Immune gene correlation networks differentiate both chronic lung allograft dysfunction and survival

Abstract

Chronic lung allograft dysfunction (CLAD) is the major barrier for long-term survival in lung transplant recipients (LTRs). CLAD remains a diagnosis of exclusion with poor responses to therapies. A molecular diagnostic for CLAD is needed to risk-stratify LTRs for prognosis and identify new targets to mitigate CLAD progression. We used weighted gene correlation network analysis on the airway brush-derived airway transcriptome to identify immune pathways and markers relevant to CLAD. Weighted gene correlation network analysis was performed on RNA sequencing from airway brushings of 37 LTRs with CLAD compared with 37 stable LTRs. We analyzed gene coexpression networks (modules) for their biological significance and association with CLAD. Three gene modules were positively correlated with CLAD, its severity, allograft dysfunction, and survival. These enriched components of the acute phase response, type 1 adaptive immunity, and innate immunity, respectively. A fourth module correlated with protection and was inversely correlated with the other modules. We validated our findings by identification of downstream protein and eicosanoid levels in the bronchoalveolar lavage, and an external validation cohort where module expression differentiated LTRs with CLAD and correlated with worse survival. The CLAD airway transcriptome enriches for coexpression networks associated with network modules that correlate with allograft dysfunction and survival.

1. Introduction

Lung transplantation is the final therapeutic option for select patients with end-stage lung disease. Chronic lung allograft dysfunction (CLAD) is the most significant barrier for long-term survival in lung transplant recipients (LTRs), leading to a median 5-year survival of 55%.^1,2 There are 2 clinical phenotypes of CLAD, the predominant bronchiolitis obliterans syndrome (BOS) and the restrictive allograft syndrome (RAS), with the latter portending a worse prognosis with decreased survival.^3,4 CLAD remains primarily a diagnosis of exclusion based on reductions in lung allograft spirometry in the forced expiratory volume in 1 second (FEV₁), in the absence of other etiologies such as acute cellular rejection (ACR) or infection. The immunopathogenesis of CLAD remains incompletely understood and the clinical syndrome is heterogenous⁵ in regard to time of onset and the kinetics of progression; hence, there is a major need to identify molecular pathways and therapeutic targets to mitigate CLAD progression. The development of a molecular diagnostic for CLAD would advance the field in diagnosis, prognosis, and potentially identify targets to mitigate disease.

Prior studies have evaluated the lung allograft transcriptome to identify signatures in ACR and CLAD, using bronchoalveolar lavage (BAL)-derived cells, transbronchial biopsies, and airway brushes.^6–8 Using single-cell RNA sequencing (RNA-seq) analysis, Snyder et al⁸ reported upregulation of CRIP1 and NME2 in CD8⁺ T cell clones that persisted in the BAL months after corticosteroid therapy and histologic resolution of ACR. Parkes et al,⁹ using transbronchial biopsy mRNA signatures found that CLAD-selective transcripts corrected for time reflected tissue injury genes—HIF1A, IFG1, and SERPINE2. We have previously reported an airway brush-derived transcriptomic signature that associated with CLAD and was marked by type 1 immunity and endogenous immune regulation, particularly indoleamine 2, 3-dioxygenase 1 and other genes downstream of interferon (IFN) gamma.¹⁰ Dugger et al¹¹ observed a lymphocytic bronchitis metagene score that predicted CLAD and outperformed transbronchial biopsy signatures. Mohanty et al¹² described a brush-derived airway inflammation 2 score comprised 32 genes that predicted CLAD and graft failure in 2 separate cohorts.

In this study, we hypothesized that gene networks would differentiate CLAD. We used weighted gene correlation network analysis (WGCNA), a statistical method for investigating complex relationships with large gene expression data sets by identifying coexpression networks as they relate to clinical traits.¹³ We performed WGCNA on airway brush transcriptomes from 74 LTRs and observed 4 gene modules significantly correlating with impending CLAD, CLAD stage, and FEV₁ decline. We found key components of these gene modules could be validated at the protein level and that these modules performed well in a validation cohort from a separate institution.

