Small airway brush gene expression predicts chronic lung allograft dysfunction and mortality

Abstract

Background: Chronic lung allograft dysfunction (CLAD) limits survival following lung transplant, but substantial lung damage occurs before diagnosis by traditional methods. We hypothesized that small airway gene expression patterns could identify CLAD risk before spirometric diagnosis and predict subsequent graft failure.

Methods: Candidate genes from 4 rejection-associated transcript sets were assessed for associations with CLAD or graft failure in a derivation cohort of 156 small airway brushes from 45 CLAD cases and 37 time-matched controls with >1-year stable lung function. Candidate genes not associated with CLAD and time to graft failure were excluded, yielding the Airway Inflammation 2 (AI2) gene set. Area under the receiver operating curve (AUC) for CLAD and competing risks of death or graft failure were assessed in an independent validation cohort of 37 CLAD cases and 37 controls.

Results: Thirty-two candidate genes were associated with CLAD and graft failure, comprising the AI2 score, which clustered into 3 subcomponents. The AI2 score identified CLAD before its onset, in early and late post-CLAD brushes, as well as in the validation cohort (AUC 0.69-0.88). The AI2 score association with CLAD was independent of positive microbiology, CLAD stage, or CLAD subtype. However, transcripts most associated with CLAD evolved over time from CLAD onset. The AI2 score predicted time to graft failure and retransplant-free survival in both cohorts (p ≤ 0.03).

Conclusions: This airway inflammation gene score is associated with CLAD development, graft failure, and death. Future studies defining the molecular heterogeneity of airway inflammation could lead to endotype-targeted therapies.

Keywords: biomarker; chronic lung allograft dysfunction; gene expression; lung transplant; molecular diagnostic; transcriptome.

Figure 1:. Study enrollment flow (CONSORT) diagram.

Figure 2:. Candidate gene scores distinguish participants with CLAD from control participants in airway brushes before and after CLAD onset.

(A) FEV₁ as a smoothed function of time from airway brush is shown for the before, early-CLAD, and late-CLAD cohorts, with cases in red and controls in blue. (B) Metagene scores were calculated for 4 candidate gene scores in CLAD cases and controls for the three time groups. Unadjusted Mann–Whitney test p-values are shown.

Figure 3:. Derivation of the Airway Inflammation 2 (AI2) metagene set.

(A) The union of the Common Rejection Module, UCSF Lymphocytic Bronchitis, the Renal Rejection versus Everything Else, and Pitt CLAD score yielded 81 candidate genes. Each gene was assessed for association with CLAD and time to graft failure, and 49 genes were excluded because of a P-value >0.05 in at least one model. (B) A metagene score, or sum of normalized gene counts, from the 32 AI2 genes distinguished CLAD from controls in all three timepoints from the derivation cohort. P-values comparing AI2 scores between cases and controls (first row) were determined by Mann–Whitney test. Score stability over time was assessed across timepoints, with lines corresponding to individual participants and Pearson correlation coefficients (r) shown. Parentheses indicate 95% confidence intervals. (C) AI2 metagene scores are shown in cases and controls stratified by the presence of viral or bacterial pathogens on clinical BAL microbial testing. (D) The AI2 metagene score is shown stratified by CLAD stage at the time of brush in cases and control brushes from the derivation cohort. CLAD stage 0 cases correspond to the before-CLAD group, whereas stages 1, 2, and 3 derive from the early and late CLAD groups. (E) AI2 scores are shown for airway brushes stratified by CLAD type: obstructive (BOS), mixed, or restrictive (RAS). Differences between groups in C–E were determined by GEE-adjusted regression to account for repeat brushes on the same participant. * denotes P<0.05 and *** denotes P<0.001.

Figure 4:. AI2 subcomponent analysis.

A) Hierarchical clustering of cross-correlation within the 32 AI2 genes identified 3 separate clusters of co-expressed genes, which were identified as associated with antigen presentation, cytokine signaling, and lymphocyte activation by Gene Ontology enrichment analysis. (B) A metagene score for each cluster was compared in cases and controls in before CLAD, early CLAD, and late CLAD timepoints of the derivation cohort. Unadjusted Mann–Whitney test p-values are shown.

Figure 5:. Changes in classifier genes across the three timepoints.

(A) Receiver operating curves (ROC) are shown for leave one out cross-validated logistic lasso regression models derived from all observed genes in the before, early, and late CLAD timepoints of the derivation cohort. The area under the ROC curve (AUC) and 95% confidence intervals are shown. (B) Differential gene expression statistics for each gene were compared between the before CLAD (x-axis) and late CLAD (y-axis) cohorts. Genes were grouped based on likely cell of origin (18), with 95% confidence intervals shown by ellipse, change in mean value by cell type assignment indicated by arrows, and significant genes are shown as points and selectively labeled.

Figure 6:. Associations with CLAD, subsequent graft failure, and survival in derivation and validation cohorts.

(A) Receiver Operating Curve (ROC) analysis for the association of AI2 scores with CLAD case versus control assignments both in (A) UCSF derivation and (B) Pitt validation cohorts. Area under ROC curve (AUC) for before CLAD was 0.80 (95% CI 0.68–0.92), early CLAD AUC was 0.88 (95% CI 0.79–0.97), and late CLAD AUC was 0.69 (95% CI 0.54–0.84). AUC for CLAD was 0.79 (95% CI 0.68–0.89) in the validation cohort. (C–D) Time to event analyses for graft failure, defined as death with CLAD or retransplantation, as a function of years from airway brush, stratified by tertile of AI2 score are shown for the (C) derivation cohort and (D) validation cohort. (E–F) Kaplan-Meier curves for retransplant-free survival versus time from airway brush stratified by tertile of AI2 score in (E) derivation cohort and (F) validation cohort. Confidence intervals in A–B reflect DeLong method. P-values in C–F are based on log-rank test.