An Exhaled microRNA Panel Interrogated for Lung Cancer Case-Control Discrimination

Background: An exhaled microRNA-based lung cancer case-control discriminant biomarker strategy is reported. Methods: A microRNA-seq discovery effort compared paired tumor to non-tumor tissue, was reconciled with analogous TCGA and published literature-based tissue-discriminant microRNA data, yielding a candidate panel of 24 microRNAs that are upregulated in either adenocarcinomas and/or squamous cell carcinomas. The technical feasibility of microRNA-PCR assays in exhaled breath condensate (EBC) was tested. The airway origin of exhaled microRNAs was then topographically “ngerprinted”, using paired EBC and bronchoscopic samples. For initial EBC testing, a clinic-based case-control set of 351 individuals (166 NSCLC cases, 185 non-cancer controls) was interrogated with the 24-candidate microRNA panel by qualitative RT-PCR, and curated by melt curve analysis. Data were analyzed by both logistic regression (LR), and by random-forest (RF) models, validated by iterative resampling. Results: Both feasibility of exhaled microRNA detection, and its origins in part from lower airway sources, were conrmed. LR models adjusted for age, sex, smoking status, pack years, quit-years, and underlying lung disease identied exhaled miR-21, 33b, 212 (p.adj,=0.019, 0.018, 0.033, resp.) as case-control discriminant. For the RF analysis, the combined clinical + microRNA models showed modest added discrimination capacity (1.1–2.5%) beyond the clinical models alone: by subgroup, all subjects 1.1% (p = 8.7e-04)); former smokers 2.5% (p = 3.6e-05); early stage 1.2% (p = 9.0e-03). Sensitivity, specicity, positive- and negative-predictive values of the clinical + microRNA models for the entire cohort were 71%-76%. Conclusion: This work suggests that exhaled microRNAs are measurable qualitatively; reect in part lower airway Clinical case dened COPD, dened clinically radiographically, and/or pathologically in medical records; Pack-yrs – Quit-yrs, in former smokers, a constructed variable combining cumulative dose years) minus proximity NOS, not specied. Adeno = adenocarcinoma; Squamous = squamous cell carcinoma; NSCLC-Undifferentiated non-small cell lung cancer; Small cell = small cell carcinoma; Mets/Other = metastases from other organs to lung or other tumor histologies. ULD = Underlying (chronic) lung disease.

This initial report describes an exhaled microRNA approach to non-invasive interrogation of the lung for the purpose of developing a lung cancer risk biomarker that is both airway compartment-derived, and population-applicable. As a rst step, we derived a candidate microRNA pool, including a total of 20 upregulated microRNAs that differentiated frank human lung NSCLC tumors versus paired non-tumor tissue from the same individual surgical resections, using RNAseq efforts from our own sample sets (Supplemental data, Fig. 1), and veri ed them with analogous data from the TCGA [15,16]. We also added 5 microRNAs of additional interest from the published lung cancer literature [17][18][19][20]. Next we tested the technical feasibility of detecting microRNAs in exhaled breath condensate (EBC), and then assessed the initial exhaled microRNA performance in discriminating those with non-small cell lung cancer (cases), and those without lung cancer (controls) drawn from the same clinical population of individuals destined for bronchoscopy or lung resection surgery. Starting with a robust base clinical model in the discovery set, for the three primary analyses (all categories, formers smokers, early stage), the exhaled microRNA biomarker data yield a modest 1.1-2.5% increment in case-control discrimination attributable to the addition of qualitative exhaled microRNA data, over and above clinical factor models alone.

