Plasma cell-free RNA signatures of inflammatory syndromes in children

Abstract

Inflammatory syndromes, including those caused by infection, are a major cause of hospital admissions among children and are often misdiagnosed because of a lack of advanced molecular diagnostic tools. In this study, we explored the utility of circulating cell-free RNA (cfRNA) in plasma as an analyte for the differential diagnosis and characterization of pediatric inflammatory syndromes. We profiled cfRNA in 370 plasma samples from pediatric patients with a range of inflammatory conditions, including Kawasaki disease (KD), multisystem inflammatory syndrome in children (MIS-C), viral infections, and bacterial infections. We developed machine learning models based on these cfRNA profiles, which effectively differentiated KD from MIS-C-two conditions presenting with overlapping symptoms-with high performance [test area under the curve = 0.98]. We further extended this methodology into a multiclass machine learning framework that achieved 80% accuracy in distinguishing among KD, MIS-C, viral, and bacterial infections. We further demonstrated that cfRNA profiles can be used to quantify injury to specific tissues and organs, including the liver, heart, endothelium, nervous system, and the upper respiratory tract. Overall, this study identified cfRNA as a versatile analyte for the differential diagnosis and characterization of a wide range of pediatric inflammatory syndromes.

Keywords: cell-free RNA; diagnostics; inflammation; machine learning; pediatrics.

Fig. 1.

Sample overview. (A) Sample counts and distribution of hospital of origin for each disease group. “Other” indicates other hospitalized pediatric controls (B) Age distribution, (C) C-reactive protein (CRP) levels, (D) sex distribution, and (E) ICU status distribution for each sample group.

Fig. 2.

Shared cfRNA signatures among different inflammation syndromes. (A) Differential abundance analysis using DESeq2 was performed between healthy controls and each other group individually. Vertical columns indicate the number of overlapping genes that are significantly differentially abundant between groups and controls (BH adjusted P-values < 0.05). Dots below bars indicate the groups being intersected. Horizontal columns indicate the total number of DAGs between groups and controls. (B) Significantly enriched pathways in the set of genes found to be differentially abundant between healthy controls and each disease group. Average P-value and fold change used for pathway analysis (Qiagen, IPA).

Fig. 3.

cfRNA distinguishes KD and MIS-C. (A) Overview of sample set and modeling scheme. (B) Volcano plot of differentially abundant transcripts between MIS-C and KD. Analysis was performed using the training dataset (DESeq2). (C) Adjusted P-value, base mean, and gene ROC AUC distributions for all significant genes from the training KD vs. MIS-C comparison. The red dashed line represents the threshold cutoff used for filtering prior to model training. The green shaded area indicates genes used. (D) ROC AUC values for the 14 machine learning classification models applied to training and validation sets. (E) ROC curves of the training, validation, and test sets using the GLMNET with LASSO regression algorithm. (F) Violin plots of the classifier scores from the GLMNET with LASSO regression algorithm.

Fig. 4.

Multiclass classification of pediatric disease using cfRNA. (A) Overview of machine learning framework used for multiclass classification. (B) ROC-AUC plots for each one-vs.-one model trained, along with train (Top) and test (Bottom) classifier score distributions. (C) Confusion matrix of reference and predicted diagnoses for train and test samples. Color indicates fraction of samples in each category.

Fig. 5.

cfRNA as a clinical decision support tool. (A) Median scaled cell type of origin fractions for samples separated by liver damage, cardiac function, COVID-19 severity, and endothelial damage, and healthy controls (Materials and Methods). Stars indicate statistically significant differences between groups in comparison (Wilcoxon rank-sum test, BH adjusted P-value < 0.05). (B–E) Case studies of patients with example clinical decision support tool results from the multiclass algorithm, along with measurements of endothelium, heart, liver, lung, and neuronal damage from the deconvolution data. Sample shown in red and healthy donor samples in gray. Z-scores calculated relative to healthy donor sample distributions. Endothelium refers to endothelial cell, heart to cardiac muscle cells, liver to hepatocyte, lung to club cell and type I pneumocyte, and neuronal to Schwann cell cfRNA cell type of origin fractions.