Based on metabolomics and fourier transforms near infrared spectroscopy characterization of Lanxangia tsaoko chemical profile differences among fruit types and development of rapid identification and nutrient prediction models

Abstract

The complex and diverse environments of Lanxangia tsaoko (LT) have given rise to a wide range of fruit types, however, there are some differences in chemical information between the different fruit types. The phenotypic data in this study showed that dry weight and long axis length are somewhat positively correlated with soluble sugar. Further UPLC-MS/MS-based broadly targeted metabolomics results showed that phenolic acids, flavonoids, amino acids and derivatives, terpenoids, and alkaloids differed most among the seven fruit types. Using fourier transform near-infrared spectroscopy (FT-NIRS) and two-dimensional correlation spectroscopy (2DCOS) to characterize the chemical composition of different fruit types of LT and the 5400-4000 cm⁻¹ region was identified as the characteristic band for the different fruit types. Subsequently, three identification models, support vector machine (SVM), partial least squares discriminant analysis (PLS-DA), residual convolutional neural network (ResNet), were established in the fruit type recognition all show excellent performance. In particular, ResNet. The accuracies of synchronized 2DCOS images in full and single bands (10000-7600 cm⁻¹, 7600-5400 cm⁻¹ and 5400-4000 cm⁻¹) are 100% for both training and test sets. Compared to traditional machine learning models, ResNet does not require complex preprocessing and is a potentially fast way to identify different fruit types in LT. Finally, a partial least squares regression (PLSR) model was developed to predict soluble sugars and soluble proteins in LT with optimal RPD values of 1.5393 and 1.4649, respectively.