Molecular Endotypes of Idiopathic Pulmonary Fibrosis: A Latent Class Analysis of Two Multicenter Observational Cohorts

Abstract

Rationale

Idiopathic pulmonary fibrosis (IPF) causes irreversible fibrosis of the lung parenchyma. Although antifibrotic therapy can slow IPF progression, treatment response is variable. There exists a critical need to develop a precision medicine approach to IPF.

Objectives

To identify and validate biologically driven molecular endotypes of IPF.

Methods

Latent class analysis (LCA) was independently performed in prospectively recruited discovery (n = 875) and validation (n = 347) cohorts. Twenty-five plasma biomarkers associated with fibrogenesis served as class-defining variables. The association between molecular endotype and 4-year transplant-free survival was tested using multivariable Cox regression adjusted for baseline confounders. Endotype-dependent differential treatment response to future antifibrotic exposure was then assessed in a pooled cohort of patients naive to antifibrotic therapy at the time of biomarker measurement (n = 555).

Measurements and Main Results

LCA independently identified two latent classes in both cohorts (P < 0.0001). WFDC2 (WAP four-disulfide core domain protein 2) was the most important determinant of class membership across cohorts. Membership in class 2 was characterized by higher biomarker concentrations and a higher risk of death or transplant (discovery, hazard ratio [HR], 2.02; 95% confidence interval [CI], 1.64–2.48; P < 0.001; validation, HR, 1.95; 95% CI, 1.34–2.82; P < 0.001). In pooled analysis, significant heterogeneity in treatment effect was observed between endotypes (P = 0.030 for interaction), with a favorable antifibrotic response in class 2 (HR, 0.64; 95% CI, 0.45–0.93; P = 0.018) but not in class 1 (HR, 1.19; 95% CI, 0.77–1.84; P = 0.422).

Conclusions

In this multicohort study, we identified two novel molecular endotypes of IPF with divergent clinical outcomes and responses to antifibrotic therapy. Pending further validation, these endotypes could enable a precision medicine approach for future IPF clinical trials.

At a Glance Commentary

Scientific Knowledge on the Subject

Idiopathic pulmonary fibrosis (IPF) is characterized by a heterogeneous disease course, and response to antifibrotic therapy is difficult to predict. Several novel plasma biomarkers have been associated with IPF survival in a recent unbiased proteomic analysis. We hypothesized that, using a subset of these biomarkers, we could identify molecular endotypes of IPF with unique prognostic and therapeutic insights.

What This Study Adds to the Field

In this multicenter observational study, we identified and validated two novel molecular endotypes of IPF using a panel of 25 plasma biomarkers. Endotype classification offered unique prognostic information compared with routine clinical data and was associated with differential therapeutic response to antifibrotic therapy in pooled analysis. These endotypes appear to represent biologically homogeneous and clinically relevant subgroups and could thereby enable a means to study precision medicine in IPF.

Idiopathic pulmonary fibrosis (IPF) is a heterogeneous disease of unknown etiology characterized by progressive, irreversible fibrosis of the lung parenchyma (1). Mortality rates in IPF are high, with a median survival time from diagnosis of 3–5 years (1, 2). Two antifibrotic medications, pirfenidone and nintedanib, have been shown to slow IPF progression (3, 4), but treatment response is variable and difficult to predict.

IPF has long been recognized as a heterogeneous disease with an unpredictable disease course (5–10), and this heterogeneity may, at least in part, be implicated in the relative paucity of therapeutic options (6–8). As in asthma (11), cancer (12), and acute respiratory distress syndrome (13), a precision medicine approach to IPF that leverages novel “omics” data has been advocated to identify biologically homogeneous subgroups that may share a common therapeutic response to medical interventions (7–10). Such an approach could also offer mechanistic clues into IPF pathobiology and serve as a platform for future clinical trial enrichment.

In this study, our primary objective was to identify and validate molecular endotypes in IPF using baseline plasma protein data generated in two multicenter prospective registry cohorts. Our secondary objective was to evaluate whether these biological endotypes have differential response to antifibrotic therapy. To define the endotypes, we applied latent class analysis (LCA), a well-validated statistical modeling technique that identifies unobserved (or latent) subgroups within a larger population (14). We selected proteins that have been previously associated with IPF pathobiology as inputs into the LCA, hypothesizing that endotypes of IPF derived from molecular biomarkers would have divergent clinical outcomes and differential responses to antifibrotic therapy. Some of the results of these studies have been previously reported in the form of an abstract (15).

