HDGGCN: Heterogeneous Disease-Gene Network Representation Learning using Similarity-based Adjacency Matrix Generation.
TL;DR
This paper proposes a heterogeneous disease-gene network representation using similarity-based adjacency matrix generation (HDGGCN) to realize disease-gene prediction and presents a novel data partitioning mechanism to achieve high-performance prediction.
Abstract
The discovery of disease-related genes is crucial for understanding disease mechanisms, which can effectively improve clinical diagnosis and treatment and ultimately realize precision medicine. However, due to the sparse and complex characteristics of bioinformatics data, it is difficult to fuse multiple-source information and extract the features of high-dimensional sparse data to achieve satisfactory prediction performance. In this paper, we propose a heterogeneous disease-gene network representation using similarity-based adjacency matrix generation (HDGGCN) to realize disease-gene prediction. First, we present a cosine similarity-based adjacency matrix generation strategy to reconstruct the disease-gene-GO heterogeneous network. Then, the reconstructed adjacency matrices and the feature matrices are taken as the inputs of the graph convolutional neural network (GCN), and the low-dimensional node representations will be generated. Finally, a novel data partitioning mechanism is presented to achieve high-performance prediction. The HDGGCN algorithm's effectiveness has been verified by four representative evaluation metrics, including Precision, Recall, F1-score, and Association Precision (AP). Specifically, compared to state-of-the-art baseline models, HDGGCN yields minimum performance improvements of $0.6\%-2.4\%$ in Precision, $1.0\%-1.4\%$ in Recall, and $1.0\%-1.3\%$ in F1-score across TOP-5 to TOP-20 cutoffs, alongside at least a $3.0\%$ increase in AP.