A method that enables diagnosis Covid-19 In about 20 minutes – at a low cost and without the need for imported reagents – Brazilian researchers described it in an article published on the medRxiv platform, still without a peer review.
The system uses artificial intelligence algorithms that are able to recognize a pattern of molecules characteristic of the disease in blood plasma samples from patients. According to the authors, it is also possible to identify, among confirmed cases, people at higher risk of developing serious manifestations such as respiratory failure.
The project is supported by Fapesp (São Paulo Research Foundation) and, in addition to employees in the Amazon, also includes researchers from the State University of Campinas (Unicamp) and the University of São Paulo (USP).
“In the tests to validate the methodology, we were able to differentiate between positive and negative samples with an accuracy of more than 90%. We also differentiated between severe and light cases with an accuracy of around 82%. Now we start the certification process with Anvisa [Agência Nacional de Vigilância Sanitária]Unicamp professor Rodrigo Ramos Catharino, research coordinator, told Agência Fapesp.
According to him, the investigation could cost on the farm approximately R $ 40 per sample, about half the cost of RT-PCR, a method that is considered the gold standard for diagnosing Covid-1
The work was developed at the Innovare Laboratory of Biomarkers during the PhD by Jeany Delafiori and integrates a research line that combines metabolomics techniques and machine learning to search for markers used to diagnose diseases such as zika, dengue hemorrhagic fever and fibrosis Cystic diseases, diabetes and other metabolic disorders can help.
The group works with the Complex Data Inference Laboratory (Recod) of the Unicamp Institute of Computing (IC), which is coordinated by Professor Anderson Rocha and in which his employee Luiz Claudio Navarro is involved.
“The project involved 728 patients, of whom 369 were diagnosed with clinically confirmed Covid-19 and using RT-PCR. For comparison, samples from non-infected people were used as a kind of control group. A second blood sample was taken from some patients who had complications and had to be hospitalized. In general, there were people with mild and severe symptoms among the confirmed cases, ”says Delafiori.
All samples were analyzed using equipment known as a mass spectrometer that is able to distinguish the substances present in body fluids. As the researchers explain, this set of molecules in the blood plasma shows the various metabolic processes that are active in the body.
“We focus on small molecules such as amino acids, small peptides and lipids. They occur in the last part of the metabolic processes and are therefore more directly linked to the symptoms that the patient showed at the time of the withdrawal, ”explains Delafiori.
Some of the samples were then used by the IC-Unicamp team to teach an artificial intelligence method to identify patterns of metabolites found in positive and negative cases, and to distinguish patterns from mild to severe cases. The other part was used in a blind test, the aim of which was to assess the final accuracy of the analysis performed by the system.
According to the article, the method in the blind test achieved a specificity of 97.6% and a sensitivity of 83.8% for the diagnosis of the disease. With regard to the risk analysis for severe manifestations, the specificity was 76.2% and the sensitivity 87.2%.
“Sensitivity [também conhecido como sensitividade] is the parameter that indicates how sensitive the method is to determine the presence or absence of Covid-19. Specificity, on the other hand, has to do with the ability to distinguish Covid-19 from other health conditions. Together, these two parameters determine the hit rate, ”explains Delafiori. “We’re still working to improve the test’s success rate as our staff take new patient samples. “”
According to Rocha, the algorithm developed can incorporate knowledge when analyzing new samples, which tends to result in an improvement in performance over time. “If it has a hit rate of around 90% today, it will likely hit even more if it reaches thousands of analyzed patients,” says the researcher.
The IC-Unicamp team also created software to automate the entire analysis process and ended up producing a report that informs the doctor whether the patient has Covid-19 and the risk of complications.
“These biomarkers, which predict the progression of the disease, can help, for example, the family doctor to decide whether the patient who is tested positive can be kept isolated at home or brought to a more complex center,” comments Rinaldo Focaccia Siciliano, assistant doctor in the department for infectious and parasitic diseases of the Hospital das Clínicas (HC-FMUSP) and the Department of Infection Control of the Instituto do Coração (InCor), one of the co-authors of the article.
In Siciliano’s assessment, the method performed well to detect both mild cases in the first few days of symptoms and the most advanced of patients who were short of breath when admitted to the hospital. “The advantage of having several centers with different profiles in the project is the variability of the samples. This enables the methodology to be used in different scenarios, both outpatient and in the hospital, ”he says.
Another advance that the researcher has pointed out is the ability to diagnose the disease early on with a blood sample that is easier to take than the nasal secretion used in the RT-PCR test. “The collection with swab [cotonete comprido inserido no fundo do nariz] requires well-trained personnel and adequate space, as there is a risk of the spread of aerosols contaminated with the virus. And the currently available blood test can detect antibodies only a few days after the symptoms appear. “”
While most laboratory tests analyze the concentration of a few substances in the blood, the computer system developed by the Unicamp team can examine thousands of variables simultaneously and extract direct and crossed connections between them: for example which substances are elevated and decreased in people with a certain disease.
“To make this possible, we have worked over the past three years to develop an explainable mathematical model that allows us not only to make a correct prediction but also to know which variables the system is looking for for this prediction. This makes it possible to select the most important ones after the identification of a first set of biomarkers and to optimize the analysis process. In addition, the generated data from the metabolomics area can be used to determine the mechanism of the disease, ”explains Navarro.
In the case of Covid-19, the group reached a set of approximately 30 metabolites that act as a signature of the disease. For example, according to Delafiori, the positive diagnosis was associated with a decrease in the level of lysophosphatidylcholine – glycerol-derived phospholipids that contain phosphate in their structure.
“These molecules are precursors to lung surfactants [compostos que reduzem a tensão superficial dentro do alvéolo pulmonar, prevenindo o colapso durante a expiração] and protect the organ from opportunistic infections. The decrease in this species has been reported in patients with severe acute respiratory syndrome, ”he says.
In positive cases, a decrease in cholesterol derivatives was also observed, which was even more pronounced in patients with a severe form. “Some studies report a reduction in cholesterol levels when the patient with Covid-19 comes to a negative result,” says the researcher.
Glycerolpid levels – previously reported as unregulated in severe acute respiratory syndrome – were increased in samples from patients with this disease.
“After this phase of biochemical validation of the biomarkers, which, for example, made it possible to discard molecules associated with the use of an anti-inflammatory drug that is unrelated to the disease, we combined the remaining variables in pairs. This new technique, which we are introducing into the model, increases the accuracy of the analysis and enables it to be carried out with various mass spectrometry devices, ”says Navarro.
Catharino believes the methodology could be used in any public or private laboratory equipped with a mass spectrometer. By filing the registration with Anvisa, the researchers intend to further increase the variety of samples analyzed as part of the research to improve the performance of the system.
The group works with researchers from the State University of Amazonas (UEA), the Foundation for Tropical Medicine Doutor Heitor Vieira Dourado, the Foundation for Health Monitoring in Amazonas, Fiocruz Amazônia and several hospitals associated with the project.
In addition to the new diagnostic methodology, the project involves researching the mechanisms involved in blood clotting disorders – including changes in platelet aggregation ability – that have been linked to Covid-19. This part of the investigation is coordinated by USP professor José Carlos Nicolau. The work described in the article is also supported by Fapesp through help for USP professor Ester Sabino and the Unicamp teachers Wagner José Fávaro and Fabio Trindade Maranhão Costa.
The article Covid-19 automated diagnosis and risk assessment through metabolomics and machine learning can be read at www.medrxiv.org/content/10.1101/2020.07.24.20161828v1.
This text was published by Agência Fapesp.