Inferring Disease-Associated Piwi-Interacting RNAs via Graph Attention Networks

Abstract

Motivation: PIWI proteins and Piwi-Interacting RNAs (piRNAs) are commonly detected in human cancers, especially in germline and somatic tissues, and correlates with poorer clinical outcomes, suggesting that they play a functional role in cancer. As the problem of combinatorial explosions between ncRNA and disease exposes out gradually, new bioinformatics methods for large-scale identification and prioritization of potential associations are therefore of interest. However, in the real world, the network of interactions between molecules is enormously intricate and noisy, which poses a problem for efficient graph mining. This study aims to make preliminary attempts on bionetwork based graph mining. Results: In this study, we present a method based on graph attention network to identify potential and biologically significant piRNA-disease associations (PDAs), called GAPDA. The attention mechanism can calculate a hidden representation of an association in the network based on neighbor nodes and assign weights to the input to make decisions. In particular, we introduced the attention-based Graph Neural Networks to the field of bio-association prediction for the first time, and proposed an abstract network topology suitable for small samples. Specifically, we combined piRNA sequence information and disease semantic similarity with piRNA-disease association network to construct a new attribute network. In the experiment, GAPDA performed excellently in five-fold cross-validation with the AUC of 0.9038. Not only that, but it still has superior performance compared to methods based on collaborative filtering and attribute features. The experimental results show that GAPDA ensures the prospect of the graph neural network on such problems and can be an excellent supplement for future biomedical research.

DLDTI: a learning-based framework for drug-target interaction identification using neural networks and network representation

“…Graph embedding-based feature for drugs and targets Graph data is rich in behavioral information about nodes, and behavioral information can be used as a descriptor to describe drugs and targets that can be more comprehensive description of the characteristics [25]. So how do we map a high-dimensional dense matrix like graph data to a low-density vector?…”

Identification of piRNA disease associations using deep learning

“…In this regard, several piRNA databases [9] , [26] , [27] , [28] , together with efficient and cost-effective web-server based computational predictors for identifying piRNA and their functions, are available [29] , [30] , [31] ; whereas, research regarding human disease-associated piRNAs is in its early stages. Recently, the development of piRDisease v1.0 [32] , which is a collection of various experimentally verified piRNA-disease associations, allows researchers to develop robust and cost-efficient computational methods in order to identify piRNA-associated diseases [33] , [34] , [35] , [36] , [37] . Thus, Wei et.…”

PDA-PRGCN: identification of Piwi-interacting RNA-disease associations through subgraph projection and residual scaling-based feature augmentation

“…Third, a dual-loss mechanism was introduced, which can optimize the discrimination of binary samples especially for data containing unbalanced positive and negative samples. Most methods of heterogeneous graph/network-based EAP via GCN modeling, by their very nature, mainly focus on how to construct the subgraph and initially represent the node [12,19,20]. It should be noted that similarity can be implemented in two alternative ways: construct subgraph as similarity graph; characterize the node feature as similarity profile.…”

“…Afterward, iPiDi-PUL extracted key features and conducted dimension reduction by principal component analysis over feature vector based on positive unlabeled learning [11]. GAPDA treated each known piRNA-disease association pair as a node in their reconstructed graph and employed graph attention network to make representation learning [12]. SPRDA applied piRNA/ disease similarity network to form a duplex network, then predicted PDAs as a matrix completion problem by structural perturbation algorithm [13].…”