Classification tools for carotenoid content estimation in manihot esculenta via metabolomics and machine learning

Abstract

Cassava genotypes (Manihot esculenta Crantz) with high pro-vitamin A activity have been identified as a strategy to reduce the prevalence of deficiency of this vitamin. The color variability of cassava roots, which can vary from white to red, is related to the presence of several carotenoid pigments. The present study has shown how CIELAB color measurement on cassava roots tissue can be used as a non-destructive and very fast technique to quantify the levels of carotenoids in cassava root samples, avoiding the use of more expensive analytical techniques for compound quantification, such as UV-visible spectrophotometry and the HPLC. For this, we used machine learning techniques, associating the colorimetric data (CIELAB) with the data obtained by UV-vis and HPLC, to obtain models of prediction of carotenoids for this type of biomass. Best values of R2 (above 90%) were observed for the predictive variable TCC determined by UV-vis spectrophotometry. When we tested the machine learning models using the CIELAB values as inputs, for the total carotenoids contents quantified by HPLC, the Partial Least Squares (PLS), Support Vector Machines, and Elastic Net models presented the best values of R2 (above 40%) and Root-Mean-Square Error (RMSE). For the carotenoid quantification by UV-vis spectrophotometry, R2 (around 60%) and RMSE values (around 6.5) are more satisfactory. Ridge regression and Elastic Network showed the best results. It can be concluded that the use colorimetric technique (CIELAB) associated with UV-vis/HPLC and statistical techniques of prognostic analysis through machine learning can predict the content of total carotenoids in these samples, with good precision and accuracy.