EBC donor recruitment and sample collection
Subject recruitment: A series of 351 consenting individuals destined for lung sampling for clinical purposes (bronchoscopy or thoracic surgery) were enrolled under a protocol approved by the Einstein-Monte ore institutional review board (IRB). This observational series work was STROBE compliant [21]. This study included 166 cases of lung cancer and 185 controls without lung cancer (Table 1). It also included 4 lab volunteers and EBC was collected from every volunteer at three different timepoints. EBC (and other non-invasive airway specimen) collection occurred immediately prior to the planned bronchoscopy/thoracic surgery, to preclude procedure-induced spillage of lung materials into the EBC (and mouthwash) samples. Clinical data was obtained by direct interview in advance of any clinicallyindicated bronchoscopic/surgical procedure (and therefore in advance of tissue diagnosis), and veri ed manually in the clinical electronic medical record. Inclusions were: age > 21; tness for the clinicallyindicated (bronchoscopy/surgical) procedure; capacity and willingness to consent. Exclusions were: acute respiratory illness, contraindications to additional brushings/bronchoalveolar lavage (coagulopathy/known poorly controlled uremia); lack of capacity for consent. As such, subjects entailed a diversity of ages, ethnicities, smoking histories, clinical diagnoses, and underlying chronic lung diseases, which were accounted for in the models.

RNA extraction
For total RNA extraction, EBC was concentrated by ethanol precipitation and then was puri ed by Trizol (Invitrogen) per manufacturer protocol and lab optimized protocol. The following components were added into a capped polypropylene tube and thoroughly mixed, including 100-400 ul of EBC sample, 40 ul of 3M sodium acetate (pH 5.5), 5 ul of 5 ug/ul glycogen carrier, and 1100 ul of 100% cold ethanol. The mixture was chilled at -80 o C for 30 min and then centrifuged at 14,000 rpm for 20 min at 4 o C. Then, the supernatant was discarded and the pellet was rinsed with cold 70% ethanol twice, and air-dried. The pellet was then dissolved in 0.5 ml of Trizol®. Total RNA was puri ed per the Trizol® manufacturer protocol.
The RNA pellet was dissolved in 15 ul of RNase-free water.

microRNA PCR analysis
The overall strategy was to amplify mature microRNAs by a previously published lab protocol involving poly-A tailing using a one-base anchored and tagged oligo-dT-RT strategy, and a microRNA-speci c forward primer coupled to a universal, unique tag-speci c reverse primer, in aggregate precluding false gDNA ampli cation [23][24][25]. Individual steps and details follow.
Cell culture samples used for microRNA PCR development For positive controls in microRNA-PCR assay development, a set of cell lines including NHBE, HBEC, A549, Hela, HTB-119 and CRL-1995 was RNA extracted in conventional column (RNEasy, Qiagen), and provided a stock solution of total RNA for initial testing of microRNA-speci c primers.

Poly(A) Tailing
The Poly(A) Tailing Kit (Ambion) was used to polyadenylate the 3' termini of microRNA. First, ATP was diluted to 1% of the original concentration. Then, the following components were added into a PCR tube and thoroughly mixed, including 2 ul of 5x buffer, 0. Since microRNAs are all of near-identical size, base composition/melting temperature was a major distinguishing feature. The criteria for including or excluding a micro-RNA-derived PCR product as present were extracted from the melting curves. If a sample had the same melting curve maximum temperature (Tm) as the positive control from cell lines for that microRNA primerset, it was called "positive". If a reaction sample had no visible melting curve, or the visible melting curve displayed greater than +/-1.5 o C different Tm from the melting curve from the positive, individual miR-speci c control, it was called "negative". We used one convention for overall scoring of samples -at least one of two replicates must be positive. Random Forests (RF): Two types of Random Forest [31,32] classi ers were built for comparison, using R package random forest [32].First, an RF classi er was built on the clinical variables alone: age, gender, smoking status, pack-years, quit-years, underlying lung disease (type), tumor histology, stage. Two-fold cross-validation was repeated 20 times to gauge the accuracy of this classi er, and its sensitivity, speci city, positive and negative predictive value. Second, an RF classi er was built on the clinical variable plus the microRNA variables together. To compare the performance of the two types of RF classi ers, we further generated 100 resampled ROC curves for each one and compared the average area under the curve (AUC) between the two models using a two-independent sample t-test. A resampled ROC curve was generated by repeatedly splitting the dataset into 50% training, 50% testing (100 times), building the two random forest models (clinical and clinical + microRNA), and predicting the outcomes of the testing split.
Airway topography similarity statistic A subset of 12 EBC donors provided bronchoscopic samples of deep alveolar (BAL) and major airway (bronchial, BB) levels, as well as sputum, mouthrinse and other specimens. The pilot sub-study (Supplemental Table 4) was designed to evaluate if an individual microRNA pro le from EBC retains the distinct features of the microRNA pro le from deep lung (bronchial brushings or bronchoalveolar lavage), or alternately resembles contaminating upper airway/mouthwash tissues. This was done by applying an arbitrary panel of 13-microRNAs interrogated by qualitative RT-PCR against samples from 12 individuals, each donating ve airway level samples for comparison [bronchoalveolar lavage (BAL), bronchial brush (BB), sputum (SP), mouthrinse (MW), EBC]. To statistically test the surrogacy of EBC-microRNA for deeper lung specimens (bronchial brushings and bronchoalveolar lavage), we developed a similarity statistic of two tissue types based on Hamming distance. That is, where and are (binary) miRNA pro les from two tissue types of the same individual i. The Hamming distance H gives the total number of miRNAs for which the two pro les d and d' are discordant. The smaller the statistic SH is, the more similar are two tissue types in miRNA pro les within each subject. If the two tissue types from the same individual are not closer than two tissue types from two random individuals, then there is no information in one of the tissues to infer the miRNA pro le of the other tissue. To test that the two tissues from the same individuals are closer than two random individuals, we performed a permutation test that permutes the miRNA pro les within each tissue type among individuals.

