A multi-platform approach to identify a blood-based host protein signature for distinguishing between bacterial and viral infections in febrile children (PERFORM): a multi-cohort machine learning study

BACKGROUND

Differentiating between self-resolving viral infections and bacterial infections in children who are febrile is a common challenge, causing difficulties in identifying which individuals require antibiotics. Studying the host response to infection can provide useful insights and can lead to the identification of biomarkers of infection with diagnostic potential. This study aimed to identify host protein biomarkers for future development into an accurate, rapid point-of-care test that can distinguish between bacterial and viral infections, by recruiting children presenting to health-care settings with fever or a history of fever in the previous 72 h.

METHODS

In this multi-cohort machine learning study, patient data were taken from EUCLIDS, the Swiss Pediatric Sepsis study, the GENDRES study, and the PERFORM study, which were all based in Europe. We generated three high-dimensional proteomic datasets (SomaScan and two via liquid chromatography tandem mass spectrometry, referred to as MS-A and MS-B) using targeted and untargeted platforms (SomaScan and liquid chromatography mass spectrometry). Protein biomarkers were then shortlisted using differential abundance analysis, feature selection using forward selection-partial least squares (FS-PLS; 100 iterations), along with a literature search. Identified proteins were tested with Luminex and ELISA and iterative FS-PLS was done again (25 iterations) on the Luminex results alone, and the Luminex and ELISA results together. A sparse protein signature for distinguishing between bacterial and viral infections was identified from the selected proteins. The performance of this signature was finally tested using Luminex assays and by calculating disease risk scores.

FINDINGS

376 children provided serum or plasma samples for use in the discovery of protein biomarkers. 79 serum samples were collected for the generation of the SomaScan dataset, 147 plasma samples for the MS-A dataset, and 150 plasma samples for the MS-B dataset. Differential abundance analysis, and the first round of feature selection using FS-PLS identified 35 protein biomarker candidates, of which 13 had commercial ELISA or Luminex tests available. 16 proteins with ELISA or Luminex tests available were identified by literature review. Further evaluation via Luminex and ELISA and the second round of feature selection using FS-PLS revealed a six-protein signature: three of the included proteins are elevated in bacterial infections (SELE, NGAL, and IFN-γ), and three are elevated in viral infections (IL18, NCAM1, and LG3BP). Performance testing of the signature using Luminex assays revealed area under the receiver operating characteristic curve values between 89·4% and 93·6%.

INTERPRETATION

This study has led to the identification of a protein signature that could be ultimately developed into a blood-based point-of-care diagnostic test for rapidly diagnosing bacterial and viral infections in febrile children. Such a test has the potential to greatly improve care of children who are febrile, ensuring that the correct individuals receive antibiotics.

FUNDING

European Union's Horizon 2020 research and innovation programme, the European Union's Seventh Framework Programme (EUCLIDS), Imperial Biomedical Research Centre of the National Institute for Health Research, the Wellcome Trust and Medical Research Foundation, Instituto de Salud Carlos III, Consorcio Centro de Investigación Biomédica en Red de Enfermedades Respiratorias, Grupos de Refeencia Competitiva, Swiss State Secretariat for Education, Research and Innovation.

What is proteomics?

Proteomics is the study of a large number of proteins in biological systems. We aim to introduce the complex field to paediatricians and present some recent examples of applications to paediatric problems. Various approaches have been used to study proteomes. The current mainstay is tandem mass spectrometry of enzymatically digested proteins ('bottom-up proteomics'), and we describe the experimental and computational approach further. Proteomics can offer advantages over transcriptomics by giving direct information about proteins rather than RNA; however, typically data are obtained at lower depth and the confident identification of mass spectra can be challenging. Proteomics frequently complements transcriptomics and other -omics. Used effectively, proteomics offers promise to help answer important clinical and biological questions.

Kawasaki Disease: The Role of Immune Complexes Revisited

Kawasaki disease (KD) is an inflammatory disease in children associated with vasculitis affecting predominantly the coronary arteries and is now the most common cause of acquired heart disease in children in developed countries. The etiology of KD is unknown but epidemiological studies implicate an infectious agent or toxin, which causes disease in genetically predisposed individuals. The presence of immune complexes (ICs) in the serum of children with KD was established in numerous studies during the 1970s and 80s. More recent genetic studies have identified variation in Fcγ receptors and genes controlling immunoglobulin production associated with KD. In this review we link the genetic findings and IC studies and suggest a key role for their interaction in pathophysiology of the disease.

