Prognostic biomarkers for idiopathic pulmonary fibrosis: findings from ISABELA clinical trials

Abstract

Background

Idiopathic pulmonary fibrosis (IPF) is characterised by progressive loss of pulmonary function and poor survival. Although biomarkers for disease progression and mortality exist, their reliability in large studies remains unproven. This study investigates prognostic biomarkers from the ISABELA trials, the largest IPF cohort to date, to identify those predicting worse clinical outcomes.

Methods

Plasma from 1280 IPF patients in ISABELA 1 and 2 (NCT03711162, NCT03733444) was analysed for 17 circulating soluble disease-related biomarkers at multiple time-points and for the MUC5B (rs35705950_T) genotype. Statistical learning algorithms investigated biomarker levels/status with disease progression (≥10% decline in forced vital capacity (FVC) or mortality within 1 year) and pharmacotherapy.

Results

Patients with ≥10% annual decline in FVC had higher median baseline of matrix metalloproteinase-7 (MMP-7) versus those with <10% decline (5.5 versus 4.2 µg·L⁻¹; p<0.005). Patients with baseline MMP-7 ≥5.2 μg·L⁻¹ and/or C-C motif chemokine ligand 18 (CCL18) ≥75.2 μg·L⁻¹ had increased risk of mortality (p<0.0001); with patients having both elevated biomarkers at an even greater risk. Machine learning identified CCL18 changes by week 26 as a predictor of disease progression. The rs35705950_T genotype predicted neither mortality nor disease progression.

Conclusions

We provide new insights into the prognostic value of MMP-7 and CCL18 in identifying high-risk IPF patients in the largest cohort to date. The combination of high baseline MMP-7 and CCL18 levels, along with longitudinal changes in CCL18, has the potential to enhance risk stratification and support efficacy assessment and monitoring in clinical trials.

Shareable abstract

In the largest IPF cohort to date, MMP-7 and CCL18 are recognised as prognostic biomarkers for disease progression and mortality. Their combination and longitudinal CCL18 changes may enhance risk stratification and support treatment efficacy, respectively. https://bit.ly/4oqmz05

Introduction

Idiopathic pulmonary fibrosis (IPF) is a chronic, fibrotic lung disease of unknown aetiology characterised by pulmonary function loss, poor survival and patient heterogeneity. While forced vital capacity (FVC) decline is an accepted marker of IPF progression and mortality in clinical trials [1], predicting patient progression remains challenging.

Recent studies highlight significant structural lung changes in IPF, including distal lung proximalisation, alveolar epithelial cell loss, fibroblast and macrophage shifts, and aberrant basaloid cell emergence [2–4]

Peer-reviewed studies consistently identify recurring blood protein biomarkers linked to IPF progression. These include markers of inflammation, epithelial damage and fibrosis with prognostic value. Inflammation markers include C-C motif chemokine ligand (CCL) 18, intracellular adhesion molecule 1 (ICAM1) and monocyte count [5], with CCL18 predicting FVC decline [6–8] and ICAM1 indicating poor transplant- and progression-free survival [9]. CCL18 reflects monocyte-derived alveolar macrophages [6, 7], while ICAM1 is expressed in epithelial and endothelial cells [10]. Elevated monocytes correlate with increased risks of FVC decline hospitalisation and mortality [5]. Fibrosis markers periostin and matrix metalloproteinase MMP-7 associated with fibroblasts and aberrant basaloid cells [2, 9] respectively, are linked to progression and poor survival outcomes [6, 9, 11, 12].

Further, MUC5B promoter rs35705950 T/G gene polymorphism increases IPF susceptibility but may slow its progression [13, 14].

Predicting IPF progression remains challenging, but incorporating prognostic biomarkers may improve trial design. Although ziritaxestat, an autotaxin inhibitor, showed no efficacy in the ISABELA 1 and 2 studies [15], the analysis of these exceptionally large IPF cohorts highlighted MMP-7 and CCL18 as strong prognostic biomarkers, supporting their clinical relevance.

Methods

Study design

The study designs of ISABELA 1 and 2 (NCT03711162, NCT03733444) have been described previously [15]. Patients were randomised 1:1:1 to placebo, ziritaxestat 200 mg or ziritaxestat 600 mg, and stratified by use of standard of care (SoC) (nintedanib, pirfenidone or neither).

