Rationale: Usual interstitial pneumonia (UIP) is the histopathologic hallmark of idiopathic pulmonary fibrosis. Although UIP can be detected by high-resolution computed tomography of the chest, the results are frequently inconclusive, and pathology from transbronchial biopsy (TBB) has poor sensitivity. Surgical lung biopsy may be necessary for a definitive diagnosis.
Objectives: To develop a genomic classifier in tissue obtained by TBB that distinguishes UIP from non-UIP, trained against central pathology as the reference standard.
Methods: Exome enriched RNA sequencing was performed on 283 TBBs from 84 subjects. Machine learning was used to train an algorithm with high rule-in (specificity) performance using specimens from 53 subjects. Performance was evaluated by cross-validation and on an independent test set of specimens from 31 subjects. We explored the feasibility of a single molecular test per subject by combining multiple TBBs from upper and lower lobes. To address whether classifier accuracy depends upon adequate alveolar sampling, we tested for correlation between classifier accuracy and expression of alveolar-specific genes.
Results: The top-performing algorithm distinguishes UIP from non-UIP conditions in single TBB samples with an area under the receiver operator characteristic curve (AUC) of 0.86, with specificity of 86% (confidence interval = 71–95%) and sensitivity of 63% (confidence interval = 51–74%) (31 test subjects). Performance improves to an AUC of 0.92 when three to five TBB samples per subject are combined at the RNA level for testing. Although we observed a wide range of type I and II alveolar–specific gene expression in TBBs, expression of these transcripts did not correlate with classifier accuracy.
Conclusions: We demonstrate proof of principle that genomic analysis and machine learning improves the utility of TBB for the diagnosis of UIP, with greater sensitivity and specificity than pathology in TBB alone. Combining multiple individual subject samples results in increased test accuracy over single sample testing. This approach requires validation in an independent cohort of subjects before application in the clinic.
Interstitial lung disease (ILD) represents a heterogeneous group of diffuse parenchymal disorders with similar clinical presentation, but varying rates of disease progression, treatment response, and survival (1). Internationally recognized guidelines recommend the multidisciplinary evaluation of clinical, radiologic, and pathologic disease features in the diagnosis and management of ILD (1, 2). The patchy interstitial fibrosis, collagen deposition, and architectural distortion characteristic of the usual interstitial pneumonia (UIP) pathologic pattern, evident by surgical pathology or at the macro scale by high-resolution computed tomography (HRCT), is generally associated with poor prognosis and survival (3–6). The identification of UIP of unknown cause or association defines a diagnosis of idiopathic pulmonary fibrosis (IPF), the most common and severe of the ILDs (3, 7, 8). Newly available antifibrotic drugs pirfenidone and nintedanib have shown promise in slowing disease progression in patients with IPF (2).
Pathologic pattern diagnoses derived from less invasive bronchoscopic biopsies are helpful to exclude sarcoidosis, lymphangitic carcinoma and infection (9). However, the reliable identification of UIP pathology in transbronchial biopsies (TBBs) is challenged by the difficulty of sufficient sampling of alveolated lung parenchyma and heterogeneous disease distribution (7, 10). In retrospective studies with high TBB sampling adequacy rates, UIP was confirmed in 30–43% of patients with clinical and radiographic features consistent with UIP (11, 12), with a third and a fourth study reporting confirmation rates of 10% or less (13). This has led many to evaluate alternate bronchoscopic biopsy methods that may provide greater alveolar sampling (14, 15). These are currently limited by availability and a lack of large multicenter studies (16). There is a clear need for a more robust method of detecting UIP using bronchoscopy.
We hypothesized that a genomic classifier could detect a gene expression signature of UIP from TBBs with high accuracy in a diverse population of patients with ILDs. We used a machine learning methodology on exome-enriched transcriptional data to train an algorithm (classifier) to differentiate UIP from among a wide variety of ILDs encountered in clinical practice. We then demonstrated that this approach accurately predicts the presence of UIP in a multicenter validation cohort. This diagnostic test using genomic techniques could potentially reduce the need for surgical lung biopsy (SLB) in the diagnosis of ILD, and inform the diagnosis and treatment of patients with IPF.