2. Materials and methods

2.1. Study cohort

This was a 2 center, cross-sectional study of airway brush samples to investigate the airway transcriptome in CLAD using RNA-seq analyses. The study was approved by the institutional review board of the University of Pittsburgh (PRO14110014), for the derivation cohort and by University of California, San Francisco institutional review board (13–10738) for the validation cohort. Derivation participants were selected from the Pitt lung transplant registry biorepository (with ongoing enrollment of consented LTRs since 2015 and transbronchial brushes since 2017). Patients are consented separately for the biorepository and for transbronchial brushings. A bronchial cytologic brush sample (ConMed) was collected for each LTR at the time of routine surveillance bronchoscopy (assessed every 3–4 months for the first 2 years posttransplant) or subsequent for-cause biopsy.

Our cohort of patients with transbronchial brushings were screened for patients with CLAD at the time of brushing (n = 37). We also identified an equal number of LTRs without CLAD at time of brushing, stable lung allograft function for at least 1 year prior, and toward the end of their 2-year institutional surveillance period. Each patient only contributed 1 brush to the analyses. LTRs were classified as having CLAD based on the 2019 International Society of Heart and Lung Transplant (ISHLT) guidelines if they had a persistent >20% decline in FEV₁ from posttransplant baseline, absent alternative explanations for this decline, as adjudicated by 2 physicians.¹ Only patients with CLAD consistent with the BOS phenotype (persistent FEV₁ without a corresponding decline in forced vital capacity [FVC]) were included. CLAD staging based on percent decline in posttransplant FEV₁ was also determined using these guidelines.

A validation cohort of LTRs from a separate center (University of California, San Francisco) were assessed using the WGCNA-derived modules identified from the Pittsburgh cohort for significant associations between our identified modules and outcomes (CLAD vs stable control), CLAD severity, and survival and to generate a receiver-operating characteristic (ROC) curve. The validation cohort consisted of small airway brushings from 22 LTRs with CLAD at time of brushing and 37 controls with >1-year stable lung function at time of brushing.¹² As with the derivation cohort, each patient in the validation cohort only contributed a single brushing to the analyses. For additional details, see Supplementary Materials and Methods.

3. Results

To evaluate the airway transcriptome in CLAD, airway brushes from a cross-sectional cohort of 74 LTRs were collected during clinical bronchoscopies, processed, and analyzed, using bulk RNA-seq analyses (Table 1). LTRs with CLAD (n = 37) were adjudicated according to the 2019 ISHLT criteria,¹ along with non-CLAD age-matched controls (n = 37). The CLAD LTRs all had the BOS phenotype, with 24 (64.9%) CLAD stage 1, 5 (13.5%) stage 2, and 8 (21.6%) stage 3. No statistically significant demographic differences between the control and CLAD groups were noted in age at transplant, sex, transplant diagnosis, procedure (single- vs double-lung transplant), or cytomegalovirus serostatus. No differences in the incidence of antidonor-specific antibodies or ACR were found between the 2 groups. There was no concomitant antibody-mediated rejection in either group per ISHLT consensus definition.¹⁴ The CLAD group had a greater proportion of positive bacterial cultures at the time of sampling (P = .01). The median time to CLAD was 1310 days posttransplant. The median time to brush differed between the 2 groups, with control group brushes predominantly obtained late during the institution’s 2-year surveillance period.

Table 1.

Derivation cohort characteristics.

Characteristic	CLAD	Control	P
n	37	37
Age at transplant (y)	55.0 (33.9–61.6)	55.7 (36.5–64.1)	.62
Time to brush posttransplant (d)	1497 (875–2486)	595 (497–701)	<.01
Female	16 (43.2)	18 (48.6)	.82
Transplant diagnosis			.92
CF	8 (21.6)	10 (27.0)
COPD	8 (21.6)	9 (24.3)
IPF	13 (35.1)	13 (35.1)
Redo	4 (10.8)	3 (8.1)
Other	4 (10.8)	2 (5.4)
Double-lung transplant	28 (75.7)	32 (86.5)	.37
Cytomegalovirus serostatus			.64
D−/R−	7 (18.9)	8 (21.6)
D−/R+	7 (18.9)	8 (21.6)
D+/R−	5 (13.5)	8 (21.6)
D+/R+	18 (48.6)	13 (35.1)
Anti-HLA donor-specific antibody	2 (5.4)	7 (18.9)	.15
Antibody-mediated rejection	0 (0)	0 (0)	1
Acute cellular rejection	11 (29.7)	6 (16.2)	.27
Acute infection	20 (54.1)	5 (13.5)	<.01
Bacterial	14 (37.8)	4 (10.8)	.01
Viral	3 (8.1)	1 (2.7)	.61
Fungal	3 (8.1)	1 (2.7)	.61
Nontuberculous mycobacterial	3 (8.1)	1 (2.7)	.61
Days to CLAD	1310 (628–2090)
CLAD stage
CLAD 1	24 (64.9)
CLAD 2	5 (13.5)
CLAD 3	8 (21.6)
CLAD 4	0 (0)
CLAD-directed treatments within 3 mo prior to brush
Total	8 (21.6)
Methylpred	5
Thymoglobulin	2
Basiliximab	1

Values are n (%) or median (IQR).