Patients and samples

All 1306 randomised patients, pooled from the ISABELA 1 and 2 clinical trials, consented to plasma biomarker analysis; 929 patients consented to optional genetic sequencing. Whole blood was collected at baseline and K2EDTA plasma was collected at baseline and weeks 12, 26 and 52.

Clinical parameters

Clinical variables (haemoglobin-corrected percent predicted diffusing capacity of the lung for carbon monoxide (D_LCO) (% pred), FVC (% pred), body mass index (BMI), 6-min walk test distance (6MWD), as well as adjudicated major events, all-cause mortality, disease progression, respiratory-related hospitalisation (RRH) and acute IPF exacerbation) for the ISABELA trials have been described previously [15].

Assay methodology

17 biomarkers were immunoassayed and 11 with acceptable quality-control (QC) criteria (CCL2, CCL18, C-X-C motif chemokine ligand 13 (CXCL13), ICAM-1, MMP-1, MMP-7, platelet-derived growth factor AA (PDGF-AA), periostin, surfactant protein D (SP-D), chitinase-3-like protein 1 (YKL-40) and autotaxin) were kept for analyses.

Whole blood for genetic sequencing was analysed as described in the supplementary materials. Due to its strong relevance in IPF, rs35705950_T (MUC5B) was also analysed.

Statistical analysis

Given the log-likelihood profiles of Box–Cox power transformation, biomarker values were log₁₀ transformed to normalise their residual distributions. The significance of the change from baseline of the log₁₀ value of biomarkers was estimated using the Wilcoxon signed-rank test [16].

Biomarker readouts were compared between subgroups using Kruskal–Wallis one-way ANOVA [17] (when comparing >2 groups), followed by pairwise comparisons using Dunn [18] or Mann–Whitney U tests [19] (when comparing two groups). Baseline analyte values were compared between Asian and White patients using a linear regression model adjusted for BMI. The analysis of biomarker baseline levels and FVC decline categories in figure 1 was based on data available up to week 52, as follows: a) 659 of 1266 patients up to week 52; b) 75 of 239 patients; and c) 546 of 969 patients.

FIGURE 1.

Baseline matrix metalloproteinase-7 (MMP-7) levels by category of annual forced vital capacity (FVC) decline (absolute decline of ≥10% or <10% in FVC % predicted). a) All patients. b) Asian patients. c) White patients. Number of patients in each treatment arm: a) placebo (n=222), ziritaxestat 200 mg (n=231), ziritaxestat 600 mg (n=206); b) placebo (n=22), ziritaxestat 200 mg (n=27), ziritaxestat 600 mg (n=26); c) placebo (n=191), ziritaxestat 200 mg (n=191), ziritaxestat 600 mg (n=164). µ^: median. ***: p<0.001.

Random survival forest (RSF) analyses were conducted using the randomForestSRC-R package (version 3.2.0). Missing values were imputed and 20 000 trees were grown with default settings for splitting variables, terminal node size and split points. Prediction error was calculated as 1-Harrell concordance index (c-index) on out-of-bag data, with variable importance assessed via permutation. The c-index is a generalisation of the area under the curve for censored time-to-event data, with values of 0.5 for random prediction and 1 for perfect prediction. Association of biological parameters with time to clinical outcomes was also tested using Cox proportional hazards models [20].

Kaplan–Meier curves were produced and log-rank tests performed to compare time to clinical worsening in patient subgroups.

Results

Patient disposition and baseline characteristics

In the pooled dataset, 1266 patients had ≥1 baseline value for one soluble biomarker (497 and 769 patients in ISABELA 1 and 2, respectively) (table E1 in the online data supplement). The disease characteristics of this subpopulation were similar to those of the main study (mean FVC % pred of 79.5% and 77.3% in ISABELA 1 and 2, respectively) [15], fewer patients were not on SoC (377 versus 439 receiving nintedanib and 450 receiving pirfenidone). The ISABELA 2 subset had a higher proportion of Asian patients than the ISABELA 1 subset (28.1% versus 5.8%).