Methods

Study Populations

In this observational, multicenter cohort study, we analyzed baseline clinical and protein data from patients with IPF recruited as part of four prospective patient registries: the Pulmonary Fibrosis Foundation Patient Registry (PFF; March 2016 to June 2018) and site-specific registries at the University of Chicago (March 2007 to April 2017); University of California, Davis (July 2016 to April 2021); and University of Virginia (September 2018 to March 2022). Patients from the PFF registry comprised the discovery cohort (n = 875), and patients from the Universities of Chicago; California, Davis; and Virginia comprised the external validation cohort (n = 347; Table 1). Patients participating in the PFF registry from the Universities of Chicago and Virginia were excluded from the validation cohort to ensure no overlap between cohorts. The University of California, Davis, did not enroll patients into the PFF registry. Patients without baseline lung function testing performed within 12 months of blood draw were excluded. Treatment data were collected as part of standardized case report forms completed at each visit for patients in the PFF registry, with treatment start/stop times relative to enrollment used to ascertain exposure time. Treatment data for the Davis, Virginia, and Chicago registries were collected by manual review of the electronic medical record with actual dates of treatment initiation/cessation used to ascertain exposure time. The study was approved by the institutional review boards at each participating site.

Table 1.

Baseline Patient Characteristics and the Primary Outcome of Death or Transplantation within 4 Years of Blood Draw in the Discovery and Validation Cohorts

Characteristic	Discovery (n = 875)	Validation (n = 347)
Age, yr	71 ± 8	73 ± 8
Female sex	215 (25%)	82 (24%)
White race	823 (94%)	310 (89%)
Smoking history (yes/no)	562 (64%)	218 (63%)
AF use at blood draw (yes/no)	519 (59%)	67 (19%)
FVC, % predicted	68 ± 17	72 ± 18
DlCO, % predicted	40 ± 15	51 ± 18
GAP stage
I	186 (21%)	132 (38%)
II	482 (55%)	154 (44%)
III	206 (24%)	61 (18%)
4-yr death or transplantation	425 (49%)	143 (41%)
Death	295 (34%)	126 (36%)
Transplantation	130 (15%)	17 (4.9%)

Definition of abbreviations: AF = antifibrotic; GAP = Gender/Age/Physiology.

Data presented as mean ± standard deviation where applicable. The discovery cohort comprised patients in the Pulmonary Fibrosis Foundation patient registry, and the validation cohort comprised patients from site-specific registries at the University of California, Davis; University of Chicago; and University of Virginia.

The primary outcome was 4-year transplant-free survival, defined as time from the blood draw to death or lung transplant. Patients were censored at 4 years or earlier if lost to follow-up. Clinical variables included demographic characteristics, smoking history, baseline percent predicted FVC, percent predicted Dl_CO, antifibrotic agent exposure (pirfenidone or nintedanib) at the time of the blood draw, and antifibrotic agent exposure (pirfenidone or nintedanib) during the subsequent 4-year study period. The overall analysis plan is shown in Figure E1 in the online supplement.

Biomarker Assay Procedure and Selection

The biomarker measurement protocol has previously been described (16). Briefly, peripheral blood ethylenediaminetetraacetic acid samples were collected and stored at −80°C at each site. Protein data were generated as a single batch using the semiquantitative Olink Explore 3072 panel (Olink) and log₂ transformed for normality. To identify input proteins for the latent class analysis, three expert senior investigators with a research focus in basic pulmonary biology (B.B.M., R.L.Z., and P.J.W.) were provided a list of 140 proteins recently associated with IPF survival (16) and an additional 60 identified through literature review (17). The experts were asked in round 1 to select as many as 50 proteins they believed were most likely to have a putative association with IPF pathobiology. Proteins with at least two votes in round 1 were advanced to round 2, in which the experts were asked to select as many as 40 proteins. After round 2, 25 proteins had received at least two votes and were selected for inclusion in LCA modeling (Table 2). See online supplement for results of the two-step voting process for all 200 proteins.

Table 2.

