Leveraging Multi-Level Biomarkers Using Machine Learning: Identifying Physiological and Skin Microbial Dynamics in Bd-Resistant Amphibians.

Abstract

Amphibians worldwide are declining due to various anthropogenic and environmental stressors. One of the most important threats is large-scale epidemics of chytridiomycosis, which is caused by Batrachochytrium dendrobatidis (Bd). Unlike in other continents, amphibian species in South Korea, such as Pelophylax nigromaculatus, are resistant to Bd, making it difficult to discern its detailed effects. This study determined the dynamics of Bd infection in P. nigromaculatus by integrating physiological, microbiological, and morphological data and applying state-of-the-art machine learning methodologies. Data are presented on Bd prevalence, body size, weight, and physiological stress responses, including corticosterone (CORT) levels and innate immune functions determined using bacterial killing assays and skin microbiome composition. Significant physiological differences between infected and non-infected animals were observed regarding elevated CORT levels and changes in bacterial killing capacity. Skin microbiome analysis indicated a subtle variation in the microbial composition, but the alpha and beta diversities did not significantly differ between infected and non-infected animals. To balance the intrinsic class imbalance of the dataset, several machine learning methods were coupled with different data-augmentation techniques. Using the Light Gradient Boosting Machine resulted in the best predictive performance when considering conditional generative adversarial networks-augmented datasets. Among the predictors, CORT level and bacterial killing ability were chosen for classifying the infection status. Machine learning can be used to complement the contrasting sensitivities of multi-level biomarkers due to differences in disease resistance or infection loads. This integrated approach may be essential for understanding the impacts of multiple threats to amphibians.