David Lauer1, Luca Romano Kolly1, Hubert Gabrys2, Matthias Brunner1, Malgorzata Anna Maciukiewicz1, Thomas Frauenfelder3, Stephanie Tanadini-Lang2, Anne-Christine Uldry4, Manfred Heller4, Kerstin Klein5, Oliver Distler6, Janine Gote-Schniering1 and Britta Maurer7, 1University Hospital Bern, Bern, Switzerland, 2University Hospital Zurich, Zurich, Switzerland, 3Dept of Radiology, University Hospital Zurich, University of Zurich, Zurich, Switzerland, 4University of Bern, Bern, Switzerland, 5Department of Rheumatology and Immunology, University Hospital Bern, Bern, Switzerland, 6Department of Rheumatology, University Hospital Zurich, University of Zurich, Zurich, Switzerland, 7University Hospital Bern, University Bern, Bern, Switzerland
Background/Purpose: Quantification of response to anti-fibrotic drugs in patients with fibrosing interstitial lung disease (ILD) relies on repeated pulmonary function tests (PFT) and visual evaluation of HRCT scans, which require long observation periods to achieve conclusive results. HRCT-derived radiomic features, capturing in-depth information of whole-organ properties without bias, may present a more sensitive approach for quantification of treatment response. Here, we studied if radiomic feature changes can classify response to treatment in experimental and human ILD.
Methods: Mice with bleomycin-induced ILD received nintedanib (n=10) or vehicle-only (n=13) from day 7 to 21. Lung µCT scans were acquired before and after treatment for calculation of delta radiomic features (n=1'386). Phosphoproteome profiles were measured in subsets to validate nintedanib target engagement. Whole-lung proteomics was performed in all sample for correlation with delta radiomics. Selected disease markers were measured by gene expression and immunohistochemistry analysis. Delta radiomic features classifying response in mice, as evaluated in univariate analysis, were validated in a nintedanib-treated ILD cohort (n=10), where pre- and post-treatment HRCT scans and PFT data were available. Response in patients was classified by FVC change ≤5% from baseline, resulting in a 5:5 class distribution. All statistical analyses were performed in R.
Results: Molecular readouts in mice indicated that response to nintedanib was heterogeneous, despite a clear separation on µCT-derived lung tissue density and on phosphoproteome level compared to vehicle treatment. Unsupervised hierarchical clustering of delta radiomic features in nintedanib-treated mice revealed two stable clusters. On molecular level, these clusters exhibited differences on expression of fibrosis (Col1a1, Col3a1, Fn1) and drug-related (Tgfb1, Timp1, Cxcl1) targets, in addition to different levels of α-SMA+ myofibroblast and F4/80+ macrophage infiltration. As such, these clusters indicated presence of different treatment response profiles. To define the biological basis underlying cluster separation, we performed correlation analysis of the top discriminative delta radiomic features (AUC≥0.9, n=44) with matched proteomics profiles. Reactome pathway analysis of correlated proteins (ρ≥0.6, p< 0.05) revealed significant enrichment for multiple pathways involved in ILD pathophysiology, including extracellular matrix remodeling and collagen/elastic fiber formation. Validation of the predictive radiomic features in our human cohort demonstrated good performance for treatment response classification (AUC≥0.6 in n=24 features), with two of the best performing features (AUC≥0.8) showing a strong association with fibrotic pathway activation in mice. Our results highlight that molecular changes are paralleled on radiomics level in experimental lung fibrosis and that they are translatable to human ILD.
Conclusion: HRCT-derived delta radiomic signatures may provide a powerful measure for early and accurate classification of patients benefiting from anti-fibrotic therapy.