key: cord-0848516-vdsvyzrt
authors: Trahtemberg, U.; Rottapel, R.; Dos Santos, C. C.; Di Battista, A. P.; Slutsky, A. S.; Baker, A. J.; Fritzler, M. J.; Group, COLOBILI - COVID19 Longitudinal Biomarkers of Lung Injury Study
title: COVID-19 associated autoimmunity is a feature of severe respiratory disease - a Bayesian analysis.
date: 2021-02-19
journal: nan
DOI: 10.1101/2021.02.17.21251953
sha: c319a710ea19a80ec62b91e4466fb1d4908425e4
doc_id: 848516
cord_uid: vdsvyzrt

Background: Serological and clinical features with similarities to systemic autoimmunity have been reported in severe COVID-19, but there is a lack of studies that include contemporaneous controls who do not have COVID-19. Methods: Observational cohort study of adult patients admitted to an intensive care unit with acute respiratory failure. Patients were divided into COVID+ and COVID- based on SARS-CoV-2 PCR from nasopharyngeal swabs and/or endotracheal aspirates. No COVID-19 specific interventions were given. The primary clinical outcome was death in the ICU within 3 months; secondary outcomes included in-hospital death and disease severity measures. Measurements including autoantibodies, were done longitudinally. ANOVA and Fisher's exact test were used with alpha=0.05, with a false discovery rate of q=0.05. Bayesian analysis was performed to provide credible estimates of the possible states of nature compatible with our results. Results: 22 COVID+ and 20 COVID- patients were recruited, 69% males, median age 60.5 years. Overall, 64% had anti-nuclear antibodies, 38% had antigen-specific autoantibodies, 31% had myositis related autoantibodies, and 38% had high levels of anti-cytokine autoantibodies. There were no statistically significant differences between COVID+ and COVID- for any of the clinical or autoantibody parameters. A specific pattern of anti-nuclear antibodies was associated with worse clinical severity for both cohorts. Conclusions: Severe COVID+ patients have similar humoral autoimmune features as comparably ill COVID- patients, suggesting that autoantibodies are a feature of critical illness regardless of COVID-19 status. The clinical significance of autoimmune serology and the correlation with severity in critical illness remains to be elucidated.

The immune response in SARS-CoV2 infection shows evidence of immune dysregulation in 58 patients with severe clinical manifestations 1 . While there is continued debate about the nature 59 of the heightened inflammatory responses in severe COVID-19 patients 2, 3 there is mounting 60 Nomenclature. AAB is a general term that encompasses the autoimmune humoral responses 128 assayed. The HEp-2 IFA are commonly referred as anti-nuclear antibodies (ANA) even though 129 cytoplasmic and cell cycle patterns were included in the analysis. Their classification was 130 according to the International Consensus on Autoantibody Patterns (ICAP) nomenclature 131 (www.anapatterns.org: last accessed 1/29/2021). The AAB test results that identified specific, 132 named target antigens (see details above), were called collectively specific AAB (spAAB). We 133 have further separated them into myositis-related and non-myositis-related AAB. Anti-cytokine 134 AAB are referred to directly. 135 Data analysis. All the data was organized by UT and analyzed by UT and ADB. ANOVA was used 136 for continuous variables and Fisher's exact test was used for categorical variables at α=0.05, 137 adjusted for multiple comparisons as indicated in the text using the false discovery rate at 138 q=0.05. Bayesian analysis: a Gibbs Markov chain Monte Carlo sampling was employed to 139 estimate posterior distributions of the difference between COVID + and COVID -, using non-140 informative priors. The mean difference and 95% high density intervals (HDI) were calculated. A 141 result was considered significant for a given variable if the 95% HDI of the difference between 142 COVID + vs. COVIDpatients did not enclose zero. An exploratory analysis was conducted to 143 estimate the effect different priors would have on the posterior distributions; the priors 144 represent varying pre-existing assumptions on the prevalence of differences between COVID + vs. The demographic and clinical characteristics, including past medical history are shown in Table 1  151   (see also Table S1 ). No statistically significant clinical differences were noted between the two 152 patient groups. Importantly, both patient groups experienced comparable disease severity 11, 12 153 as measured in ICU days, mechanical ventilation days and mortality rates, as well as surrogate 154 severity scores (Table 1) . Age, sex and ethnicity were not correlated with the presence of AAB 155 (not shown). 156

