Circulating proteins to predict COVID-19 severity

Abstract

Predicting COVID-19 severity is difficult, and the biological pathways involved are not fully understood. To approach this problem, we measured 4701 circulating human protein abundances in two independent cohorts totaling 986 individuals. We then trained prediction models including protein abundances and clinical risk factors to predict COVID-19 severity in 417 subjects and tested these models in a separate cohort of 569 individuals. For severe COVID-19, a baseline model including age and sex provided an area under the receiver operator curve (AUC) of 65% in the test cohort. Selecting 92 proteins from the 4701 unique protein abundances improved the AUC to 88% in the training cohort, which remained relatively stable in the testing cohort at 86%, suggesting good generalizability. Proteins selected from different COVID-19 severity were enriched for cytokine and cytokine receptors, but more than half of the enriched pathways were not immune-related. Taken together, these findings suggest that circulating proteins measured at early stages of disease progression are reasonably accurate predictors of COVID-19 severity. Further research is needed to understand how to incorporate protein measurement into clinical care.

Figure 1

Overall Study design. Schematic of training and testing stages of this study. Severe COVID-19 is defined as death or use of any form of oxygen supplementation. Critical COVID-19 is defined as death or severe respiratory failure (non-invasive ventilation, high flow oxygen therapy, intubation, or extracorporeal membrane oxygenation).

Figure 2

AUC score results. (a) L1 regularized logistic regression training and testing results for severe COVID-19 and (b) critical COVID-19. Blue and red are used to represent the protein model and baseline model, respectively, while solid and dotted lines represent the testing and training performance, respectively. Shaded areas denote the 95% confidence intervals for the training cohort. (c) Two-by-two contingency table results from the test set are shown for predicting severe (top left, bottom left) and critical COVID-19 (top right, bottom right) using the protein model (blue) and baseline model (red). The threshold for predicting cases was determined during training using Youden’s J statistic which selects a threshold that maximizes the sum of the sensitivity and specificity score. PPV positive predictive value, NPV negative predictive value.

Figure 3

Feature importance and correlation of SOMAmers selected in the protein model to predict severe and critical COVID-19. (Left) Coefficient values of the 92 nonzero SOMAmer reagents in the final trained L1 regularized logistic regression protein model fitted to predict severe COVID-19. The original data contained 4984 SOMAmer reagents and 4 other variables: age, sex, sample processing time, and hospital site. 92 SOMAmer reagents remained within the model along with age and sample processing time which are not shown. The model was trained on the entire BQC19 cohort using lambda = 10.0 (log₁₀ lambda = 1.0) which was the best lambda value found from the hyperparameter search. (Bottom) Coefficient values of the 67 nonzero SOMAmer reagents of the final trained L1 regularized logistic regression protein model fitted to predict critical COVID-19. The original data contained 4984 SOMAmer reagents and 4 variables age, sex, sample processing time, and hospital site. 67 SOMAmer reagents remained within the model along with age and sample processing time which are not shown. The model was trained on the entire BQC19 cohort using lambda = 10.0 (log₁₀ lambda = 1.0) which was the best lambda value found from the hyperparameter search. (Right) Spearman’s rank correlations between the 92 proteins associated with severe COVID-19 and the 67 proteins associated with critical COVID-19. These results suggest that while there were 14 overlapping proteins (SFTPD, CXCL10, RAB3A, NAGPA, CDH5, IFNA7, ZNRF3, CBS, CCL7, SETMAR, TNXB, CDHR1, CXCL13, and CBLN1), in general, the protein levels were uncorrelated with one another. Out of 6164 total correlations (92 × 67), 6150 correlations (99.8%) had a Spearman's absolute ρ < 0.8.