Digital multiplexed analysis of circular RNAs in FFPE and fresh non-small cell lung cancer specimens

Digital multiplexed analysis of circular RNAs in FFPE and fresh non-small cell lung cancer specimens Carlos Pedraz-Valdunciel 1,2 , Stavros Giannoukakos 3 , Nicolas Potie 4 , Ana Gim enez-Capit an 5 , Chung-Ying Huang 6 , Michael Hackenberg 3 , Alberto Fernandez-Hilario 4 , Jill Bracht 2,5 , Martyna Filipska 1,2 , Erika Aldeguer 5 , Sonia Rodr ıguez 5 , Trever G. Bivona 7 , Sarah Warren 6 , Cristina Aguado 5 , Masaoki Ito 8 , Andr es Aguilar-Hern andez 9 , Miguel Angel Molina-Vila 5 and Rafael Rosell 1,9,10 1 Germans Trias I Pujol Research Institute, Badalona, Spain 2 Department of Biochemistry, Molecular Biology and Biomedicine, Autonomous University of Barcelona, Spain 3 Department of Genetics, University of Granada, Spain 4 Andalusian Research Institute in Data Science and Computational Intelligence, University of Granada, Spain 5 Laboratory of Oncology, Pangaea Oncology, Barcelona, Spain 6 NanoString Technologies, Seattle, WA, USA 7 UCSF Helen Diller Family Comprehensive Cancer Center, University of California San Francisco, CA, USA 8 Department of Surgical Oncology, Research Institute for Radiation Biology and Medicine, Hiroshima University, Japan 9 Oncology Institute Dr. Rosell, IOR, Quir on-Dexeus University Institute, Barcelona, Spain 10 Autonomous University of Barcelona, Spain Keywords biomarkers; cancer; circRNA; diagnosis; nCounter; NSCLC Correspondence C. Pedraz-Valdunciel, Pangaea Oncology, Calle de Sabino Arana, 5, 08028 Barcelona, Spain Tel: +34 935 46 01 19 E-mail: [email protected] and R. Rosell, Germans Trias I Pujol Research Institute, Cam ı de les Escoles, s/n, 08916 Badalona, Spain Tel: +34 930330520 E-mail: [email protected] (Received 16 August 2021, revised 22 November 2021, accepted 19 January 2022) doi:10.1002/1878-0261.13182 Although many studies highlight the implication of circular RNAs (circRNAs) in carcinogenesis and tumor progression, their potential as cancer biomarkers has not yet been fully explored in the clinic due to the limitations of current quantification methods. Here, we report the use of the nCounter platform as a valid technology for the analysis of circRNA expression patterns in non-small cell lung cancer (NSCLC) specimens. Under this context, our custom-made circRNA panel was able to detect circRNA expression both in NSCLC cells and formalin-fixed paraffinembedded (FFPE) tissues. CircFUT8 was overexpressed in NSCLC, contrasting with circEPB41L2, circBNC2, and circSOX13 downregulation even at the early stages of the disease. Machine learning (ML) approaches from different paradigms allowed discrimination of NSCLC from nontumor controls (NTCs) with an 8-circRNA signature. An additional 4-circRNA signature was able to classify early-stage NSCLC samples from NTC, reaching a maximum area under the ROC curve (AUC) of 0.981. Our results not only present two circRNA signatures with diagnosis potential but also introduce nCounter processing following ML as a feasible protocol for the study and development of circRNA signatures for NSCLC. Abbreviations AUC, area under the curve; circRNA, circular RNA; FFPE, formalin-fixed paraffin-embedded; GBM, gradient boosting machines; KNN, k-nearest neighbors; LOOCV, leave-one-out cross-validation; miRNA, micro RNA; ML, machine learning; NPV, negative predictive value; NSCLC, non-small cell lung cancer; PCR, polymerase chain reaction; PPV, positive predictive value; RF, random Forest; RFE, recursive feature elimination; RNAseq, RNA sequencing; ROC, receiver operating characteristic; RT-qPCR, quantitative reverse transcription PCR; SD, standard deviation. 1Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. 