Molecular endotypes predict differential response to immunosuppressant therapy in non–idiopathic pulmonary fibrosis interstitial lung disease

Abstract

Rationale:

Interstitial lung diseases (ILDs) are a clinically and biologically diverse group of disorders characterized by varying inflammation and fibrosis of the lung parenchyma. Immunosuppressant therapy is commonly used to treat non-idiopathic pulmonary fibrosis (non-IPF) ILD, but treatment response is variable and difficult to predict.

Objective:

Identify and validate molecular endotypes of non-IPF ILD.

Methods:

Twenty plasma proteins associated with inflammation were used to perform latent class analysis in 2 observational non-IPF ILD cohorts (discovery n = 676; validation n = 585). Proteins were measured using a semi-quantitative Olink Explore 3072 platform. The primary outcome was 3-year transplant-free survival. Weighted Cox regression was used to assess differential response to mycophenolate or azathioprine in each cohort according to molecular endotype classification.

Results:

A 2-class model best fit both cohorts (P <0.01), with Class 2 comprising ~30% of patients. Compared to Class 1, Class 2 was associated with significantly lower 3-year transplant-free survival in both discovery (78% vs 36%, P <0.001) and validation (83% vs 46%, P <0.001) cohorts. Significant interaction between molecular endotype and immunosuppressant exposure was observed in both cohorts (discovery P_interaction = 0.022; validation P_interaction = 0.019), with survival benefit seen only in Class 2. In pooled analysis, similar trends were observed irrespective of ILD subtype. Pathway analysis supported enrichment of inflammatory signatures in Class 2.

Conclusion:

In this multicenter observational cohort study, we identified and validated 2 distinct molecular endotypes of non-IPF ILD with divergent outcomes and response to immunosuppressant therapy. These endotypes could inform precision medicine strategies and clinical trial design in ILD.

Introduction

Interstitial lung disease (ILD) comprises a biologically and clinically heterogeneous group of disorders characterized by inflammation and/or fibrosis of the lung parenchyma. Idiopathic pulmonary fibrosis (IPF), the prototypical fibrotic ILD, is highly progressive and often fatal.¹ Progressive non-IPF ILDs, including connective tissue disease (CTD)-associated ILD, fibrotic hypersensitivity pneumonitis (fHP), and idiopathic interstitial pneumonias (IIP), exhibit similarly poor prognoses but often involve a more prominent inflammatory component.^2,3 Current therapeutic approaches in ILD primarily involve antifibrotic and immunosuppressive medications.⁴ Antifibrotics slow lung function decline in IPF and other fibrotic ILDs, whereas immunosuppressants may benefit patients with a predominantly inflammatory component.^5–10 However, immunosuppressive therapy is generally not advised in fibrotic-predominant ILD, as it can lead to increased mortality and hospitalization.¹¹ Determining which patients benefit from each therapeutic class remains a critical unmet need and underscores the urgency for precision medicine approaches in ILD.

Our group recently identified molecular endotypes of IPF using fibrosis-associated plasma proteins and latent class analysis (LCA), a form of unsupervised clustering.¹² The 2 endotypes identified demonstrated divergent survival and suggested a differential response to antifibrotic therapy, highlighting the potential for biology-driven molecular classification to inform clinical care. However, while IPF is well-studied, it accounts for only one-quarter of all ILDs.¹³ For the broader population of non-IPF ILDs, treatment decisions between antifibrotics and immunosuppressants remain largely empiric. The lack of established endotyping approaches in non-IPF ILDs represents a key gap in advancing precision medicine in this population.

The objective of this study was to apply LCA to baseline proteomic data from 2 independent multicenter observational cohorts to identify molecular endotypes of non-IPF ILDs. We hypothesized that an inflammatory protein-informed LCA would reveal distinct molecular endotypes of non-IPF ILD with divergent clinical outcomes and differential responses to immunosuppressive therapy. Some of the results of these studies have been previously reported in the form of an abstract.¹⁴

Methods

Study populations

In this observational, multicenter study, we analyzed proteomic data from patients with 3 common non-IPF ILDs enrolled in 6 prospective patient registries.¹⁵ Patients with CTD-ILD, fHP, and IIP were included; patients with IPF were excluded. Patients from the University of California San Francisco (UCSF; October 2004 to February 2019), University of California Davis (UCD; May 2016 to February 2021), and University of Texas Southwestern (UTSW; March 2007 to August 2019) comprised the discovery cohort (n = 676). Patients from the Pulmonary Fibrosis Foundation Patient Registry (PFF; March 2016 to June 2018), University of Chicago (UChicago, March 2007 to July 2017), and University of Virginia (UVA; September 2018 to November 2021) comprised the validation cohort (n = 585).

