Discovery and systematic assessment of early biomarkers that predict progression to severe COVID-19 disease

Abstract

Background: The clinical course of COVID-19 patients ranges from asymptomatic infection, via mild and moderate illness, to severe disease and even fatal outcome. Biomarkers which enable an early prediction of the severity of COVID-19 progression, would be enormously beneficial to guide patient care and early intervention prior to hospitalization.

Methods: Here we describe the identification of plasma protein biomarkers using an antibody microarray-based approach in order to predict a severe cause of a COVID-19 disease already in an early phase of SARS-CoV-2 infection. To this end, plasma samples from two independent cohorts were analyzed by antibody microarrays targeting up to 998 different proteins.

Results: In total, we identified 11 promising protein biomarker candidates to predict disease severity during an early phase of COVID-19 infection coherently in both analyzed cohorts. A set of four (S100A8/A9, TSP1, FINC, IFNL1), and two sets of three proteins (S100A8/A9, TSP1, ERBB2 and S100A8/A9, TSP1, IFNL1) were selected using machine learning as multimarker panels with sufficient accuracy for the implementation in a prognostic test.

Conclusions: Using these biomarkers, patients at high risk of developing a severe or critical disease may be selected for treatment with specialized therapeutic options such as neutralizing antibodies or antivirals. Early therapy through early stratification may not only have a positive impact on the outcome of individual COVID-19 patients but could additionally prevent hospitals from being overwhelmed in potential future pandemic situations.

Fig. 1. Patient and sample characteristics of the 1^st cohort.

Characteristics (age, sex) of individuals included in this study and timepoints of sample collection after symptom onset. Letters a–p indicate the individual patients. Male patients are depicted in blue, female patients in red. Circles represent samples of mild or moderate (MM) patients, squares those of critical or severe (CS) patients. Acute phase: 14 MM and 3 CS samples; intermediate phase: 13 MM and 7 CS samples; late phase: 10 and 6 samples. Dotted vertical lines indicate 10 days and 21 days after onset of symptoms.

Fig. 2. Venn diagram and volcano plots illustrating the number, degree and significance of differential protein expression in the 1^st cohort.

The volcano plots visualize the p values (adjusted for multiple testing) and corresponding log-fold changes (logFC) of the identified protein biomarker candidates. A significance level of adj. p value = 0.05 is indicated as a horizontal red line. Absolute logFC cutoffs of |logFC | > 0.5 are indicated as vertical lines. 53 plasma samples from 16 COVID-19 patients were analyzed and divided into an acute (a) an intermediate (b) and a late stage (c), for which we included 18, 20 and 16 samples, respectively. Proteins with a positive logFC had a higher abundance in CS samples, proteins with a negative value in MM samples. d: Venn diagram listing the differential proteins and their numbers in the respective phases. Green numbers and protein IDs indicate proteins more abundant in CS patients and red numbers and IDs proteins with higher abundance in MM patients.

Fig. 3. Stripcharts representing individual array values for all proteins selected as top candidates in the 1^st cohort.

Each protein is measured by four replicate spots per array and is represented by their mean. The y-axis illustrates the log2 ratio of the individual samples and a reference sample while the x-axis is divided based on clinical course of disease of the patient (CS and MM) as well as phase of infection. 53 plasma samples from 16 COVID-19 patients were analyzed and divided into an acute (A) an intermediate (M) and a late stage (L), for which we included 18, 20 and 16 samples, respectively. Acute CS and MM samples are highlighted in red and blue respectively. Diamonds indicate arithmetic sample group means. Whiskers indicate one standard deviation, calculated based on arithmetic means. Empty circles indicate the group coefficients fitted by the linear model with additional factors sex, age and comorbidities, which were used for logFC and p value calculation.

Fig. 4. Protein biomarker candidates from the 2^nd cohort to predict a severe or critical disease in acute phase.

Plasma samples from 94 COVID-19 patients during the acute phase of disease were analyzed on antibody microarrays to identify differentially abundant proteins between patients with either a mild or moderate (MM, n = 47) or a critical or severe (CS, n = 47) disease course. The volcano plot visualizes the p values and corresponding log-fold changes (logFC). A significance level of adj. p value = 0.05 is indicated as a horizontal red line. Absolute logFC cutoffs of |logFC | > 0.5 are indicated as vertical lines. Proteins with a positive logFC had a higher abundance in CS samples, proteins with a negative value in MM samples. Two different antibodies against S10A8/A9 were included on the microarray, with both antibodies showing significant differences between CS and MM samples.

Fig. 5. Biomarkers to predict a severe / critical disease in both study cohorts.

The y-axis illustrates log2(sample / reference) values after subtracting the group mean of the respective MM samples per cohort/protein, thus setting the mean value of MM samples as a baseline. Within the 1^st cohort 18 samples (MM = 17; CS = 3) were analyzed, while 94 sample (MM = 47; CS = 47) were analyzed within the 2^nd cohort. The x-axis is divided based on clinical course of disease of the patient (CS and MM). Only acute CS and MM samples are shown. Diamonds indicate arithmetic sample group means. Whiskers indicate one standard deviation, calculated based on the arithmetic means. Empty circles indicate the group coefficients fitted by the linear model with additional factors sex, age and comorbidities, which were used for logFC and p value calculation.

Fig. 6. Validation of microarray data.

Plasma samples from 94 COVID-19 patients during the acute phase of disease (2^nd cohort) were analyzed by ELISA, to validate differentially abundant proteins between patients with either a mild or moderate (MM, n = 47) or a critical or severe (CS, n = 47) disease course. Stripcharts representing individual S100A8/A9 (a) and CRP (b) ELISA measurements. The y-axis displays the log₂ of the measured protein concentration while the x-axis is divided based on the later clinical course of disease of the patient (CS and MM). Triangles and whiskers indicate means and one standard deviation of the sample groups with critical/severe or mild/moderate course of the disease respectively. Possible cut-offs with a sensitivity of 89% are indicated by dotted grey lines. c, d Scatter plots demonstrate a high correlation between discovery antibody microarray data (y-axis) and ELISA validation (x-axis) with Pearson’s r > 0.9.

Fig. 7. Estimation of diagnostic accuracy for disease severity prediction.

For the individual protein biomarkers S100A8/A9 and CRP, ROC curves of ELISA data as well as coherent multiplex antibody array data were aligned. Marker combinations of 2–4 proteins were selected from linSVM models, which outperform the individual biomarker candidates and exhibit a high accuracy with an AUC of up to 0.928. Area under the ROC curve (AUC) is presented for each biomarker or combination.