Reconhecimento de padrões em microbioma e predição de variáveis ambientais contínuas

Abstract

With the advancement of sequencing technologies, the development of tools capable of processing this type of data and extracting actionable knowledge for industry has become increasingly necessary. A promising application is the identification of bacteria correlated with agricultural productivity, with the potential to predict crop yields in future harvests. This type of application has a direct impact on the agricultural industry and significant economic potential. Although there are tools that integrate physical, chemical, and biological soil data to support agricultural management, they are not designed to handle the compositional nature of sequencing data and, therefore, cannot identify the bacteria responsible for producing these elements. Selecting bacteria, rather than just soil elements, is essential, as microbial communities respond dynamically to environmental changes—such as rainfall or drought—offering more sensitive and predictive insights into soil health. To address this gap, we developed CODARFE (COmpositional Data Analysis with Recursive Feature Elimination), a tool capable of associating microbiome data with continuous environmental variables and making predictions on new samples — a capability not previously available in tools from this field. CODARFE was developed in collaboration with the company SUPERBAC, which provided its proprietary microbiome database for model training and validation. The tool was able to predict soybean yield in clay soils with a mean absolute error of 248 kg/ha, equivalent to only 7% deviation from the national average, by integrating selected bacterial abundances with physical and chemical soil variables. Furthermore, during a 10-month research exchange at the European Bioinfor matics Institute (EMBL-EBI), under the Dr. Robert Finn’s supervision, through the CAPES (PDSE) program, the method was refined and tested on a variety of public datasets, leading to its integration into the MGnify platform. As a result, CODARFE was made available in five accessible formats (Python class, Windows executable, Linux command line, interactive notebook, and via MGnify) and published in the journal GigaScience, enhancing its visibility and accessibility to the scientific community. The tool has two software registrations: a public registration, ensuring access for the research community, and a private registration tied to the partnership with SUPERBAC, protecting the commercial usage rights of the model developed with proprietary data.