The primary outcome was 3-year transplant-free survival, defined from the time of blood draw to death or lung transplantation. Patients were censored at 3 years or earlier if lost to follow-up. Clinical covariates included demographic data, smoking status, ILD subtype, baseline lung function, prior rituximab or cyclophosphamide exposure, baseline immunosuppressant (prednisone, mycophenolate, azathioprine, rituximab, or cyclophosphamide) and antifibrotic (pirfenidone or nintedanib) use, and exposure to either mycophenolate/azathioprine or antifibrotic agents during the 3-year follow-up.

Biomarker measurement protocols are described in the parent study.¹⁵ Briefly, proteomic data were generated in separate batches for the discovery and validation cohorts using the Olink Explore 3072 platform (Olink, Uppsala, Sweden), which generates semi-quantitative measurements using a proximity extension assay. A total of 2925 analytes were measured in the discovery cohort and 2921 of these 2925 analytes in the validation cohort. Further details regarding data collection and biomarker measurement protocols are provided in the Supplementary material.

Latent class analysis

LCA was performed independently in the discovery and validation cohorts using 20 plasma candidate proteins as input variables. Candidate proteins were selected from 44 prognostic biomarkers previously associated with transplant-free survival in the parent study and were chosen based on literature supporting potential involvement in diverse pathways associated with inflammation, with an emphasis on relevance to pulmonary disease.¹⁵ To capture biological heterogeneity and preserve statistical validity for LCA, only proteins with limited collinearity (|ρ|<0.6) were included (Table S1).

Models ranging from 1 to 5 classes were evaluated. The optimal number of latent classes was determined based on the largest decrease in Bayesian Information Criterion (BIC), entropy, Vuong-Lo-Mendell-Rubin likelihood ratio test, and size of the smallest class. Class assignment was based on the highest posterior probability. Additional methodological details are provided in the Supplementary material.

To confirm consistency of the classes across cohorts, a logistic regression classifier was trained on the top 5 differentially expressed biomarkers between latent classes in the discovery cohort and tested in the validation cohort. As proteomic measurements were performed in separate batches, biomarker values were z-scaled within each cohort prior to model fitting. Classifier performance was evaluated using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC).

Baseline clinical characteristics, including demographics, ILD subtype, and lung function, were compared between latent classes (subsequently referred to as molecular endotypes). The association between endotype and transplant-free survival was evaluated using a multivariable Cox model adjusted for age, sex (female, male), race (White, non-White), smoking history (ever, never), ILD subtype (CTD-ILD, fHP, IIP), baseline lung function, and treating center. Subgroup analysis was performed to assess endotype association with transplant-free survival for common ILD subtypes.

Testing for differential treatment response to immunosuppressant therapy

Differential response to immunosuppressant therapy between endotypes was independently evaluated in each cohort using weighted Cox proportional hazards regression with robust variance estimation. Molecular endotype and immunosuppressant exposure were included as the main effects and their product as the interaction term. To address confounding, inverse probability of treatment weighting (IPTW) using propensity scores was applied separately in each cohort to reduce indication bias and estimate the average treatment effect of immunosuppressive therapy.¹⁶ Propensity scores were calculated in each cohort using multivariate logistic regression to estimate the likelihood of exposure to immunosuppressant therapy during the study period, including observed covariates known or suspected to influence treatment probability or survival: age, sex (female, male), race (White, non-White), smoking history (ever, never), ILD subtype (CTD-ILD, fHP, IIP), treating center, baseline lung function, and prednisone use at blood draw (yes, no).

Immunosuppressant therapy was defined as exposure to mycophenolate or azathioprine and modeled as a time-dependent variable to minimize immortal time bias. Patients already receiving immunosuppressive therapy at blood draw were considered exposed from the start of follow-up. To reduce confounding from antifibrotic treatment, patients who started antifibrotics before blood draw were excluded, and follow-up was censored at anti-fibrotic initiation if it occurred during the study period. Stabilized weights for IPTW were calculated by dividing the marginal probability of treatment by the propensity score for treated individuals, and one minus the marginal probability by one minus the propensity score for untreated individuals. Covariate balance before and after weighting was assessed using standardized mean differences (SMD). The proportional hazards assumption was confirmed using Schoenfeld residuals.

