Machine‐learning analysis identifies “elite” viral controllers with increased survival and homeostatic responses in critical COVID‐19

Résumé fourni par la source

Dear Editor, The outcome of COVID-19 disease is strongly related to the interaction between the virus and the host immune response, which may become dysregulated in critically ill patients. This dysregulated response is characterized by elevated levels of inflammatory mediators, an overactivation of the innate immune system,1 lymphopenia,2 delayed antibody and interferon responses,3 and a massive dissemination of viral components into the blood,4 all of which contribute to severity and increased mortality.5-7 These immune and non-immune parameters can be integrated into so-called combitypes8 to identify subgroups of patients with different immune profiles and outcomes, helping to guide clinical strategies. In a previous study we used viral RNA levels in plasma to categorize a multicentre cohort of critically ill COVID-19 patients into three subgroups with different mortality rate.4 In this study, we combined virological data (SARS-CoV-2 N1 RNA plasma load and N-antigenemia) and 32 host response biomarkers to improve classification of critically ill COVID-19 patients, with the objective to identify biological clues explaining survival. We conducted a prospective cohort study in 785 critically ill COVID-19 patients with a plasma EDTA sample collected at intensive care unit (ICU) admission. The detailed methods and the biological parameters measured are summarized in the Supporting Information. The biological characteristics of 90-day survivors compared to non-survivors (Table S1) indicated that non-survivors were more likely to exhibit the presence of SARS-CoV-2 N antigen, along with higher viral RNA load in plasma, higher tissue damage (RNase P RNA), lower lymphocyte counts, and higher neutrophils levels. Additionally, non-survivors exhibited increased concentrations of multiple biomarkers involved in endothelial dysfunction (angiopoietin 2, endothelin-1, ICAM-1 and VCAM-1), inflammation (TNF-α, IL-15 and IL-6), coagulation (D-dimmer), chemotaxis (CXCL10, CCL2, and IL-8), immunosuppression (IL-10, PD-L1, and IL1-RA), T-cell biology (CD27), apoptosis (Fas) and innate immune-related proteins (EGF and SP-D). Based on these biological characteristics, XGBoost algorithm was employed to develop a model for predicting 90-day mortality (AUROC of 0.80) (Supplementary Figure 1) and SHAP values were obtained to evaluate the influence of each biological feature on the outcome variable (Figure 1). Levels of SARS-CoV-2 N1 RNA was the parameter ranking the first to predict 90-day mortality, following by endothelin-1, IL-15, IL-8, neutrophils, IL-6, TREM-1, CCL2, CD27, SP-D, myeloperoxidase, IL-10, D-dimer, PTX-3, CXCL10, RNase P and VCAM-1, suggesting that viral control, endothelial dysregulation, pro-inflammatory mechanisms and chemotaxis are key biological functions in determining 90-day mortality in critical COVID-19 disease. On the contrary, high levels of the cytokine RANTES, anti-SARS-CoV-2 S IgM and anti-SARS-CoV-2 S IgG antibodies represented a protective factor against mortality. We further classified the patients into three groups or combitypes with different 90-day mortality rate, using a partitional clustering method based on the biological characteristics (Figure 2A, B). The Combitype-1 group was the most common (41.5%) and showed the lowest mortality rate at day 90 after ICU admission (7.7%), followed by the Combitype-2 group (21.5%) with a 90-day mortality rate of 25.4%. The 90-day mortality dramatically increased to 65.9% in the Combitype-3 group, who represented 36.9% of the cohort. Survival mean time in the first 90 days in each group was as follows [days (lower limit—upper limit)]: Combitype-1 [84.7 (82.7–86.8)], Combitype-2 [73.0 (68.5–77.6)] and Combitype-3 [44.2 (40.2–48.2)] (Figure 2C). The three groups of 90-day mortality risk exhibited different biological characteristics (Figure 3 and Table S2). The Combitype-1 group had the lowest viral RNA load in plasma, the lowest prevalence of antigenemia, the highest concentration of anti-SARS-CoV-2 S IgG and IgM antibodies, and a homeostatic response to infection, with reduced levels of all pro-inflammatory cytokines and chemoattractant proteins tested (except RANTES). Thus, Combitype-1 could be considered a group of “elite” viral controllers within the population of patients admitted to the ICU. In contrast, the Combitype-2 and -3 groups had a higher viral RNA load and higher prevalence of SASR-CoV-2 N antigen in plasma. The overall biomarker profile in the Combitype-2 and Combitype-3 groups indicated a broad dysregulation of the host response to infection, but with striking differences between these two groups. While the Combitype-2 had moderate viral RNA load along with intermediate levels of inflammatory and endothelial dysfunction biomarkers, the Combitype-3 showed the highest concentration in plasma of lipocalin-2, MPO, VCAM-1, PTX-3, IL-10, CXCL10, angiopoietin-2, IL-6, IL-15, endothelin-1, IL-8, and TREM-1, indicating an exacerbated pro-inflammatory profile coupled with higher endothelial dysregulation and very high viral RNA load in plasma. These three immune signatures were linked to significant clinical differences (Table 1). Patients in the Combitype-1 group were younger and presented better respiratory function (PaO2/FiO2 ratio), and lower organ dysfunction (SOFA score) at ICU admission, together with lower frequency of hypertension, diabetes, chronic kidney disease, and chronic neurological disease as comorbidities. On the contrary, the Combitype-3 group had the highest prevalence of diabetes and immunosuppression. In terms of complications during hospital admission, the Combitype-1 group needed less often invasive mechanical ventilation and showed a lower frequency of secondary infections, acute kidney injury and septic shock, while the Combitype-3 group suffered more frequently acute liver failure, acute kidney injury, coagulation disorders and septic shock. As mentioned earlier, the Combitype-3 group was the one who presented the highest levels of viral RNA load and pro-inflammatory mediators. Taken together, these results point to the important role of uncontrolled viral replication in the development of multiorgan failure and the extremely high mortality rate observed in this group. In line with these results, a previous investigation has shown a novel mechanism for propagating inflammation, which involves SARS-CoV-2 fragments,9 which could underlie the extrapulmonary pathologies observed in critical COVID-19 patients, particularly in the Combitype-3 group, which exhibited a very high SARS-CoV-2 RNA load in plasma. In conclusion, this is the first study combining SARS-CoV-2 RNA levels with host response data to develop a 90-day mortality prediction model by an XGBoost algorithm and employing SHAP values to evaluate the influence of each biological feature on the outcome variable. Our results showed that SARS-CoV-2 RNA load was the most important biological factor influencing 90-day mortality among COVID-19 patients admitted to the ICU and revealed that endothelin-1 and IL-15 had a higher influence on COVID-19 mortality than other pro-inflammatory cytokines, like IL-6. This prediction model confirmed our previous findings demonstrating that viral N1 RNA load was a predictor of 90-day mortality.4 However, the current clustering analysis considering 33 biological features on top to viral RNA load enabled better classification of patients with different severity (Figure 4), revealing the existence of the group showing a better prognosis within critically ill COVID-19 patients, the “elite” viral controllers. This group represented the largest group of our cohort and exhibited a robust antibody response that prevent uncontrolled viral replication and/or propagation, leading to more homeostatic immune responses to infection and increased survival. These results could help to understand the factors leading to survival not only in severe SARS-CoV-2 infection, but also in the infections ca

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.