An approach in developing graphical feature maps derived from machine learning and its application in loquat juice classification

This study introduces a new methodology for developing graphical feature maps using weighted artificial neural networks (w-ANNs) and demonstrates its application in the classification of loquat juice varieties (namely loquat_baisha and loquat_hongsha) by convolutional neural network (CNN) implemented with TensorFlow (TF). During the feature map generation phase, headspace gas chromatograph-ionic mass spectroscopy (HS-GC-IMS) analysis identified two distinct groups of key compounds, each comprising seven highly responsive chemicals specific to loquat_baisha and loquat_hongsha. The SHapley Additive exPlanations (SHAP) analysis further revealed the most influential molecular feature descriptors (MFDs), with Kappa2, Gasteiger charge, and LogP emerging as the top three MFDs for loquat_baisha, while Kappa2, Kappa3, and Fraction_SP3 were the most significant for loquat_hongsha. Graphical feature maps were subsequently constructed. Additionally, a comprehensive chemical library for loquat was developed by retrieving structurally similar compounds from PubChem. Performance evaluations indicated that the methodology's effectiveness is context-dependent and may vary across different application scenarios.