Robust classification of bacterial and viral infections via integrated host gene expression diagnostics

Abstract

Improved diagnostics for acute infections could decrease morbidity and mortality by increasing early antibiotics for patients with bacterial infections and reducing unnecessary antibiotics for patients without bacterial infections. Several groups have used gene expression microarrays to build classifiers for acute infections, but these have been hampered by the size of the gene sets, use of overfit models, or lack of independent validation. We used multicohort analysis to derive a set of seven genes for robust discrimination of bacterial and viral infections, which we then validated in 30 independent cohorts. We next used our previously published 11-gene Sepsis MetaScore together with the new bacterial/viral classifier to build an integrated antibiotics decision model. In a pooled analysis of 1057 samples from 20 cohorts (excluding infants), the integrated antibiotics decision model had a sensitivity and specificity for bacterial infections of 94.0 and 59.8%, respectively (negative likelihood ratio, 0.10). Prospective clinical validation will be needed before these findings are implemented for patient care.

Conflict of interest statement

Competing interests: The seven-gene set and the IADM have been disclosed for possible patent protection to the Stanford Office of Technology and Licensing by T.E.S. and P.K. T.E.S. also serves as a scientific advisor to Multerra Bio, which had no role in this manuscript. H.R.W. declares that he has no competing interests.

Copyright © 2016, American Association for the Advancement of Science.

Figures

Fig. 1. Summary ROC curves for discovery and direct validation data sets for the bacterial/viral meta-score

Summary ROC curve is shown in black, with 95% CIs in dark gray.

Fig. 2. Bacterial/viral score in COCONUT-conormalized whole-blood validation data sets

The global AUC across all whole-blood discovery data sets is 0.93. Top: Score distribution by data set (blue, bacterial; red, viral). Middle: Individual gene expression (exp.). Bottom: Housekeeping genes (grayscale). The dotted line at the top shows a possible global threshold for discriminating infection type.

Fig. 3. IADM across COCONUT-conormalized public gene expression data that matched inclusion criteria

(A) IADM schematic. (B) Distribution of scores and cutoffs for IADM in COCONUT-conormalized data. SIRS, systemic inflammatory response syndrome. (C) Confusion matrix for diagnosis. Bacterial infection sensitivity, 94.0%; bacterial infection specificity, 59.8%; viral infection sensitivity, 53.0%; viral infection specificity, 90.6%.

Fig. 4. Targeted NanoString gene expression data for children with SIRS/sepsis from the GPSSSI cohort never tested with microarrays

Total n = 96, of which SIRS = 36, bacterial sepsis = 49, and viral sepsis = 11. (A) Breakdown of infected patients by organism type. (B and C) ROC curves for the SMS and the bacterial/viral metascore. (D) Distribution of scores and cutoffs for IADM. (E) Confusion matrix for IADM. Bacterial infection sensitivity, 89.7%; bacterial infection specificity, 70.0%; viral infection sensitivity, 54.5%; viral infection specificity, 96.5%.