Machine Learning-Based Identification of Key Bacterial Biomarkers during Anaerobic Succession in Rice Paddy Soil

Characterizing microbial community succession in anaerobic environments is essential for elucidating ecosystem dynamics and establishing predictive ecological models. However, identifying specific bacterial biomarkers that drive these temporal transitions remains a significant challenge, as traditional univariate statistical methods often struggle with the high dimensionality and small sample sizes typical of microbiome datasets. In this study, a Random Forest (RF) classifier was applied to high-throughput sequencing data from anaerobically incubated paddy soil across three time points (days 1, 15, and 56) to identify core predictive operational taxonomic units (OTUs) that govern microbial succession. From a total of 13,497 OTUs, the RF model successfully identified 30 key OTUs whose combined feature importances enabled temporal classification. Leave-one-out cross-validation (LOOCV) demonstrated a classification accuracy of 66.67% using the top 30 OTUs, compared to 0% when using all OTUs. The high scored taxa such as Deltaproteobacteria and Rhodospirillales provided consistent results with transition to anaerobic conditions. This study suggests that machine learning-based feature selection provides a robust complement to typical statistical methods for biomarker discovery in time-series microbiome research.