A transcriptional signature accurately identifies Aspergillus Infection across healthy and immunosuppressed states

Abstract

Invasive aspergillosis (IA) is a major cause of critical illness in immunocompromised (IC) patients. However, current fungal tests are limited. Disease-specific gene expression patterns in circulating host cells show promise as novel diagnostics, however it is unknown whether such a 'signature' exists for IA and the effect of iatrogenic immunosuppression on any such biomarkers. Male BALB/c mice were separated into 6 experimental groups based on Aspergillus fumigatus inhalational exposure and IC status (no immunosuppression, cyclophosphamide, and corticosteroids). Mice were sacrificed 4 days postinfection. Whole blood was assayed for transcriptomic responses in peripheral white blood cells via microarray. An elastic net regularized logistic regression was employed to develop classifiers of IA based on gene expression. Aspergillus infection triggers a powerful response in non-IC hosts with 2718 genes differentially expressed between IA and controls. We generated a 146-gene classifier able to discriminate between non-IC infected and uninfected mice with an AUC of 1. However, immunosuppressive medications exhibited a confounding effect on this transcriptomic classifier. After controlling for the genomic effects of immunosuppression, we were able to generate a 187-gene classifier with an AUC of 0.92 in the absence of immunosuppression, 1 with cyclophosphamide, and 0.9 with steroids. The host transcriptomic response to IA is robust and conserved. Pharmacologic perturbation of the host immune response has powerful effects on classifier performance and must be considered when developing such novel diagnostics. When appropriately designed, host-derived peripheral blood transcriptomic responses demonstrate the ability to accurately diagnose Aspergillus infection, even in the presence of immunosuppression.

Figure 1.. Experimental Design.

Breakdown of animals by inhalational Aspergillus exposure (present or absent) and type of immunosuppression (none, cyclophosphamide, and corticosteroids).

Figure 2.. Clinical outcomes of mice in the setting of inhalational Aspergillus fumigatus exposure and immunosuppression with cyclophosphamide and corticosteroids.

a. Mean mouse weights in grams at day +4 after inhalational Aspergillus exposure based on experimental group (no Aspergillus/no immunosuppression, no Aspergillus/corticosteroids, no Aspergillus/cyclophosphamide, Aspergillus/no immunosuppression, Aspergillus/corticosteroids, Aspergillus/cyclophosphamide). b. Mean mouse weights in grams at day +4 comparing those with inhalational Aspergillus exposure and those without. c. Mean mouse weights in grams at day +4 comparing those mice who received immunosuppression to those who did not. d. Mean mouse lung fungal burden, measured in colony-forming units/gram (CFU/g) at time of day+4 after inhalational Aspergillus infection comparing those with no suppression, cyclophosphamide immunosuppression, and corticosteroid immunosuppression. Error bars represent standard deviation. AF = Aspergillus fumigatus, IS = immunosuppression, Cyclo = cyclophosphamide, CA = corticosteroids.

Figure 3.. Control Classifier (Classifier 1) Performance and Robustness.

a. Full model predictions for control (no drug) data; all samples were used in model training. The predictions represent a best-case scenario as the entire control data set was used to train the model. b. Leave-one-out cross-validation results of the control data set showing model robustness and expected outcome if the model was run on a validation cohort. c. Heatmap of the mean expression value of genes included in the full model. For probes without an annotated gene, the Probe ID is indicated in parentheses. Asterisks (*) indicate that the probe has been annotated for more than one gene; the first gene symbol in the list has been provided. Combined infection status (NoInfection/Infection) and immunosuppressive therapy category (NoDrug/Cyclo/Steroids) is indicated by color.

Figure 4.. Complete Classifier (Classifier 2) Performance and Robustness.

a. Full model predictions for each drug therapy; all samples were used in model training. All predictions represent a best-case scenario as the same data was used in training and testing the model. b. Ten-fold cross-validation results showing model robustness and expected outcome if the model was run on a validation cohort. c. Heatmap showing the z-score transformed normalized expression value of genes included in the full model. Clustering of genes was done using a correlation distance with complete linkage. The column on the left of the heatmap indicates the number of times a particular probe was included during our cross-validation analysis. For probes without an annotated gene, the Probe ID is indicated in parentheses. For cases where different probes correspond to the same gene, the Probe ID is also included in parentheses. Asterisks (*) indicate that the probe has been annotated for more than one gene; the first gene symbol in the list has been provided. Combined infection status (NoInfection/Infection) and immunosuppressive therapy category (NoDrug /Cyclo/Steroids) is indicated by color. d. ROC curves of cross-validation performance of the model in different immunosuppressive settings, with an AUC of 0.92 in the setting of no immunosuppression, 1 for cyclophosphamide exposure, and 0.9 for corticosteroid exposure.