Protein Biomarkers (n = 25) Used as Class-defining Variables in the LCA to Derive IPF Molecular Endotypes

Protein	Abbreviation
Amphiregulin	AREG
C-C motif chemokine 17	CCL17
C-C motif chemokine 18	CCL18
C-C motif chemokine 7	CCL7
Collagen α-1(XXIV) chain	COL24A1
Collagen α-1(IV) chain	COL4A1
Granulocyte–macrophage colony–stimulating factor curve	CSF2
Collagen triple helix repeat-containing protein 1	CTHRC1
Growth/differentiation factor 15	GDF15
Intercellular adhesion molecule 1	ICAM1
Interleukin-10	IL10
Interleukin-6	IL6
Integrin β-6	ITGB6
Keratin, type I cytoskeletal 19	KRT19
Leucine-rich α-2-glycoprotein	LRG1
Matrilysin	MMP7
Urokinase plasminogen activator surface receptor	PLAUR
Periostin	POSTN
Periplakin	PPL
R-spondin-3	RSPO3
Secretoglobin family 3A member 2	SCGB3A2
Syndecan-1	SDC1
Osteopontin	SPP1
Tumor necrosis factor	TNF
WAP four-disulfide core domain protein 2	WFDC2

Definition of abbreviations: IPF = idiopathic pulmonary fibrosis; LCA = latent class analysis.

Biomarkers were measured in peripheral plasma using a semi-quantitative proteomic assay (Olink). Biomarkers were selected a priori to the LCA procedure based on putative association with IPF pathobiology.

LCA

LCA was performed independently in the discovery and validation cohorts. After a priori selection of the 25 input biomarkers described above, LCA was performed using only protein data. We built models comprising one to five classes. The best-fitting model was determined based on the largest decrease in the Bayesian information criterion (BIC), entropy, Vuong-Lo-Mendell-Rubin likelihood ratio test, and size of the smallest class. Class assignment was assigned based on the highest probability. See online supplement for additional details of the LCA procedure.

To confirm that the latent classes in the discovery and validation cohorts represented shared underlying biology, we built a logistic regression classifier model trained on the top five predictive biomarkers in the discovery cohort and tested this model in the validation cohort. Model performance was evaluated using area under the receiver operating characteristic curve and area under the precision–recall curve.

When class membership had been assigned, we compared baseline clinical characteristics, including demographic characteristics, lung function, and Gender/Age/Physiology (GAP) stage (18), a widely used prognostic scoring system in IPF. The association between latent class membership and the primary outcome of 4-year transplant-free survival was assessed using a multivariable Cox regression model adjusted for age, sex, race, smoking history, and baseline FVC and Dl_CO.

Testing for Differential Treatment Response

We then sought to evaluate whether molecular endotypes respond differentially to antifibrotic therapy. To study this, we pooled cohorts and analyzed patients who were naive to antifibrotic therapy at the time of the blood draw (n = 555). This pooled approach was facilitated by biomarker processing in a single batch. Endotype class assignments in the pooled cohort were carried forward from the LCA results from each individual discovery/validation cohort.

As a result of the observational nature of the cohorts, we applied inverse probability of treatment weighting (IPTW) using propensity scores to minimize indication bias and estimate the average treatment effect of antifibrotic therapy across the entire population. Propensity scores were calculated in antifibrotic therapy–naive patients at the time of biomarker measurement using multivariate logistic regression to estimate the conditional probability of pirfenidone or nintedanib exposure during the study period. The propensity score model was adjusted for measured confounders that are known or suspected to influence survival or probability of treatment, including age, sex, race, smoking history, treating center, FVC, and Dl_CO. Stabilized weights were then estimated for antifibrotic-exposed individuals by dividing the crude probability of treatment (e.g., propensity model without covariates) by the propensity score and for nonexposed individuals by dividing 1 minus the crude probability of treatment by 1 minus the propensity score (19). Confounder balance before and after weighting was assessed using the standardized mean difference.

