Machine learning assisted prediction of pulcherrimin production through pul gene expression in Metschnikowia pulcherrima

Metschnikowia pulcherrima is a widespread non-Saccharomyces yeast whose ability to produce pulcherrimin, an iron-chelating red pigment, contributes to biocontrol activity. In this study, we investigated pulcherrimin production in 49 M. pulcherrima strains isolated from different ecological niches. Pulcherrimin production showed a pronounced strain-dependent variability that enabled classification of strains into low, intermediate, high and very high producers. The expression of key genes involved in pulcherrimin biosynthesis and transport (pul1, pul2, pul3, and pul4) was evaluated by RT-qPCR, highlighting heterogeneous transcriptional profiles that were not strictly associated with strain origin. Notably, pul2 showed the widest range of expression and was consistently elevated in high-producing strains. In order to better capture genotype-phenotype relationships, machine learning-based regression of pulcherrimin production relying on gene expression was implemented. A feature importance analysis demonstrated that the pul2 gene was the key transcriptional predictor to the regression model for pulcherrimin production. A 1D cluster analysis through Kernel Density Estimation showed that pulcherrimin production is highly variable, a strain-specific trait shaped by regulatory diversity. The combined experimental and machine-learning framework proposed could be useful for the rational selection of strains with enhanced pulcherrimin-producing potential for biocontrol applications.