A Microbe-Drug Association Prediction Model Based on Graph Attention Network and Rotating Forest

Background In recent years, drug abuse has led to the increase of microbial resistance, therefore exploring possible connections between drugs and microorganisms becomes more and more important. However, since traditional biological experiments are very expensive and time-consuming, then creating effective computational models to predict possible associations between microbes and drugs turns to be a crucial and challenging work. Result In this paper, we proposed a new calculative model GATROF based on graph attention networks and rotating forests to infer potential microbe-drug correlations. In GATROF, we first constructed a heterogeneous microbe-drug network by combining multiple microbe, drug and disease similarity measures. And then, based on different characteristics of microbes and drugs, we further built two original feature matrices of microbe and drug. Subsequently, we inputted the heterogeneous microbe-drug network together with these two original feature matrices into the graph attention network to extract low dimensional feature representations for microbes and drugs separately. Finally, we further inputted these two low dimensional feature representations together with these two original feature matrices of microbes and drugs into a rotating forest classifier to infer latent associations between drugs and microorganisms. Conclusion Experimental results and case studies indicated that GATROF can achieve better performance in microbe-drug association prediction than existing advanced methods, which means that GATROF may make a satisfactory contribution to the field of medicine in the future.