Integrated host-microbe plasma metagenomics for sepsis diagnosis in a prospective cohort of critically ill adults

Abstract

We carried out integrated host and pathogen metagenomic RNA and DNA next generation sequencing (mNGS) of whole blood (n = 221) and plasma (n = 138) from critically ill patients following hospital admission. We assigned patients into sepsis groups on the basis of clinical and microbiological criteria. From whole-blood gene expression data, we distinguished patients with sepsis from patients with non-infectious systemic inflammatory conditions using a trained bagged support vector machine (bSVM) classifier (area under the receiver operating characteristic curve (AUC) = 0.81 in the training set; AUC = 0.82 in a held-out validation set). Plasma RNA also yielded a transcriptional signature of sepsis with several genes previously reported as sepsis biomarkers, and a bSVM sepsis diagnostic classifier (AUC = 0.97 training set; AUC = 0.77 validation set). Pathogen detection performance of plasma mNGS varied on the basis of pathogen and site of infection. To improve detection of virus, we developed a secondary transcriptomic classifier (AUC = 0.94 training set; AUC = 0.96 validation set). We combined host and microbial features to develop an integrated sepsis diagnostic model that identified 99% of microbiologically confirmed sepsis cases, and predicted sepsis in 74% of suspected and 89% of indeterminate sepsis cases. In summary, we suggest that integrating host transcriptional profiling and broad-range metagenomic pathogen detection from nucleic acid is a promising tool for sepsis diagnosis.

Fig. 1. Study overview.

a, Study flow diagram. Patients studied were enrolled in the EARLI cohort. Sepsis adjudication performed following hospital discharge was based on ≥2 SIRS criteria plus clinical suspicion of infection and was used to delineate five patient subgroups. Following QC, whole blood was subjected to RNA-seq, and plasma to RNA-seq and DNA-seq. WBC, white blood cell count. b, Analytic approaches. Host transcriptional sepsis diagnostic classifiers were trained and tested on RNA-seq data from whole blood (n = 221) or plasma (n = 110), with a goal of differentiating patients with microbiologically confirmed sepsis (Sepsis^BSI + Sepsis^non-BSI) from those without clinical evidence of infection (No-sepsis). Viral infections were identified via a secondary host transcriptomic classifier. Sepsis pathogens were detected from plasma nucleic acid using mNGS followed by an RBM. Finally, an integrated host + microbe model for sepsis diagnosis was developed and evaluated.

Fig. 2. Host gene expression differentiates patients with sepsis from those with non-infectious critical illnesses.

a, Heat map of top 50 differentially expressed genes from whole-blood transcriptomics comparing patients with microbiologically confirmed sepsis (Sepsis^BSI + Sepsis^non-BSI) versus those without evidence of infection (No-sepsis). The heatmap colour range represents the row Z-score of the normalized gene expression values ranging from +4 (red) to −4 (blue). b, GSEA of the differentially expressed genes demonstrating pathways enriched in patients with sepsis. Source data including enriched genes and pathway P values (hypergeometric test) are provided in Supplementary Data 2a and in the Source Data file. c, ROC curve demonstrating performance of the bSVM classifier for sepsis diagnosis from whole-blood transcriptomics (n = 221). The AUC and s.d. (in parentheses, when applicable) are listed for cross validation (CV) in the training set (red line: average over 10 random splits; red shaded area: ±1 s.d.), the held-out validation set (dashed grey line) and over 10 randomly generated validation sets (solid grey line: average; grey shaded area: ±1 s.d.). d, Plasma RNA-seq expression differences of selected differentially expressed genes previously identified as sepsis biomarkers, with Sepsis patients in maroon (n = 73) and No-sepsis patients in grey (n = 37). Adjusted P values (Benjamini–Hochberg method) from DESeq2 noted above boxplot. Expression data are presented as boxes extending from the 25th to the 75th percentiles, with whiskers extending to the 5th and 95th percentiles, and a central horizontal line at the median. Source data are provided in the Source Data file. e, ROC curve demonstrating performance of the bSVM classifier for sepsis diagnosis from plasma RNA (n = 110). The AUC and s.d. are listed for CV in the training set (red line: average over 10 random splits; red shaded area: average ± 1 s.d.), the held-out validation set (dashed grey line) and over 10 randomly generated validation sets (solid grey line: average; grey shaded area: average ± 1 s.d.). Source data

Fig. 3. Plasma mNGS for detecting sepsis pathogens.

