In this study, response surface methodology (RSM) was used to optimize in vitro regeneration of the Brazilian micro sword (Lilaeopsis brasiliensis) aquatic plant, followed by data prediction and validation using machine learning algorithms. The basal salt, sucrose and Benzyaminopurine (BAP) concentrations were derived from Box-Behnken design of RSM. The response surface regression analysis revealed that 1.0 g/L MS + 0.1 mg/L BAP + 25 g/L sucrose was optimized for maximum regeneration (100%), shoot counts (63.2), and fresh weight (1.382 g). The RSM-based predicted scores were fairly similar to the actual scores, which were 100% regeneration, 63.39 shoot counts, and 1.44 g fresh weight. Pareto charts analysis illustrated the significance of MS for regeneration and fresh weight but remained insignificant. Conversely, MS × BAP was found to be the most crucial factor for the shoot counts, with MS coming in second and having a major influence. The analysis of the normal plot ascertained the negative impact of elevated MS concentration on shoot counts and enhanced shoot counts from the combination of MS × BAP. Results were further optimized by constructing contour and surface plots. The response optimizer tool demonstrated that maximum shoot counts of 63.26 and 1.454 g fresh weight can be taken from the combination of 1.0 g/L MS + 0.114 mg/L BAP + 23.94 g/L. Using three distinct performance criterias, the results of machine learning models showed that the multilayer perceptron (MLP) model performed better than the random forest (RF) model. Our findings suggest that the results may be utilized to optimize various input variables using RSM and verified via ML models.
Key message
Optimization of in vitro whole plant regeneration of Brazilian sword wood using response surface methodology
Data analysis through ANOVA, response surface regression anlaysis and machine learning
Graphical presentation of data via Pareto charts, normal plots, contour plots and surface plots for optimization
Better performance of ANN-based MLP model compared to decision tree based RF model
Graphical abstract