Graph Neural Network-Based Method For Predicting The Function Of Splicing Isoforms

Alternative splicing plays an important role in gene regulation,which greatly increases the complexity of protein types and functions.At present,gene functions have been extensively studied,and there are multiple public databases with gene function annotations.However,due to the high experimental cost and the large number of splice isoforms,extensive experimental analysis of the function of splice isoforms is currently impossible.Little is known about function at the splice isoform level.Prediction by algorithms provides us with an alternative way to obtain functional annotations of splice isoforms.This paper proposes different prediction methods to obtain functional annotations at the splice isoform level.The main contributions of this paper are as follows.First,this paper proposes Graph Iso Fun,a computational method for predicting isoform function by combining the expression information of genes and splice isoforms with a graph neural network.The method first constructs a heterogeneous co-expression network with genes and splice isoforms as nodes.The network integrates the associations between splice isoforms and genes,which can help predict the function of splice isoforms.Then,Graph Iso Fun exploits the graph neural network to mine the information of the co-expression network to obtain the predicted results of the function of each splicing isoform.Experiments on three datasets show that Graph Iso Fun outperforms some existing methods on different evaluation metrics.Second,this paper proposes Cross Iso Fun,a data fusion method based on splice isoform expression and domain data.The method firstly constructs a splice isoform-gene interaction network based on expression features and domain features,respectively.Cross Iso Fun then employs multiple graph convolutional layers to learn initial prediction scores for specific omics data.Subsequently,a multi-layer fully-connected network is used to identify a two-dimensional matrix consisting of initial scores of expression data and domain data,and trained to obtain functional prediction scores for unknown splicing isoforms.Experiments on three datasets show that the prediction performance of Cross Iso Fun is significantly improved compared to the comparison methods.