Advancing Food Analysis: Amino Acid Quantification by NMR and Deep Learning

Nuclear magnetic resonance (NMR) is a strong tool for the quantification of metabolites inside bio- systems. During last years, a new approach for NMR spectral quantification has been suggested based on convolutional neural network (CNN).[1][2] In this work, we present a Deep Learning model able to quantify mixtures of three amino-acids (Alanine, Aspartic acid and Glycine). A matrix of 64 random mixtures of these amino-acids was created by design of experiment (DoE) to maximize the concentration distribution in a selected interval to be close to bio- matrices.[3] From this 64 sample set, spectra where acquired to be used as test for the model. Two simulated data sets, made of up to 100000 spectra, were constructed, based on the deconvolution results on one selected spectrum, to be used as training and validation of the deep learning model. The model is a collection of 1D convolutional, inception and fully connected layers with 16348 inputs and 4 outputs representing the concentration of the three amino acids and TSP standard. The model shows a good linearity between the expected and predicted values either on simulated and experimental spectra. Actually, the model is capable to predict, starting for the spectrum, the concentration of a compound in a mixture with a low error level. In addition, quantification of many spectra can be achieved simultaneously.