As a secondary analysis, cohorts were pooled, and treatment heterogeneity was evaluated according to ILD subtype (CTD-ILD vs non-CTD fibrotic ILD) to explore whether the observed endotype-treatment interaction varied by clinical diagnosis. Multiple sensitivity analyses were performed to confirm robustness of the findings (Supplementary material).

Pathway analysis

To explore the biological characteristics of the endotypes, pathway analysis was performed separately in each cohort using all available proteomic data (~3,000 analytes). Differential protein expression between classes was assessed using t-test and log-fold change. Proteins with Benjamini-Hochberg-adjusted P <0.05 were considered differentially expressed and were visualized in a volcano plot according to the magnitude and direction of fold-change and statistical significance. Proteins with a Benjamini-Hochberg-adjusted P <0.05 were then analyzed in Ingenuity Pathway Analysis (QIAGEN Inc., Redwood City, CA) to identify overrepresented molecular pathways in Class 2. Enrichment was determined using Fisher’s exact test with Benjamini-Hochberg correction.

Statistical analysis

Details regarding power calculations and missing data are described in the Supplementary material. Statistical comparisons between groups were conducted using t-test, Wilcoxon rank sum test, or chi-square test, as appropriate. Transplant-free survival was visualized using Kaplan-Meier survival curves. Hazard ratios (HR) from regression modeling are presented with 95% CIs. Statistical analyses were performed using R (version 4.5.0) and Mplus (version 8.10).

Results

Cohort characteristics

Demographics, clinical characteristics, and transplant-free survival were largely similar across both cohorts (Table 1). The average age was 63 years, with female predominance. CTD-ILD was the most common ILD subtype, with rheumatoid arthritis, systemic sclerosis, and idiopathic inflammatory myopathies each comprising approximately one-quarter of cases (Table S2). Radiographic classifications were available only in the discovery cohort, where non-specific interstitial pneumonia was the most common pattern both overall and among patients with CTD-ILD (Table S3). Baseline lung function was moderately impaired, and a minority of patients were receiving immunosuppressive or antifibrotic therapy at blood draw.

Table 1.

Baseline patient characteristics and 3-year transplant-free survival between discovery and validation cohorts.

Characteristic	Discovery	Validation
	N = 676	N = 585
Age (years)	63 (±13)	63 (±12)
Diagnosis
CTD-ILD	310 (46%)	240 (41%)
fHP	197 (29%)	140 (24%)
IIP	169 (25%)	205 (35%)
Sex (female)	415 (61%)	320 (55%)
Race/Ethnicity
White	474 (70%)	472 (81%)
Black	48 (7.1%)	79 (14%)
Hispanic	92 (14%)	5 (0.9%)
Other/Unknown	62 (9.2%)	29 (5.0%)
Smoking history	255 (38%)	293 (50%)
FVC (% predicted)	66 (±19)	67 (±20)
DLCO (% predicted)	46 (±18)	47 (±18)
Prior RTX/CYC exposure	44 (6.5%)	8 (1.4%)
Prednisone use at blood draw	260 (38%)	119 (20%)
IS use at blood draw	156 (24%)	126 (22%)
AF use at blood draw	20 (3.0%)	26 (4.4%)
3-year transplant-free survival	445 (66%)	421 (72%)

Data are shown as mean (±SD) and n (%). Abbreviations: AF, antifibrotic; CTD-ILD, connective tissue disease-associated interstitial lung disease; CYC, cyclophosphamide; DLCO, diffusion capacity of the lung for carbon monoxide; fHP, fibrotic hypersensitivity pneumonitis; FVC, forced vital capacity; IIP, idiopathic interstitial pneumonia; IS, immunosuppressant medication (mycophenolate, azathioprine, rituximab, or cyclophosphamide); RTX, rituximab.

Identifying 2 distinct molecular endotypes of non-IPF ILDs

In the discovery cohort, a 2-class solution was optimal, based on the largest decrease in BIC, high entropy, and adequate class size (Table S4; Figure S2). The 2-class model fit significantly better than the 1-class model (P <0.01); adding additional classes did not improve model fit. Class 1 comprised 70.3% (475/676) of patients and Class 2 comprised 29.7% (201/676). Median posterior probability of class membership was 1.0 (IQR 1.0–1.0) for Class 1 and 1.0 (IQR 0.96–1.0) for Class 2, indicating strong class separation (Figure S3).