Subjects with suspected ILDs with planned, clinically indicated lung biopsy procedures were eligible for enrollment in an institutional review board–approved multicenter sample collection study (Bronchial Sample Collection for a Novel Genomic Test [BRAVE]) after informed consent (see the online supplement for additional detail). When available, HRCT scans were reviewed by an expert radiologist (D.A.L.). Pathologic diagnoses were made by expert lung pathologists (A.-L.A. Katzenstein [retired], T.V.C., J.L.M., and S.D.G.) from SLB (BRAVE-1), TBB (BRAVE-2), or cryobiopsy (BRAVE-3) specimens, according to a centralized review process (19). Biopsy slides prepared by study sites and tissue site-of-origin information were provided to two pathologists for blinded, independent review. Each pathologist made a pathological diagnosis, either a specific pathologic diagnosis or a description of the findings present, for each lobe sampled. Diagnoses were evaluated for categorical concordance (see the online supplement for additional details). In the event of agreement, a categorical UIP or non-UIP “truth” label was assigned. In the event of disagreement, blinded review was performed by a third pathologist. If two of three diagnoses showed concordance, a truth label was assigned; otherwise, consultation among the three pathologists was used to achieve a consensus diagnosis. Laboratory and analytical personnel remained blinded to the pathology diagnoses and labels of the test set during algorithm development and lock, until test set scoring.
Up to five paired TBB samples were collected concurrently from each subject during the clinically indicated diagnostic procedure described previously here and placed into RNAprotect (Qiagen, Valencia, CA) preservative for molecular testing. Although the lung tissue sampling was performed at the discretion of the proceduralist, guidance was provided to collect two upper and three lower lobe TBB samples. Reference labels of UIP or non-UIP were assigned to TBB samples based on the diagnostic pathology determined for the same lobe of origin. Within-subject TBB mixtures were composed of RNA from all TBBs sampled for that patient, whether or not pathology diagnoses were available for all lobes. Reference labels for subject mixtures were inferred from diagnostic lobe labels according to the following rules: if any lobe was diagnostic for UIP, the subject mixture was given a UIP label; if any lobe was diagnostic for non-UIP, all other lobes were considered to be non-UIP or nondiagnostic in order to assign a non-UIP label to that subject mixture. Subjects with nondiagnostic or unclassifiable fibrosis in all lobes remained without reference labels.
We generated exome enriched RNA sequence data from TBB samples (see the online supplement for sample extraction, library preparation, and sequencing details). After data quality filtering, expression count data for 17,601 Ensembl genes were normalized with respect to library size and transformed to log2 scale by variance-stabilizing transformation and input to machine learning algorithms. Model feature selection and parameter estimation were performed by logistic regression with elastic net penalty (20). Parameter tuning and training performance was evaluated by leave-one-patient-out cross-validation. All features and parameters used by the classifier, to include the score threshold (decision boundary) that separates a UIP call from a non-UIP call, were determined using only the training subjects, and locked before the scoring of the test set.
All statistical analyses were performed with the use of R software, version 3.2.3 (https://www.r-project.org). All confidence intervals (CIs) are 95% exact binomials.
We used DESeq2 (21) to identify differential expression between UIP and non-UIP TBBs for the 84 subjects with central pathology reference labels. Ensembl genes significantly up-regulated in UIP (n = 926) and in non-UIP (n = 1,330) at false discovery rate adjusted P values of 0.05 or less were used as input to the PANTHER (Protein ANalysis THrough Evolutionary Relationships) classification system for pathway overrepresentation analysis (www.pantherdb.org; Web version 11.0, released 2016-07-15) (22). PANTHER pathways were curated to remove general or redundant pathway classifications, and ordered by significance.
A group of 140 subjects were evaluated for use in developing the genomic classifier (Figure 1). We excluded 27 subjects with insufficient tissue sampling such that local or central pathology returned a nondiagnostic reference label, leaving 113 subjects (496 TBB samples) from 18 clinical sites for laboratory processing and further pathology review (see Table E1 in the online supplement for subject information). We subsequently excluded 18 additional subjects with nondiagnostic pathology or unclassifiable fibrosis, 6 subjects with histopathology requiring unblinded review (i.e., conferral) to achieve a diagnosis, and 1 subject diagnosed with lung cancer, resulting in 88 subjects with diagnostic UIP or non-UIP pathology in at least 1 lung lobe.
We extracted total RNA and generated RNA-seq data for the 496 TBB samples collected from all 113 subjects. Of the 88 subjects with usable reference labels, 4 subjects lacked total RNA and/or sequence data of sufficient quality for use in the study, and were thus excluded. We ultimately generated usable UIP/non-UIP labels and high-quality sample data for 84 subjects, 60 (72%) of whom had pathology diagnoses determined from SLB, 17 (20%) from cryobiopsy, and 7 (8%) from TBB (Table 1).
