Artificial Intelligence-Based Prediction of Marine Microbial Biogeochemical Activity Using Multi-Omics Data

Abstract

Marine microorganisms play a crucial role in global biogeochemical cycles, and accurately predicting their biogeochemical activities is essential for understanding changes in marine ecosystems. Addressing the challenges of high dimensionality, strong heterogeneity, and complex nonlinear relationships in multi-omics data, this paper proposes a deep learning prediction framework based on multi-omics fusion: AE-GAT-AFNet. This method first utilizes an autoencoder (AE) to perform feature compression and representation learning on high-dimensional multi-omics data, reducing redundant information and enhancing feature representation capabilities. Then, a microbial relationship graph is constructed, and a graph attention network (GAT) is used to model the potential structural relationships among microbial communities. Based on this, an attention fusion mechanism (AF) is introduced to adaptively weight and integrate different omics modalities, and the final prediction task is completed using a multilayer perceptron (MLP). Experimental results show that the proposed method outperforms several baseline models in both prediction accuracy and stability. Further ablation experiments validate the effectiveness of each key module in the model. The results show that this method can effectively integrate multi-omics information and capture the structural characteristics of microbial communities, providing a feasible intelligent analysis framework for predicting marine microbial biogeochemical activities.