Baseline soluble biomarker levels

Soluble biomarker levels did not differ between treatment arms, but varied among SoC-treated patients. Median baseline CCL18 and MMP1 levels were higher in nintedanib-treated patients than in those without SoC or on pirfenidone (CCL18: 85.1 (95% CI 82.3–88.1) versus 77.3 (95% CI 72.4–81.7) or 73.4 (95% CI 70.2–76.9) µg·L⁻¹, p<0.0002; MMP1: 2.55 (95% CI 2.40–2.75) versus 1.48 (95% CI 1.36–1.60) or 1.42 (95% CI 1.31–1.51) µg·L⁻¹, p<10⁻²²) (figure E1A,D).

Median baseline MMP-7 levels were higher without SoC than with nintedanib or pirfenidone (5.14 (95% CI 4.85–5.56) versus 4.59 (95% CI 4.36–4.75) or 4.64 (95% CI 4.31–4.92) µg·L⁻¹, p=0.04) (figure E1G). When analysed by ethnicity (Asian and White), these differences were not significant (figure E1).

Irrespective of SoC, baseline levels of several biomarkers (including CCL18 and MMP-7) differed between ethnicities (Asian versus White) (figure E2).

BMI adjustment had minimal impact on these findings (data not shown).

Patient genotype

A subgroup of 896 patients had ≥1 single nucleotide polymorphism identified for any gene assessed; their disease characteristics resembled the main study (e.g. mean FVC % pred 78.6% and 77.4% in ISABELA 1 and 2, respectively) (table E2).

In this subpopulation, MUC5B rs35705950_T allele frequency was 33.1% (table E3), consistent with IPF data in White/European populations [21].

Frequency differed by ethnicity, at 5.69% in Asian patients and 40.2% in White patients (Table E3), matching 1000 Genomes data [22].

Baseline MMP-7 predicts lung function decline

Baseline plasma biomarkers were assessed by categorising patients with FVC % pred data to week 52 into ≥10% or <10% annual decline. Median baseline MMP-7 levels were 1.31-fold higher in the ≥10% group (5.5 versus 4.19 µg·L⁻¹, p=0.0000192) (figure 1a), with consistent differences in Asian and White patients (figure 1b,c). Other analytes showed no significant differences (data not shown). In the placebo arm, baseline MMP-7 levels were 1.2-fold higher in the ≥10% group, though not statistically significant (p=0.06) (figure E3).

D_LCO, MMP-7 and CCL18 have distinct prognostic value for mortality and disease progression

An RSF analysis identified baseline predictors of all-cause mortality using clinical variables (sex, age, FVC % pred and D_LCO % pred, BMI, 6MWD, treatment, SoC stratum and proton pump inhibitor use) and QC-passed biomarkers (CCL2, CCL18, CXCL13, ICAM1, MMP-1, MMP-7, PDGF-AA, periostin, SP-D, YKL40 and autotaxin). D_LCO % pred was the strongest predictor of mortality, followed by MMP-7, CCL18 and FVC % pred (c-index: 0.763) (figure 2a). Using a Cox model with the top three variables, a 10% higher MMP-7 or CCL18 level was associated with 4.4% or 8.7% higher mortality risk, respectively (log₁₀ MMP-7 hazard ratio (HR) 2.85 (95% CI 1.13–7.17), p<0.05; log₁₀ CCL18 HR 7.53 (95% CI 2.13–26.61), p<0.01), whereas a 10% decrease in D_LCO % pred increased mortality risk by 67%.

FIGURE 2.

Random survival forest of baseline analytes predicting a) all-cause mortality for adjudicated events and b) disease progression for adjudicated events. The random forest variable importance (VIMP) can be interpreted as the increase of the standardised mean squared error in percentage when the corresponding predictor is randomly permutated into a noise variable. Positive VIMP values identify variables that are predictive after adjusting for all the other variables. 6MWD: 6-min walk test distance; BMI: body mass index; CCL: C-C motif chemokine ligand; CXCL13: C-X-C motif chemokine ligand 13; D_LCOHCPP: diffusing capacity of the lung for carbon monoxide, corrected for haemoglobin concentration (% predicted D_LCO); FVCPP: forced vital capacity % predicted; ICAM1: intracellular adhesion molecule 1; MMP: matrix metalloproteinase; PDGF-AA: platelet-derived growth factor AA; PPI: proton pump inhibitor; SP-D: surfactant protein D; YKL-40: chitinase-3-like protein 1.