|Training Set||Test Set||Total|
|No. of subjects||53||31||84|
|Age, median (range)||63.5 (31–88)||62 (18–78)||63 (18–88)|
|Male sex, n (%)||26 (49)||14 (45)||40 (48)|
|Smoking history, yes, n (%)||34 (64)||19 (61)||53 (63)|
|UIP prevalence by pathology|
|By surgical lung biopsy, n UIP (%)||26 of 38 (68)||17 of 22 (77)||43 of 60 (72)|
|Classic UIP, n||11||9||20|
|Difficult UIP, n||3||2||5|
|Favor UIP, n||0||2||2|
|By cryobiopsy, n UIP (%)||6 of 11 (55)||2 of 6 (33)||8 of 17 (47)|
|Favor UIP, n||5||1||6|
|By transbronchial biopsy, n UIP (%)||1 of 4 (25)||0 of 3 (0)||1 of 7 (14)|
|Difficult UIP, n||1||0||1|
|Total UIP prevalence, n UIP (%)||33 of 53 (62)||19 of 31 (61)||52 of 84 (62)|
|UIP prevalence by radiology|
|Definite UIP, n||5||2||7|
|Probable UIP, n||2||3||5|
|Possible UIP, n||1||0||1|
|Total UIP prevalence, n UIP (%)||8 of 52 (15)||5 of 27 (19)||13 of 79 (16)|
We prospectively defined an algorithm training group of 53 subjects and a separate test group of 31 subjects. To initiate training of the algorithm, we assigned a group of subjects that became the core of the training set. As more subjects were accrued, we utilized a third party to randomly assign additional subjects to the training group and to the test group, with periodic rebalancing of pathology subtypes between the two groups as that information became available. Thus by design, UIP prevalence was balanced between the training and testing groups (Table 1). The prevalence of UIP in our study is higher in subjects with SLB than in bronchoscopic biopsies (72% vs. 47% [cryobiopsy] vs. 14% [TBB]), with definitive UIP most commonly identified in SLBs (Table 1). Central expert radiologist review of HRCTs performed as a part of routine clinical care provides an independent estimate of radiologic UIP prevalence in this study. The prevalence of HRCT UIP pattern in our study is 16%, compared with 62% by all pathology biopsy types (Table 1), highlighting the lower sensitivity of HRCT in identifying UIP.
We evaluated multiple normalization schemes, feature selection, and machine learning algorithms on our training set of 170 TBB samples from 53 subjects, using a variety of genomic and clinical features. Within the training set, we observed the highest and most stable classification performance from a logistic regression model with elastic net penalty trained on expression count data, which uses 169 genes as features (see Table E2 for the list of classifier genes). This model achieves a receiver operator characteristic (ROC) area under the curve (AUC) in cross-validation on the training set of 0.85 when each TBB sample is scored separately (Figure 2A). We defined a decision boundary targeting high specificity (92% [CI = 81–97%]), and observed a corresponding sensitivity of 65% [CI = 55–74%] in training. On the independent test set of 113 samples from 31 subjects, the classifier shows an ROC-AUC of 0.86, with sensitivity of 63% (CI = 51–74%) and specificity of 86% (CI = 71–95%) when each TBB sample was scored (Figure 2B). Cross-validation performance that generalizes to a validation cohort suggests that robust training was achieved, despite the relatively modest cohort size.
Multiple TBB samples were collected per subject, typically two to three per lung lobe, to mitigate possible disease and sampling heterogeneity effects that could result in training error or false test calls. For most subjects, our sample-level classifier correctly detected disease in more than one sample of the available TBBs per subject, consistent with overall high sample-level test accuracy (Figure 2). This raised the possibility that the detection of UIP might be most accurate if multiple TBB samples from each patient are pooled for testing.
We separately processed the TBBs from both training and test group subjects as pools. The pools were composed of total RNA from each available TBB, mixed by equal RNA mass. We required a minimum of three TBBs per subject for pooling, based on a simulation of mixtures containing two to five TBBs per subject (see the online supplement for more details). Thus, 38 of the 53 training subjects (72%) and 27 of the 31 test set subjects (87%) had sufficient numbers of samples for pooling. One test subject was pooled and processed, but the resulting sequence data did not meet quality criteria, and were thus excluded. We then evaluated classification performance on the subjects with pooled samples.
Overall, classification performance on pooled TBB samples is as high or higher than performance on individual TBB samples. ROC-AUCs of 0.91 and 0.92 are achieved in pooled samples on the training and test sets, respectively (Figures 2A and 2B). In pooled TBBs, the classifier achieves specificity of 93% and sensitivity of 74% in the training subjects, and specificity of 100% and sensitivity of 59% in the test subjects (Figure 3). Substantially identical performance is observed when the three training subjects and one test subject with reference labels determined from TBB (e.g., the BRAVE-2 subjects) are removed from the pooled sample performance determination (Figure 3). Thus, a single test result derived from a pool of three to five TBBs per subject may provide the greatest and most accurate separation of subjects with UIP from subjects without UIP.
