Scalable and Accurate Drug–target Prediction Based on Heterogeneous Bio-linked Network Mining

Abstract

Motivation:Despite the existing classification-and inference-based machine learning methods that show promising results in drug-target prediction, these methods possess inevitable limitations, where: 1) results are often biased as it lacks negative samples in the classification-based methods, and 2) novel drug-target associations with new (or isolated) drugs/targets cannot be explored by inference-based methods. As big data continues to boom, there is a need to study a scalable, robust, and accurate solution that can process large heterogeneous datasets and yield valuable predictions. Results: We introduce a drug-target prediction method that improved our previously proposed method from the three aspects: 1) we constructed a heterogeneous network which incorporates 12 repositories and includes 7 types of biomedical entities (#20,119 entities , # 194,296 associations), 2) we enhanced the feature learning method with Node2Vec, a scalable state-of-art feature learning method, 3) we integrate the originally proposed inference-based model with a classification model, which is further fine-tuned by a negative sample selection algorithm. The proposed method shows a better result for drug-target association prediction: 95.3% AUC ROC score compared to the existing methods in the 10-fold cross-validation tests. We studied the biased learning/testing in the network-based pairwise prediction, and conclude a best training strategy. Finally, we conducted a disease specific prediction task based on 20 diseases. New drug-target associations were successfully predicted with AUC ROC in average, 97.2% (validated based on the DrugBank 5.1.0). The experiments showed the reliability of the proposed method in predicting novel drug-target associations for the disease treatment.

Predicting compound-protein interaction using hierarchical graph convolutional networks

“…However, structural information regarding the molecular interaction between the protein and compound is not always available, limiting the scope of these methods. Recently, Zhao et al [ 6 ] and Zong et al [ 29 ] built association networks among drugs and targets, allowing them to learn features for drugs, targets or drug-target pairs using graph embedding learning algorithms such as graph convolutional networks (GCN) [ 30 ] or node2vec [ 31 ]. The drawback of these network-based methods is that they require retraining when new nodes are inserted and cannot predict associations for any unseen nodes.…”

GADTI: Graph Autoencoder Approach for DTI Prediction From Heterogeneous Network

“…Meanwhile, DTINet (Luo et al, 2017) first uses the random walk to obtain the low-dimensional feature vector of each drug and protein, projects the drug vector and protein vector into the same space, and then discovers new interactions through matrix completion. Encouraged by the DeepWalk (Perozzi et al, 2014) model, some researchers have combined the random walk with shallow neural networks (Zong et al, 2017(Zong et al, , 2019Zhu et al, 2018). These methods first construct a heterogeneous network based on multiple data sources, and then apply DeepWalk, node2vec (Grover and Leskovec, 2016), and other algorithms to the network to obtain the embedding vectors of drug nodes and target nodes.…”