The RSF model predicting disease progression ranked MMP-7, 6MWD, SoC stratum and D_LCO % pred as top predictors; FVC % pred ranked eighth (c-index: 0.609) (figure 2b). Associations were confirmed in Cox models despite the RSF's limited predictive power. A 10% higher MMP-7 level was associated with 3% greater progression risk (log₁₀ MMP-7 HR 2.02 (95% CI 1.35–3.03); p<0.001). A 10% lower BMI and a 10-point lower D_LCO % pred corresponded to 13% and 11% higher risk, respectively. A 50-m shorter 6MWD increased risk by 7%. Nintedanib and pirfenidone reduced progression risk by 49% and 26%, respectively, compared to patients receiving no antifibrotic (AF) (nintedanib HR 0.51 (95% CI 0.38–0.69), p<0.001; pirfenidone HR 0.74 (95% CI 0.56–0.97), p<0.05).

We cross-validated the cohorts, using ISBAELA 1 and 2 as reciprocal test sets. Validation with ISABELA 2 yielded c-indices of 0.75 (mortality) and 0.59 (progression); ISABELA 1 validation gave 0.78 and 0.61, respectively. The predictive performance was stronger for mortality (c-index: >0.75) than progression (c-index: ∼0.60).

Median MMP-7 (5.2 μg·L⁻¹) and CCL18 (75.2 μg·L⁻¹) thresholds were derived from baseline and placebo data without SoC to confirm prognostic value.

Kaplan–Meier curves stratified patients by baseline MMP-7 (<5.2 µg·L⁻¹ (low) or ≥5.2 µg·L⁻¹ (high)) and baseline CCL18 levels (<75.2 µg·L⁻¹ (low) or ≥75.2 µg·L⁻¹ (high)). High MMP-7 was associated with greater mortality (log-rank p<0.0001) (figure 3a) and progression risk (log-rank p<0.01) (figure E4A), with significant differences maintained in patients receiving nintedanib or pirfenidone (p<0.01), but not in those without SoC (figure E5). High CCL18 was associated with higher mortality (p<0.0001) but not progression (figures 3b and E4B–C). Both biomarkers predicted increased RRH risk (p<0.01) (figures E4D–E) and patients with high levels of both had greater mortality (p<0.00001) and RRH risks (figures 3c and E4F).

FIGURE 3.

Kaplan–Meier curves for all-cause mortality for adjudicated events with a) patients stratified by matrix metalloproteinase-7 (MMP-7) (cut-off: 5.2 µg·L⁻¹), b) patients stratified by C-C motif chemokine ligand-18 (CCL18) (cut-off: 75.2 µg·L⁻¹) and c) patients stratified by MMP-7 (cut-off: 5.2 µg·L⁻¹) and CCL18 (cut-off: 75.2 µg·L⁻¹). The log-rank p-values from the Cox model compare each curve to the reference (red curve) and include an overall test for group impact (black p-value).

Monocyte count is not prognostic for mortality or disease progression

The prognostic value of monocytes was assessed by generating Kaplan–Meier curves for each time-to-event outcome in the whole patient cohort using previously reported cut-off values [5]. Considering all patients, no association between monocyte count and mortality, progression, or hospitalisation was observed [5] (figure E6).

MUC5B (rs35705950_T) genotype is not prognostic for mortality or disease progression

The prognostic value of genotypes was assessed by generating Kaplan–Meier curves for each time-to-event outcome in the whole patient cohort. The analysis showed no association between MUC5B rs35705950_T allele and mortality, disease progression, or acute IPF exacerbation in all or only White patients (figure E7).

Investigating the treatment effect of ziritaxestat

To determine treatment effect on biomarker levels, median changes from baseline were plotted over time. MMP-7 levels increased to week 52, with no treatment differences (figure 4 and table E4). Autotaxin increased in all arms, with greater increases with ziritaxestat; periostin increased only with ziritaxestat (figures E8A and B and table E4). Week 12 changes in autotaxin or periostin did not correlate with week 52 FVC changes (data not shown).

FIGURE 4.