1. Introduction circRNAs are a newly re-defined type of endogenous RNA molecules originated by a noncanonical process called ‘back-splicing’. Through this mechanism, the 50 splice donor covalently links to the 30end of an upstream exon, resulting in a single-stranded circular structure which can include one or different exonic/intronic regions [1]. This particular assembly lacking a poly(A) tail makes them very stable and resistant to exonuclease-mediated degradation when compared to their linear counterparts [2]. The existence of circRNAs has been acknowledged for more than 45 years. First evidence was reported in 1976 with the first description of viroids as ‘singlestranded and covalently closed circular RNA molecules’ [3], and their discovery in humans followed almost two decades later [4]. However, it is not until recently that their role has been clarified, evolving from abnormally spliced unfunctional ‘scrambled’ transcripts to circular RNA molecules with a marked role in homeostasis [5,6]. CircRNAs have been classified as noncoding RNA for many years, due to the lack of a 50cap structure and their inability to bind to ribosomes. However, recent studies reported that some circRNAs can be translated into small functional peptides in a capindependent manner [7]. Other functions may include serving as protein decoys, scaffolds, and/or recruiters [8], or regulating the canonical transcription by competing with the formation of linear cognates via backsplicing [9,10]. Nonetheless, the most well-studied function is their interaction with miRNAs. A single circRNA can have several miRNA-binding sites through which targeted miRNAs get ‘sponged’, thereby blocking their activity [11]. It is throughout this mechanism how they predominantly exert their role as cell proliferation regulators targeting mediators of classical signaling pathways, such as MAPK/ERK, PI3K/AKT, and WNT/b-catenin, or cell cycle checkpoint regulators [11]. Because of their implication in the abovementioned processes, dysregulation of circRNA expression can be associated to the development of different malignancies, including lung cancer. CircRNAs are significantly associated with tumorigenesis, proliferation, migration, and sensitivity to lung cancer therapies [12] and, as a result, have been presented in many recent studies as novel biomarkers to assess disease status. However, the number of studies focusing on the development of circRNA signatures with either diagnostic or prognostic value in human malignancies is rather small, probably due to the lack of standardized circRNA quantification methods, which in turn is hampering the development of clinically applicable assays. RT-qPCR is widely used as a quantification tool for circRNA expression studies. While its sensitivity and short turnaround time proves beneficial for circRNA research, several events such as template switching, rolling circle amplification, or the bias attached to this technique may hinder the results [13]. In addition, it does not allow high-throughput analysis, which is necessary for biomarker discovery. Microarrays or RNAseq may overcome these limitations; however, the first have a limited range of detection disregarding those targets with either very low or high expression, while the latter not only results rather expensive but also includes other restrictions such as the use of long time-consuming protocols, or complex data analysis [14,15]. The nCounter technology allows multiplex analysis of up to 800 transcripts by direct capture and counting of individual targets [16]. With a short turnaround time and minimal hands-on work, it provides results in <48 h with the use of an intelligible software. However, despite the growing number of laboratories using this platform, it still gets mostly restricted to mRNA analysis. In this proof-of-concept study, we retrospectively analyzed the circRNA expression profiles in NSCLC cell lines and FFPE tissues by using a custom-designed 78 circRNA nCounter panel. Our data demonstrate that nCounter can be employed not only for basic circRNA research but also for the development of clinically useful circRNA signatures. 