Deep learning analysis of amino acids Raman spectra for the determination of their structural features
Lyudmila A. Bratchenko,1 Elena N. Tupikova,2 Vitaliya A. Belova,2 Sahar Zead,1 Valery P. Zakharov,1 Ivan A. Bratchenko,1
1 Department of Laser and Biotechnical Systems, Samara University, Samara, Russia
2 Department of Chemistry, Samara University, Samara, Russia
Abstract
Amino acids play an important physiological role in the formation of proteins and various tissues of the body. The analysis of the structural features of amino acids is of certain interest in the goals of metabolomics. In our work, we investigated the possibility of using a combination of Raman techniques and deep learning to analyze the structural features of amino acids. The experimental setup includes a spectrometric system (EnSpectr R785, Spektr-M, Chernogolovka, Russia) and a microscope (ADF U300, ADF, China). Focusing the exciting radiation and collecting the scattered radiation were implemented using 50x Objective LMPlan. The stimulation of collected spectra was performed by the laser module with central wavelength 785 nm. Raman characteristics of 25 amino acids were obtained. This study includes the analysis of the following amino acids: glycine (Gly), alanine (Ala), valine (Val), leucine (Leu), tyrosine (Tyr), β-phenyl-α-alanine (Phe), isoleucine (Ile), tryptophan (Trp), proline (Pro), histidine (His) hydrochloride, threonine (Thr), serine (Ser), methionine (Met), cysteine (Cys), glutamic acid (Glu), aspartic acid (Asp), glutamine (Gln), asparagine (Asn), arginine (Arg) hydrochloride, lysine (Lys) hydrochloride, ornithine hydrochloride, norvaline, norleucine, cysteine hydrochloride, β-phenyl-β-alanine (from Reakhim Ltd, Russia, reagent grade > 99%). Data analysis is performed on the basis of a pretrained one-dimensional convolutional neural network and a basic solution without the involvement of deep learning based on the method of projections onto latent structures. It was revealed that the regression model based on a one-dimensional convolutional network is more accurate than a model based on projection on latent structures. Convolutional network model is stable, according to the results of cross-validation, and has the potential to reveal nonlinear and complex dependences of the spectral characteristics of amino acids on their structural features. The combination of spectral analysis based on Raman with deep learning allowed the determination of structural features with an R2 value for chain length > 0.8, for a degree of unsaturation of a chain > 0.8, for determining the number of CH2 group > 0.9, for determining the number of CH3 group > 0.9, for determining the amount of NH bond > 0.8. The investigated approach can be extended to the analysis of the structural features of other organic compounds in metabolic profiling problems.
File with abstract
Speaker
Lyudmila A. Bratchenko
Samara University
Russia
Discussion
Ask question