In the validation cohort, a 2-class solution also provided the best model fit, based on the largest decrease in BIC, high entropy, and adequate class size. The 2-class model fit significantly better than the 1-class model (P <0.01), while additional classes did not improve model fit. Class 1 included 69.9% of patients (409/585), and Class 2 included 30.1% (176/585). Median posterior probability of class membership was 1.0 (IQR 1.0–1.0) for Class 1 and 1.0 (IQR 0.93–1.0) for Class 2, supporting strong class separation.

In both cohorts, Class 2 was characterized by higher levels of Macrophage Mannose Receptor 1 (MRC1), Amphiregulin (AREG), and Syndecan-1 (SDC1), and lower levels of Tetranectin (CLEC3B) and Chondroadherin (CHAD; Figure 1). Patterns of other biomarkers between classes were similar across cohorts and remained consistent within ILD subtypes (Figure S4). A logistic regression classifier trained on the top 5 predictive biomarkers in the discovery cohort (MRC1, AREG, SDC1, Interleukin-1 receptor-like 1 [IL1RL1], and Thrombospondin-2 [THBS2]) achieved AUROC of 0.97 (95% CI 0.96–0.98) and AUPRC of 0.94 (0.91–0.96) in the validation cohort, suggesting stability and reproducibility of the endotypes.

Figure 1.

Mean standardized biomarker values stratified by endotype in discovery (solid lines) and validation (dotted lines) cohorts. Abbreviations: AREG, Amphiregulin; CEACAM6, carcinoembryonic antigen-related cell adhesion molecule 6; CHAD, Chondroadherin; CLEC3B, tetranectin; GDF15, growth/differentiation factor 15; IL1RL1, interleukin-1 receptor-like 1; LTBP2, latent-transforming growth factor beta-binding protein 2; MAMDC2, MAM domain-containing protein 2; MRC1, Macrophage Mannose Receptor 1; MSTN, growth/differentiation factor 8; OSMR, Oncostatin-M-specific receptor subunit beta; PPL, periplakin; RBFOX3, RNA binding protein fox-1 homolog 3; SDC1, Syndecan-1; SERPINA3, alpha-1-antichymotrypsin; SFTPA1, pulmonary surfactant-associated protein A1; SIGLEC8, sialic acid-binding Ig-like lectin 8; TFF2, Trefoil factor 2; TGFBR1, TGF-beta receptor type-1; THBS2, thrombospondin-2.

Clinical characteristics were largely similar between endotypes in both cohorts (Table 2), though patients in Class 2 were slightly older and had lower baseline lung function. Notably, ILD subtype did not differ between endotypes in either cohort, nor were there any differences in sex, race/ethnicity, smoking history, prior rituximab or cyclophosphamide exposure, or baseline immunosuppressant and antifibrotic use between cohorts. Among patients with CTD-ILD, the distribution of underlying rheumatologic diagnoses did not differ significantly between endotypes (Table S5). In the discovery cohort where radiographic data were available, radiographic patterns were similar between endotypes, including in the subset of patients with CTD-ILD (Table S6). MUC5B polymorphism status and peripheral blood leukocyte telomere length were available for approximately half of the patients in the validation cohort. Among these patients, there were no differences in MUC5B polymorphisms or short telomeres (defined as bottom tenth percentile) between molecular endotypes (Table S7).

Table 2.

Baseline clinical characteristics and outcomes between endotypes for discovery and validation cohorts.

Characteristic	Discovery			Validation
	Class 1	Class 2	P-value	Class 1	Class 2	P-value
	N = 475	N = 201		N = 409	N = 176
Age (years)	62 (±13)	65 (±11)	0.033	63 (±12)	65 (±11)	0.045
Diagnosis			>0.9			>0.9
CTD-ILD	220 (46%)	90 (45%)		167 (41%)	73 (41%)
fHP	138 (29%)	59 (29%)		99 (24%)	41 (23%)
IIP	117 (25%)	52 (26%)		143 (35%)	62 (35%)
Sex (female)	296 (62%)	119 (59%)	0.4	224 (55%)	96 (55%)	>0.9
Race/Ethnicity			0.064			0.11
White	322 (68%)	152 (76%)		329 (80%)	143 (81%)
Black	36 (7.6%)	12 (6.0%)		57 (14%)	22 (13%)
Hispanic	65 (14%)	27 (13%)		1 (0.2%)	4 (2.3%)
Other/Unknown	52 (11%)	10 (5.0%)		22 (5.4%)	7 (4.0%)
Smoking history	170 (36%)	85 (42%)	0.11	199 (49%)	94 (53%)	0.3
FVC (% predicted)	68 (±19)	60 (±19)	<0.001	70 (±20)	60 (±17)	<0.001
DLCO (% predicted)	49 (±17)	38 (±16)	<0.001	50 (±19)	38 (±14)	<0.001
Prior RTX/CYC exposure	31 (6.5%)	13 (6.5%)	>0.9	5 (1.2%)	3 (1.7%)	0.7
Prednisone use at blood draw	152 (32%)	108 (54%)	<0.001	78 (19%)	41 (23%)	0.2
IS use at blood draw	105 (23%)	51 (26%)	0.4	80 (20%)	46 (26%)	0.076
AF use at blood draw	16 (3.4%)	4 (2.0%)	0.3	18 (4.4%)	8 (4.5%)	>0.9
3-year transplant-free survival	372 (78%)	73 (36%)	<0.001	340 (83%)	81 (46%)	<0.001