2. Materials and methods 2.1. Patient samples and cell lines This study was carried out in accordance with the principles of the Declaration of Helsinki, under an approved protocol of the institutional review boards of Quir on Hospitals, and the IGTP-HUGTP Biobank. FFPE lung cancer tissues were retrospectively collected from 27 early-stage and 26 late-stage cancer patients from the different Quir on hospitals (Table 1). FFPE tissue samples from 16 donors were collected as controls from the IGTP-HUGTP Biobank. Most controls did not present any type of cancer, except for four samples which were extracted from the nontumorigenic region of the lung from a cancer patient. Individuals with different pathologies were also included to ensure 2Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. nCounter for circRNA detection in NSCLC samples C. Pedraz-Valdunciel et al. the development of signatures specific of lung cancer (Table S1). All collected samples were assessed for tumor and lymphocyte infiltration by a pathologist (Table S2). Written informed consent was obtained from all patients and further documented; samples were deidentified for patient confidentiality. Clinical information collected from each patient was limited to gender, age, smoking status, tumor histology, driver mutation, and stage. A panel of seven human lung cancer cell lines harboring different mutations was selected along with two normal epithelial cell lines (Table 2). Cell lines were maintained following standard culture conditions [17] in RPMI-1640 or DMEM (Gibco, Life Technologies, Carlsbad, CA, USA) supplemented with 10% fetal bovine serum (Gibco). All cell lines were tested for mycoplasma infection. 2.2. RNA extraction RNA extraction was performed following previously published methods [18,19]. RNA from fresh cell lines was isolated using the Allprep DNA/RNA/miRNA universal kit (Qiagen, Hilden, Germany). FFPE cells and tissues were deparaffined with xylene. After the removal of xylene using ethanol, RNA was extracted using the High Pure FFPET RNA isolation Kit (Roche, Rotkreuz, Switzerland). RNA quantification was performed using the Qubit 4 Fluorometer (Invitrogen, Carlsbad, CA, USA) with the Qubit RNA HS Assay Kit (Invitrogen). RNA integrity was assessed with the 2100 Bioanalyzer system (Agilent Technologies, Santa Clara, CA, USA) using the RNA 6000 Nano kit (Agilent Technologies). 2.3. Rnase-R treatment 5µg of total RNA was either treated or mock-treated with RNase-R (Lucigen, Madison, WI, USA). RNA Table 1. Clinicopathologic characteristics of enrolled patients (n=69). NSCLC, non-small cell lung cancer. Clinicopathological characteristics Lung cancer patients (n=53) Noncancer controls (n=16) Gender—no. (%) Male 28 (52.8) 10 (62.5) Female 25 (47.2) 6 (37.5) Age—years Median 66 59 Range 32–85 29–76 Smoking status—no. (%) Exor current smoker 40 (75.5) 9 (56.25) Never smoker 11 (20.8) 5 (31.25) Not information 2 (3.7) 2 (12.5) Histological type Adenocarcinoma 43 – Squamous carcinoma 1 – Other NSCLC 9 – Driver mutation EGFR 6 – Exon19 3 – Exon21 1 – Exon20-21 1 – Exon21 and amplification 1– KRAS 12 – G12A 2 – G12C 3 – G12V 4 – G12R 1 – Other 2 – BRAF 1 – ROS 1 – RET 2 – ALK 1 – MET (exon14 mutation) 1– Other alterations 5 Not information 24 – Tumor stage—no. (%) I 16 (30.2) – II 4 (7.5) – IIIA 7 (13.2) – IIIB 3 (5.6) – IV 23 (43.4) – Table 2. Characteristics of the lung cell lines included in the study. AD, adenocarcinoma; ATCC, American Type Culture Collection; NE, normal epithelial; UCSF, University California San Francisco; UTSW, University of Texas Southwestern. Cell line Histology Gene Mutation Origin A549 AD KRAS G12S ATCC HOP-62 G12C ATCC PC9 EGFR E746_A750 DL Hoffmann-La Roche, with the authorization of Dr. Mayumi Ono HCC-827 E746_A750 DL ATCC NCI-H1666 BRAF G466V ATCC NCI-H2228 ALK EML4-ALK, variant 1 ATCC NCI-H3122 EML4-ALK, variant 3 ATCC AALE NE –wt Dr. Trever Bivona Lab, UCSF HBEC30KT Dr. Minna Lab, UTSW 3Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. C. Pedraz-Valdunciel et al.nCounter for circRNA detection in NSCLC samples samples were denatured at 95 °C for 30 s following addition of a master mix containing RNase-R (or molecular grade water in the case of mock-treated samples), 109RNase-R buffer adjusted to the final volume, and molecular grade water. Samples were incubated 160 min at 40 °C and kept at 4 °C prior RNA quantification and subsequent nCounter hybridization. 