Development of a machine-learning model to identify the impacts of pesticides characteristics on soil microbial communities from high-throughput sequencing data.

Abstract

High-throughput sequencing (HTS) of soil environmental DNA provides an advanced insight into the effects of pesticides on soil microbial systems. However, the association between the properties of the pesticide and its ecological impact remains methodically challenging. Risks associated with pesticide use can be minimized if pesticides with optimal structural traits were applied. For this purpose, we merged the 20 independent HTS studies, to reveal that pesticides significantly reduced beneficial bacteria associated with soil and plant immunity, enhanced the human pathogen and weaken the soil's ecological stability. Through the machine-learning approach, correlating these impacts with the physicochemical properties of the pesticides yielded a random forest model with good predictive capabilities. The models revealed that physical pesticide properties such as the dissociation constant (pKa), the molecular weight and water solubility, determined the ecological impact of pesticides to a large extent. Moreover, this study identified that eco-friendly pesticides should possess a value of pKa > 5 and a molecular weight in the range of 200-300 g/mol, which were found to be conducive to bacteria related to plant immunity promotion and exerted the lowest fluctuation of human opportunistic pathogen and keystone species. This guides the design of pesticides for which the impacts on soil biota are minimized.