Data are shown as mean (±SD) and n (%). Abbreviations: AF, antifibrotic; CTD-ILD, connective tissue disease-associated interstitial lung disease; CYC, cyclophosphamide; DLCO, diffusion capacity of the lung for carbon monoxide; fHP, fibrotic hypersensitivity pneumonitis; FVC, forced vital capacity; IIP, idiopathic interstitial pneumonia; IS, immunosuppressant medication (mycophenolate, azathioprine, rituximab, or cyclophosphamide); RTX, rituximab.

Patients in Class 2 had significantly worse transplant-free survival in both cohorts (P <0.001 for both; Figure 2). Three-year transplant-free survival for Class 1 patients was 78% (372/475) in the discovery cohort and 83% (340/409) in the validation cohort. By contrast, 3-year transplant-free survival for Class 2 patients was 36% (73/201) and 46% (81/176) in the discovery and validation cohorts, respectively. In multivariate Cox regression, Class 2 was associated with higher risk of death or transplantation (discovery: HR 2.67 [1.99–3.58], P <0.001; validation: HR 2.74 [1.96–3.84], P <0.001). In pooled subgroup analysis, survival was lower in Class 2 across ILD subtypes (Figure S5), with the highest relative risk of death or transplantation observed among patients with CTD-ILD (Figure S6).

Figure 2.

Three-year transplant-free survival curves stratified by molecular endotype in (A) discovery and (B) validation cohorts. P-value was calculated using log-rank test. Shaded regions indicate 95% CIs, calculated using Greenwood’s formula.

Testing for differential treatment response to immunosuppressant therapy

Unadjusted baseline patient characteristics and outcomes, stratified by endotype and immunosuppressant exposure, are presented for each cohort in Tables S8 and S9. After applying IPTW, all covariates achieved SMD ≤0.15, indicating adequate covariate balance (Figure S7). In both cohorts, the association between immunosuppressant exposure and transplant-free survival differed by endotype (Figure 3). In the discovery cohort, patients in Class 2 had a significantly reduced risk of death or lung transplant when exposed to immunosuppressive therapy (HR 0.51 [0.33–0.79]; P = 0.003), whereas no benefit was observed in Class 1 (HR 1.14 [0.67–1.95]; P = 0.62). Formal interaction testing confirmed that class assignment modified the effect of immunosuppressant response (P_interaction = 0.022). Similar findings were observed in the validation cohort, where patients in Class 2 experienced reduced risk of death or lung transplant when exposed to immunosuppressants (HR 0.52 [0.31–0.87]; P = 0.013), while those in Class 1 did not (HR 1.60 [0.75–3.39]; P = 0.22). Interaction testing again confirmed effect modification by endotype (P_interaction = 0.019).

Figure 3.

Forest Plot showing the association between immunosuppressant (IS) exposure (mycophenolate or azathioprine) and risk of death or lung transplantation within 3 years, stratified by molecular endotype (class 1 = blue; class 2 = red) and by cohort (discovery = top; validation = bottom). IS exposure was modeled as a time-varying covariate to mitigate immortal time bias. Inverse probability of treatment weighting was applied to adjust for indication bias, accounting for covariates known or suspected to influence treatment probability or survival: age, sex (female, male), race (White, non-White), smoking history (ever, never), ILD subtype (CTD-ILD, fHP, IIP), treating center, baseline lung function, and prednisone use at blood draw (yes, no). Hazard ratios (HRs) and 95% CIs were estimated using Cox proportional hazards models with robust variance estimation. P-values were derived from the Wald test. Abbreviations: CTD-ILD, connective tissue disease-associated interstitial lung disease; fHP, fibrotic hypersensitivity pneumonitis; HR, hazard ratio; IIP, idiopathic interstitial pneumonia; ILD, interstitial lung disease; IS, immunosuppressant; P-int. = P-value for the interaction term.

