Targeted metabolomics by CZE-UV associated with deep learning applied to the investigation of organic acid profile in human urine of COVID-19 patients.

Deep learning is a modeling method capable of making trained decisions. Over the last years, the use of Artificial Neural Networks (ANN) combined with NMR, MS, and IR acquired data have reemerged into the metabolomics field. ANNs are inspired in the logic behind the human brain, working as a complex data network that results in a model that can process information and make decision through them. In this work, we proposed the use of capillary electrophoresis associated to UV detection for analyzing urine samples divided into a test group, patients with COVID 19 confirmed by the RT-PCR, against a control group, negative for COVID 19. The protocol was based in the adaptation of a previously optimized method that indirectly detects dicarboxylic organic acids. Those analytes are an important molecular group in regard of common metabolites found in urine with relevance for diagnostics/prognosis purposes. Sample collection was authorized by the ethics committee and were submitted to a triage process to inactive any virus particles and followed by deproteinization. This process was performed by preparing each sample in authentic triplicates. After the analysis, the total dataset consisted of 300 raw electropherograms with 1.22% RSD split into 80% training and 20% test dataset. The pre-processed training data set consisted of 84 with positive results for COVID-19 and 153 negatives. A PCA was primarily used to select the number of variables followed by ANN modeling that resulted in a 3 layer-model with 20, 13, and 5 neurons validated with 77.9 % accuracy, 79.1 % of sensitivity, and 75.0 % specificity. The model was applied in the test dataset and predicted correctly 34 out of the 38 negative-assigned samples, and 12 out of the 21 positively ones. Overall, the aim of this study is to show the applicability of CZE-UV as a targeted metabolomics tool and as an automated high-throughput instrumentation to get reliable spectra-like input data for classificatory-driven deep learning approaches.