Proteome analysis develops novel plasma proteins classifier in predicting the mortality of COVID-19

Abstract

COVID-19 has been a global concern for 3 years, however, consecutive plasma protein changes in the disease course are currently unclear. Setting the mortality within 28 days of admission as the main clinical outcome, plasma samples were collected from patients in discovery and independent validation groups at different time points during the disease course. The whole patients were divided into death and survival groups according to their clinical outcomes. Proteomics and pathway/network analyses were used to find the differentially expressed proteins and pathways. Then, we used machine learning to develop a protein classifier which can predict the clinical outcomes of the patients with COVID-19 and help identify the high-risk patients. Finally, a classifier including C-reactive protein, extracellular matrix protein 1, insulin-like growth factor-binding protein complex acid labile subunit, E3 ubiquitin-protein ligase HECW1 and phosphatidylcholine-sterol acyltransferase was determined. The prediction value of the model was verified with an independent patient cohort. This novel model can realize early prediction of 28-day mortality of patients with COVID-19, with the area under curve 0.88 in discovery group and 0.80 in validation group, superior to 4C mortality and E-CURB65 scores. In total, this work revealed a potential protein classifier which can assist in predicting the outcomes of COVID-19 patients and providing new diagnostic directions.

FIGURE 1

Study design and patients. (A) Overview of blood samples collection from COVID‐19 patients, including the discovery cohort from centre 1 and the validation cohort from the other centre. The workflow for processing the proteomic data was shown, including the plasma separation, protein digestion, LC–MS/MS analysis, database search and further computational analyses. (B) The gender distribution between 28‐day survival and 28‐day death groups of the discovery cohort and independent validation cohort. (C) The age distribution. Data points indicate the data of single patient at each time point and are presented as median with interquartile range. The centre line within each box shows the median, and the top and bottom of each box represent the 75th and 25th percentile values, respectively. (D) The severity distribution.

FIGURE 2

Proteome profiling of plasma samples from COVID‐19 patients of discovery cohort. (A,B) The distribution of numbers of quantified peptides and proteins between 28‐day survival and 28‐day death groups of the discovery cohort. (C) The distribution of peptide numbers of quantified proteins. The orange bar represents protein count with relevant number of peptides. The blue curve represents accumulative proportion of proteins from those with more than 19 peptides per protein to one peptide. (D) Pearson correlation coefficients between each pair of technical replicates and biological replicates from discovery cohort. (E) Unsupervised clustering of whole proteomic data among different batches of discovery cohort. (F) The heatmap of proteomic data of the discovery cohort.

FIGURE 3

Proteomic alternations between 28‐day survival and 28‐day death groups of the discovery cohort. (A) Dysregulated proteins of plasma collected at week 1, 2 and 3 between two groups. The top five proteins with most significant B‐H adjusted p value were labelled with protein names. Fold change represents protein intensity of 28‐day death versus 28‐day survival. (B) Summary of numbers of dysregulated proteins. (C) The overlapping of dysregulated proteins among three times frames. (D) Enriched pathways for overall dysregulated proteins using Metascape. (E) The network of dysregulated proteins by String and their involved functions.

FIGURE 4

Identification of protein classifier to distinguish 28‐day survival and death groups using a XGBoost strategy. (A) The workflow to build a protein classifier for classification of 28‐day survival and death groups of COVID‐19 patients, including feature selection, parameter optimization and predictive model validation. (B) The receiver operating characteristic curves of the internal validation dataset across five‐fold cross‐validation and independent validation cohort using the five‐protein classifier. (C) The confusion matrix of the five‐protein classifier in the independent validation cohort. (D) The Uniform Manifold Approximation and Projection (UMAP) analysis of the five proteins among different plasma samples of the discovery cohort. (E) UMAP analysis of the five proteins among different plasma samples of the independent validation cohort.