Interoperability of machine learning classifiers across LC-MS platforms for non-targeted authentication of honey...

The botanical origin of honey largely defines its taste, appearance, and nutritive properties. This has led to honey being one of the most frequently frauded foods today, whereby honey from less valuable flowers is packaged as honey originating from a more valuable flower and sold at a higher price. Non-targeted LC-MS analysis is a powerful tool for quantifying the chemical fingerprint of honey. However, this high sensitivity also leads to difficulty in reproducing results, especially across different LC-MS platforms, limiting the application of non-targeted analysis in a regulatory environment.

This work presents an algorithm to transform the LC-MS spectra of one instrument, to make it appear as though it came from another instrument. This allows a machine learning model trained on one set of samples using one LC-MS platform to classify new honey samples on a different LC-MS platform. This cross-instrument classification is demonstrated using a dataset of 262 monofloral honey samples across 4 LC-MS platforms using a logistic regression classifier model that predicts whether a sample is blueberry, buckwheat, clover, or "other" monofloral honey. A Matthews correlation coefficient (MCC) of 0.77-0.84 was obtained for same-instrument classification, whereas a relatively high MCC of 0.71-0.80 was still observed for cross-instrument classification. In 75% of cases, there was no significant difference (p > 0.05) between same-instrument and cross-instrument performance.

This method enables greater collaboration between research groups and helps facilitate the usage of non-targeted analysis in regulatory honey authentication by achieving a reproducible outcome across multiple LC-MS platforms.