In pooled analysis stratified by ILD subtype, this pattern remained consistent (Figure 4). Among patients with CTD-ILD, immunosuppressive therapy was associated with reduced risk of death or transplant in Class 2 (HR 0.41 [0.25–0.68]; P <0.001), but not in Class 1 (HR 1.29 [0.51–3.28]; P = 0.59; P_interaction = 0.038). Similarly, among patients with fHP and IIP, Class 2 showed decreased risk (HR 0.64 [0.42–0.97]; P = 0.038) but not Class 1 (HR 1.29 [0.79–2.11]; P = 0.31; P_interaction = 0.034). In sensitivity analyses, differential risk by endotype remained robust. Consistent findings were observed when endotype was modeled as a continuous variable, when the exposed cohort was restricted to patients initiating immunosuppressants within 1 year of blood draw, when it was expanded to include patients receiving rituximab or cyclophosphamide, and in the pooled cohort of immunosuppressant-naïve patients. Furthermore, a consistent treatment effect across both cohorts was not observed with lung function or individual biomarkers (Table S12).

Figure 4.

Forest Plot showing the association between immunosuppressant (IS) exposure (mycophenolate or azathioprine) and risk of death or lung transplantation within 3 years, stratified by molecular endotype (class 1 = blue; class 2 = red) and ILD subtype (CTD-ILD = top; IIP and fHP = bottom) in the pooled discovery and validation cohorts. IS exposure was modeled as a time-varying covariate to mitigate immortal time bias. Inverse probability of treatment weighting was applied to adjust for indication bias, accounting for covariates known or suspected to influence treatment probability or survival: age, sex (female, male), race (White, non-White), smoking history (ever, never), ILD subtype (CTD-ILD, fHP, IIP), treating center, baseline lung function, and prednisone use at blood draw (yes, no). Hazard ratios (HRs) and 95% CIs were estimated using Cox proportional hazards models with robust variance estimation. P-values were derived from the Wald test. Abbreviations: CTD-ILD, connective tissue disease-associated interstitial lung disease; fHP, fibrotic hypersensitivity pneumonitis; HR, hazard ratio; IIP, idiopathic interstitial pneumonia; ILD, interstitial lung disease; IS, immunosuppressant; P-int. = P-value for the interaction term.

Pathway analysis

A volcano plot of differentially expressed proteins between endotypes, derived from the full proteomic panel of ~3,000 analytes, is shown in Figure S13. Pathway analysis revealed consistent enrichment of immune and fibrotic pathways across cohorts, suggesting a biologically coherent endotype signature (Figure S14). Top pathways included neutrophil degranulation, hepatic fibrosis/hepatic stellate cell activation, and pathogen-induced cytokine storm signaling. Given that MRC1 was the most predictive biomarker of endotype assignment in both cohorts, we further assessed enrichment of macrophage-associated pathways. Although ranked lower (possibly due to the smaller number of proteins annotated to these pathways), several pathways relating to macrophage biology were significantly enriched in Class 2 across cohorts, including Macrophage Classical Activation Signaling (−log(P) = 14.0 [discovery], 11.2 [validation]), Macrophage Alternative Activation Signaling (−log(P) = 4.39, 3.62), and IL-12 Signaling and Production in Macrophages (−log(P) = 7.55, 7.4).

Discussion

In this multicenter, observational study, we identified 2 distinct molecular endotypes of non-IPF ILD using a panel of 20 plasma proteins associated with inflammation. These endotypes were reproducible across independent cohorts and showed divergent clinical outcomes even after adjusting for demographics, ILD subtype, and baseline lung function. Notably, this framework identified a subset of patients with survival benefit to immunosuppressive therapy, an observation that was not apparent using conventional clinical data. The endotypes were agnostic to ILD subtype, suggesting they reflect underlying biology that transcends traditional diagnostic categories. Moreover, the endotypes do not appear to be readily identifiable using routinely available clinical data, underscoring the added value of novel molecular diagnostics. Taken together, these findings add to a growing body of evidence that molecular endotyping represents a promising strategy to inform treatment selection and advance precision care in ILD.