Median change from baseline in matrix metalloproteinase-7 (MMP-7) by treatment arm: placebo (grey), ziritaxestat 200 mg (red), ziritaxestat 600 mg (blue). All p-values are provided in supplemental table E4. Significance is indicated for change from baseline per treatment arm. SoC: standard of care. The n value above each time-point represents the number of patients. p-values were calculated using the Wilcoxon signed-rank test. *: p<0.05. **: p<0.01. ***: p<0.001.

Although SoC assignment was not randomised, biomarker changes were analysed by stratum. ICAM1 levels increased to week 52 in placebo patients without SoC but not with SoC (figure E8C and table E4). No similar trends were seen for other analytes (data not shown).

Week 26 change in CCL18 has distinct prognostic value for disease progression and acute IPF exacerbation

RSF analysis was conducted to assess if biomarker changes predict all-cause mortality, disease progression, or acute IPF exacerbation. Fold-change to week 26 in each analyte and baseline clinical covariates (D_LCO % pred, FVC % pred, 6MWD and BMI) were included. The top soluble biomarkers predicting mortality were fold-changes in CCL18 and SP-D, ranking below baseline D_LCO % pred and FVC (c-index: 0.752) (figure E9A). Disease progression was best predicted by fold-change in CCL18, followed by ICAM1 and CCL2, ranking below several clinical covariates (c-index: 0.59) (figure E9B). Acute exacerbation was predicted by fold-changes in CCL18, CCL-2 and CXCL13, ranking below baseline D_LCO % pred (c-index: 0.67) (figure E9C).

Discussion

IPF progression is highly variable, with some patients declining rapidly while others slowly [23]. This variability reflects lung function, tissue remodelling, cellular dysfunction, genetics and comorbidities; factors not fully captured by current tools.

Using longitudinal blood samples from the largest IPF trials to date (ISABELA 1 and 2), we confirmed MMP-7 and CCL18 as key prognostic biomarkers [6, 9, 11, 24]. Baseline levels of MMP-7, CCL18 and CCL18 change over time, may support clinical stratification and monitoring.

Identifying patients at risk of major FVC decline or death is crucial for innovating trial design in IPF. We investigated baseline biomarker associations with FVC decline, mortality and progression. Bauer et al. [11] previously associated baseline MMP-7 levels with FVC decline over 4 months, although BUILD3 patients did not receive current SoC.

Baseline MMP-7 levels identified patients with greater FVC decline by week 52, with higher levels in those with ≥10% decline, regardless of SoC. This supports a meta-analysis of eight studies (1383 patients) linking baseline MMP-7 to 12-month progression [25].

FVC decline is crucial for risk stratification in IPF and as a surrogate for mortality but its variability limits predictive accuracy. We used RSF to assess baseline biomarkers for mortality and disease progression to week 52. The model predicted mortality well and had limited accuracy for disease progression.

The top baseline mortality predictors were D_LCO % pred, MMP-7 and CCL18. MMP-7 was the strongest progression predictor, confirmed by Cox models for both outcomes. Prior studies linked MMP-7 to disease worsening (defined as confirmed decrease from baseline absolute FVC of ≥10% and absolute D_LCO ≥15%, or acute exacerbation up to end of study) [11, 25] and our findings validate its prognostic value in a large IPF cohort on current SoC. However, MMP-7's ability to predict time to RRH requires confirmation in an independent cohort.

Our findings confirmed D_LCO as a mortality predictor [26] and reduced progression risk for patients receiving nintedanib (49%) or pirfenidone (26%) [27, 28]. Baseline CCL18 and monocyte count were weakly correlated (r=0.16, p<0.00001; data not shown), and monocyte count did not predict mortality or progression [5], possibly due to prior AF treatment. High baseline MMP-7 levels were associated with increase mortality and progression risks, regardless of treatment or SoC.

For SoC, optimal disease progression cut-offs were similar to or above the median computed threshold (4.88 and 6.67 µg·L⁻¹, respectively). A higher MMP-7 threshold in SoC patients, as shown by Adegunsoye et al. [29], aligns with higher baseline prognostic biomarker levels in AF-treated patients, linked to transplant-free survival risk. In patients not on SoC, the optimal MMP-7 cut-point for disease progression (3.9 µg·L⁻¹) was lower than the median threshold but consistent with Bauer et al. [11] and Richards et al. [9]. These findings underscore the need for further clinical validation of these thresholds within a standardised clinical assay [30].