The primary objective was to determine whether class membership was associated with differential antifibrotic response in relation to the primary endpoint of 4-year transplant-free survival. We tested for differential treatment response using weighted Cox regression models with robust variance estimation, with molecular endotype and antifibrotic agent exposure as the main effects and their product as the interaction term. Model weights were determined using the IPTW procedure described above. To minimize immortal time bias, antifibrotic agent exposure was modeled as a time-varying covariate such that patients were considered unexposed until the initiation of antifibrotic therapy. The assumption of proportional hazards was confirmed by inspection of the Schoenfeld residuals. See online supplement for a description of the multiple sensitivity analysis to confirm treatment interaction.

Statistical Analysis

Between-group differences were assessed using a Student’s t test, Wilcoxon rank sum test, or χ² test as appropriate. Kaplan-Meier estimator curves were plotted to represent the probability of transplant-free survival censored at 4 years. Hazard ratios (HRs) from regression modeling are presented with 95% confidence intervals (CIs). Missing data are presented in Table E1, and sample size calculations are presented in Figure E2. Protein–protein correlations are shown in Figures E3 and E4. All statistical analyses were performed in R (version 4.2.3) using RStudio interface (version 2023.09) and Mplus (version 8.10). Statistical significance was set at P < 0.05.

Results

LCA

In the discovery cohort, a two-class model was determined to be the best fit based on the largest decrease in the BIC, entropy, and the number of patients in the smallest class (Table 3, upper). The greatest decrease in BIC occurred with the two-class model (Figure E5). Furthermore, the two-class model was a statistically better fit than the one-class model (P < 0.0001), and three or more classes did not significantly improve model fit. Class 1 comprised 58.7% of patients (514 of 875), and class 2 comprised 41.3% (361 of 875). The median probabilities of class assignment were 1.0 in class 1 (IQR, 0.99–1.0) and class 2 (IQR, 0.95–1.0), suggesting good class separation (Figure E6).

Table 3.

Fit Statistics for LCA Models with 1–5 Classes for the Discovery and Validation Cohorts

Class	BIC	Entropy	P Value	N 1	N 2	N 3	N 4	N 5
Discovery cohort (n = 875)
1	59677	—	—	875	—	—	—	—
2	57541	0.841	<0.0001	514	361	—	—	—
3	56855	0.905	0.44	552	299	24	—	—
4	56426	0.863	0.29	401	346	105	23	—
5	56071	0.887	0.75	387	344	102	22	20
Validation cohort (n = 347)
1	23806	—	—	347	—	—	—	—
2	23070	0.858	<0.0001	201	146	—	—	—
3	22942	0.891	0.58	170	158	19	—	—
4	22952	0.868	0.44	147	146	45	9	—
5	22956	0.892	0.72	145	140	46	9	7

Definition of abbreviations: BIC = Bayesian information criterion; LCA = latent class analysis.

BIC is a statistical metric for model selection, with lower values suggesting improved model fit. Entropy is a measure of class separation, with values ⩾0.8 suggesting adequate separation. P value is determined using the Vuong-Lo-Mendell-Rubin likelihood ratio test, which tests the improvement in fit of a k class model with k – 1 class model. N_i represents the number of individuals in the ith class.

In the validation cohort, a two-class model was also determined to be the best fit (Table 3, lower). As in the discovery cohort, the greatest decrease in BIC occurred with the two-class model (Figure E7). A two-class model was a statistically better fit than a one-class model (P < 0.0001), and adding more classes did not improve model fit. The proportions of patients assigned to each class were strikingly similar to the discovery cohort, with 57.9% (201 of 347) assigned to class 1 and 42.1% (146 of 347) assigned to class 2. Median probability of class membership was similarly high at 1.0 in class 1 (IQR, 0.99–1.0) and class 2 (IQR, 0.97–1.0; Figure E8).

Class 2 in both cohorts was driven by higher levels of proteins associated with pulmonary epithelial cell dysfunction, extracellular matrix maintenance, and dysregulated response to cellular injury, such as WFDC2 (WAP four-disulfide core domain 2), GDF15 (growth/differentiation factor 15), PLAUR (plasminogen activator urokinase receptor), SDC1 (syndecan-1), and SPP1 (osteopontin). Overall, the most informative biomarkers defining the classes were similar between cohorts (Figure 1). A logistic regression model trained to predict LCA classes using the top five predictive biomarkers in the discovery cohort achieved high performance in the validation cohort (Table E2), with an area under the receiver operating characteristic curve of 0.94 (95% CI, 0.91–0.96; Figure E9) and area under the precision–recall curve of 0.93 (95% CI, 0.89–0.95; Figure E10), suggesting that the same biological endotypes were independently identified in each cohort.