Given the recognized disease heterogeneity in the lungs of patients with ILD (4, 23–25), the finding that strong classifier performance can be obtained with variable sampling of the lungs prompted the question of whether adequate alveolar sampling is necessary when using a genomic approach. We hypothesized that, if accurate classification of UIP versus non-UIP required gene signals from alveolar cells, then those samples with a paucity of alveolar cells should give rise to more classifier errors (particularly false negatives) than samples with greater alveolar content. To address this question, we developed two semiquantitative genomic measures of alveolar content, one for type I and one for type II alveolar cells, based on gene expression (see the online supplement for methods and analysis). We then determined the average level of type I and type II gene expression in individual TBB samples, grouped by classification correctness (true negatives, false negatives, true positives, and false positives; Table 2). Using an ANOVA, we note no significant association between classification correctness and type I alveolar content. Although type II content is correlated with classification accuracy, the direction is such that lower alveolar type II expression is seen among true UIP samples, not false negatives (Table 2).
|True Negatives Mean (SD)||False Negatives Mean (SD)||True Positives Mean (SD)||False Positives Mean (SD)||P Value*|
|Type I alveolar genes||16.6 (1.5)||16.6 (1.3)||16.9 (1.4)||16.3 (2.2)||0.3|
|Type II alveolar genes||49.3 (16.1)||47.3 (14.2)||41.6 (15.5)||50.9 (10.9)||0.002|
|RIN||4.9 (1.4)||5.0 (1.7)||4.5 (1.4)||4.7 (1.5)||0.08|
|DV200||73.5 (13.6)||73.9 (14.5)||69.3 (12.3)||76.8 (9.9)||0.04|
|RNA yield (ng)||48.5 (53.7)||35.4 (41.6)||32.5 (27.3)||53.7 (34.4)||0.03|
There is an association between classifier accuracy and sample RNA quality and yield as well (Table 2), where higher sample quality in non-UIP samples correlates with a stronger non-UIP score (data not shown). Taken together, the minimal dependency of this method on sampling adequacy represents a clear advantage over pathology, where sampling of alveolated tissue is frequently required for diagnosis.
In developing UIP classifiers, we generated differential gene expression data. We conducted pathway overrepresentation analysis, and found that TBBs with UIP are significantly enriched for the expression of markers of cellular metabolism, adhesion, and developmental processes (Table 3). In contrast, the comparatively heterogeneous non-UIP group shows evidence of immune activation, lipid metabolism, stress responses, and cell death (Table 3). Aberrant reactivation of developmental pathways and cellular proliferation are known hallmarks of IPF (26–29).
|Biological Process||No. of Genes Expected||No. of Genes Observed||Fold Enrichment||FDR-Adjusted P Value|
|Overrepresented in UIP|
|Cellular component morphogenesis||23||53||2.3||<0.0001|
|Nervous system development||29||63||2.2||<0.0001|
|RNA metabolic process||88||144||1.6||<0.0001|
|Nucleobase metabolic process||135||189||1.4||0.0002|
|Nitrogen compound metabolic process||86||129||1.5||0.0008|
|Overrepresented in non-UIP|
|Antigen processing and presentation||4||20||5.3||<0.0001|
|Cellular defense response||13||39||3.0||<0.0001|
|Lipid metabolic process||33||68||2.1||<0.0001|
|Immune system process||79||131||1.7||<0.0001|
|Cholesterol metabolic process||5||19||3.8||0.0003|
|Steroid metabolic process||11||30||2.7||0.0004|
|Phosphate-containing compound metabolism||77||114||1.5||0.0076|
|IκB kinase/NF-κB cascade||4||15||3.4||0.0157|
|Response to stress||53||83||1.6||0.0172|
|Transmembrane tyrosine kinase signaling||12||28||2.3||0.0213|
The diagnosis of ILD remains quite challenging given the complexity of diffuse parenchymal disorders. Diagnostic approaches have emphasized multidisciplinary evaluation of clinical, radiological, and pathological data. The challenge in obtaining adequate sampling of lung tissue to support a pathological diagnosis has traditionally led to invasive and riskier surgical procedures. We hypothesized that it would be possible to obtain diagnostic information from lung samples obtained without the need for surgery.