2.4. RT-qPCR and Sanger sequencing analysis RT-qPCR and Sanger sequencing of circRNA junction sites were performed as previously described [18]. 10 µL of total RNA was converted into cDNA using the M-MLV reverse transcriptase enzyme and random hexamers (Invitrogen). A 1 : 3 dilution of cDNA was performed, and 2.5 µL were added to the Taqman Universal Master Mix (Applied Biosystems) in a 12.5 µL reaction containing a specific pair of primers and probe for each gene. Three replicas of each sample were run for the quantification of the expression of each assessed circRNA. Three replicas of each sample were run for the quantification of the expression of each assessed circRNA. Divergent primers and probe sets were designed using Primer Express 3.0 Software (version 3.0.1, Applied Biosystems) with the latter spanning the circRNA junction site (Table 3). Quantification of gene expression was performed using the QuantStudioTM 6 Flex System (Applied Biosystems) and calculated according to the comparative Ct method. In all quantitative experiments, a sample was considered not evaluable when the standard deviation of the Cq values was >0.30 in two of the three independent analyses (n=3). For Sanger sequencing, 10 µL of each PCR product was loaded on a Precast Agarose HT-1gel and visualized under UV light (E-Gel TM Safe Imager TM RealTime Transilluminator, Invitrogen) after electrophoresis (E-Gel TM iBase TM Power System, Invitrogen). Five microliters of each cDNA sample were purified using the PCR ExoSAP-IT Product Clean up Reagent (Applied Biosystems). Sequencing PCRs were set up using the BigDye Terminator v3.1 Cycle Sequencing Kit (Applied Biosystems), forward primer, cDNA and water in a final volume of 20 µL. Sequencing PCR was performed using a Verity 96 well thermal cycler (Applied Biosystems). After sequencing amplification, samples were loaded into a 96-well plate and subjected to Sanger sequencing using the 3130 Genetic Analyzer (Applied Biosystems). 2.5. miRNA prediction and circRNA-miRNA network construction MiRNAs targeted by the differentially expressed circRNAs found in early-stage FFPE lung cancer tissues were predicted using the CIRCINTERACTOME tool (https://circinteractome.nia.nih.gov). circRNA-miRNA interaction network was built using CYTOSCAPE (v3.8.2; https://cytoscape.org). Association of miRNAs with cancer-associated downstream signaling pathways was investigated using the miRCancer database (https:// mircancer.ecu.edu). 2.6. NanoString nCounter panel design and sample processing A custom-made panel of 78 circRNAs was produced, including both highly and lowly expressed circRNAs that could be related to lung cancer (Table S3). Each probe was designed to target a flanking exonic sequence between 35–55 nucleotides of the circRNA junction site. They also contain a complementary region to capture and reporter probes, conforming a precise configuration that allows specific recognition of circular transcripts (Fig. S1). In addition, six linear reference genes (GAPDH, MRPL19, PSMC4, RPLP0, SF3, and UBB) and four mRNAs of FAM13B, Table 3. Primer and probe design for circRNA validation by RT-qPCR. In blue marked the junction site. circRNA circEPB41L2 (hsa_circRNA_0001640) Forward GAAGACCAAAACTGTCCAGTGTAAAG Reverse CACTTCAGACACAGAGCCTACTTCA Probe TGACCTGGAGCATAAG circSOX13 (hsa_circRNA_0004777) Forward CAGTGACTGGAAGGAGAGGTTTC Reverse CTGGGCAGAGATGGGGCT Probe AAAGATGTCAAAGGATGTCCATGA circBNC2 (hsa_circ_0086414) Forward GTCTGCACAGTGGCTGGTTG Reverse GGTGATGATTTCCTCTTCTCGAG Probe AGACAGGATGCTGCTG 4Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. nCounter for circRNA detection in NSCLC samples C. Pedraz-Valdunciel et al. HIPK3, MGA, and UBXN7 genes were included (Table S2). Sample processing in the nCounter was performed as previously described [18] following NanoString’s guidelines (Fig. S2). 