The coronavirus disease 2019 (COVID-19) pandemic has caused a pandemic for hospitals around the world, with many reporting shortages in intensive care units (ICUs) and beds.
While large-scale vaccination schemes have reduced hospitalizations in developed countries, many countries are still battling the disease.
Researchers at the Georgia Institute of Technology have created a model that could help determine disease outcome in patients in the ICU.
A preprint version of the study is available at biorxiv* Server while the article undergoes peer review.
The researchers used their platform to characterize and quantify serum antibodies against canonical antigens from blood taken from 21 severe COVID-19 patients. These patients were diagnosed using a standard nasopharyngeal swab followed by PCR and admitted to the ICU ~6 days after symptom onset.
Serum samples were taken within 24 h of ICU admission. The scientists aimed to determine whether the antibody profile of these samples taken at ICU admission could predict disease outcome. After noticing that those who survived and who did not have higher levels of specific antibodies than health controls before the pandemic, the researchers decided to pursue a multivariate machine learning approach involving different aspects of the antibody response. Did.
A two-stage machine learning approach was used initially with feature selection using LASSO and L1 regularization to prevent overfitting, followed by classification using down-selected features. The models they predicted were predictive results measured in K-fold cross-validation with permutation test, and were based on anti-spike IgA, anti-spike RBD IgA2 and RBD-directed antibody galactosylation. Partial least squares discriminant analysis (PLS-DA) using these three down-selected features helped visualize stratification and successfully demonstrated that they could stratify survivors and non-survivors, with all three characteristics being survivors. are high in
To ensure that the model was reliable, its performance was tested on a group of critically ill ICU patients. The generated model was still fairly predictable, although it performed slightly worse. This was explained by a lack of available information on antigen-specific antibody glycosylation measurements, which was one of the major factors dependent on the model. Despite this, the results were still accurate enough to validate the model’s effectiveness.
Next, the scientists attempted to examine the levels of antibody responses directed against the non-canonical antigens nsp3, nsp13, orf3a and orf8 in patients with severe COVID-19, to find out whether they were independent. were predicting the outcome. They detected antibody responses in both living and non-living, but could not identify any significant differences between them, so constructed a model similar to the one described above.
Unsurprisingly, this model was just as good at predicting outcomes as the previous model. It selected four features to be analysed: anti-Orf8 IgA, anti-NSP13 IgG3, anti-M antibody FCR3A binding and anti-M antibody galactosylation. PLS-DA analysis showed that these four characteristics could differentiate between survivors and non-survivors – generally showing higher outcomes in survivors.
The results indicate that higher antibody titers to isotopes directed against both canonical and non-canonical antigens may be associated with increased survival. The model also helped to detect that increased galactosylation of RBD- and M-specific antibodies is associated with favorable outcomes, and this was confirmed by building a predictive model combining the two. To test whether the proportions of antibodies directed against certain outcomes were predictive of survival, the researchers performed a post-hoc analysis focusing on the ratio of IgA/IgG antibodies to features identified in the previous two models. . Multivariate PLS-DA visualization demonstrated that these could successfully discriminate between survivors and non-survivors.
During prior analyses, the scientists observed levels of reactivity for nsp13 and nsp3 in health control patients, and that these were different from antibodies directed against canonical antigens. As previous studies have shown that antibodies against the SARS-CoV-2 antigen can be generated through infection with different species of coronavirus, they decided to test whether this was the case here. They analyzed the sequence similarity of the spike protein and nsp13 against the respective antigens in previously transmitted coronaviruses from SARS-CoV-2, and found that nsp13 had high similarity, whereas the spike protein did not.
Next, they built a 3-way multivariate machine learning model to respond to the possibility that pre-existing cross-reactive antibodies against SARS-CoV-2 are likely to be reactive. It performed significantly better than previously built models and included five features drawn from all previous analyses.
The authors highlight that they have successfully created a model that can predict with high reliability the outcome of severe COVID-19 infection in patients from samples obtained immediately after admission to the ICU.
This model could be extremely useful for healthcare workers, and potentially used in triage for severe outbreaks, as well as helping to focus where it is most needed.
bioRxiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be considered conclusive, guide clinical practice/health-related behavior, or be regarded as established information.