Handheld Raman Spectroscopy and Machine Learning Monitoring of Fungal-Plant Symbioses

Raman spectroscopy with machine learning models can non-destructively predict plant-fungal symbiosis in soil.

Abstract Text:

Plant growth depends heavily on the composition of surrounding soil, specifically nutrient, microbiota, and water concentrations. The presence of arbuscular mycorrhizal fungi (AMF) in soil allows for a symbiotic relationship to be formed with most terrestrial plants, improving nutrient uptake, stress resistance, photosynthetic efficiently, and ultimate growth. Current methods for analysis require complete plant harvest with lengthy root cleaning and staining, destroying both plant and fungal complexes, preventing both in-field analysis and longitudinal colonization studies. Monitoring these growth differences via a non-destructive method is possible via the Raman-active species found in plant tissues, mainly carotenoids and chlorophylls. Across several growing seasons, tomato plants were grown for five weeks in various conditions prior to imaging the leaves with a handheld Raman spectrometer. These conditions included sterile or AMF inoculated, water control or drought, and soap control or soapy acid (Jasmonic or salicylic). Sample spectra were preprocessed by baselining and normalizing to the cellulose signal (1440 cm^-1) to help account for the seasonal variability. Analysis included splitting the data set in half for machine learning calibration and validation to determine if the changes in Raman signal can predict sample conditions. A univariate analysis of the main pigment peak (1525 cm^-1) was insufficient for comparison across seasons, so a supervised multivariate technique was employed. Projection to Latent Structures - Discriminant Analysis (PLS-DA) decreased the dimensionality of the data set, exposed latent variables, and maximized class separations to accurately predict class labels. Despite supplemental greenhouse lighting, sunlight hours were identified as a confounding factor in analysis from one growing season to the next. While PLS-DA modelling was unable to classify the application of soapy acidic stressors or drought conditions, it was able to model the presence of AMF in plant root systems without destructive harvesting with 70% accuracy.