Deep learning meta-analysis for predicting plant soil-borne fungal disease occurrence from soil microbiome data

Accurately predicting soil-borne fungal diseases linked to plant diseases through the analysis of soil microbial communities is advantageous for early disease detection and monitoring. In this study, a meta-analysis was conducted to establish a classification model for two soil-borne plant fungal diseases, Fusarium and Verticillium wilt disease, based on soil microbiome datasets. The study integrated a scalable denoising method and an imbalanced data processing strategy for processing imbalanced data. The findings reveal a substantial enhancement in model performance when employing denoised and balanced datasets as opposed to the original dataset. Overall, the model based on bacterial ASV features outperformed the model based on fungal ASV features, achieving an accuracy of over 90 % in predicting Fusarium and Verticillium wilt disease on the independent test set. Some bacteria, such as those classified as the Chitinophagaceae, Nocardioides, and Sphingomonas, have been identified as biomarkers for distinguishing between healthy and diseased soils. Despite this achievement, the models exhibited suboptimal classification precision, underscoring the necessity for additional training sets or more comprehensive environmental information to augment disease prediction capabilities. Our analysis highlights the importance of microbiome-based deep learning (DL) models to make plant disease predictions based on microbiome characteristics.