2.7. Data normalization and differential expression analysis Raw count values were exported to Microsoft Excel (version 16.40, Microsoft, Redmond, WA, USA) using nSolver Analysis Software (version 4.0.70, NanoString Technologies, Seattle, WA, USA). For each of the circRNAs included in the panel, raw counts lower than the cut-off value established as background were automatically excluded from further analysis. Background was calculated for each sample by using the mean of the negative probe counts plus two times the standard deviation. Only circRNAs with a value >10 counts after background subtraction were considered as expressed. Subsequent circRNA-specific counts were normalized by dividing this number by the total number of counts for this sample. Resulting number was multiplied by 10 000 (units expressed in counts per 10 000). Further differential expression analysis of raw nCounter data was carried out with R(version 4.0.2; R Core Team and the R Foundation for Statistical Computing, Vienna, Austria) and R studio (version 1.3.1056; RStudio PBC, Boston, MA, USA). Technical variability correction, normalization, and differential expression analysis was performed using the RUVSEQ (version e1.24.0; Bioconductor Core Team, Buffalo, NY, USA) and DESEQ2 (version 1.30.0; Bioconductor Core Team) packages (RUVSEQ-DESEQ2, Bioconductor Core Team). Firstly, the RUVg function was used to estimate the unwanted variation among samples based on the positive controls. The positive controls used in the NanoString panel are Spike-In control sequences; therefore, analogous constant expression of positive controls is expected across all samples. Secondly. DESeq2 was used to perform the normalization of the data, while accommodating the estimated factors provided by the RUVg function. Finally, DESeq2 was used to perform hypothesis testing in order to identify differentially expressed circRNAs. Shrunken log 2 foldchange (log 2 FC) was then reported by DESeq2 along with adjusted P-values. Batch effect was considered during normalization using RUVSeq-DESeq2. The normalized data were employed for ML techniques. Volcano plots were used to visualize log 2 FC on the x-axis and log 10 adjusted P-values on the y-axis. 2.8. Machine learning classification Recursive feature elimination (RFE) was used to perform feature selection and the LOOCV algorithm was applied on the full panel of circRNA transcripts. The number of features to select was set by default at 4, 8, 16, and 78. The number of features that yielded best performance after cross-validation was automatically selected. To test whether generated data had enough discriminative information to build a robust model for the classification of cancer samples from controls, different paradigms of classification models were tested to provide the most accurate results. Under this context, three classification approaches were performed with the selected features: an ‘instance-based’ model (KNN). This model uses the distances among samples to obtain a predictive label; and two different ensemble mechanisms with decision trees—bagging (RF) and boosting (GBM). For the analysis of early-stage lung cancer samples versus control samples, GBM was excluded due to the high volume of samples is required for this model. The model with the highest ROC AUC value was then selected as the final model. A confidence threshold of 0.5 was considered for the calculation of PPV and NPV. Additional statistical indicators such as accuracy, sensitivity, and specificity were also calculated. 