Predicting functional bioactivities in fermented milk using deep learning on liquid chromatography-mass spectrometry metabolomics

ABSTRACT: Fermented dairy products are increasingly valued not only for their nutritional content but also for their potential health-promoting properties. However, assessing these functional benefits often requires time-consuming chemical assays that limit scalability. In this study, we investigated whether deep learning (DL) could offer a faster, more efficient alternative. Using liquid chromatography (LC)-MS quadrupole time-of-flight metabolomics, we analyzed 18 fermented milk samples (derived from camel and bovine milk fermented with different bacterial strains) and measured their bioactivity across 9 in vitro assays, including antioxidant capacity (2,2′-azino-bis(3-ethylbenzothiazoline-6-sulfonic acid), 2,2-diphenyl-1-picrylhydrazyl), enzyme inhibition (angiotensin-converting enzyme, degree of hydrolysis), and anticancer activity (HT-29, MDAMB). To address the challenge of limited sample size, we implemented a robust preprocessing pipeline including outlier detection, robust scaling, and data augmentation techniques. We trained a one-dimensional convolutional neural network (1D-CNN; DL) architecture with regularization strategies to predict these bioactivity scores from preprocessed LC-MS data. The model achieved strong performance with a mean absolute error of 0.548 ± 0.089 across all outputs through 3-fold cross-validation, demonstrating effective generalization despite the small dataset. Principal component analysis revealed biologically meaningful structure in the metabolomic data, distinguishing samples by milk type and fermentation condition. Together, these results demonstrate that DL with appropriate regularization and data augmentation can accurately predict the functional bioactivity of fermented milk products from metabolomic signatures, offering a promising path toward scalable, DL-assisted screening in functional food development, even with limited training data.