Integrating Non-Targeted Mass Spectrometry and Machine Learning for the Classification of Organic and Conventionally Grown Agricultural Products: A Case Study on Tomatoes

Rising demand for organic agricultural products has made the verification of their authenticity a critical concern. Traditional classification approaches for mass spectrometry using full-scan high-resolution mass spectrometry data often emphasize feature selection, which can lead to information loss and limited model robustness. Here, an artificial intelligence-based classification framework was developed to differentiate between organically and conventionally grown agricultural products, utilizing mass spectrometry data. Tomatoes were used as a representative case. The proposed method, incorporating 1000 mass-to-charge ratio (<i>m</i>/<i>z</i>) binning and an Adaptive Boosting model, achieved high predictive performance, with an Area Under the Receiver Operating Characteristic Curve exceeding 0.990 and an accuracy of 0.997. Beyond model performance, interpretability is enhanced through Shapley Additive Explanations-based importance analysis, followed by structural annotation of key <i>m</i>/<i>z</i> fragments. Notably, Gibberellin A4 and (<i>Z</i>)-resveratrol were identified as potential chemical markers of organically grown tomatoes. This enables the identification of chemically meaningful ions associated with different cultivation practices. The proposed framework provides a scalable and interpretable solution for the authentication of organic agricultural products.