CF, cystic fibrosis; CLAD, chronic lung allograft dysfunction; COPD, chronic obstructive pulmonary disease; D, donor; HLA, human leukocyte antigen; IPF, idiopathic pulmonary fibrosis; IQR, interquartile range; R, recipient.

3.1. WGCNA identifies 4 modules associated with worsening CLAD severity

We hypothesized that distinct groups of related genes would be differentially expressed in the airway transcriptome of LTRs with CLAD compared with controls. WGCNA of bulk RNA sequencing from 74 LTRs identified 4 modules of coexpressed genes that were significantly correlated (Benjamini-Hochberg–corrected P < .05) with the presence of CLAD (Fig. 1A). Three of these gene modules (red, blue, and turquoise) demonstrated significant positive correlation with increased CLAD stage (worsening CLAD severity) (Fig. 1A). In contrast, a fourth gene module (green) demonstrated significant inverse correlation with CLAD stage, with expression more predominant in the control group (Fig. 1A). Significant differences in module expression were also observed in a pairwise comparison between patients in each CLAD stage (Fig. 1B). No significant differences in module expression were identified between the CLAD LTRs with and without ACR at time of brushing (Supplementary Fig. S1A). Although the CLAD group had a greater proportion of positive bacterial cultures than controls, we did not observe significant differences in module expression in CLAD LTRs with or without infection (Supplementary Fig. S1B).

Figure 1.

Weighted gene correlation network analysis identifies 4 modules associated with worsening chronic lung allograft dysfunction (CLAD) severity. (A) Three gene coexpression networks (modules) were significantly positively correlated with the presence of CLAD and worsening CLAD severity (ie, stage). A fourth was significantly inversely correlated. (B) Module expression by CLAD stage (pairwise significance as measured via Wilcoxon rank sum). (C) Top enriched pathways for each module (via ingenuity pathways analysis). *P < .05; **P < .005; ***P < .0005; P values adjusted for false discovery.

We performed pathways analysis of the genes (Fig. 1C) and assessed top gene membership (Fig. 2) for the respective modules. The blue module demonstrated enrichment for elements of the acute phase response including activation of complement, coagulation, and arachidonic acid pathways (Figs. 1C and 2A). The red module enriched for components of type 1 adaptive immunity, including elements downstream of IFN gamma signaling, antigen presentation, cytotoxic molecules, and type 1 chemokines (Figs. 1C and 2B). The turquoise module enriched for pathways associated with innate immunity including interleukin (IL)-8 (the major chemoattractant for neutrophils), tumor necrosis factor (TNF)α, IL-1β, IL-18 receptor signaling, macrophage activation, and matrix metalloproteinases associated with tissue remodeling (Fig. 1C and 2C). Genes in the green module demonstrated enrichment for a broader variety of pathways that, overall, appear to relate to normal cellular homeostasis with molecules associated with regulation of inflammation-(WNT2b), protection from oxidative molecules (SCARA3), and normal physiology (CADH7 and BCAM) (Figs. 1C and 2D).

Figure 2.

Top module genes by membership and their significance for correlation with the presence of chronic lung allograft dysfunction (CLAD). (A-D) Module gene significance for CLAD vs module membership for the blue (acute phase response) module (A), red (type 1 adaptive immunity) module (B), turquoise (innate immunity) module (C) and green (non-CLAD) module (D).

3.2. Module correlation with CLAD severity is time independent

Recent work has suggested that gene expression changes associated with CLAD, might be more accurately associated with CLAD after adjusting for time since transplant (time-associated vs CLAD-associated transcripts).⁹ Allograft surveillance protocols at our institution include routine bronchoscopies for 2 year posttransplant. Control LTRs primarily had brush samples obtained during this window, whereas most patients with CLAD had brushes obtained during for-cause bronchoscopies significantly later posttransplant (Fig. 3A). However, when patients were stratified by CLAD stage, these differences were less pronounced for patients with CLAD stage 1 compared with advanced CLAD (stage 2 or 3). We observed patients in all CLAD stages significantly overlapping, having been identified both early and late posttransplant (Fig. 3B). Together, this variable distribution of CLAD stages with respect to time underscores the heterogeneity of CLAD kinetics. Stratified univariate analyses did not identify a significant relationship between time to brush and module expression in any subgroup, except for the blue module in the control subgroup (Fig. 3C, D). Our analyses support that module expression association with CLAD and CLAD severity is time independent from when the brush was obtained.