Characterisation of immune complexes in Kawasaki Disease and other infectious diseases by protein sequencing

Background

Kawasaki Disease is a paediatric inflammatory disease associated with self-limiting vasculitis. It has a world-wide distribution with an ethnic bias towards East Asian populations and is the most common cause of acquired heart disease in children in developed countries. The immunopathogenesis of KD remains poorly understood. The presence of antibody-antigens (pathogen) known as immune complexes in the blood of children with KD was established in numerous studies leading to the hypothesis that immune complexes contribute to the damages of the coronary artery.

Methodology

Using proteomic technologies we characterised the composition of immune complexes in KD and compare the immune complexes in KD with those in children with other febrile illnesses, TB and healthy children. Immune complexes were precipitated from the blood of 80 children with KD, 80 children with other febrile conditions, 30 with TB and 30 healthy children. We used Lumos Orbitrap mass spectrometry to identify the recovered proteins. Using the bioinformatics program PEAKS, we performed database searches and compared the protein abundances between the different comparator groups.

Conclusions

Immune complexes isolated from children with KD are different from those recovered from other febrile illnesses in terms of the nature of proteins within the complex. The pattern of proteins in the immune complexes containing immunoglobulins, complement proteins as well as other serum proteins, provides insight into the nature of the unique inflammatory response in KD versus the other inflammatory diseases.

Diagnosis of Kawasaki Disease Using a Minimal Whole-Blood Gene Expression Signature

IMPORTANCE

To date, there is no diagnostic test for Kawasaki disease (KD). Diagnosis is based on clinical features shared with other febrile conditions, frequently resulting in delayed or missed treatment and an increased risk of coronary artery aneurysms.

OBJECTIVE

To identify a whole-blood gene expression signature that distinguishes children with KD in the first week of illness from other febrile conditions.

DESIGN, SETTING, AND PARTICIPANTS

The case-control study comprised a discovery group that included a training and test set and a validation group of children with KD or comparator febrile illness. The setting was pediatric centers in the United Kingdom, Spain, the Netherlands, and the United States. The training and test discovery group comprised 404 children with infectious and inflammatory conditions (78 KD, 84 other inflammatory diseases, and 242 bacterial or viral infections) and 55 healthy controls. The independent validation group comprised 102 patients with KD, including 72 in the first 7 days of illness, and 130 febrile controls. The study dates were March 1, 2009, to November 14, 2013, and data analysis took place from January 1, 2015, to December 31, 2017.

MAIN OUTCOMES AND MEASURES

Whole-blood gene expression was evaluated using microarrays, and minimal transcript sets distinguishing KD were identified using a novel variable selection method (parallel regularized regression model search). The ability of transcript signatures (implemented as disease risk scores) to discriminate KD cases from controls was assessed by area under the curve (AUC), sensitivity, and specificity at the optimal cut point according to the Youden index.

RESULTS

Among 404 patients in the discovery set, there were 78 with KD (median age, 27 months; 55.1% male) and 326 febrile controls (median age, 37 months; 56.4% male). Among 202 patients in the validation set, there were 72 with KD (median age, 34 months; 62.5% male) and 130 febrile controls (median age, 17 months; 56.9% male). A 13-transcript signature identified in the discovery training set distinguished KD from other infectious and inflammatory conditions in the discovery test set, with AUC of 96.2% (95% CI, 92.5%-99.9%), sensitivity of 81.7% (95% CI, 60.0%-94.8%), and specificity of 92.1% (95% CI, 84.0%-97.0%). In the validation set, the signature distinguished KD from febrile controls, with AUC of 94.6% (95% CI, 91.3%-98.0%), sensitivity of 85.9% (95% CI, 76.8%-92.6%), and specificity of 89.1% (95% CI, 83.0%-93.7%). The signature was applied to clinically defined categories of definite, highly probable, and possible KD, resulting in AUCs of 98.1% (95% CI, 94.5%-100%), 96.3% (95% CI, 93.3%-99.4%), and 70.0% (95% CI, 53.4%-86.6%), respectively, mirroring certainty of clinical diagnosis.

CONCLUSIONS AND RELEVANCE

In this study, a 13-transcript blood gene expression signature distinguished KD from other febrile conditions. Diagnostic accuracy increased with certainty of clinical diagnosis. A test incorporating the 13-transcript disease risk score may enable earlier diagnosis and treatment of KD and reduce inappropriate treatment in those with other diagnoses.