Immunophenotypes of anti-SARS-CoV-2 responses associated with fatal COVID-19

Background The relationship between anti-SARS-CoV-2 humoral immune response, pathogenic inflammation, lymphocytes and fatal COVID-19 is poorly understood. Methods A longitudinal prospective cohort of hospitalised patients with COVID-19 (n=254) was followed up to 35 days after admission (median, 8 days). We measured early anti-SARS-CoV-2 S1 antibody IgG levels and dynamic (698 samples) of quantitative circulating T-, B- and natural killer lymphocyte subsets and serum interleukin-6 (IL-6) response. We used machine learning to identify patterns of the immune response and related these patterns to the primary outcome of 28-day mortality in analyses adjusted for clinical severity factors. Results Overall, 45 (18%) patients died within 28 days after hospitalisation. We identified six clusters representing discrete anti-SARS-CoV-2 immunophenotypes. Clusters differed considerably in COVID-19 survival. Two clusters, the anti-S1-IgGlowestTlowestBlowestNKmodIL-6mod, and the anti-S1-IgGhighTlowBmodNKmodIL-6highest had a high risk of fatal COVID-19 (HR 3.36–21.69; 95% CI 1.51–163.61 and HR 8.39–10.79; 95% CI 1.20–82.67; p≤0.03, respectively). The anti-S1-IgGhighestTlowestBmodNKmodIL-6mod and anti-S1-IgGlowThighestBhighestNKhighestIL-6low cluster were associated with moderate risk of mortality. In contrast, two clusters the anti-S1-IgGhighThighBmodNKmodIL-6low and anti-S1-IgGhighestThighestBhighNKhighIL-6lowest clusters were characterised by a very low risk of mortality. Conclusions By employing unsupervised machine learning we identified multiple anti-SARS-CoV-2 immune response clusters and observed major differences in COVID-19 mortality between these clusters. Two discrete immune pathways may lead to fatal COVID-19. One is driven by impaired or delayed antiviral humoral immunity, independently of hyper-inflammation, and the other may arise through excessive IL-6-mediated host inflammation response, independently of the protective humoral response. Those observations could be explored further for application in clinical practice.


Study participants.
We conducted this study of COVID-19 at the University Clinic of Respiratory and Allergic Diseases Golnik, a quaternary, acute care hospital in Slovenia. The first hospitalized patient was included on September 4 th and the last on December 12 th , 2020. All information presented in this report is based on a data cut-off of January 9 th , 2021. During a follow-up of 3 to 26 days after presentation, 45 patients (18%) died (8 after intubation), with December 16th, 2020, as the time of the last endpoint event. During a follow-up of 3 to 35 days after presentation, 209 patients had survived hospital discharge (7 after successful cessation of invasive mechanical ventilation), with December 17, 2020, as the time of discharge of the last patients. Six of those patients were discharged more than 28 days after the presentation (during a follow-up of 29 to 35 days; one after successful cessation of invasive mechanical ventilation). None of the patients was transferred to anothe For non-hospialized control subjects we included 40 adult individuals with prior COVID-19 confirmed by a nasopharyngeal swab PCR test (Supplement Table E1). In first control subject SARS-CoV-2 infection was confirmed on April 19th, 2021 and in the last on February 18th, 2022. The first blood sampling was done on July 12th, 2021 and the last on March 16th, 2022. Consequently a single post-blood sample was collected median of 60 days (range 20 to 175 days) after infection. None of the controls subjects was vaccinated neither before infection nor before sampling.

Data analysis
First, we performed a standard Cox proportional hazards analysis. The analysis was done with all the variables and in a stepwise manner. We analysed the variables in an univariate fashion and if the variable correlated significantly with the risk of death in univariate analysis it was further used as predictor in multivariable analysis (Supplement Table E1).
Regression analysis (including Cox regression) is considered a supervised machine learning method, since it fits the predictors to the outcome of interest (i.e. fatal COVID-19). To search for patterns in immune response that are not fitted to the outcome of interest and thus possibly reflect inherent biological processes, we next performed cluster analysis (unsupervised machine learning method) of immune variables. We used Gaussian mixture model algorithm as previously

IL-6 measurements in sera, and longitudinal dynamic of circulating B T cells, C CD4 and D CD8 subpopulations of T cells, E B and F
NK cell and G IL-6 in sera of hospitalized COVID-19 patients per day after admission according to whether patients progressed to the primary endpoint of death at 28 days or recovered to hospital discharge. Non-hospitalized controls represent adult individuals with prior COVID-19 confirmed by a nasopharyngeal swab PCR test, post-sampling median 60 days after infection.

Figure E3. The lowest T cells, CD4 and CD8 subpopulations of T cells, B and NK cells absolute
counts, and the highest serum IL-6 after admission to the hospital according to whether COVID-19 patients progressed to the primary endpoint of death at 28 days or recovered to hospital discharge. Non-hospitalized controls represent adult individuals with prior COVID-19 confirmed by a nasopharyngeal swab PCR test, post-sampling median 60 days after infection.  Table E2. Associations between demographic and clinical factors, antibody response, lymphocyte subsets, IL-6, and the endpoint of death at 28 days in 254 hospitalized subjects.
*SARS-CoV-2 anti S1 IgG antibodies were measured median 4 days after admission to the hospital. †Decreases in circulating lymphocytes were defined as the lowest T cell, CD4, and CD8 subpopulations of T cells, B and NK cell absolute counts recorded after admission. ‡IL-6 increase was defined as the maximal serum level recorded after admission. §BMI data were missing for 34 patients. The decision to prescribe glucocorticoids was at the discretion of the treatment team for each patient. In the multivariable Cox model, only significant predictors (P<0.05) from univariate Cox analysis were retained and only CD3 T-cell counts were used (since they reflect the presence of CD4 as well as CD8 cells). HR-hazard ratio.