ABSTRACT
Introduction. We aimed to develop and validate machine learning (ML) -based algorithms to predict COVID-19 diagnosis as well as to identify new biomarkers associated with the disease. Methods. Initially, 96 blood samples of patients diagnosed with COVID-19 (Thaizhou Hospital, China) were analyzed through liquid chromatography coupled to mass spectrometry. Samples of patients presenting other pneumonias or severe acute respiratory syndrome, but with negative RT-PCR for SARS-CoV-2, were used as positive controls. Samples from healthy volunteers were used as negative controls. The final database included around 1000 metabolites. Exploratory analyses for the development of ML-based models using principal component analysis (PCA) were performed. Leverage plot versus studentized residuals method was used to detect outliers. Three supervised ML-based models were developed: discriminant analysis by partial least squares (PLS-DA), artificial neural networks discriminant analysis (ANNDA) and k-nearest neighbors (KNN). Samples for the training (70%) and testing sets (30%) were randomly selected using the Kenrad Stone algorithm. Models' performance was evaluated considering accuracy, sensitivity and specificity. Analyses were conducted in SOLO (Eigenvector-Research). Results. The PCA model was able to distinguish the three classes of patients' samples (positive for COVID-19, negative controls, positive controls) with an overall accumulated variance of 94.27 percent. The PLS-DA model presented the best performance (accuracy, sensitivity, and specificity of 93%, 98% and 88%, respectively). Increased levels of the biomarkers uridine (linked to glucose homeostasis, lipid, and amino acid metabolisms), 4-hydroxyphenylacetoylcarnitine (metabolite from the tyrosine metabolism;probably associated with anorexia) and ribothymidine (resulting from oral and fecal microbiota alterations) were significantly associated with COVID-19. Conclusions. Three different and updated ML-based algorithms were developed to predict COVID-19 diagnosis;PLS-DA led to the most accurate results. High levels of some metabolites were found as potentially predictors of the disease. These biomarkers should be further evaluated as potential therapeutic targets in well-designed clinical trials. These ML-based models can help the early diagnosis of COVID-19 and guide the development of tailored interventions.
ABSTRACT
Objectives: Despite great advancements in COVID-19 immunization, the development of therapeutic interventions is urgent to control the ongoing pandemic, especially infected patients. The spike protein (S1) of SARS-Cov-2 virus plays a major role in attachment to the host and further series of events. We aimed to identify natural bioactive compounds (NBC) that act as potential inhibitors of S1 by means of in silico assays. Methods: NBCs with proved biological in vitro activities were obtained from the ZINC database (https://zinc.docking.org) and analyzed through virtual screening and molecular docking to identify those with higher affinity to the S1. Machine learning models of principal component analysis (PCA), artificial neural networks (ANN), discriminant analysis by partial least squares (PLS-DA) and decision tree (DT) were used to validate the Results: Selected NBCs were submitted to drug-likeness analysis using the Lipinsk and Vebber's five rule. The prediction of pharmacokinetic parameters (i.e. absorption, metabolism, distribution, elimination) and toxicity (e.g. hepatotoxicity, cardiotoxicity, carcinogenicity, immunotoxicity) were performed (ADMET). The influence of the NBC’s stereoisomeric, tautomeric and protonation states at physiological pH on the pharmacodynamics, pharmacokinetics and toxicity analyses were also evaluated. Results: A total of 170,906 compounds were analyzed. Of these, only 36 showed greater affinity with the S1 (affinity energy <0.8 kcal/mol). The PCA and PLS-DA models were able to reproduce the results of the virtual screening and docking analyzes with an accuracy of 97.5%. Of these 36 CNBs, only 12 (33.33%) were drug-likeness. The ADMET analysis showed that the natural compound phaselol (7-[[(1R,4aS,6R,8aR)-6-hydroxy-2,5,5,8a-tetramethyl-1,4,4a,6,7,8-hexahydronaphthalen-1-yl]methoxy]chromen-2-one) was the most promising in inhibiting the SARS-COV-2 spike. Conclusions: Machine learning-based research is efficient for retrieving novel approaches to diseases’ treatment. We identified 12 compounds as possible inhibitors of S1;phaselol was the most promising candidate for treating COVID-19. In vitro, preclinical studies and clinical trials are now needed to confirm these findings.