Given that pathological analysis of TBBs shows poor sensitivity for UIP, there has been increasing interest in the development of molecular markers in samples, such as bronchoalveolar lavage and/or serum, that could serve as diagnostic tools (19, 26, 30–32). These methods, although potentially informing a specific diagnosis or disease progression, do not currently function as substitutes for pathology. To be clinically useful, a diagnostic test for ILD needs to distinguish UIP from among similar, but pathologically distinct, disease processes. Consequently, we chose expert pathology review as the reference standard for the presence of UIP. Despite known issues of interoperator disagreement (4, 33, 34), we achieved blinded agreement between 2 expert pathologists in 89 of the 140 subjects (63.6%) enrolled in our study. The current study demonstrates proof of the concept that a UIP genomic classifier using a gene expression signature can effectively distinguish UIP from other forms of ILD in TBB samples.
The classifier does not detect UIP in every subject with UIP pathology; there were false negatives. This is partially by design, as a rule-in test with imperfect class separation would necessarily favor specificity at the expense of sensitivity. The variable scoring of individual TBBs in some subjects with UIP nevertheless suggests that sampling effects, either insufficient tissue sampling or disease heterogeneity, are a potential source of false negatives. We did not observe a systematic reduction in alveolar content in false-negative samples, ruling out inadequate sampling of alveolated tissue in those samples as the cause of misclassification. We also did not observe an association between false negatives and technical factors related to single TBB sample quality or size. These data provide important clinically practical information for clinicians who will use genomic diagnostic techniques based on bronchoscopic sampling.
Mixing samples from the same subject improved overall detection. In the reduced set of subjects in which pooling was possible, there is evidence of improved class separation between subjects with UIP and those without UIP. We hypothesize that the pooling of at least three TBBs from each subject partially mitigates sampling effects, thus improving detection and accuracy. We also expect that pooled sample test results will prove sufficiently accurate to inform patient care on an intention-to-diagnose basis when pathology results are inconclusive or not available. The final proof of this testing approach requires a future validation in a larger independent cohort of patients.
There are several limitations to this study. Our cohort is limited in size, and was reduced further by attrition due to our pooling requirements. Our cohort has sparse representation of several other ILD subtypes, and potentially difficult-to-categorize conditions, such as chronic (fibrotic) hypersensitivity pneumonitis, fibrotic nonspecific interstitial pneumonia, and connective tissue disease–associated ILD. Our study requirement for diagnostic pathology as a reference standard likely biased our training and validation cohorts in favor of patients with ILD with nondiagnostic chest imaging who could tolerate an SLB and for whom an unambiguous pathologic pattern diagnosis was achievable. On the other hand, these are the patients that prove particularly challenging for clinicians.
The lower diagnostic rate and UIP prevalence that we observe in the subjects with pathology obtained from bronchoscopic biopsies is consistent with previous reports (11–13). Although diagnostic pathology from TBB is controversial, we chose to include such subjects in classifier training to avoid further potential cohort bias and reduction in study size. Although we also evaluated subjects with ambiguous pathologic diagnoses during the development of this test (e.g., unclassifiable chronic interstitial fibrosis), we are unable to inform on test performance in this cohort, due to the absence of a comparable reference standard. Additional evaluation will be required to address the value of our genomic approach within the context of all of the clinicopathologic characteristics considered by multidisciplinary committees.
Importantly, the low sensitivity of radiology for the presence of pathologic UIP, observed in this study and previously (5, 35, 36), suggests that this genomic classifier may have significant utility in patients in whom the HRCT does not provide a definitive pattern diagnosis (37, 38). This is likely to be the most relevant patient population in whom clinicians require additional diagnostic information. As the primary endpoint of our study was the determination of pathologic pattern at the time of diagnosis, the ability of the classifier to predict prognosis was not addressed.
Our data suggest that a genomic classifier, properly trained and validated, can differentiate pathologic UIP in TBB from non-UIP conditions where fibrosis is associated with inflammatory and immune responses. In the context of a clinical evaluation, such a test may serve as a proxy for a histopathology diagnosis, thus more confidently informing the diagnostic scheme and treatment decisions without requirement for a more invasive SLB. Future work will be aimed toward clinical validation of the genomic classifier on a larger, independent, prospective cohort, collecting clinical follow-up data to determine the classifier’s ability to predict prognosis and, ultimately, to build genomic classifiers that inform on response to therapy.