Figure 3.

Module correlation with chronic lung allograft dysfunction (CLAD) severity is independent of time since transplantation. (A, B) Posttransplant time point when the airway brush was obtained (ie, time to brush) by group (A) and CLAD stage (B) (CLAD stage 0 = control). (C, D) Pearson correlation between module expression and time to brush by group (C) and CLAD stage (D).

3.3. Module expression is associated with active and impending allograft dysfunction

Because the diagnosis of CLAD is based predominantly on spirometric parameters (FEV₁ and FVC), we next determined whether module gene expression correlated with allograft function during periods preceding (12 months prior) or surrounding brush sample procurement (6 months prior to or after brush). Allograft function was calculated for each LTR as the percent change over time from the maximum posttransplant spirometry parameter (FEV₁, FVC in liters, FEV₁/FVC, FEV₁ percent predicted, and FVC percent predicted). A rate of change was then calculated for each spirometric measurement as the slope of the least squares linear regression for this trend over the 12-month period, and this value was then used in downstream analyses.

The rate of change for all spirometric measurements demonstrated statistically significant correlation with gene expression for 2 or more modules except for the FEV₁/FVC ratio (Fig. 4A). Highly significant associations with module expression were observed for the FEV₁ and FEV₁ percent predicted rates of change spanning the 6 months before or after the brush (Fig. 4A). After adjusting the period of analysis to the 12 months preceding brush, we observed that these associations with FEV₁ either weakened or disappeared for most modules, suggesting that airway gene expression in the blue, red and turquoise modules were associated with either ongoing or developing, but not antecedent, allograft dysfunction (Fig. 4B). We next investigated module expression within individual LTRs and found that upregulation of the blue, red and turquoise modules occurred in tandem with downregulation of the green module, in conjunction with declining FEV₁ slope (from peak baseline) during the 6-month presample/postsample window (Fig. 4C).

Figure 4.

Module expression is associated with active and impending allograft dysfunction. (A) Heatmap of correlation and statistical significance between module expression and percent decline from posttransplant maximum of FEV₁, FVC, FEV₁/FVC ratio, FEV₁ percent predicted, and FVC percent predicted during a time of 6 months before and after airway brushing. (B) Heatmap of correlation and statistical significance between module expression and percent decline from posttransplant maximum of FEV₁ for two 12-month periods: 6 months before and after airway brushing and 12 months preceding airway brushing. (C) Representative plots of FEV₁ trajectories (plotted as percentage of posttransplant baseline, ie, maximum) alongside module expression for patients in each quartile of FEV1 rate of change. (D) Heatmaps of mean scaled expression of module hub genes by group, chronic lung allograft dysfunction (CLAD) stage, and quartile of FEV₁ decline. FEV₁, forced expiratory volume in 1 second; FVC, forced vital capacity.

We assessed gene expression for the top hub genes within each module and demonstrated similar patterns of coordinated and enhanced gene expression with the diagnosis of CLAD, worsening CLAD stage, and worsening FEV₁ decline (Fig. 4D). These findings point to an intensification of coordinated module expression (blue, red and turquoise) with progressive CLAD stage and allograft dysfunction, countered by the loss of the green module. A similar trend was observed for module expression at the patient level based on CLAD stage (Supplementary Fig. S2). Our findings indicate that differential module expression correlates not only with CLAD stage but also changes in the FEV₁ in the time window 6 months prior to, and after, brush samples.

3.4. Upstream type 1 immunity regulates global gene module expression

We used the ingenuity pathway analysis platform to determine the major upstream mediators driving the blue (acute phase response), turquoise (innate immunity), and red (type 1 immunity) modules positively correlated with CLAD and FEV₁ decline (Fig. 5A). The hallmark type 1 cytokine, IFNg, along with TNFa, had the highest activation Z-scores and resided at the top of the hierarchy driving downstream expression for these modules (Fig. 5A, B). Importantly, key signaling molecules downstream of these 2 cytokines, STAT1 (IFNg) and STAT3 (TNFa) were also found to be significant upstream regulators (Fig. 5B). Other major proinflammatory cytokines, IL1B and IL6 demonstrated high activation Z-scores for upstream regulation (Fig. 5A, B). Both IL1B and TNFa activate NFKB1, a master regulator of inflammatory gene expression, with the latter then regulating the production of pro-IL1B.