Figure 1.

Mean standardized biomarker values stratified by endotype between (A) discovery and (B) validation cohorts. AREG = amphiregulin; CCL17 = C-C motif chemokine 17; COL24A1 = collagen α-1(XXIV) chain; COL4A1 = collagen α-1(IV); CSF2 = granulocyte–macrophage colony–stimulating factor; CTHRC1 = collagen triple helix repeat-containing protein 1; GDF15 = growth/differentiation factor 15; ICAM1 = intercellular adhesion molecule 1; ITGB6 = integrin β-6; KRT19 = keratin, type I cytoskeletal 19; LRG1 = leucine-rich α-2-glycoprotein; MMP7 = matrilysin; PLAUR = urokinase plasminogen activator surface receptor; POSTN = periostin; PPL = periplakin; RSPO3 = R-spondin-3; SCGB3A2 = secretoglobin family 3A member 2; SDC1 = syndecan-1; SPP1 = osteopontin; TNF = tumor necrosis factor; WFDC2 = WAP four-disulfide core domain protein 2.

Across cohorts, patients in class 2 were older on average by 2 years (discovery, P < 0.001; validation, P = 0.038) and had lower baseline lung function (P < 0.001 for both cohorts; Table E3). Baseline GAP stage was higher for patients in class 2 in both cohorts (P < 0.001). Female sex was more prevalent in class 1 (28% [56 of 201] vs. 18% [26 of 146]; P = 0.030) in the validation cohort but not the discovery cohort. Notably, there were no statistically significant differences in ethnicity, smoking history, and baseline antifibrotic agent use at the time of biomarker measurement between classes in either cohort (P > 0.05 for all).

In both cohorts, patients in class 2 had significantly worse 4-year transplant-free survival (Figure 2). In the discovery cohort, the 4-year transplant-free survival rates were 64% (331 of 514) in class 1 versus 33% (119 of 361) in class 2 (P < 0.001). In the validation cohort, we observed a similar trend: 74% (149 of 201) in class 1 versus 38% (55 of 146) in class 2 (P < 0.001). In Cox regression modeling, this higher risk of 4-year death or transplant persisted even after adjusting for potential confounders (discovery, HR, 2.02; 95% CI, 1.64–2.48; P < 0.001; validation, HR, 1.95; 94% CI, 1.34–2.82; P < 0.001; Tables E4 and E5).

Figure 2.

Four-year transplant-free survival curves stratified by endotype in (A) discovery and (B) validation cohorts. P values were calculated using a log-rank test.

Testing for Differential Treatment Response

Unadjusted baseline patient characteristics, antifibrotic agent exposure details, and clinical outcomes for the pooled cohort of antifibrotic agent–naive patients at the time of the blood draw (n = 555) stratified by endotype and future antifibrotic agent exposure are presented in Table E6. After the IPTW procedure, the standardized mean difference for all measured covariates showed adequate balance (standardized mean difference ⩽0.10) between antifibrotic-exposed and -nonexposed patients (Figure E11).

We observed a differential risk for death or transplant with future antifibrotic agent exposure according to endotype. Survival benefit with antifibrotic therapy was observed for patients in class 2 (HR, 0.64; 95% CI, 0.45–0.93; P = 0.018), but not for those in class 1 (HR, 1.19; 95% CI, 0.77–1.84; P = 0.422), with a statistically significant effect modification (P = 0.030 for interaction; Figure 3).

Figure 3.

Forest plot showing differential risk of 4-year death or transplant for future antifibrotic exposure (pirfenidone or nintedanib) versus no exposure stratified by endotype among antifibrotic therapy–naive patients at the time of the biomarker measurement. Antifibrotic agent use was modeled as a time-dependent covariate to mitigate immortal time bias. P values were calculated using a Wald test. AF = antifibrotic; CI = confidence interval; HR = hazard ratio; p-int. = P value for the interaction term.