The authors are grateful to Anna-Luise A. Katzenstein (retired) for expert pathology review and early study participation, Mei Wong, Huimin Jiang, Manqiu Cao, Zhanzhi Hu, and Sharlene Velichko (Veracyte) for technical assistance, Richard Lanman, Jonathan Wilde, and Lyssa Friedman (Veracyte) for early contributions to strategy and study design, Sascha Ellers, Nicole Sindy, Katie Beliveau, Shella Edejer, and Hannah Neville for Bronchial Sample Collection for a Novel Genomic Test (BRAVE) clinical operations (Veracyte). The authors are especially grateful to the BRAVE investigators who contributed subjects to this study: H. Bencheqroun (Riverside, CA); S. Benzaquen (Cincinnati, OH); A. Case (Atlanta, GA); H. Farah (Hannibal, MO); T. Freudenberger (Bellevue, WA); W. Krimsky (Baltimore, MD); M. Molina Molina (Barcelona, Spain); H. Murthy (San Jose, CA); S. Nathan (Falls Church, VA); E. Oliveira (Weston, FL); N. Rai (Tacoma, WA); W. Randerath (Solingen, Germany); M. Steele (Nashville, TN); J. Stewart (Kansas City, MO); W. Tillis (Peoria, IL); T. Wiese (Louisville, KY); J. Wilson (Palm Springs, CA); and D. Wilson (Columbus, IN). Finally, the authors are indebted to the patients who made this study possible.
|1 .||Travis WD, Costabel U, Hansell DM, King TE Jr, Lynch DA, Nicholson AG, Ryerson CJ, Ryu JH, Selman M, Wells AU, et al.; ATS/ERS Committee on Idiopathic Interstitial Pneumonias. An official American Thoracic Society/European Respiratory Society statement: update of the international multidisciplinary classification of the idiopathic interstitial pneumonias. Am J Respir Crit Care Med 2013;188:733–748.|
|2 .||Raghu G, Rochwerg B, Zhang Y, Garcia CAC, Azuma A, Behr J, Brozek JL, Collard HR, Cunningham W, Homma S, et al.; American Thoracic Society; European Respiratory society; Japanese Respiratory Society; Latin American Thoracic Association. An official ATS/ERS/JRS/ALAT clinical practice guideline: treatment of idiopathic pulmonary fibrosis: an update of the 2011 clinical practice guideline. Am J Respir Crit Care Med 2015;192:e3–e19. [Published erratum appears in Am J Respir Crit Care Med 192:644.]|
|3 .||Bjoraker JA, Ryu JH, Edwin MK, Myers JL, Tazelaar HD, Schroeder DR, Offord KP. Prognostic significance of histopathologic subsets in idiopathic pulmonary fibrosis. Am J Respir Crit Care Med 1998;157:199–203.|
|4 .||Flaherty KR, Travis WD, Colby TV, Toews GB, Kazerooni EA, Gross BH, Jain A, Strawderman RL, Flint A, Lynch JP, et al. Histopathologic variability in usual and nonspecific interstitial pneumonias. Am J Respir Crit Care Med 2001;164:1722–1727.|
|5 .||Flaherty KR, Toews GB, Travis WD, Colby TV, Kazerooni EA, Gross BH, Jain A, Strawderman RL III, Paine R, Flint A, et al. Clinical significance of histological classification of idiopathic interstitial pneumonia. Eur Respir J 2002;19:275–283.|
|6 .||Flaherty KR, Thwaite EL, Kazerooni EA, Gross BH, Toews GB, Colby TV, Travis WD, Mumford JA, Murray S, Flint A, et al. Radiological versus histological diagnosis in UIP and NSIP: survival implications. Thorax 2003;58:143–148.|
|7 .||Katzenstein AL, Mukhopadhyay S, Myers JL. Diagnosis of usual interstitial pneumonia and distinction from other fibrosing interstitial lung diseases. Hum Pathol 2008;39:1275–1294.|
|8 .||Raghu G, Collard HR, Egan JJ, Martinez FJ, Behr J, Brown KK, Colby TV, Cordier J-F, Flaherty KR, Lasky JA, et al.; ATS/ERS/JRS/ALAT Committee on Idiopathic Pulmonary Fibrosis. An official ATS/ERS/JRS/ALAT statement: idiopathic pulmonary fibrosis: evidence-based guidelines for diagnosis and management. Am J Respir Crit Care Med 2011;183:788–824.|
|9 .||American Thoracic Society; European Respiratory Society. American Thoracic Society/European Respiratory Society international multidisciplinary consensus classification of the idiopathic interstitial pneumonias. This joint statement of the American Thoracic Society (ATS), and the European Respiratory Society (ERS) was adopted by the ATS board of directors, June 2001 and by the ERS executive committee, June 2001. Am J Respir Crit Care Med 2002;165:277–304.|