Figure 5.

Upstream type 1 immune activation regulates global gene module expression. (A) Representative genes from each module and their top predicted upstream regulators (in yellow). (B) Ingenuity pathways analysis activation Z-scores for individual upstream regulators by module (≥2 = significant). (C) Predicted protein–protein interactions within and between modules.

Exploration of potential protein–protein interactions among these modules identified a large network of connections between the blue, red, and turquoise modules (Fig. 5C).¹⁵ Network hubs were identified among interactions associated with antigen presentation, inflammation, and type 1 immunity. This analysis demonstrates type 1 immunity and inflammatory cytokines upstream of our interactive gene modules associated with CLAD and FEV₁ decline.

3.5. Molecular validation of components of the gene modules positively correlated with CLAD demonstrates elevated type 1 alloimmune responses, components of the arachidonic acid pathway and IL-1β in the BAL

We next wished to validate our airway transcriptome findings and perform targeted molecular validation studies in our cohort. The blue module (acute phase) showed increased expression of a member of the cytosolic phospholipase A₂ group, PLA2G4A. This enzyme catalyzes the hydrolysis of cell membrane phospholipids to release arachidonic acid, which is subsequently metabolized into eicosanoids, which include the prostaglandins (PGs) and leukotrienes (LTs). We evaluated PLA2G4A expression and observed increasing levels with CLAD severity (Fig. 6A). Elevated PLA2G4A levels also correlated with the rate of FEV₁ change across the cohort (Fig. 6B). Next, we measured BAL levels of 2 downstream eicosanoids, LTC₄ and PGD₂, at the time of the brush and found these mediators to be elevated in CLAD LTRs compared with controls (Fig. 6C, D; Supplementary Table 1). Because we observed a high and low PGD2 groups, we correlated PGD₂ expression with module expression and found statistically significant correlations with all 4 modules and strongest with the blue module (Fig. 6E). Our findings support a role for increased PLA2G4A levels, leading to activation of the arachidonic acid pathway components in CLAD.

Figure 6.

Proinflammatory activation of cytosolic phospholipase A₂ is associated with chronic lung allograft dysfunction (CLAD) and allograft dysfunction. (A) Normalized PLA2G4A gene expression by CLAD stage. (B) Pearson correlation between PLA2G4A expression and rate of FEV₁ decline during 12-month period surrounding brush (6 months before and 6 months after). (C) Bronchoalveolar lavage leukotriene (LT)C₄ concentrations (as measured by enzyme-linked immunosorbent assay [ELISA]). (D) Bronchoalveolar lavage prostaglandin (PG)D₂ concentrations (as measured by ELISA). (E) Pearson correlation between PGD₂ concentrations and module expression. *P < .05; **P < .005; ***P < .0005; ***P < .00005; P values calculated by Wilcoxon rank sum and adjusted for false discovery. FEV₁, forced expiratory volume in 1 second.

Because the red module (adaptive immunity) showed increased levels of IFN gamma–driven signaling mediators including STAT1 (signal transducer and activation of transcription 1), IRF1 (interferon regulatory factor 1), and IDO1 (indolamine 2,3-dioxygenase 1), we evaluated a subset of our LTRs for IFN gamma production. In addition, the red module showed increased expression of key cytotoxic genes GZMA (granzyme A), GZMB (granzyme B), and PRF1 (perforin). We evaluated the donor-specific alloimmune response in 5 CLAD and 5 control LTRs, using total BAL cells with/without in vitro restimulation with irradiated donor peripheral blood mononuclear cells. In these studies, CLAD LTRs demonstrated increased frequencies of BAL CD8⁺ T cells producing IFN gamma, TNFα, and CD107a, marker of cytotoxic degranulation. Our studies support lung allograft T cell alloimmune responses contributing to the gene expression detected in the red module (Fig. 7A, B; Supplementary Table 2).

Figure 7.