Results
The clinical characteristics of the 351 subjects are described in . Both logistic regression (LR) and random forest (RF) discriminant models took these clinical inter-group differences into account. For RF, this included measuring the incremental impact on case-control discrimination of microRNAs over and above these clinical factors alone.
EBC surrogacy for the lung The similarity statistic of two tissue types that were based on Hamming distance,, and 1000 permutations of miRNA pro les within each tissue type among individuals, gave an estimated P-value of 0.007, suggesting that the miRNA pro les of EBC are closer to miRNA pro les of BAL of the same individual than to miRNA pro les of BAL of random individuals. The same analysis was applied between EBC and BB (p = 0.23), EBC and SP (p = 0.18), EBC and MW (p = 0.04).
Logistic regression: LR models were created ( lung disease. For the entire data set, miR-21, 33b and 212 appeared to be somewhat informative for casecontrol status (p < 0.05), after adjustment for the above-listed clinical factors.      Table 3.
Temporal stability of EBC miRNA for an individual across time Three target miRNAs (miR-141, miR-142-3p and miR-205) in EBC of three different timepoints of four individuals were detected by realtime PCR and normalized to housekeeping miR-423-3p. It shows the EBC samples from different timepoints of the same subject were stable to a large extent (Fig. 2).

Discussion
This report entails the most comprehensive interrogation of microRNAs in exhaled breath, here uniquely performed to distinguish subjects with and without primary lung cancers [33]. Starting with a lung tissue microRNA-seq discovery effort combined with published literature-suggested microRNAs, we interrogated a panel of 25 microRNAs in exhaled breath condensate using our RNA-speci c qualitative RT-PCR. We found that: (i) microRNAs are detectable in exhaled breath condensate; (ii) there are individual exhaled microRNAs that offer case-control discrimination by logistic regression (microRNAs 21, 33b, 212), and (iii) additional RF models can be developed, using the entire microRNA panel, that also suggest some modest additional case-control discrimination, particularly in the subsets of former smoker, and early stage subjects, over and above that demonstrated in comprehensive clinical models.
Technical challenges abound in examining nucleic acids in EBC. While EBC is widely available noninvasively, this specimen entails only trace levels of microRNA template. This is perhaps because the templates are by de nition, higher in molecular weight (22 nucleotides in length) than is typically true for exhaled airstream-suspended molecules such as H 2 O 2 , 8-isoprostane, and others [22]. Nonetheless, the PCR confers capacity for detection of microRNAs at the low template copy level, as is suggested here.
The trace concentrations inherent to EBC specimens for most analytes, including nucleic acids has, to date, precluded performing discovery efforts such as microRNA next gen sequencing, directly from this matrix.
The microRNA interrogation panel choice was therefore based on: (i) a previously unpublished microRNA seq effort (GEO#: GSE33858) inter-tissue comparison of 32 lung resected bronchogenic carcinoma versus remote lung tissue. (strati ed for adenocarcinoma, squamous cell carcinoma histologies), with 10 representative overexpressed microRNAs included from each of those two histologies. The remainder came from: (ii) TCGA [15,16]; and (iii) several literature-identi ed microRNA markers of lung cancer.
We used our previously published microRNA-PCR that is micro/mRNA -speci c, as it excludes gDNA fragment false priming by employing a uniquely tagged RT-primer strategy [23,25], and in primer design precluded false ampli cation of messenger RNA fragments. We chose to treat the data as qualitative (individual miR, present/absent) because we were insu ciently con dent of robust quantitative RT-PCR data that could not be reliably scaled to a robustly quanti able internal housekeeper at these trace levels.
Performance of the uorescent intercalating (SYBR®) dye detection strategy coupled to URT-PCR on the realtime PCR platform allowed quality assurance using quantitation curve, melt-curve, melt temperature with each PCR reaction. This was superimposed on a series of other analyses invoked during primer design, using multiple positive and negative controls, described in the Methods and Additional Studies sections. We additionally piloted a commercial qPCR platform (In Vitrogen/Taqman®) without readily apparent additional precision nor microRNA-PCR sensitivity.
