Generative adversarial network integrated with metabolomics identifies potential biomarkers related to quality changes of atemoya (Annona cherimola × Annona squamosa) stored at 10 and 25 °C

Atemoya fruit deteriorates rapidly during post-harvest storage. A complete understanding of the metabolic mechanisms underlying this process is crucial for developing effective preservation strategies. Metabolomic approaches combined with machine learning offer new opportunities to identify quality-related biomarkers. This study compared atemoya quality stored at 25 °C and 10 °C using untargeted metabolomics integrated with generative adversarial network (GAN) and random forest (RF) analysis. It was found that GAN successfully amplified the metabolomic dataset 10-fold, enabling robust RF-based identification of 20 quality change-related biomarkers. These biomarkers were primarily involved in energy metabolism, reactive oxygen species regulation and primary metabolic pathways including amino acids, lipids and carbohydrates. Low-temperature storage inhibited respiration, preserved cell structure and altered specific glycerophospholipid metabolic pathways. These findings provide molecular insights into low temperature preservation mechanisms and establish a novel framework for metabolomic data analysis in postharvest research. •GAN amplified the metabolomic data 10-fold, enabling robust biomarker identification•Random forest screening identified 20 quality-related metabolic biomarkers•Identified biomarkers participated in amino acid, lipid and carbohydrate metabolism.