|10 .||Katzenstein A-LA, Myers JL. Idiopathic pulmonary fibrosis: clinical relevance of pathologic classification. Am J Respir Crit Care Med 1998;157:1301–1315.|
|11 .||Berbescu EA, Katzenstein A-LA, Snow JL, Zisman DA. Transbronchial biopsy in usual interstitial pneumonia. Chest 2006;129:1126–1131.|
|12 .||Tomassetti S, Cavazza A, Colby TV, Ryu JH, Nanni O, Scarpi E, Tantalocco P, Buccioli M, Dubini A, Piciucchi S, et al. Transbronchial biopsy is useful in predicting UIP pattern. Respir Res 2012;13:96.|
|13 .||Shim HS, Park MS, Park IK. Histopathologic findings of transbronchial biopsy in usual interstitial pneumonia. Pathol Int 2010;60:373–377.|
|14 .||Tomassetti S, Wells AU, Costabel U, Cavazza A, Colby TV, Rossi G, Sverzellati N, Carloni A, Carretta E, Buccioli M, et al. Bronchoscopic lung cryobiopsy increases diagnostic confidence in the multidisciplinary diagnosis of idiopathic pulmonary fibrosis. Am J Respir Crit Care Med 2016;193:745–752.|
|15 .||Dhooria S, Sehgal IS, Aggarwal AN, Behera D, Agarwal R. Diagnostic yield and safety of cryoprobe transbronchial lung biopsy in diffuse parenchymal lung diseases: systematic review and meta-analysis. Respir Care 2016;61:700–712.|
|16 .||Poletti V, Ravaglia C, Gurioli C, Piciucchi S, Dubini A, Cavazza A, Chilosi M, Rossi A, Tomassetti S. Invasive diagnostic techniques in idiopathic interstitial pneumonias. Respirology 2016;21:44–50.|
|17 .||Kennedy GC, Pankratz DG, Choi Y, Colby T, Myers JL, Brown KK, Flaherty K, Lynch DA, Steele MP, Imtiaz U, et al. Bridging the gap between HRCT and final IPF diagnosis: classifying the usual interstitial pneumonia pattern in transbronchial biopsies using machine learning on high dimensional genomic data [abstract]. Am J Respir Crit Care Med 2016;193:A2380.|
|18 .||Martinez FJ, Flaherty KR, Brown KK, Colby TV, Groshong SD, Myers JL, Lynch DA, Steele MP, Huang J, Choi Y, et al. Diagnosis of idiopathic pulmonary fibrosis: accurate molecular detection of usual interstitital pneumonia pattern in transbronchial biopsies using RNA sequencing [abstract]. QJM 2016:109(suppl_1):S8.|
|19 .||Kim SY, Diggans J, Pankratz D, Huang J, Pagan M, Sindy N, Tom E, Anderson J, Choi Y, Lynch DA, et al. Classification of usual interstitial pneumonia in patients with interstitial lung disease: assessment of a machine learning approach using high-dimensional transcriptional data. Lancet Respir Med 2015;3:473–482.|
|20 .||Friedman J, Hastie T, Tibshirani R. Regularization paths for generalized linear models via coordinate descent. J Stat Softw 2010;33:1–22.|
|21 .||Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15:550.|
|22 .||Mi H, Lazareva-Ulitsky B, Loo R, Kejariwal A, Vandergriff J, Rabkin S, Guo N, Muruganujan A, Doremieux O, Campbell MJ, et al. The PANTHER database of protein families, subfamilies, functions and pathways. Nucleic Acids Res 2005;33:D284–D288.|
|23 .||Katzenstein AL, Zisman DA, Litzky LA, Nguyen BT, Kotloff RM. Usual interstitial pneumonia: histologic study of biopsy and explant specimens. Am J Surg Pathol 2002;26:1567–1577.|
|24 .||Trahan S, Hanak V, Ryu JH, Myers JL. Role of surgical lung biopsy in separating chronic hypersensitivity pneumonia from usual interstitial pneumonia/idiopathic pulmonary fibrosis: analysis of 31 biopsies from 15 patients. Chest 2008;134:126–132.|
|25 .||Akashi T, Takemura T, Ando N, Eishi Y, Kitagawa M, Takizawa T, Koike M, Ohtani Y, Miyazaki Y, Inase N, et al. Histopathologic analysis of sixteen autopsy cases of chronic hypersensitivity pneumonitis and comparison with idiopathic pulmonary fibrosis/usual interstitial pneumonia. Am J Clin Pathol 2009;131:405–415.|
|26 .||Selman M, Pardo A, Barrera L, Estrada A, Watson SR, Wilson K, Aziz N, Kaminski N, Zlotnik A. Gene expression profiles distinguish idiopathic pulmonary fibrosis from hypersensitivity pneumonitis. Am J Respir Crit Care Med 2006;173:188–198.|