The presence of ANA in general were not significantly associated with disease severity (Figure  158 1a), although there was a positive correlation with disease severity that did not reach statistical 159 significance. Nevertheless, the presence of ICAP AC19 (cytoplasmic dense fine speckled) and/or 160 AC20 (cytoplasmic fine speckled) IFA patterns, specifically, were consistently associated with 161 worse severity of illness scores in both patient groups (Figure 1a ). Neither the prevalence nor 162 the IFA staining patterns were different between the COVID + and COVIDgroups, although some 163 patients demonstrated unique IFA patterns (Figure 1b) . 164 165 We compared the incidence of spAAB directed against autoantigens typically associated with 166 systemic autoimmune diseases and myositis in critically ill COVID + and COVIDpatients ( Figure  167 2). Patients who had a single, transient, low-titer spAAB result were not classified as having a 168 positive test. Figure 2 shows that the presence of spAAB was similar in both COVID + and COVID -. 169

In addition, there was no difference between the temporal development of spAAB between the 170 two groups (Figure 2) , nor was there a correlation between the emergence of AAB and SARS-171

CoV-2 seroconversion in the COVID + group (not shown). Of note, most of these AAB persisted 172 beyond 10 days (Figure 2) , with the longest measurement available up to 54 days of 173 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) (Table 2) . This difference was driven mainly by anti-interleukin (IL)-6, anti-IL-10 and anti-IL-17f 178 AAB but it did not reach statistical significance (Tables 1 and 2) . When analyzing all positive 179 results of anti-cytokine AAB (as opposed to only the high titres), the difference between COVID + 180 and COVIDremained at ~25% (Table 1) . Interestingly, COVIDpatients had a similar incidence of 181 anti-interferon-γ and anti-interferon-β AAB as COVID + (Table 2) . Anti-cytokine AAB were not 182 associated with the presence of spAAB or ANA (not shown). 183

To better understand the ramifications of our data as well as confirm our results, we performed 185 a Bayesian analysis, which is more information-rich than null hypothesis statistical testing 186 (NHST) 13 . In Figure 3a we show the distribution of posterior probabilities for the difference 187 between the COVID + and COVIDcohorts for ANA, spAAB and anti-cytokine AAB. Since all the 188 95% credible HDI cross the zero, our data is compatible with the null hypothesis, confirming the 189 results from the NHST analysis. Importantly, this analysis used a non-informative prior; that is, 190 we did not assume any previous knowledge of what the true state of nature is. This is fitting 191 since there little is known on the prevalence of autoimmune phenomena among critically ill 192 patients 5 and, to our knowledge, this is the first report on autoimmunity among severe 19 patients that has a control group of similarly ill, critical patients. We then performed an 194 exploratory analysis assigning speculated, a priori differences between the COVID + and COVID -195 populations -Bayesian prior probabilities. Figure 3b shows that, for our results to be compatible 196 with the existence of a significant difference, it would be necessary to assume a prior probability 197 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.17.21251953 doi: medRxiv preprint of the COVID + patients having a prevalence that is at least 15% higher for anti-cytokine AAB, 20% 198 higher for spAAB, and 35% higher for ANA. The resulting posterior mean differences would be in 199 the order of 15 to 20% more autoimmunity among COVID + than COVID -(summarized in Table  200 S2). Overall, close to 15,000 data points were collected between demographic variables and 210 laboratory and clinical variables over time. We observed high prevalence and titres of AAB 211 among the COVID + patients our study (68% had ANA), in agreement with some previous reports 212 of COVID +14, 15 . Yet, surprisingly, we also observed high prevalence and titres of AAB in the 213 COVIDpatient cohort (60% had ANA). In another cross-sectional study 4 , the majority of positive 214 sera had reactivity to only single nuclear antigens, whereas most of our patients showed 215 multiple reactivities (Figure 2) , suggesting a relatively widespread loss of humoral tolerance in 216 some of our patients. We found no major differences in either the autoantigen specificity, 217 temporal dynamics, or titers in COVID + vs. COVIDcohorts. These data suggest that AAB 218 production may be a feature of immune dysfunction associated with acute systemic illness 219 rather than specifically with a SARS-CoV-2 driven immune pathology. Importantly, specific ANA 220 patterns (AC19 and/or AC20) were broadly correlated with disease severity measures, 221 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) pneumonia had neutralizing IgG AAB against a spectrum of interferon proteins, which blocked 226 SARS-CoV-2 infection in vitro 16 . Of note, these AAB were not found in asymptomatic or mild 227 SARS-CoV-2 infection, but they were not studied in similarly ill, COVID-patients. Our results 228