This cross-sectional case-control design was chosen as representing a typical initial step in early development of potential risk biomarkers [34,35]. Clinical-demographic differences were observed in cases versus controls for age, smoking, pack-years, quit years, a pack-years minus quit-years composite index, underlying lung disease (COPD, in ammation/ brosis, asthma, sarcoidosis, bronchiectasis). However, these differences were equally modelled in both clinical-only models and in the clinical + microRNA combined models identically, so they should not have biased the incremental microRNAattributable risk prediction. We emphasized current and former smokers predominantly, as they are at elevated risk for lung cancer, and therefore commonly come to clinical attention for surveillance, biopsy/resection, and thus were considered appropriately e cient for enrollment in this initial study. Our case and control ascertainment was crisp, minimizing misclassi cation as subjects were all con rmed histologically by virtue of their bronchoscopic/surgical procedures, underwent further veri cation of case and control status by an additional 3-6 month period of clinical follow-up, facilitated by electronicallyretrieved clinical assessments from the engaged clinical pathologists, radiologists, surgeons, and pulmonologists on each subject. Recruited subjects with disputed case-control ascertainment (< 1% of enrolled) were excluded from the study.
In this moderate size case-control subject set, with an already selected candidate 24-microRNA panel, we initially performed logistic regression, using case-control status as the main outcome variable, and a clinical model tested with/without each individual miR on the panel. Separately, we then employed iterative cross validation by random forests to assure stability of our results, rather than separate discovery and validation sets. The RF approach iteratively and randomly splits the data, substantively cross validating in truly random fashion, and minimizing over-t.
The clinical versus clinical-microRNA incremental differences are admittedly modest (~ 0.0-3.0%). We surmise that this is, in part, due to the strength of the clinical model alone displaying ROCs ~ 0.75-80. These were unusually robust clinical models, we believe for two reasons. First is the clinical model comprehensiveness, in part attributable to inclusion of all major known substantive risk factors for lung cancer (including quit years, underlying lung disease, others). Secondly, there is positive selection inherent to enrolling clinical bronchoscopy and surgical subjects such as these (above), wherein both (case and control) sets of subjects are drawn from the same base (procedural-destined) population that is itself selected on clinical criteria to be at high risk for lung cancer. By de nition, that high risk is perceived by the clinician as su cient to warrant an invasive diagnostic/therapeutic procedure, the enrollment point for a majority of our subjects. Both of these factors (clinical model comprehensiveness, and clinical series enrollment bias) contribute to high risk in this clinical series, and imply that clinical risk model performance will be elevated. Thus, the difference between this comprehensive clinical model alone, and that for this clinical model plus microRNA could potentially be arti cially narrowed (as compared to that using conventional sparse clinical models) by virtue of the comprehensiveness of the clinical model. We believe the negative impact of such bias on the estimate of the actual contribution of exhaled microRNAs to case-control discrimination, is counter-balanced by the strength inherent in using the same (robust) clinical model when comparing clinical-only models versus combined clinical + microRNA models. Additionally, the de nitive diagnoses inherent in recruiting those destined for lung sampling/pathologic readout was another strength. Overall, then, the above considerations suggest ours is a conservative estimate of the exhaled biomarker contribution in real clinical conditions.
Among study limitations, we were forced to use a dichotomous (present/absent) signal for a given microRNA in a given EBC sample, despite being run on a realtime machine, for technical reasons. The realtime CT values, using the chosen platform, were not robust enough to generate reliable quantitative data, worth re-addressing in optimization studies, which are ongoing.
Additionally the discriminant microRNA signal may in fact be small in magnitude, as our data suggest.
This small magnitude of microRNA change in the " eld" of bronchial epithelium itself has been suggested in a comprehensive RNA-seq study of bronchial brushings in a similar case control setting [14].
Notably, of the discriminant four bronchial epithelial case-control discriminant microRNAs in that report, only one (146-5p) was interrogated in our study. While 146-5p was not individually case-control discriminant in LR models, it was contributory in the RF models for former smokers and early stage.
There is a very recent pilot report that EBC miRNAs might allow the identi cation, strati cation and monitoring of lung cancer [31].
We set out to survey the "state of the epithelium" rather than detection of a small peripheral tumor itself. This view of broad epithelial " eld" interrogation is appropriate to risk assessment, rather than that of a suspect tumor diagnostic tool. That the signal was likely from the eld of normal cell material, rather than spillage of a tumor is supported by the observation that early stage subset showed more case-control discrimination than the late stage cases, which would not be expected if the tumor itself was spilling microRNA material.

