MSc Loïc Thurre

05/2024

Supervision: Prof. Klaus Schlaeppi

Abstract:

In a fast-changing climate, the need for more sustainable agricultural practices is unprecedented. Field inoculation with arbuscular mycorrhizal fungi is a promising alternative to mineral fertilizers and pesticides but this method suffers from a high context-dependency. Models using soil physiochemical parameters and fungal communities aimed to solve this problem but never considered the contribution of soil bacteria. Additionally, microbiome sequencing is costly and threatens the democratization of such models with farmers. With this thesis we aimed to investigate the potential of Random Forest in fungal microbiome predictions and assessed the contribution of bacteria to the predictive model of mycorrhizal growth response. 

We used data produced in a large-scale field inoculation experiment to compute a finely tuned Random Forest model and used leave-one-out validation to assess the accuracy of our approach. Additionally, we performed 16S full-length soil profiling and used three different methods to find soil bacteria responsible of context-dependency and produced models made of soil parameters, fungal and bacterial communities. Finally, we conducted ten field inoculation experiments to validate the previous models in an uncovered geographic area. 

Random Forest model gave decent performance in most of the fields and revealed non-linear patterns shaping soil fungal communities. 13 soil bacteria were found to be indicative of high or low mycorrhizal growth response and their addition to the models increased the covered variance up to 95 %. Moreover, cross validation of model that could recommend inoculation to farmer resulted in an accuracy of 96 %. Nevertheless, the model validation on new fields resulted in inaccurate predictions because of model overfitting.

In this work we showed that microbial inoculation would benefit from such models to reduce the costs of sequencing and increase inoculation success.