2vec. Then, we combine the topological features with the target protein information of drug-cancer cell lines, to generate drug features and cancer cell line features, and utilize extreme gradient boosting (XGBoost) to predict the synergistic relationship between drug combinations and cancer cell lines. We apply our model on two recently developed datasets, the Oncology-Screen dataset (Oncology-Screen) and the large drug combination dataset (DrugCombDB). The experimental results show that NEXGB outperforms five current methods, and it effectively improves the predictive power in discovering relationships between drug combinations and cancer cell lines. This further demonstrates that the network information is valid for detecting combination therapies for cancer and other complex diseases.

Keywords: cancer; drug combination; drug synergy prediction; network embedding.

Figure 1

This figure shows the ROC curve and AUC value of the 32-dimensional, 64-dimensional, and 128-dimensional features of the node under the five-fold cross-validation. Among them, the blue curve is the average ROC and the corresponding mean, and the mean is around 0.85. Other colors correspond to the ROC curve and AUC value of each fold.

Figure 2

Classification method comparison. The five colors in the figure represent the five respective methods, the horizontal axis represents the performance index, and the vertical axis represents the corresponding value. Of the six indicators, XGBoost shows its superiority.

Figure 3

Cell-line-specific and tissue-specific prediction performances of NEXGB. (a) Cell-line-specific performance. The average ROC-AUC value in the 29 cell lines is around 0.75, with the highest value of 0.835 for A2780 cell line and the lowest value of 0.587 for ES2 cell line. (b) Tissue-specific performance. Among the six types of tissues, the prostate tissue had the lowest performance in general.

Figure 4

Different cell lines are subjected to t-SNE analysis to visualize the results. High-dimensional cell line vector representations are projected into 2D space with the first two t-SNE components. (a) Different colors indicate clusters per cell line assigned by the density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm. (b) Different colors indicate different tissues of each cell line.

Figure 5

The framework of the NEXGB. The PPI network serves as the input of the model, where red represents the drug target protein, blue is the cell line target protein, and purple is the drug–cell line interaction protein. The struc2vec component perform feature extraction for the protein nodes in the input PPI network. We further obtain the features of drugs and cell lines through the drug–protein and cell–protein relationships. Labels 1 and 0 in the drug–drug–cell line combination matrix indicate synergistic and antagonistic effects, respectively. The drug characteristics and cell line characteristics are then concatenated and input into XGBoost. The output of XGBoost is the combined synergistic probability.