Machine Learning-Aided Optimization of In Vitro Tetraploid Induction in Cannabis

Polyploidy, characterized by an increase in the number of whole sets of chromosomes in an organism, offers a promising avenue for cannabis improvement. Polyploid cannabis plants often exhibit altered morphological, physiological, and biochemical characteristics with a number of potential benefits compared to their diploid counterparts. The optimization of polyploidy induction, such as the level of antimitotic agents and exposure duration, is essential for successful polyploidization to maximize survival and tetraploid rates while minimizing the number of chimeric mixoploids. In this study, three classification-based machine learning algorithms-probabilistic neural network (PNN), support vector classification (SVC), and k-nearest neighbors (KNNs)-were used to model ploidy levels based on oryzalin concentration and exposure time. The results indicated that PNN outperformed both KNNs and SVC. Subsequently, PNN was combined with a genetic algorithm (GA) to optimize oryzalin concentration and exposure time to maximize tetraploid induction rates. The PNN-GA results predicted that the optimal conditions were a concentration of 32.98 µM of oryzalin for 17.92 h. A validation study testing these conditions confirmed the accuracy of the PNN-GA model, resulting in 93.75% tetraploid induction, with the remaining 6.25% identified as mixoploids. Additionally, the evaluation of morphological traits showed that tetraploid plants were more vigorous and had larger leaf sizes compared to diploid or mixoploid plants in vitro.