|27 .||Lockstone HE, Sanderson S, Kulakova N, Baban D, Leonard A, Kok WL, McGowan S, McMichael AJ, Ho LP. Gene set analysis of lung samples provides insight into pathogenesis of progressive, fibrotic pulmonary sarcoidosis. Am J Respir Crit Care Med 2010;181:1367–1375.|
|28 .||Selman M, Pardo A. Revealing the pathogenic and aging-related mechanisms of the enigmatic idiopathic pulmonary fibrosis: an integral model. Am J Respir Crit Care Med 2014;189:1161–1172.|
|29 .||Bauer Y, Tedrow J, de Bernard S, Birker-Robaczewska M, Gibson KF, Guardela BJ, Hess P, Klenk A, Lindell KO, Poirey S, et al. A novel genomic signature with translational significance for human idiopathic pulmonary fibrosis. Am J Respir Cell Mol Biol 2015;52:217–231.|
|30 .||Jonigk D, Izykowski N, Rische J, Braubach P, Kühnel M, Warnecke G, Lippmann T, Kreipe H, Haverich A, Welte T, et al. Molecular profiling in lung biopsies of human pulmonary allografts to predict chronic lung allograft dysfunction. Am J Pathol 2015;185:3178–3188.|
|31 .||Buendía-Roldán I, Ruiz V, Sierra P, Montes E, Ramírez R, Vega A, Salgado A, Vargas MH, Mejía M, Pardo A, et al. Increased expression of CC16 in patients with idiopathic pulmonary fibrosis. PLoS One 2016;11:e0168552.|
|32 .||White ES, Xia M, Murray S, Dyal R, Flaherty CM, Flaherty KR, Moore BB, Cheng L, Doyle TJ, Villalba J, et al. Plasma surfactant protein-D, matrix metalloproteinase-7, and osteopontin index distinguishes idiopathic pulmonary fibrosis from other idiopathic interstitial pneumonias. Am J Respir Crit Care Med 2016;194:1242–1251.|
|33 .||Nicholson AG, Fulford LG, Colby TV, du Bois RM, Hansell DM, Wells AU. The relationship between individual histologic features and disease progression in idiopathic pulmonary fibrosis. Am J Respir Crit Care Med 2002;166:173–177.|
|34 .||Walsh SL, Wells AU, Desai SR, Poletti V, Piciucchi S, Dubini A, Nunes H, Valeyre D, Brillet PY, Kambouchner M, et al. Multicentre evaluation of multidisciplinary team meeting agreement on diagnosis in diffuse parenchymal lung disease: a case–cohort study. Lancet Respir Med 2016;4:557–565.|
|35 .||Raghu G, Mageto YN, Lockhart D, Schmidt RA, Wood DE, Godwin JD. The accuracy of the clinical diagnosis of new-onset idiopathic pulmonary fibrosis and other interstitial lung disease: a prospective study. Chest 1999;116:1168–1174.|
|36 .||Sumikawa H, Johkoh T, Colby TV, Ichikado K, Suga M, Taniguchi H, Kondoh Y, Ogura T, Arakawa H, Fujimoto K, et al. Computed tomography findings in pathological usual interstitial pneumonia: relationship to survival. Am J Respir Crit Care Med 2008;177:433–439.|
|37 .||Flaherty KR, King TE Jr, Raghu G, Lynch JP III, Colby TV, Travis WD, Gross BH, Kazerooni EA, Toews GB, Long Q, et al. Idiopathic interstitial pneumonia: what is the effect of a multidisciplinary approach to diagnosis? Am J Respir Crit Care Med 2004;170:904–910.|
|38 .||Tominaga J, Sakai F, Johkoh T, Noma S, Akira M, Fujimoto K, Colby TV, Ogura T, Inoue Y, Taniguchi H, et al. Diagnostic certainty of idiopathic pulmonary fibrosis/usual interstitial pneumonia: the effect of the integrated clinico-radiological assessment. Eur J Radiol 2015;84:2640–2645.|
Supported by funding from Veracyte, Inc., South San Francisco, California.
Author Contributions: F.J.M. and G.C.K. directed the study. D.G.P., Y.C., U.I., G.M.F., J.D.A., N.M.B., P.S.W., J.H., and G.C.K. designed, managed, performed, analyzed, and interpreted the results of the study; T.V.C., J.L.M., D.A.L., K.K.B., K.R.F., M.P.S., S.D.G., G.R., and F.J.M. performed clinical patient review and provided key study feedback and direction; D.G.P., Y.C., G.C.K., and F.J.M. prepared the manuscript; all authors contributed to the development and intellectual refinement of the work; D.G.P., Y.C., U.I., G.M.F., J.D.A., N.M.B., P.S.W., J.H., and G.C.K. are employees of the study sponsor; all authors have given final approval of the manuscript.
This article has an online supplement, which is accessible from this issue’s table of contents at www.atsjournals.org