Cancer is a multifaceted disease that results from co-mutations of multi biological molecules. A promising strategy for cancer therapy involves in exploiting the phenomenon of Synthetic Lethality (SL) by targeting the SL partner of cancer gene. Since traditional methods for SL prediction suffer from high-cost, time-consuming and off-targets effects, computational approaches have been efficient complementary to these methods. Most of existing approaches treat SL associations as independent of other biological interaction networks, and fail to consider other information from various biological networks. Despite some approaches have integrated different networks to capture multi-modal features of genes for SL prediction, these methods implicitly assume that all sources and levels of information contribute equally to the SL associations. As such, a comprehensive and flexible framework for learning gene cross-network representations for SL prediction is still lacking. In this work, we present a novel Triple-Attention cross-network Representation learning for SL prediction (TARSL) by capturing molecular features from heterogeneous sources. We employ three-level attention modules to consider the different contribution of multi-level information. In particular, feature-level attention can capture the correlations between molecular feature and network link, node-level attention can differentiate the importance of various neighbors, and network-level attention can concentrate on important network and reduce the effects of irrelated networks. We perform comprehensive experiments on human SL datasets and these results have proven that our model is consistently superior to baseline methods and predicted SL associations could aid in designing anti-cancer drugs.

Computational drug repositioning which accelerates the process of drug development is able to reduce the cost in terms of time and money dramatically which brings promising and broad perspectives for the treatment of complex diseases. Heterogeneous networks fusion has been proposed to improve the performance of drug repositioning. Due to the difference and the specificity including the network structure and the biological function among different biological networks, it poses serious challenge on how to represent drug features and construct drug-disease associations in drug repositioning. Therefore, we proposed a novel drug repositioning method (ATDR) that employed attention transfer across different networks constructed by the deeply represented features integrated from biological networks to implement the disease-drug association prediction. Specifically, we first implemented the drug feature characterization with the graph representation of random surfing for different biological networks, respectively. Then, the drug network of deep feature representation was constructed with the aggregated drug informative features acquired by the multi-modal deep autoencoder on heterogeneous networks. Subsequently, we accomplished the drug-disease association prediction by transferring attention from the drug network to the drug-disease interaction network. We performed comprehensive experiments on different datasets and the results illustrated the outperformance of ATDR compared with other baseline methods and the predicted potential drug-disease interactions could aid in the drug development for disease treatments.

Computational methods to discover potential synthetic lethality (SL) pairs has become a promising strategy for targeted cancer therapy and cancer medicine development. Despite many computational methods by integrating multiple biological networks were proposed to improve the identification performance. It is essential to propose feature representation approach via embedding latent biological variables in various networks into a unified feature space. Therefore, we propose a method to identify synthetic lethality genes by modeling latent space with embedding variables resulting from the potential interpretation of synthetic lethality on integrating heterogeneous networks (LSTF) to obtain gene representation. Meanwhile, manifold subspace regularization is applied to capture the geometrical manifold structure in the latent space with gene PPI functional and GO semantic embeddings. Subsequently, SL gene pairs are identified by the reconstruction of the associations with gene representations in the latent space. The comprehensive experimental results illustrate that LSTF is superior to other state-of-the-art methods. Case study demonstrates the effectiveness of the identified potential SL genes.

Synthetic lethality (SL) is a potential cancer therapeutic strategy and drug discovery. Computational approaches to identify synthetic lethality genes have become an effective complement to wet experiments which are time consuming and costly. Graph convolutional networks (GCN) has been utilized to such prediction task as be good at capturing the neighborhood dependency in a graph. However, it is still a lack of the mechanism of aggregating the complementary neighboring information from various heterogeneous graphs. Here, we propose the Multiple Attention Graph Convolution Networks for predicting synthetic lethality (MAGCN). First, we obtain the functional similarity features and topological structure features of genes from different data sources respectively, such as Gene Ontology data and Protein-Protein Interaction. Then, graph convolutional network is utilized to accumulate the knowledge from neighbor nodes according to synthetic lethal associations. Meanwhile, we propose a multiple graphs attention model and construct a multiple graphs attention network to learn the contribution factors of different graphs to generate embedded representation by aggregating these graphs. Finally, the generated feature matrix is decoded to predict potential synthetic lethal interaction. Experimental results show that MAGCN is superior to other baseline methods. Case study demonstrates the ability of MAGCN to predict human SL gene pairs.