COGCN: A multi-omics phenotype prediction method for maize based on graph convolutional networks and cross-omics tensors

Genomic prediction is an effective way to accelerate the improvement of traits in crops such as maize, but it relies solely on genomic information, which cannot cover the regulatory effects of important downstream mechanisms such as transcription and translation on traits. Therefore, multi-omics prediction models have been proposed as superior alternatives to genomic prediction models. However, most of the existing multi-omics-based plant phenotype prediction methods only combine different omics data in tandem or consider them as multiple predictive factors to perform weighted linear combinations. This lacks an in-depth exploration of the complex interactions among different omics data and may introduce the risk of bias towards specific omics data. To address this problem, we propose a multi-omics phenotype prediction method based on graph convolutional neural networks and cross-omics tensor. The method combines omics-specific learning and cross-omics interactive learning to achieve better maize phenotype prediction performance. First, a graph convolutional neural network is utilized to learn the specificity of each omics data set after principal component analysis, obtaining the higher-order features of different omics data in a high-dimensional nonlinear space. Then, the cross-omics tensor is constructed using the higher-order features of each omics data to explicitly model the outer product interactions between different omics data. Finally, the cross-omics tensor is transported to the cross-omics higher-order prediction network to learn the interaction information of each omics and obtain the final maize phenotype prediction results. The experimental results of predicting different traits using multi-omics data on maize dataset show that COGCN has better prediction performance than other multi-omics phenotype prediction methods.