基于深度学习的微生物群落数据分析与预测

Microbial Community Data Analysis and Prediction Using Deep Learning

摘要: 微生物群落与代谢组相互作用共同影响着宿主健康与环境稳定，随着多组学整合研究逐步发展，之前的方法在捕捉微生物与代谢物之间复杂的非线性关系和异质性方面存在局限。本文提出了一种新的深度学习框架，基于双编码器–解码器架构，通过共享潜在空间实现微生物组和代谢组的双向跨模态预测，并结合数据增强策略提升了模型的稳健性，同时在框架中嵌入诊断分类器以预测疾病的诊断结果。试验结果表明我们的架构在多个真实数据集中的预测精度优于现有方法(如SparseNED、MiMeNet等)，模型在三个数据集上预测代谢物的平均斯皮尔曼相关系数均得到有效提升，同时数据扩增策略使得模型诊断预测的AUC值从0.8425提升至0.8895。

Abstract: Microbial communities and metabolomes interactively influence host health and environmental stability. With the progressive development of multi-omics integration studies, conventional methods exhibit limitations in capturing the complex nonlinear relationships and heterogeneity between microorganisms and metabolites. This study proposes a novel deep learning framework based on a dual encoder-decoder architecture, which enables bidirectional cross-modal prediction between microbiome and metabolome data through a shared latent space. The framework incorporates a data augmentation strategy to enhance model robustness and embeds a diagnostic classifier to predict disease outcomes. Experimental results demonstrate that our architecture outperforms existing methods (e.g., SparseNED, MiMeNet) in prediction accuracy across multiple real-world datasets. The average Spearman correlation coefficient for metabolite prediction improved significantly across all three datasets, while the data augmentation strategy increased the AUC value of diagnostic prediction from 0.8425 to 0.8895.