Machine-learning-enhanced Optical Photothermal infrared (O-PTIR) spectroscopy for rapid simultaneous identification of multiple metabolites to monitor biotransformations: proof of concept on Brewer's spent grain (BSG).

چکیده اصلی

BACKGROUND: Monitoring fermentation in chemically complex matrices remains analytically challenging because multiple metabolites evolve simultaneously. Conventional assays are targeted and sequential, and only partial annotation is often available in real samples. In this study, we present a masked multi-label analytical workflow combining optical photothermal infrared spectroscopy (O-PTIR) with a one-dimensional convolutional neural network (1D-CNN) for rapid screening of brewer's spent grain fermentations. The methodology couples a library of standard compound spectra with transfer learning on partially labelled fermentation spectra, while masked optimisation avoids penalising unannotated classes during fine-tuning. RESULTS: A library of ten reference compounds and their mixtures was first used to train the model, which was then adapted to fermented samples using qualitative labels for starch, glucose, lactic acid and kojic acid. Supernatants from non-inoculated, separated and sequential bacterial/fungal fermentations were analysed after minimal sample preparation. The pretrained model showed excellent internal validation on the standard library. The fine-tuned 1D-CNN retained high discriminative performance (macro-F1 = 0,950). Benchmarking against one-vs-rest PLS-DA and SVM models using the same held-out BSG validation split showed that the nonlinear models outperformed the linear chemometric baseline. Predicted temporal profiles were consistent with assay-derived phase transitions, notably glucose depletion and lactic acid accumulation during the bacterial stage, followed by starch disappearance and kojic acid appearance during the fungal stage. SIGNIFICANCE: These results support the use of multi-label O-PTIR workflow to provide rapid and label-free identification of key metabolic phases in partially labelled, heterogeneous matrices, and highlight its potential as a transferable analytical strategy for bioprocess monitoring.

متن کامل اصلی