In sensitivity analyses (see online supplement), we observed a similar effect modification when using a “landmark analysis” approach to mitigate immortal time bias, with a differential treatment response between endotype and antifibrotic agent exposure initiated within 6 months of the blood draw (P = 0.037 for interaction; Figure E14). This effect modification also persisted when modeling endotype membership as a probability rather than a binary indicator (P = 0.023 for interaction; Figure E15). Notably, we did not identify a treatment interaction between antifibrotic agent exposure and FVC, Dl_CO, or GAP stage (all P > 0.05 for interaction; Tables E7–E9) or with any of the individual top five endotype-defining biomarkers (all P > 0.05 for interaction; Tables E10–E14), suggesting that differential treatment response is captured only with the composite endotype schema.

Discussion

In this multicohort observational study, we identified two novel molecular endotypes of IPF derived using a panel of 25 protein biomarkers from four prospective patient registries. The endotypes had divergent clinical trajectories even after accounting for baseline demographic characteristics and lung function, indicating that they are not simply a reflection of disease severity. Importantly, in the pooled analysis, we observed a differential survival response to antifibrotic agent exposure when stratifying by endotype but not by routinely available physiological data, suggesting that there is unique biological information captured within the endotype schema. Patient characteristics were largely similar between endotypes with overlapping physiological function, increasing confidence that these endotypes represent truly latent (i.e., not readily identifiable using routine clinical data) rather than manifest classes. Taken together, these molecular endotypes appear to identify a distinct, clinically relevant, and previously unobserved group of biologically homogeneous patients with IPF. Pending further validation, these endotypes could offer a means to a precision medicine approach for future clinical trials and a platform to study new therapies in IPF.

The biomarkers that define the endotypes, such as GDF15, WFDC2, and PLAUR, provide clues into their underlying biology. GDF15, a member of the TGF-β (transforming growth factor) family, has been directly implicated in fibroblast activation and differentiation and is also a marker of cellular senescence (20–23). Increased circulating levels of GDF15 have been causally associated with several fibrotic disorders, including IPF and systemic sclerosis (24–26). Higher circulating levels of PLAUR have been similarly linked to progressive interstitial lung disease (ILD) (17) and systemic sclerosis (27), perhaps as a result of dysregulated fibrinolysis at the pulmonary epithelial surface (28, 29). Although WFDC2, also known as HE4 (human epididymis protein 4), is best known clinically as a diagnostic biomarker for ovarian cancer (30), it has also been associated with fibrotic diseases such as progressive pulmonary fibrosis and glomerulosclerosis (31–33). In murine models, WFDC2 has been identified as a mediator of increased myofibroblast activity, resulting in decreased degradation of type I collagen (34). The patterns of these biomarkers, along with other profibrotic proteins such as SDC1 (35), KRT19 (keratin-19) (36), and AREG (amphiregulin) (37) depict class 2 as a “fibroproliferative” IPF endotype characterized by highly abnormal fibroblast activation, persistently dysregulated cellular healing, and, ultimately, accelerated fibrosis.

Several aspects of our analysis warrant further consideration. First, although each of the prognostic biomarkers above has individually been implicated in IPF pathogenesis, a differential treatment response was not observed between individual biomarkers and antifibrotic therapy in our analysis. Given the heterogeneity of IPF, an individual biomarker in isolation may not adequately account for the multiple biological axes that drive disease progression or treatment response. Indeed, pirfenidone and nintedanib are thought to modulate a variety of biological pathways in IPF, including fibroblast proliferation and signaling, angiogenesis, and oxidative and inflammatory processes (1). The presented endotype schema thus underscores the strengths of multivariate, data-driven approaches that capture and distill complex biological phenomena into clinically meaningful and practical insights. Second, some biomarkers with known association with IPF such as ITGB6 (integrin subunit β6) (16, 38) did not feature prominently as endotype-defining variables, perhaps as a result of collinearity with other, more informative proteins. Finally, these endotypes were derived using an expert-curated list of 25 biomarkers with positive associations with IPF progression. Whether a similar result may have been found with the use of data-driven approaches for feature selection remains unknown and warrants testing in future studies. Furthermore, the impact of including biomarkers with an inverse correlation with IPF progression, such as FASLG (Fas ligand) (16, 17), also remains unknown.