Chronic lung allograft dysfunction (CLAD) demonstrates elevated type 1 alloimmune responses and interleukin (IL)-1β in bronchoalveolar lavage (BAL). (A) Representative gating strategy plots for CD107a, interferon (IFN) gamma, and tumor necrosis factor (TNF)α. (B) BAL CD8⁺ T cells frequency of CD107a, IFN gamma, and TNF-α (%). (C) IL-1β concentrations of BALs (as measured by enzyme-linked immunosorbent assay [ELISA]). *P < .05; **P < .005; P values calculated by Wilcoxon rank sum and adjusted for false discovery.

Given the enrichment of innate immunity in the turquoise module and upstream regulation by the inflammatory cytokine IL-1β, we measured BAL levels and found significantly increased protein levels in CLAD vs controls (Fig. 7C; Supplementary Table 3), supporting IL-1β as an important inflammatory cytokine in CLAD.

3.6. CLAD modules predict disease and survival in a validation cohort

An external institutional cohort of LTRs was used to validate the WGCNA-derived modules from our study cohort. Characteristics of the 22 CLAD and 37 control LTRs in the validation cohort are described in Table 2. The validation cohort had a higher proportion of patients with interstitial lung disease, but characteristics were otherwise similar as our derivation cohort. Our 3 positively correlated modules (red, blue, and turquoise) and the protective green module from our study cohort were similarly correlated with CLAD and CLAD stage in the validation cohort, with blue, turquoise, and green modules reaching statistical significance (Fig. 8A). Pairwise comparisons revealed significant differences in module expression between control LTRs and those with CLAD stage 1 in all but the turquoise module (Fig. 8B). Overall, all 4 modules demonstrated significant associations with CLAD in at least 1 analysis in the validation cohort.

Table 2.

Validation cohort characteristics.

Characteristic	CLAD	Control	P
n	22	37
Age at transplant (y)	58.5 (47.0–64.8)	59.0 (50.0–67.0)	.428
Time to brush posttransplant (d)	1789 (1347–2489)	544 (369–668)	<.0001
Female	6 (27.2)	16 (43.2)	.343
Transplant diagnosis			.112
CF/bronchiectasis	1 (4.5)	3 (8.1)
COPD	7 (31.8)	3 (8.1)
ILD	14 (63.6)	30 (81.1)
PH	0 (0.0)	1 (2.7)
Double-lung transplant	20 (90.9)	34 (91.9)	1
Cytomegalovirus serostatus			.051
D−/R−	4 (18.2)	2 (5.4)
D−/R+	5 (22.7)	13 (35.1)
D+/R−	2 (9.1)	11 (29.7)
D+/R+	11 (50.0)	9 (24.3)
Unknown	0 (0.0)	2 (5.4)
Anti-HLA donor-specific antibody	2 (9.1)	1 (2.7)	.64
Acute rejection	3 (13.6)	0 (0.0)	.065
Acute infection	11 (50.0)	11 (29.7)	.201
Bacterial	5 (22.7)	9 (24.3)	1
Viral	3 (13.6)	1 (2.7)	.28
Fungal	4 (18.2)	2 (5.4)	.261
Nontuberculous mycobacterial	1 (4.5)	0 (0.0)	.791
Days to CLAD	1022 (558–2317)
CLAD stage
CLAD 1	7 (31.8)
CLAD 2	9 (40.9)
CLAD 3	6 (27.3)
CLAD 4	0 (0)

Values are n (%) or median (IQR).

CLAD, chronic lung allograft dysfunction; CF, cystic fibrosis; COPD, chronic obstructive pulmonary disease; D, donor; HLA, human leukocyte antigen; ILD, interstitial lung disease; IQR, interquartile range; PH, pulmonary hypertension; R, recipient.

Figure 8.

Chronic lung allograft dysfunction (CLAD) gene modules predict disease and survival in a validation cohort. (A) Correlation and significance of module expression with group and CLAD stage in validation cohort. (B) Module expression by CLAD stage in validation cohort (pairwise significance as measured via Wilcoxon rank sum). (C, D) Receiver-operating characteristic area under the curve of the blue (acute phase response) module expression in the derivation cohort (C) and validation cohort (D). (E, F) Kaplan–Meier curve for survival by blue (acute phase response) module expression in derivation cohort (E), and by green (non-CLAD) module expression in derivation cohort (F). (G) Kaplan–Meier curve for survival by blue (acute phase response) module expression in validation cohort. AUC, area under the curve; CI, confidence interval.

