Predicting severity in COVID-19 disease using sepsis blood gene expression signatures

Abstract

Severely-afflicted COVID-19 patients can exhibit disease manifestations representative of sepsis, including acute respiratory distress syndrome and multiple organ failure. We hypothesized that diagnostic tools used in managing all-cause sepsis, such as clinical criteria, biomarkers, and gene expression signatures, should extend to COVID-19 patients. Here we analyzed the whole blood transcriptome of 124 early (1-5 days post-hospital admission) and late (6-20 days post-admission) sampled patients with confirmed COVID-19 infections from hospitals in Quebec, Canada. Mechanisms associated with COVID-19 severity were identified between severity groups (ranging from mild disease to the requirement for mechanical ventilation and mortality), and established sepsis signatures were assessed for dysregulation. Specifically, gene expression signatures representing pathophysiological events, namely cellular reprogramming, organ dysfunction, and mortality, were significantly enriched and predictive of severity and lethality in COVID-19 patients. Mechanistic endotypes reflective of distinct sepsis aetiologies and therapeutic opportunities were also identified in subsets of patients, enabling prediction of potentially-effective repurposed drugs. The expression of sepsis gene expression signatures in severely-afflicted COVID-19 patients indicates that these patients should be classified as having severe sepsis. Accordingly, in severe COVID-19 patients, these signatures should be strongly considered for the mechanistic characterization, diagnosis, and guidance of treatment using repurposed drugs.

Figure 1

Functional characterization of the gene expression differences between severity groups and eventual mortality in Quebec COVID-19 patients. Shown are the top 10 enriched pathways/hallmarks for each comparison, direction, and gene set database assessed (i.e., lowest p values). Severity groups were assessed using the respiratory support required at time of sampling with 54 severe, 28 intermediate, and 42 moderate patients. There were 20 patients who died and 100 patients who survived. The severe vs. moderate, severe vs. intermediate, and intermediate vs. moderate, died vs. survived, comparisons yielded 2671 (1794 up, 877 down), 694 (535 up, 159 down), 169 (136 up, 33 down), and 1708 (1006 up, 702 down) DE genes, respectively (displaying ≥  ± 1.5-fold change; adjusted p ≤ 0.05). Age, sex, batch, and time of sampling relative to hospital admission were included in the underlying DESeq2 model to correct for their contribution to gene expression. A post-hoc power analysis (performed with the online tool RnaSeqSampleSize) using parameters estimated from previous RNA-Seq studies, indicated our study was sufficiently powered (p = 0.8) to detect DE gene differences in gene expression (at least 15% of transcriptome) between patient groups.

Figure 2

Predictive performance of the reduced CR, Organ Dysfunction, and Mortality signatures among patients of the COVID-19 patient cohorts. (a) Accuracy/AUCs across a tenfold repeated cross validation scheme (i.e., total 100 folds) using logistic regression. (b) Kaplan Meier curves of patients with High or Low Mortality signature enrichment in all patients (i.e., early and late sampled patients). Time to mortality was assessed over 30 days, with the baseline time being the day of sampling. Patients were stratified into either High (score ≥ 0; n = 60) or Low (score < 0; n = 64) signature expression based on GSVA enrichment statistics, and this was used as a predictor of 30-day survival.

Figure 3

Heatmap depicting GSVA endotype signature enrichment scores and endotype classification COVID-19 patients. (a) Endotype classification of early sampled patients. (b) Endotype classification of late sampled patients. Endotype enrichment scores are indicated in the bottom five rows in each subfigure with red blocks indicating positive enrichment. Various metrics are indicated in the top five bars in each subgraph including endotype assignment, patient severity (as assessed by respiratory support requirement at sampling), impending mortality, and ICU admission, as well as Early (top subfigure) or Late (bottom subfigure) sampling. In both the early and late groups, the NPS endotype displayed the worst patient severity and highest mortality rates.

Figure 4

Function-based, minimally connected first order protein:protein interaction networks of the top endotype specific DE genes. (a) NPS endotype network. (b) INF endotype network. The top 200 endotype-specific DE genes (by absolute value fold changes) were input into NetworkAnalyst. Each DE gene set formed a well-connected minimum connected first-order network, indicating that the genes involved are functionally related and likely collectively regulate or play key roles in one or more related biological mechanisms. Red and green nodes are genes with increased and decreased expression specific to the endotype, respectively, while grey nodes are interconnecting first-order interaction nodes. The size of the nodes indicates their connectivity (hub degree) within the network, i.e., how well interconnected any given node is to other nodes in the Network (NB. hubs are key molecules in signalling since they are highly interconnected; they receive and integrate multiple signals and pass them on to downstream nodes). Lines represent edges that indicate known (experimentally-determined) protein:protein interactions derived from www.innatedb.ca (version 5.4).