Abstract:
The Gold Fyfe dataset is the largest collection of daily chromatograms of one plant (n = 29 979, covering over 27 years). Despite its high chronobiological relevance, it has so far been only partially explored. In this study, we tested a new characterization method of the chromatograms, based on qualitative features perceptible to human observers (i.e. gestalt recognition through kinaesthetic engagement) using deep learning. A trained human evaluator identified three gestalts and labelled chromatograms accordingly. Three deep learning models (one learning from-scratch and two transfer learning approaches) were developed and evaluated across three data partitions for training validity, performance, and reliability. An explainability analysis was performed on one model, and the models’ descriptive ability was compared with traditional image analysis descriptors previously used. A preliminary chronobiological application was tested on the dataset (PSDs and cross-correlation analysis). Evaluator consistency was 100%. The transfer learning models outperformed the learning-from-scratch model across all evaluation phases. The explainability analysis confirmed that the models relied on the same morphological features identified by the evaluator. Few correlations with traditional descriptors were found, suggesting an expansion of the descriptive palette. The preliminary chronobiological analysis revealed a correlation between one class and seasonal weather parameters. The proposed deep learning models represent a first attempt to achieve a qualitative characterization of metabolomic fingerprinting images guided by human perception. The models successfully extended the descriptive palette of the Gold Fyfe dataset with human-relevant descriptors, and their preliminary chronobiological application yielded promising results, warranting further investigation.