a, Microbial plasma DNA mass differences between sepsis groups. Data are presented with a centre horizontal bar at the median, and error bars representing the interquartile ranges. Pairwise comparisons between groups were performed with a two-sided Mann–Whitney test. Sample sizes are as follows for each group: Sepsis^BSI n = 42, Sepsis^non-BSI n = 31, Sepsis^suspected n = 19, Indeterminate n = 9, No-sepsis n = 37, Control n = 18. Source data and P values for comparisons between samples, including water controls, are provided in Supplementary Data 8 and in the Source Data file. b, Graphical depiction of the RBM for sepsis pathogen detection from two different exemplary cases. The RBM identifies established pathogens with disproportionately high abundance compared with other commensal and environmental microbes in the sample. c, Concordance between plasma DNA mNGS for detecting bacterial pathogens in Sepsis^BSI patients and bacterial bloodstream infections compared to a reference standard of culture. d, Sensitivity of plasma nucleic acid mNGS for detecting pathogens in Sepsis^non-BSI patients with sepsis from non-bloodstream, peripheral sites of infection. LRTI = lower respiratory tract infection; UTI = urinary tract infection; CDI = Clostridium difficile infection. Clinical microbiology and metagenomics data are tabulated in Supplementary Data 9. Source data

Fig. 4. Detection of viral sepsis based on host gene expression.

a, GSEA of differentially expressed genes from whole-blood RNA-seq (n = 129) demonstrating pathways enriched in patients with viral sepsis. The top five most significant pathways by P value (hypergeometric test) are plotted. Source data including enriched genes and pathway P values are provided in Supplementary Data 12a and in the Source Data file. b, GSEA of differentially expressed genes from plasma RNA-seq (n = 73) demonstrating pathways enriched in patients with viral sepsis. All identified pathways are plotted. Source data including enriched genes and pathway P values (hypergeometric test) are provided in Supplementary Data 12b and in the Source Data file. c, ROC curve demonstrating performance of the bSVM classifier for detecting viral sepsis from whole-blood RNA-seq (n = 129). The AUC and s.d. (in parentheses, when applicable) are listed for CV in the training set (red line: average over 10 random splits; red shaded area: ±1 s.d.), the held-out validation set (dashed grey line) and over 10 randomly generated validation sets (solid grey line: average; grey shaded area: ±1 s.d.). d, ROC curve demonstrating performance of the bSVM classifier for detecting viral sepsis from plasma RNA-seq (n = 73). The AUC and s.d. are listed for CV in the training set (red line: average over 10 random splits; red shaded area: ±1 s.d.), the held-out validation set (dashed grey line) and over 10 randomly generated validation sets (solid grey line: average; grey shaded area: ±1 s.d.). Source data

Fig. 5. Integrated host-microbe model for sepsis diagnosis from plasma mNGS.

a–d, Host criteria for positivity can be met by a sepsis transcriptomic classifier probability >0.5 (maroon bars, dotted line). Microbial criteria can be met on the basis of either: (1) detection of a pathogen by mNGS and a sample microbial mass (grey bars) >20 pg (dashed line), or (2) viral transcriptomic classifier probability >0.9 (blue circles, dotted line). Host and microbial metrics are highlighted for patients with sepsis due to bloodstream infections (Sepsis^BSI) (a), peripheral infection (Sepsis^non-BSI) (b), patients with non-infectious critical illness (No-sepsis) (c), patients with suspected sepsis but negative microbiological testing (Sepsis^suspected) (d, left) and patients with indeterminate sepsis status (Indeterm) (d, right). Maroon cross, sepsis-positive based on model; blue circles, virus predicted from plasma RNA secondary viral host classifier; filled blue circles, virus also detected by clinical respiratory viral PCR. Cases with <20 pg microbial mass are indicated by lighter grey shading. Samples with mNGS-detected pathogens have the microbe(s) listed below the sample microbial mass. Raw values for plots and original training/test split assignments are tabulated in Supplementary Data 16 and provided in the Source Data file. Source data

Extended Data Fig. 1. Plasma host gene expression differentiates patients with sepsis from those with non-infectious critical illnesses.

Plasma host gene expression differentiates patients with sepsis from those with non-infectious critical illnesses. (a) Heatmap of top 50 differentially expressed genes from whole blood transcriptomics comparing patients with microbiologically confirmed sepsis (Sepsis^BSI + Sepsis^non-BSI) versus those without evidence of infection (No-sepsis). (b) Gene set enrichment analysis of the differentially expressed genes. All identified pathways are plotted. Source data including enriched genes and pathway P values (hypergeometric test) are provided in Supplementary Data 2b. Source data

Extended Data Fig. 2. Overlap of significant genes in the differential expression analyses between the Sepsis and No-Sepsis groups for whole blood and plasma samples.

Overlap of significant genes in the differential expression analyses between the Sepsis and No-Sepsis groups for whole blood and plasma samples. Scatter plot of -log10(adjusted p-value) for individual genes from the differential expression analyses comparing patients with microbiologically confirmed sepsis (Sepsis^BSI + Sepsis^non-BSI) versus those without evidence of infection (No-sepsis), from whole blood (x-axis) and plasma (y-axis). P-values (two-sided) are derived from DESeq2 and incorporate Benjamini–Hochberg adjustment for multiple testing. Dashed grey lines indicate the threshold of adjusted p-value <0.1. Selected, significant, differentially expressed genes highlighted in blue. Source data