Given the strength of correlation between the blue module (acute phase) and presence of CLAD, CLAD severity, and allograft dysfunction, we hypothesized the blue module would perform well at differentiating LTRs with/without CLAD. An ROC curve was generated and showed good ability to distinguish CLAD LTRs from control with an ROC area under the curve of 0.82 in our cohort and 0.88 in the validation cohort (Fig. 8C, D).

Within our cohort, we found all 4 modules had a significant association with survival in a Cox proportional hazard model, with age at transplant, recipient sex, recipient ethnicity, surgical procedure (double- vs single-lung transplant), and transplant diagnosis as covariates (Table 3). The blue module also demonstrated the most significant association with, and greatest hazard ratio against survival, while the green module was protective (Fig. 8E, F). The survival impact of the blue module was also significant in our validation cohort (Fig. 8G). Together, these data support our WGCNA-derived modules as a potential molecular diagnostic for CLAD and survival.

Table 3.

Modules and impact on survival.

Module	Hazard ratio	95% CI	P
Acute phase response (blue)	23.37	4.01–136.30	.00046
Type 1 adaptive immunity (red)	9.69	2.07–45.24	.0049
Innate immunity (turquoise)	8.44	2.05–34.69	.00312
Non-CLAD (green)	0.057	0.009163–0.3529	.00208

CI, confidence interval; CLAD, chronic lung allograft dysfunction.

4. Discussion

We identified 3 gene correlation network modules (blue, red, and turquoise) within the airway transcriptome that positively correlated with CLAD and a fourth (green) that inversely correlated with CLAD. Importantly, we found that these modules intensified with advancing CLAD stage, supporting a role for these molecular pathways in disease progression. Our analyses also indicate significant interactions among the acute phase (blue), innate immunity (turquoise), and adaptive immunity (red) modules, with key upstream cytokines IFN gamma (adaptive), TNFα (innate and adaptive), and IL-1β (innate), playing key regulatory roles leading to overlapping inflammatory signaling pathways in CLAD. These findings support our earlier findings using differential gene expression revealing a type 1 immunity signature–driven predominantly by these cytokines.¹⁰ Our studies point to a persistent T cell–driven alloimmune response as a central element shaping the CLAD airway transcriptome, with IFN gamma, TNFα and cytotoxic (CD107a, GZMB, and PRF1) responses, as represented in the red module. Additionally, key type 1 immunity chemokines CXCL9, CXCL11 (CXCR3 family), and CCL5 (CCR5 family) were upregulated in the red module, which we and others previously found elevated in BOS.^10,16,17 Importantly, IL-1β, a key component of the inflammasome along with IL-18 signaling,¹⁸ were both represented in the innate immunity module and can be induced downstream of type 1 alloimmunity. Several approved IL-1β inhibitor drugs (canakinumab and gevokizumab)¹⁹ are used for autoimmune disease although neither have been tested in lung transplant. Our studies also found upregulation of the arachidonic acid pathway in the blue module showing increased expression of PLA2G4A, which induces LTC₄, and PGD₂, and were elevated in the BAL of CLAD LTRs. Phospholipase A₂ induction has been reported to be associated with asthma and obstructive lung disease, along with eosinophil and mast cell activation.^20–22 Corticosteroids inhibit both membrane-bound phospholipase A₂ and the downstream prostaglandins such as PGD₂, and to a lesser extent LTC₄, although paradoxical stimulation has also been reported.^23,24 Additionally, certain LT antagonists can target LTC₄,²⁵ while aspirin and other nonsteroidal anti-inflammatory drugs can block PGD₂; however, the latter is contraindicated in LTRs. Together, our findings identify 3-airway transcriptome modules that point to several inflammatory pathways where existing therapies exist, and new targets could be developed.

Our findings also demonstrate that all 4 modules were found to correlate with changes in the FEV₁, with the green module protective and remaining elevated during FEV₁ stability, while the red, blue, and turquoise modules correlated with FEV₁ decline. Importantly, we found the airway transcriptome correlated best with the rate of change in FEV₁ and %FEV₁, compared with all other spirometry parameters, during a time window 6 months before and after airway brushing. Module expression did not correlate as strongly with spirometry change in the 12 months preceding airway brushing, suggesting the airway transcriptome provides better prognostic utility for active or impending allograft changes vs those that have already occurred. This needs to be further evaluated in future studies using serial brush samples prior to, and after, onset of CLAD, to determine how early CLAD-associated changes develop.