show that matched COVIDcontrols have a similar incidence of anti-interferon-γ and anti-229 interferon-β AAB as COVID + patients. While we did not study if the anti-cytokine AAB in our 230 study had neutralizing activity, we did not find that the presence of these AAB differed between 231 COVID + and COVID -. The small sample size and the large absolute difference for anti-IL-6, anti-IL-232 10 and anti-IL-17f raise the possibility of a type-2 error. We addressed this concern using 233 Bayesian analysis, which confirmed the results of the NHST. 234

Our data and analysis support that, humoral autoimmunity, as described in detail using an 236 extensive panel of ANA, spAAB and anti-cytokine AAB, longitudinally, does not seem to be a 237 particular characteristic of COVID+ patients but rather of critically ill patients with respiratory 238 failure. The collection of data over time significantly increases the reliability of our results by a) 239 reducing the biases of different, arbitrary sampling times between different patients, b) 240 providing test-retest internal validity, and c) characterizing the development of autoimmunity in 241 a physiological context -over time. Although there is no direct evidence to assume a priori that 242 COVID + have more autoimmunity than COVID -, some clinicians may consider the immune 243 characteristics of severe COVID-19 enough reason to warrant such assumptions. Our analysis 244 suggests that, in that case, the expected differences would be in the order of a modest 15-20% 245 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Our study's main shortcoming is the small sample size. We account for this by employing 267

Bayesian analysis to explore the boundaries of the possible states of nature underlying our 268 results. Another shortcoming is the need to consolidate the different autoimmune serologies 269 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Table 1 for 373 details on the variables. AC19 is shown together with AC20 given their similarity, as is the 374 common usage. The variables were compared in a pairwise fashion using ANOVA for continuous 375 variables and Fisher's exact test for categorical variables at α=0.05, followed by the false 376 discovery rate at q=0.05; green squares indicate lack of association, red squares indicate 377 statistically significant associations. Y/N denotes dichotomization into whether they were 378 present "yes or no" (i.e. titre above 1:80). Similar to table 1, the ANA represent the results 379 during longitudinal sampling up to day 10 of ICU admission, for standardization among patients. 380

The clinical outcomes were measured for up to 3 months. B) Characterization of autoantibodies. 381

Heatmap representation of the highest ANA titers during the first 3 months of admission. 382

Includes more patients than in Table 1 The left panel shows the specific AAB serology of COVID+ patients, the right panel shows the 391 specific AAB serology of COVID-patients, and the lower panel shows specific AAB that were 392 detected on only one patient, shown separately for clarity. The titres were classified into low, 393 medium and high levels using cutoffs established by the laboratory, shown in light grey, dark 394 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. distributions. The data was generated using the same procedure as in "A", with the X axis 410 representing the difference in the posterior distribution between the COVID + and COVID -. The 411 flat prior probability was exchanged varying levels of prior bias; the box and whiskers represent 412 the 95% HDI at each level of prior bias. The positive Y axes show the results assuming that 413 COVID + have a higher incidence of positivity than COVID -, while the negative Y axes show the 414 opposite; the zero Y indicates a prior that assumes no bias, with an equal probability of 50% 415 positivity rate for both COVID + and COVID -. A 95% HDI that does not include the zero on the X 416 axis indicates a significant difference between the COVID + and COVIDposterior probability 417 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ;  distributions at that level of prior bias, indicated with a red dot. The simulations were performed 418 on 5% intervals; the results are shown for 10% intervals for clarity. 419 420 421 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ; Mean of SOFA score for 1st 3 days mean (CI) 9.6 (10.7-8.5) 9.3 (11-7.7) 9.9 (11.6-8.3) There was no statistically significant difference between COVID + and COVIDpatients for all 440 variables, using ANOVA for continuous variables and Fisher's exact test for categorical variables 441 at α=0.05, followed by the false discovery rate at q=0.05. 442 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. cytokine antibodies sum to more than "ALL" since some patients had more than one high titre 450 anti-cytokine antibody. Once adjusted for multiple comparisons, there were no statistically 451 significant differences between COVID+ and COVID-for any of the results (Fisher's exact test at 452 α=0.05 followed by the false discovery rate at q=0.05). The following anti-cytokine AAB did not 453 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. Abbreviations: ANA, anti-nuclear antibodies; AC, International Consensus on Autoantibody Patterns anticellular pattern nomenclature; APACHE, Acute Physiology and Chronic Health Evaluation (score); Max, end point titer; ICU, intensive care unit; MV, mechanical ventilation; SOFA, sequential organ failure assessment (score). For more details on the nomenclature used for the autoimmune assays please refer to the online materials and methods.

