Dual-Channel Heterogeneous Graph Neural Network for Predicting microRNA-Mediated Drug Sensitivity

Abstract

Many studies have confirmed that microRNAs (miRNAs) are mediated in the sensitivity of tumor cells to anticancer drugs. MiRNAs are emerging as a type of promising therapeutic targets to overcome drug resistance. However, there is limited attention paid to the computational prediction of the associations between miRNAs and drug sensitivity. In this work, we proposed a heterogeneous network-based representation learning method to predict miRNA-drug sensitivity associations (DGNNMDA). An miRNA-drug heterogeneous network was constructed by integrating miRNA similarity network, drug similarity network, and experimentally validated miRNA-drug sensitivity associations. Next, we developed a dual-channel heterogeneous graph neural network model to perform feature propagation among the homogeneous and heterogeneous nodes so that our method can learn expressive representations for miRNA and drug nodes. On two benchmark datasets, our method outperformed other seven competitive methods. We also verified the effectiveness of the feature propagations on homogeneous and heterogeneous nodes. Moreover, we have conducted two case studies to verify the reliability of our methods and tried to reveal the regulatory mechanism of miRNAs mediated in drug sensitivity. The source code and datasets are freely available at .

A comprehensive checklist of computational resources and methodologies for noncoding RNAs in systems medicine

“…In terms of model architecture, the multi‐channel model may contribute to obtaining high‐quality embedding representations of multiple types of nodes in heterogeneous graphs. Existing works have attempted to address the heterogeneous graph issue by using multi‐channel models, such as DGNNMDA, which leverages dual channels to represent ncRNAs and drugs, respectively (Deng, Fan, et al, 2022). In addition, the attention mechanism can increase the interpretability of the model, it is thus often used in AI models of ncRNA‐drug association prediction.…”

“…Therefore, it is necessary to randomly sample the negative samples equal to the number of positive samples before training models. The sampling strategy that ranks samples according to the similarity matrix and selects the least similar ones as negative samples is more reliable than the random sampling strategy (Deng, Fan, et al, 2022).…”

“…The dual and multi-channel GNN models take heterogeneity into account when aggregating the neighboring node features. DGNNMDA is a dual-channel GNN model for miRNA-drug response prediction (Deng, Fan, et al, 2022). A miRNA-drug heterogeneous network can be constructed by integrating miRNA and drug similarity networks, as well as experimentally validated miRNA-drug sensitivity associations.…”