Cord blood proteomics identifies biomarkers of early-onset neonatal sepsis

Abstract

BACKGROUNDSymptoms of early-onset neonatal sepsis (EOS) in preterm infants are nonspecific and overlap with normal postnatal physiological adaptations and noninfectious pathologies. This clinical uncertainty and the lack of reliable EOS diagnostics results in liberal use of antibiotics in the first days to weeks of life, leading to increased risk of antibiotic-related morbidities in infants who do not have an invasive infection. METHODSTo identify potential biomarkers for EOS in newborn infants, we used unlabeled tandem mass spectrometry proteomics to identify differentially abundant proteins in the umbilical cord blood of infants with and without culture-confirmed EOS. Proteins were then confirmed using immunoassay, and logistic regression and random forest models were built, including both biomarker concentration and clinical variables to predict EOS. RESULTSThese data identified 5 proteins that were significantly upregulated in infants with EOS, 3 of which (serum amyloid A, C-reactive protein, and lipopolysaccharide-binding protein) were confirmed using a quantitative immunoassay. The random forest classifier for EOS was applied to a cohort of infants with culture-negative presumed sepsis. Most infants with presumed sepsis were classified as resembling infants in the control group, with low EOS biomarker concentrations.CONCLUSIONThese results suggest that cord blood biomarker screening may be useful for early stratification of EOS risk among neonates, improving targeted, evidence-based use of antibiotics early in life. FUNDINGNIH, Gerber Foundation, Friends of Prentice, Thrasher Research Fund, Ann & Robert H. Lurie Children's Hospital, and Stanley Manne Children's Research Institute of Lurie Children's.

Keywords: Bacterial infections; Biomarkers; Immunology; Infectious disease; Proteomics.

Figure 1. Proteomic identification of differentially abundant proteins in early-onset sepsis cord blood.

(A) Diagram of workflow. (B) Hierarchically clustered heatmap of protein abundance in cord blood for EOS and control specimens. Specimens are clustered by gestational age category, sex, and sample type. Missing values were imputed. (C) Plot of mean abundance of proteins in EOS and control specimens. Black points were significant by Mann-Whitney U test with Benjamini-Hochberg FDR adjustment (P < 0.05).

Figure 2. Details of differentially abundant proteins.

(A) PCA and (B) clustered heatmap of EOS and control specimen values for the 5 differentially abundant proteins. For B, missing protein abundance values were imputed. (C) Distribution of protein abundance in EOS and control specimens for each protein. Box plots show median and interquartile range. Whiskers extend to the last point within 1.5× interquartile range of the box. Bee swarms show individual samples. Comparisons were significant by Mann-Whitney U test with Benjamini-Hochberg FDR adjustment.

Figure 3. Quantitative multiplex immunoassay detection of potential EOS biomarkers.

(A) Diagram of experimental procedure. (B) Distribution of protein concentration in mg/mL in EOS and control specimens for each protein. Box plots show median and interquartile range. Whiskers extend to the last point within 1.5× interquartile range of the box. Bee swarms show individual samples. Comparisons were significant by Mann-Whitney U test with Benjamini-Hochberg FDR adjustment. (C) PCA of EOS and control specimen protein abundance based on MSD data.

Figure 4. Modeling of EOS risk using biomarkers.

(A) Model fit parameters for random forests models trained with (amber) or without (black) cord blood biomarker concentrations as a factor. Metrics are calculated with EOS as the positive class. Points represent performance for a single run of the model. Box plots show the median and interquartile range. Whiskers extend to the last point within the 1.5× interquartile range of the box. (B) Permutation variable importance for variables in the random forest model with cord blood biomarker concentrations included. Biomarker concentrations are shown on the left. Variables included in both models are shown on the right. Box plots show median and interquartile range. Whiskers extend to the last point within 1.5× interquartile range of the box. Points represent performance for a single run of the model.

Figure 5. Presumed sepsis cases categorized by random forest model.

Biomarker concentrations in cord blood from presumed sepsis cases were measured by immunoassay; cases were then categorized as either predicted EOS (PS-pEOS) or predicted control (PS-pControl). (A) PCA plot of biomarker concentrations in cases colored by status and predicted status. (B) Concentration of biomarker proteins in predicted EOS and predicted control cases. Box plots show median and interquartile range. Whiskers extend to the last point within the 1.5× interquartile range of the box. Points represent individual samples. Wide red lines show the median value for ascertained EOS cases (excluding outliers); wide blue lines show the median value for ascertained controls. Comparisons were significant by Mann-Whitney U test with Benjamini-Hochberg FDR adjustment.