Legend: A) Association between anti-nuclear antibodies and outcomes. Refer to the legend of Table 1 for details on the variables. AC19 is shown together with AC20 given their similarity, as is the common usage. The variables were compared in a pairwise fashion using ANOVA for continuous variables and Fisher's exact test for categorical variables at α=0.05, followed by the false discovery rate at q=0.05;

Rescue days MV days ICU days ICU death SOFA 3d SOFA 7d

green squares indicate lack of association, red squares indicate statistically significant associations. Y/N denotes dichotomization into whether they were present "yes or no" (i.e. titre above 1:80). Similar to table 1, the ANA represent the results during longitudinal sampling up to day 10 of ICU admission, for standardization among patients. The clinical outcomes were measured for up to 3 months. B) Characterization of autoantibodies. Heatmap representation of the highest ANA titers during the first 3 months of admission. Includes more patients than in Table 1 and Figure 1a since some AABs developed after 10 days. This may introduce bias since not all patients had prolonged hospitalizations but is a better representation of the dynamics and spectrum of ANA detected. Only patients with titers > 1:80 are included. AC4 is shown together with AC5, and AC19 is shown together with AC20, given their similarity, as is the common usage. AC7, 26, 27 and 28 are shown together given their rarity individually.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ; Legend: The left panel shows the specific AAB serology of COVID + patients, the right panel shows the specific AAB serology of COVIDpatients, and the lower panel shows specific AAB that were detected on only one patient, shown separately for clarity. The titres were classified into low, medium and high levels using cutoffs established by the laboratory, shown in light grey, dark grey and black, respectively; white indicates negative. Arrival indicates the first sample obtained, either on the day of admission or the next morning. Within 10 d indicates during the longitudinal sampling, up to day 10 of ICU stay. After 10 d indicates later samples, collected every two weeks; not all patients had samples after 10 days. The following specific AAB were not detected in any samples: histones, Sm/U2-U6 RNP, U1-RNP, ribosomal P, topoisomerase I, Mi2-α, MDA5, NXP2, PL7, PL12, SRP, EJ, OJ, HMGCR or NT5C1 A/Mup44 and SAR1. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.17.21251953 doi: medRxiv preprint four Gibbs sampling chains employing a flat prior probability, at 5000 iterations per chain, minus 1% burn-in, resulting in 19800 iterations in total for each histogram. B) Effect of prior bias on posterior distributions. The data was generated using the same procedure as in "A", with the X axis representing the difference in the posterior distribution between the COVID + and COVID -.

The flat prior probability was exchanged varying levels of prior bias; the box and whiskers represent the 95% HDI at each level of prior bias. The positive Y axes show the results assuming that COVID + have a higher incidence of positivity than COVID -, while the negative Y axes show the opposite; the zero Y indicates a prior that assumes no bias, with an equal probability of 50% positivity rate for both COVID + and COVID -. A 95% HDI that does not include the zero on the X axis indicates a significant difference between the COVID + and COVIDposterior probability distributions at that level of prior bias, indicated with a red dot. The simulations were performed on 5% intervals; the results are shown for 10% intervals for clarity.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.17.21251953 doi: medRxiv preprint

Not just antibodies: B cells and T cells mediate immunity to COVID-314

 8  320  9  160  160  320  13 320  640  2560  16  160  19  1280 320  640 320  24  320  36  640  37  160 160  41  160 160  48  320  49  320  55 640 320 60 160 320 1280 63 320 320 160