Mortality-associated plasma proteome dynamics in a prospective multicentre sepsis cohort

Abstract

Background: Sepsis remains a leading cause of mortality in intensive care units. Understanding the dynamics of the plasma proteome of patients with sepsis is critical for improving prognostic and therapeutic strategies.

Methods: This prospective, multicentre observational cohort study included 363 patients with sepsis recruited from five university hospitals in Germany between March 2018 and April 2023. Plasma samples were collected on days 1 and 4 after sepsis diagnosis, and proteome analysis was performed using mass spectrometry. Classical statistical methods and machine learning (random forest) were employed to identify proteins associated with 30-day survival outcomes.

Findings: Out of 363 patients, 224 (62%) survived, and 139 (38%) did not survive the 30-day period. Proteomic analysis revealed significant differences in 87 proteins on day 1 and 95 proteins on day 4 between survivors and non-survivors. Additionally, 63 proteins were differentially regulated between day 1 and day 4 in the two groups. The identified protein networks were primarily related to blood coagulation, immune response, and complement activation. The random forest classifier achieved an area under the receiver operating characteristic curve of 0.75 for predicting 30-day survival. The results were compared and partially validated with an external sepsis cohort.

Interpretation: This study describes temporal changes in the plasma proteome associated with mortality in sepsis. These findings offer new insights into sepsis pathophysiology, emphasizing the innate immune system as an underexplored network, and may inform the development of targeted therapeutic strategies.

Funding: European Regional Development Fund of the European Union. The State of North Rhine-Westphalia, Germany.

Keywords: Feature importance; Machine learning; Mass spectrometry; Proteomics; Sepsis.

Fig. 1

Proteome analysis according to 30-day survival. (a) Volcano plots representing the statistical analysis of survived and deceased patients for day 1 and day 4, respectively (Death/Survival). Coloured proteins passed the significance threshold of p_FDR value ≤ 0.05 (t-test, Benjamini-Hochberg corrected). (b) Volcano plot illustrating the statistical analysis of survived and deceased patients based on day 4/day 1 ratios. Differences calculated as mean ratio death – mean ratio survival. Coloured proteins passed the significance threshold of p_FDR value ≤ 0.05 (t-test, Benjamini-Hochberg corrected). (c) Venn diagram illustrating significant proteins for day 1, day 4 and the day 4/day 1 ratios. The 14 proteins in the intersect of all comparisons are labelled with gene names in figure parts a and b. (d) Density plots representing distribution of day 4/day 1 ratios for survived and deceased patients. The variance within both patient groups was found significantly different (Levene test, p < 0.0001). (e) Linear regression of the standard deviation of day 4/day 1 ratios with the SOFA score. Blue line representing the linear fit with its confidence interval. (f) Linear regression analysis of Myoglobin (MB) with the SOFA score separately for days 1 and 4. Blue and red data points represent survived and deceased patients, respectively.

Fig. 2

Machine Learning and Functional Protein Analysis. (a) Heatmap illustrating the 25 most frequently selected ML features. Three independent Random Forest classifiers were trained 100 times each and the selected features were interpreted using Shapley Additive Explanations (SHAP). Colour representing the median importance rank (b) Functional enrichment analysis using Gene Ontology (GO) Biological Processes. Ten selected categories illustrated for analysis using all selected ML features as well as the significant proteins for day 1, day 4 and day 4/day 1 ratios. (c) Protein interaction network illustrating major groups of proteins associated with 30-day survival. All differentially abundant proteins were analysed using STRING and significantly enriched biological processes were highlighted. Only interactions with highest confidence displayed, disconnected nodes not shown, highlighted proteins labelled with gene names.

Fig. 3

Correlation of protein intensities with clinical parameters. Heatmap illustrating the linear regression analysis of baseline characteristics with protein intensities at day 1. All proteins found to be significantly differentially abundant in a univariate analysis were considered for analysis. Colours represent the R² of the linear models for each pair. Models that did not pass a significance threshold of p < 0.05 displayed in grey; hierarchical clustering using Pearson correlation and complete linkage.

Fig. 4

Comparison and validation using an independent sepsis cohort. Heatmap illustrating the comparison of this study with the study of Mi et al. Colour coded ratios of mean intensities (RoM, calculated Death/Survival) are shown for both studies. Analysis according to 30-day survival was done for days 1 and 4 (this study) and days 1, 3 and 5 (Mi et al.). Column titles indicate the respective comparisons. All displayed genes passed an p_FDR value ≤ 0.05 threshold in both studies in at least one of the three comparisons; all displayed RoM had a significant corresponding p_FDR-value.