Predictive gene expression signature diagnoses neonatal sepsis before clinical presentation

Abstract

Background: Neonatal sepsis is a deadly disease with non-specific clinical signs, delaying diagnosis and treatment. There remains a need for early biomarkers to facilitate timely intervention. Our objective was to identify neonatal sepsis gene expression biomarkers that could predict sepsis at birth, prior to clinical presentation.

Methods: Among 720 initially healthy full-term neonates in two hospitals (The Gambia, West Africa), we identified 21 newborns who were later hospitalized for sepsis in the first 28 days of life, split into early-onset sepsis (EOS, onset ≤7 days of life) and late-onset sepsis (LOS, onset 8-28 days of life), 12 neonates later hospitalized for localized infection without evidence of systemic involvement, and 33 matched control neonates who remained healthy. RNA-seq was performed on peripheral blood collected at birth when all neonates were healthy and also within the first week of life to identify differentially expressed genes (DEGs). Machine learning methods (sPLS-DA, LASSO) identified genes expressed at birth that predicted onset of neonatal sepsis at a later time.

Findings: Neonates who later developed EOS already had ∼1000 DEGs at birth when compared to control neonates or those who later developed a localized infection or LOS. Based on these DEGs, a 4-gene signature (HSPH1, BORA, NCAPG2, PRIM1) for predicting EOS at birth was developed (training AUC = 0.94, sensitivity = 0.93, specificity = 0.92) and validated in an external cohort (validation AUC = 0.72, sensitivity = 0.83, and specificity = 0.83). Additionally, during the first week of life, EOS disrupted expression of >1800 genes including those influencing immune and metabolic transitions observed in healthy controls.

Interpretation: Despite appearing healthy at birth, neonates who later developed EOS already had distinct whole blood gene expression changes at birth, which enabled the development of a 4-gene predictive signature for EOS. This could facilitate early recognition and treatment of neonatal sepsis, potentially mitigating its long-term sequelae.

Funding: CIHR and NIH/NIAID.

Keywords: Machine learning; Neonatal sepsis; Ontogeny; Predictive biomarkers; Transcriptomics.

Fig. 1

At DOL0, there were substantial gene expression differences between neonates who later developed early-onset sepsis compared to healthy, localized infection, and late-onset sepsis neonates. a: Volcano plots comparing Early-Onset Sepsis (EOS, n = 14) to Healthy (n = 32), Local Infection Only (Loc. Infect., n = 12), and Late-Onset Sepsis (LOS, n = 6) neonates at DOL0. Coloured dots indicate genes that were significant (p-adj <0.05 by Benjamini-Hochberg correction, fold change ≥1.5), and the top 5 most up- and down-regulated genes are labelled. Red dots indicate shared DEGs between the three comparisons. No DEGs were detected in any of the other comparisons between Healthy, Local Infection Only, and LOS, and are not shown. b: Significantly enriched Reactome pathways (padj <0.001 by Bonferroni correction) from up- and down-regulated DEGs at DOL0. The total number of DEGs in each comparison are shown under each label.

Fig. 2

The expression of heat shock genes HSPH1 and DNAJB1 distinguished early onset sepsis neonates at birth. a: Sparse Partial Least-Squares Discriminant Analysis (sPLS-DA) separated all DOL0 samples of neonates who later developed early-onset sepsis (EOS, n = 14), late-onset sepsis (LOS, n = 6), localized infection only (Loc. Infect., n = 12), or stayed healthy (n = 32). b:HSPH1 and DNAJB1 contributed the most in component 1 to identify EOS neonates (Figure S10), which was also reflected by gene expression patterns across healthy, localized infection (Loc. Infect.), EOS, or LOS neonates at DOL0. p-values are from the DESeq2 model, with Benjamini-Hochberg adjusted p-values. ∗ = p < 0.05, ∗∗ = p < 0.01, ∗∗∗ = p < 0.001.

Fig. 3

A 4-gene signature involving HSPH1 distinguished early onset sepsis neonates from other neonates at birth. a: Receiver operating characteristic curve of the cross-validated LASSO 4-gene signature (HSPH1, BORA, NCAPG2, PRIM1) model to discriminate early-onset sepsis (EOS) neonates from healthy, localized infection only (Loc. Infect.), and late-onset sepsis (LOS) neonates, using all DOL0 samples. Maximum sensitivity and specificity are calculated at Youden's Index (indicated by the point on the curve). Monte Carlo cross-validation average training and testing AUCs were 0.98 and 0.78, respectively (Figure S12). b: GSVA enrichment scores of all DOL0 samples using the 4-gene signature. Wilcox ranked-sum test was performed for comparisons. ∗∗ = p < 0.01, ∗∗∗ = p < 0.001, ∗∗∗∗ = p < 0.0001. c: Heatmap of scaled variance-stabilized-transformed counts of the 4-gene signature (top) and GSVA enrichment score (bottom).

Fig. 4

Early onset sepsis is associated with immune and metabolic perturbations to ontogeny across the first week of life. A subset of enriched Reactome pathways using unique or shared DEGs over time between matched EOS (n = 5) and healthy neonates (n = 5), with all enriched pathways in Figure S16. The total numbers of unique/shared DEGs are shown under each label. DEGs are from the DOL7 vs DOL0 comparison, indexed for each patient to account for individual confounders.