Although the presented study highlights the feasibility of molecular endotyping in IPF, several key steps remain before translation of these findings to clinical practice. First, LCA models are computationally complex and require prior knowledge of each biomarker’s population distribution to facilitate z-scaling. This is not possible given the semiquantitative nature of these data, which results in batch effects that preclude direct comparison with external cohorts. To establish more reliable biomarker estimates that can be tested in a prospective clinical setting, quantitative data generation is necessary. Several commercially developed high-fidelity custom arrays are now available that have the potential to advance these and other biomarkers toward the clinical arena when technical validation has been performed and generalizable point estimates have been established. The accurate performance of the five-biomarker parsimonious model suggests that a small number of biomarkers may allow for effective molecular classification of this population.

Our study adds to a growing body of literature leveraging omics approaches to enable precision medicine in IPF and other ILDs. For example, a composite 52-gene score was previously shown to improve IPF prognostication and identified patients who may experience improved lung function with antifibrotic therapy (39). Mutations in the TOLLIP gene have been associated with differential response to N-acetylcysteine (40), and a prospective trial is currently under way to evaluate N-acetylcysteine therapy in patients with this mutation (41). Telomere dysfunction has long been implicated in IPF pathogenesis (1), and patients with measurably shortened telomeres have worse outcomes with immunosuppression in IPF and non-IPF fibrotic ILDs (42, 43). Although protein measurements from peripheral circulation, as demonstrated in our study, are readily accessible and have the potential for scalability, integration with other omics modalities or measurements from other body compartments (e.g., BAL) could enhance the depth and richness of biological information captured in an endotyping schema. To this end, the relationship between the molecular endotypes identified in our study and prior omics approaches, and whether these endotypes are conserved across other ILDs and diseases of fibrosis, are an important area for future research.

Our study has several strengths. The molecular endotypes were independently validated in two sizeable cohorts of IPF totaling more than 1,200 patients across a geographically heterogeneous population in the United States. Moreover, because the baseline severity of disease differed between cohorts, the biological similarity of the two endotypes in both cohorts suggests generalizability. Additionally, we accounted for immortal time bias of antifibrotic agent exposure using multiple statistical techniques and observed similar findings of differential treatment response to antifibrotic therapy according to endotype. Finally, the endotypes were derived using a panel of biomarkers selected a priori to the LCA that were enriched for a broad range of protein pathways.

This study has important limitations. First, our study was a secondary analysis of observational cohorts, and, because of sample size considerations, the treatment effect analysis was performed in a pooled cohort. Although these results should be interpreted as hypothesis-generating, they support further research to prospectively confirm differential treatment responses between endotypes in independent cohorts. Second, because these biomarkers have been previously associated with IPF progression in the same cohorts, there may have been selection bias and potential for model overfitting. However, outcome data were not included in the LCA modeling, and many of these proteins have been implicated in IPF pathogenesis in other cohorts and laboratory models, suggesting generalizability. Furthermore, the association between these LCA-defined endotypes and response to antifibrotic therapy is a novel finding with hypothesis-generating therapeutic implications. This outcome was not previously tested and thus did not play a role in protein selection. Nonetheless, further validation of these endotypes and confirmation of differential treatment response in additional cohorts remains an essential next step. Third, we accounted for antifibrotic therapy indication bias by applying inverse probability of treatment weighting, which balanced demographic and clinical characteristics and treating center between patient subgroups. However, there may have been unmeasured confounders, such as coexisting medical conditions, for which we were unable to account. The method to determine antifibrotic treatment exposure also varied slightly between cohorts. Moreover, our study was not sufficiently powered to evaluate for differential treatment response with each individual antifibrotic drug. However, the timing of antifibrotic therapy exposure and the proportions of medication used were comparable between endotypes, suggesting that our findings were not driven by one individual antifibrotic agent. Fourth, because the proportions of antifibrotic therapy–naive patients at the time of the blood draw differed between cohorts, the validation cohort may have been relatively overrepresented in the pooled analysis. However, antifibrotic agent use did not differ statistically between endotypes in the validation cohort, and we accounted for treating center in the IPTW procedure. Finally, although our cohorts enrolled patients across the United States, women, individuals of non-White race, and patients from other countries were underrepresented in the analysis.

In conclusion, in this multicohort study, we identified two novel molecular endotypes of IPF with divergent clinical outcomes and differential responses to antifibrotic therapy. These endotypes could enable a precision medicine approach for future IPF clinical trials.