Several features of these endotypes suggest they are capturing distinct patterns of immune dysregulation. The proteins used for endotype discovery were selected based on established or putative associations with inflammation, spanning diverse processes such as macrophage activation,^17–19 cytokine signaling,^20,21 and immune checkpoint control.^22,23 This was corroborated in pathway analysis, which revealed neutrophil degranulation and granulocyte adhesion and diapedesis among the top-enriched pathways, indicating a heightened signal of innate immune cell recruitment, activation, and tissue infiltration in Class 2. Furthermore, while not among the highest-ranked canonical pathways, macrophage activation pathways were also significantly enriched, consistent with elevated levels of endotype-defining proteins such as MRC1 and AREG that are implicated in macrophage-mediated tissue remodeling and polarization.^17,19 Importantly, while these input proteins were selected based on literature supporting potential involvement in inflammatory processes, several are pleiotropic and also well-recognized for their roles in epithelial-mesenchymal signaling^24,25 and tissue remodeling.^26,27 This may explain the enrichment of fibrosis-associated pathways such as hepatic fibrosis/hepatic stellate cell activation and reflect the overlapping spectrum of chronic immune activation and fibrosis characteristic of many ILDs.^28,29 Together, these findings suggest that Class 2 may represent an endotype characterized by immune activation alongside increased fibrogenic potential, which could inform risk stratification and guide therapeutic decisions.

The differential response to immunosuppressive therapy between endotypes likely reflects this underlying immune activation and dysregulation. Both mycophenolate and azathioprine are known to impair neutrophil function and broader immune cell activity, which may in part explain the observed survival benefit in endotypes marked by potential immune dysregulation.^30,31 Moreover, treatment heterogeneity was most apparent in CTD-ILD, a subtype of ILD characterized by systemic autoimmunity and chronic inflammation.³² Patients in Class 2 with CTD-ILD had both the highest relative risk of death or transplantation and the greatest observed benefit from immunosuppressive therapy, compared to other ILD subtypes. This convergence of clinical and molecular findings in an ILD subtype with well-known chronic inflammation supports the hypothesis that Class 2 represents an immune-dysregulated, yet potentially immunosuppressant responsive, subgroup.

Importantly, these molecular endotypes were agnostic to ILD subtype. In other words, clinical diagnoses did not differ between classes, and endotype assignment was not entirely explained by conventional clinical features such as lung function or baseline treatment. Moreover, in the discovery cohort where radiographic data were available, radiographic patterns did not differ between endotypes, and the prevalence of MUC5B polymorphisms and short telomeres was likewise similar between endotypes in the validation cohort. This suggests that the biology captured by endotyping reflects a dimension of disease that cuts across conventional diagnostic labels. Notably, the immunosuppressant-responsive endotype was observed not only in CTD-ILD but also among patients with fHP and non-IPF IIPs, with similar biomarker patterns, outcomes, and differential treatment response. The presence of a shared, treatment-relevant subgroup across diverse ILD subtypes raises the possibility that future therapeutic strategies, including clinical trials, could be stratified or enriched by molecular endotype, potentially improving power and interpretability in settings where clinical heterogeneity has historically obscured treatment effects.^33–35 Collectively, these findings challenge the current paradigm of clinical diagnosis-based treatment selection and underscore the potential of molecular endotyping to reclassify ILD based on underlying biology rather than clinical phenotype alone.

While the endotypes identified in this study appear both biologically coherent and clinically meaningful, they likely represent one plausible projection of underlying biological heterogeneity rather than a definitive classification. The 20 candidate proteins were selected based on their association with inflammation and survival in non-IPF ILD along with diverse biological pathways represented, reflecting our hypothesis that immune dysregulation may underlie differential responsiveness to immunosuppressive therapy. Among the broader proteomic dataset, several other proteins putatively associated with inflammation, such as Tumor Necrosis Factor Receptor Superfamily Member 10B (TNFRSF10B),³⁶ WAP Four-disulfide Core Domain Protein 2 (WFDC2),³⁷ and V-set and Immunoglobulin Domain Containing 4 (VSIG4)³⁸ were also highly differentially expressed between classes, suggesting that similar endotypes potentially reflecting immune dysregulation could be recovered using an alternative set of biomarkers. Moreover, it is also plausible that selecting a different panel of candidate proteins altogether, especially those reflecting orthogonal biological axes, could yield distinct endotypes that are similarly valid and clinically relevant. As with our prior work in IPF, neither routine clinical features nor individual biomarkers consistently identified treatment response, further underscoring the need for multivariate data-driven approaches to molecular endotyping.¹² Ultimately, refining these endotyping strategies through validation in prospective trials and real-world clinical settings will be essential to realizing their potential as precision medicine tools.

