DeepBeta: A Deep Learning Framework for Nonlinear Feature Fusion in Microbial Beta Diversity Analysis

Abstract Understanding beta diversity is fundamental to assessing how ecological communities respond to environmental change. However, traditional linear methods often fail to capture the high-dimensional, non-linear complexities inherent in multi-source microbial data. Here, we introduce DeepBeta, an algorithmic framework for non-linear information fusionleveraging deep undercomplete autoencoders. By integrating and compressing ecological dissimilarity matrices through a constrained neural bottleneck, DeepBeta filters stochastic noise and fuses latent features that conventional methods overlook. We evaluate the fusion performance across bacterial, eukaryotic, and integrated datasets. Results demonstrate that DeepBeta consistently extracts a higher density of information, identifying stronger community separation across temporal and depth gradients compared to standard approaches. By providing superior algorithmic resolution for detecting subtle shifts, DeepBeta offers a robust, integrative fusion solution for analyzing complex high-dimensional systems. This framework establishes a new benchmark for representation learning in microbial biogeography and the analysis of community assembly.