Conclusions
In conclusion, this is one of the rst reports of exhaled microRNAs in lung cancer. Given the technical demands of this application, we plan to re ne the exhaled microRNA interrogation technique, including miR-PCR quantitation, and microRNA panel adjustments, to better serve case-control discrimination.
Assuming improved performance with these re nements, risk assessment efforts can be pursued in future prospective cohorts. Such trials could evaluate whether the biomarker platform can predict future events, the "mother lode" of risk assessment [5]. If such utility was demonstrated, it would then allow for actionable clinical interventions, such as focusing early detection, or alternately perhaps directing chemoprevention, onto those individuals at highest risk.
ROC: Test performance plots sensitivity versus speci city.

Declarations
Ethics approval and consent to participate See the attachment.

Consent for publication
Not applicable.

Availability of data and materials
All data generated or analyzed during this study are included in this published article [and its supplementary information les].

Competing interests
The author(s) declare that they have no competing interests.  (1a2), former smokers only; (1a3) early stage only. Random forests, recursive partitioning and crossvalidation were employed as described in the statistical analysis section. This ROC plot of true positives versus false positives shows borderline incremental information (Fig. 1a1, +1.2% (p= 0.07)) value of the exhaled microRNAs over and above the clinical model alone, in all subjects combining all smoking status', stages, and histologies. It particularly held true in main subgroup analyses, separating out former smokers (Fig. 1a2, +3.0% (p=6.0e-03)), and in early stage (I,II) models (Fig. 1a3, +2.2% (5.1e-03)). In combining these subgroups (1a4), early stage x former smokers combined did not show signi cant casecontrol discrimination (Fig. 1a4, NS). Model components and signi cance testing of area under curve (AUC) differences are described in Table 3. b. Current smokers only (left), cases versus non-cancer controls; Adenocarcinoma subjects only (right). Random forests, recursive partitioning and crossvalidation were employed as described in the statistical analysis section. For Current smokers (left, 1b1), value of the exhaled microRNAs over and above the clinical model alone, was apparent +3.3% (p=3.5e-02)). For adenocarcinoma case subsets (right, 1b2), the miR model detracted from the clinical+miR combined case-control discrimination, (-2.1% (p= 1.1e-02). Model components and signi cance testing of area under curve (AUC) differences are described in Table 3.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. Supplementarymaterial.docx