Molecular profiling using a range of omics technologies is increasingly being applied across the continuum of ILD care, from assessing disease susceptibility and refining diagnoses to predicting outcomes and guiding treatment decisions. While most studies to date have focused on IPF, an expanding body of work has begun to explore non-IPF ILDs, which comprise the majority of fibrotic ILDs.^13,39 For example, single-nucleotide polymorphisms in the MUC5B and TOLLIP genes, initially identified in IPF,^1,40 have also been associated with disease risk and progression in non-IPF ILDs, including fHP and rheumatoid arthritis-associated ILD.^41–43. Machine learning models trained on proteomic or metabolomic data have shown promise in differentiating non-IPF ILD subtypes^44,45 and in identifying patients both with and without IPF who are at high risk of progression.^15,46 However, relatively few studies have examined molecular approaches for therapeutic stratification in non-IPF ILDs. Among the most established is peripheral blood leukocyte telomere length, which has been associated with poor outcomes following immunosuppression across fibrotic ILD subtypes.^47,48 To our knowledge, the present study represents the largest effort to date to evaluate molecular endotypes for treatment selection in non-IPF ILD as a step toward realizing precision medicine in this heterogeneous population.

This study has several important limitations. First, this was an observational cohort study, which is subject to confounding and cannot establish causality. Rigorous prospective validation is essential prior to clinical implementation. Similarly, while pathway analysis supports biological plausibility, it does not establish causal mechanisms. The associations of endotypes with biological pathways linked to inflammation and fibrosis are hypothesis-generating and require functional validation. Second, although IPTW was used to mitigate confounding by indication, this approach can only account for observed covariates that were measured in the cohorts. Several key clinical covariates were unavailable, including ILD duration, oxygen use, line of therapy, routine laboratory tests, and comorbidities, and unmeasured confounders may have influenced class assignment, treatment decisions, or outcomes. Some patients were already receiving immunosuppressive therapy at study entry, which may have affected baseline clinical and biomarker profiles and propensity score estimation; however, baseline immunosuppressant use was balanced between endotypes, and sensitivity analyses restricted to immunosuppressant-naïve patients yielded similar results. Third, we did not differentiate between mycophenolate and azathioprine or account for variation in dosage or subsequent prednisone exposure, which may have obscured drug-specific effects. Although rituximab and cyclophosphamide exposure were assessed in sensitivity analyses, data on other immunosuppressants were not available. Fourth, the temporal stability of these classifications and their relationship to longitudinal biomarker changes remains uncertain. Finally, the proteomic platform provided semi-quantitative measurements which are limited by batch effects. Translation of these findings will require development of reliable, quantitative patient-level classifiers that can be readily deployed in clinical settings.

In conclusion, in this multicenter, observational cohort study, we identified and validated 2 novel molecular endotypes of non-IPF ILDs with divergent clinical outcomes and differential response to immunosuppressive therapy. These findings highlight the potential of molecular endotyping to inform therapeutic decision-making, advance precision medicine across clinically heterogeneous ILDs, and support the development of biomarker-guided therapeutic strategies in clinical trials.

Supplementary Material

Supplementary material is available at American Journal of Respiratory and Critical Care Medicine online.

At a Glance Commentary.

Scientific Knowledge on the Subject:

Non-idiopathic pulmonary fibrosis (non-IPF) interstitial lung diseases (ILDs) are biologically and clinically heterogeneous. Immunosuppressant therapy is commonly used to treat non-IPF ILDs, but predicting which patients will benefit remains a major challenge. Prior work in IPF has shown that plasma protein-derived molecular endotypes could potentially stratify risk and treatment response, but comparable approaches have not been established in non-IPF ILDs.

What This Study Adds to the Field:

In this multicenter observational study, we identified and validated 2 novel molecular endotypes of non-IPF ILDs using a panel of 20 plasma proteins associated with inflammation. These endotypes were agnostic to ILD subtype and conventional clinical features. In both cohorts, one endotype exhibited a survival benefit with immunosuppressive therapy, whereas the other did not. These findings support molecular endotyping as a promising strategy to guide therapeutic selection and improve trial design in ILD.