Patients with high baseline CCL18 levels had higher risks of mortality and RRH. Previous studies have associated high baseline CCL18 levels with mortality and progression [31, 32], using higher baseline thresholds (120 and 140 µg·L⁻¹), likely due to subjects with lower baseline FVC/D_LCO, different assays and the use of serum instead of plasma. Patients with high baseline levels of both MMP-7 and CCL18 had the highest risk of death. No therapeutic benefit for ziritaxestat was observed in subgroups of patients stratified by MMP-7 and CCL18 thresholds. Evaluating biomarkers related to therapeutic targets is crucial for identifying high-risk populations, enhancing treatment efficacy and reducing side-effects.

Although the MUC5B rs35705950_T allele is a known IPF risk factor [33], it did not independently predict mortality, disease progression or exacerbation in these studies. Large genome-wide association studies found no significant association between rs35705950 and disease progression or survival [33, 34], suggesting limited prognostic value. Nonetheless, certain variants may still influence treatment response.

Serial changes in prognostic biomarkers can predict mortality [35], but limited data prevented extensive exploration. To support early efficacy predictions in future studies, we explored the relationship between 6-month biomarker changes and clinical outcomes. Autotaxin, MMP-7 and periostin levels increased by week 52 across all treatments. Autotaxin increase was not associated with a lack of target engagement [36]. An increase in fibrotic biomarkers, across treatment arms or SoC strata, suggests that AF therapy does not reduce or stabilise fibrosis markers. Limited evidence exists on the effect of AF therapy on prognostic markers [29, 37] and its impact on fibrotic markers such as MMP-7 remains unproven. Additionally, changes in autotaxin, MMP-7 or periostin levels showed no distinct prognostic value for the measured clinical outcomes. The prognostic potential of autotaxin warrants further investigation. Fold-change to week 26 in CCL18 was the strongest predictor of the clinical outcomes. Although baseline CCL18 is a known prognostic marker of FVC decline [6–8], its fold-change has not been previously linked to mortality and disease progression. Moreover, building on the evidence from Schupp et al. [38], we demonstrated that increases in CCL18 predict acute exacerbations.

This biomarker study has some limitations, including incomplete 52-week data due to early termination of the ISABELA studies, a lack of randomisation by SoC limiting stratification analysis and the exclusion of patients with severe disease (e.g. FVC <45%).

The integration of prognostic biomarkers into clinical trial design, well established in oncology, is now advancing in IPF. Biomarkers such as MMP-7 and CCL18, which reflect key fibrotic and immune–epithelial processes, offer valuable tools for patient selection and stratification when used alongside clinical and imaging data.

What was once aspirational, measuring and applying blood biomarkers at scale in large, controlled trials, is now feasible. This study marks a shift from small exploratory analyses to the inclusion of over 1200 patients in a prospective and longitudinal trial setting. The innovation lies not in the novelty of the biomarkers, but in their robust and reproducible application across a large IPF population, enabling more precise and informative trial designs.

These biomarkers can be applied at several stages of clinical trials, including risk stratification at baseline, prognostic enrichment to select patients likely to progress and longitudinal monitoring to support adaptive trial designs. Their integration may increase event rates and reduce sample sizes, with potential applicability across progressive fibrotic interstitial lung diseases (ILDs).

Ongoing efforts by the PROLIFIC consortium (www.pulmonaryfibrosis.org/prolific) to standardise assays for established biomarkers such as MMP-7 and CCL18 support their clinical trial implementation. While advances in proteomics may bring forward new biomarkers, the current innovation lies in applying those we already trust, laying the foundation for biomarker-guided studies in IPF that enhance efficiency, precision and, most importantly, patient outcomes. Future work should focus on validating these markers across diverse ILD populations and disease stages, while optimising composite biomarker strategies. Such advances will enable more precise, personalised approaches to drug development across the spectrum of fibrotic lung diseases.

Supplementary material

Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author.

Supplementary material

00893-2025.SUPPLEMENT

Data availability

Data will not be made available as informed consent was not obtained for the sharing of individual patient data.