In vitro micropropagation of Chlorophytum borivilianum: A Predictive Model Employing Artificial Neural Networks trained with a range of Algorithms

Abstract

The formulation of plant tissue culture media continues to be a complex undertaking, primarily due to the intricate interplay of multiple components. Numerous factors (such as genotype, disinfectants, media pH, temperature, light, and immersion time) interact to affect the process of plant tissue culture. The artificial neural network is considered one of the most potent computational techniques that has emerged as a highly potent and valuable methodology for effectively representing intricate non-linear systems. This research paper focuses on the development of a predictive model for determining the number of shoots in response to different macronutrient compositions in the culture medium used for in-vitro micropropagation of Chlorophytum borivilianum. The study employs artificial neural networks (ANNs) trained with different algorithms to accurately predict the number of shoots and shoot length of the plant species. These algorithms include the Levenberg-Marquardt (LM), Scaled Conjugate Gradient (SCG), and Bayesian Regularisation (BR) backpropagation algorithms. A feed-forward backpropagation network was constructed with a single hidden layer consisting of ten nodes and two output units in the output layer. The input vector contained five elements. The transfer functions 'tansig' and 'purelin' were utilized for the hidden and output layers, respectively. In this study, the effectiveness of neural networks was tested by contrasting the outcomes with real-life data gathered from in-depth tissue culture experiments, which was named the target set. The comparative analysis of "Mean Square Error" and Pearson's correlation coefficient (R) were used to evaluate the effectiveness of networks for improved training initialization. The prediction ability of Levenberg-Marquardt was found superior to other training algorithms with an R-value of 9.92 also the output range of network ‘trainlm’ was closest to the empirical target range during the comparison of experimental target data ranges from wet lab practice.