The current ISHLT guidelines for CLAD remain a diagnosis of exclusion, based on FEV₁ decline from peak posttransplant baseline.¹ Whether our data suggest a potential role for using the airway transcriptome to identify patients with allograft dysfunction that is occurring due to airway inflammatory and immune responses associated with CLAD, in contrast to allograft dysfunction from alternative etiologies (eg, pulmonary edema, obesity, and functional status decline) warrants further investigation. A molecular diagnostic for CLAD may not only confirm the diagnosis but also provide important insights into prognosis, as CLAD is a heterogenous disease with variable kinetics of progression, as well as the BOS vs RAS clinical endotypes.²⁶ Importantly, our study found our modules predict survival similar to recent work from Mohanty et al.¹² Because LTRs with the RAS endotype have been shown to have shortened survival with concomitant FVC decline,³ airway transcriptome studies might not only delineate signatures associated with poor survival but also further identify CLAD molecular endotypes and pathways to target in future interventional studies. Our data also show our CLAD modules intensify with advanced CLAD stages, supporting the hypothesis that many upregulated genes in the airway transcriptome are associated with progression of disease.

Other studies have evaluated the BAL, transbronchial biopsy or explant tissue transcriptomes during ACR episodes or CLAD and reported a predominantly cytotoxic T cell or macrophage signature, with similar features to our CLAD modules.^6,27,28 A previous study assessed gene signatures in CLAD from transbronchial biopsies and found that CLAD-selective transcripts corrected for time reflected predominantly tissue injury genes including HIF1A, IGF1, and SERPINE2. In contrast, genes associated with inflammation segregated more with time but were not found to be selective for CLAD after correcting for time.⁹ However, we found our modules were time independent in correlation with CLAD. The different tissue sources and control populations between our study and this prior report might have contributed to these different findings. While our study had shorter median time from transplant in the controls than that in CLAD, our patients with CLAD demonstrated significant time overlap with respect to CLAD severity, further supporting the modules reflecting disease rather than time. Taken together, our positively correlated CLAD modules indicate important time-independent immune signatures.

There are several limitations to our study. As our study was cross-sectional in design, we cannot definitively conclude that our identified modules are causal for CLAD. Our sampling strategy led to for-cause brushes from patients with CLAD being collected significantly longer posttransplant than the control population, whose brushes were collected at the end of their 2-year surveillance period. This sampling strategy may have led to patients with CLAD having variable levels of immune activation and airway inflammation. Nevertheless, our findings overall support CLAD as being an active inflammatory process. Future studies will assess module gene expression using serial brushes in LTRs to assess whether progression occurs, and if so, what are the earliest markers heralding the CLAD state. Our separate external validation cohort supported our modules detecting CLAD and provided reasonable discriminatory value. Overall, our findings appear internally consistent with our modules biomarking the CLAD state and its severity.

In summary, using WGCNA, we found 3 immune gene modules representative of the acute phase (blue), innate immunity (turquoise) and adaptive immunity (red) that positively correlate with CLAD stage and FEV₁ decline. A fourth module (green) demonstrated an inverse correlation with CLAD and was found to be protective. In addition to differentiating CLAD and CLAD stage, our modules predicted LTR survival and were corroborated in a validation cohort from a separate institution. Our targeted molecular validation of components for each of the positively correlated CLAD modules suggest plausible pathways contributing to the underlying pathogenesis of CLAD. Our results further support the development of a molecular diagnostic that could lead to stratification of high-risk LTRs. Moreover, the airway transcriptome could be a useful biomarker tool to assess the impact of an interventional study for CLAD. Further studies are needed, and a multicenter study to investigate the role of the airway transcriptome as a molecular diagnostic for CLAD is underway.

Abbreviations:

ACR

acute cellular rejection

BAL

bronchoalveolar lavage

BOS

bronchiolitis obliterans syndrome

CLAD

chronic lung allograft dysfunction

FEV₁

forced expiratory volume in 1 second

FVC

forced vital capacity

IFN

interferon

IL

interleukin

ISHLT

International Society of Heart and Lung Transplant

LT

leukotriene

LTR

lung transplant recipient

PG

prostaglandin

RAS

restrictive allograft syndrome

ROC

receiver-operating characteristic

RNA-seq

RNA sequencing

TNF

tumor necrosis factor

WGCNA

weighted gene correlation network analysis.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.ajt.2026.01.025.

Data availability

All data generated or analyzed during this study are included in this published article.