3. Results 3.1. nCounter for circRNA detection in fresh NSCLC samples Based on a literature review, 78 circRNAs were selected according to their differential expression in lung cancer specimens for the development of an nCounter panel (Table S3). To test the reproducibility of this panel for circRNA detection, RNA from fresh PC9 cells was subjected to nCounter analysis in three independent reactions. As a result, a strong correlation was found between the normalized counts for each individual circRNA, represented by a Spearman’s rof 0.82–0.88, P<0.01 (Fig. S3). Then, RNA from the same cell line was used in an experiment with RNAase-R, an enzyme that degrades linear RNA, to elucidate if the nCounter probes bind specifically to the circRNA of the genes included in the panel (Fig. 1A). As a result, 18 new transcripts that could not be detected in mock-treated samples were observed after RNase-R treatment (Fig. 1B). In addition, among the 34 transcripts identified in both 5Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. C. Pedraz-Valdunciel et al.nCounter for circRNA detection in NSCLC samples types of samples, the counts of 28 (82.3%) increased at least 2-fold after RNAse-R treatment. CircSND1 and circBANP were found with the highest enrichment, with a 56and 33-fold change, respectively. CircCHD9, circAASDH, circVRK1, circSLC8A1, and circSMARCA5 were the only circular transcripts affected by the exonuclease activity of RNase-R, showing a lower number of counts after incubation with the enzyme (Fig. 1C). All mRNA controls, including the linear forms of FAM13B, HIPK3, MGA, and UBXN7, were found with reduced or null expression after treatment (Fig. 1D). A high correlation was found between the two replicas included for each of the conditions (Pearson’s r=0.99917; P<0.01 and r=0.9985; P<0.01 for mock-treated and RNase-R-treated samples, respectively) demonstrating the specificity of the assay (Fig. 1E). 3.2. nCounter for circRNA detection in FFPE NSCLC samples To assess the performance of our panel in FFPE samples, RNA from paired FFPE and fresh PC9 cell line was extracted and processed in the nCounter. The number of total raw counts in PC9 FFPE samples was significantly lower compared to fresh PC9 samples (771.870 versus 1.353.811). However, despite the suboptimal quality observed in the RNA extracted from the FFPE cells (Fig. S4), a statistically significant correlation was found when comparing both types of input (Fig. S5A). Next, we assessed the feasibility of RNase-R treatment in FFPE samples. As a result, overall circRNA enrichment was not achieved, in contrast to what was previously observed in RNA extracted from fresh cells. Most circRNAs were found to be degraded to different extents in RNase-R-treated replicas when compared with the controls, indicating that such treatment should be avoided when working with FFPE samples (Fig. S5B). Then, different concentrations of FFPE-derived RNA (between 250 and 2000 ng of total RNA) were tested assessing the effect on downstream nCounter analysis. As a result, saturation was not achieved with the highest concentration, suggesting that a greater RNA input could be applied. Analysis of normalized counts across all samples indicated similar performance of 250 ng compared to the rest of tested concentrations, with a Pearson’s correlation between circHOMER1 circTMEM39B circFUT8 circDUS2L circSNX25 circNUPL2 circCSPP1 circNEDD4L_1 circPIK3C2B circCLK1 circLYPLAL1 circSOX13 circRDH11 circANXA7 circTXNDC11 circCHD2 circBANP circACACA 0 100 200 300 400 Total number of raw counts (arbitrary units) Newly discovered circRNAs in RNase treated samples (PC9 cell line) PC9, mock treated PC9, RNase R treated circFARSA circHIPK3 circRUSC2 circ_C1orf116 circCHD9 circCHD1L circDENND1B circNEDD4L circAASDH circCORO1C circVRK1 circMGA circSMAD2 circSLC8A1 circCCDC134 circUBXN7 circSMARCA5 circCCNB1 circFAM13B circRUNX1 circLIN54 circSND1 circHIBADH circLYPLAL1 circPSD3 circSOX13 circYWHAZ circDNA2 circANXA7 circZCCHC6 circDHCR24 circCHD2 circBANP circACACA FAM13B GAPDH HIPK3 MGA MRPL19 PSMC4 RPLP0 SF3 UBB UBXN7 –100 –50 0 50 Fold change total number of raw counts (arbitrary units) circRNA fold change in RNase R treated samples (PC9 cell line) circRNA mRNA circFAM13B circHIPK3 circMGA circUBXN7 FAM13B HIPK3 MGA UBXN7 0 2000 4000 6000 Total number of raw counts (arbitrary units) mRNA / circRNA in RNase R / mock treated samples (PC9 cell line) PC9, mock treated PC9, RNase R treated AB C DE Fig. 1. Analysis of circRNA from RNase-R-treated samples. (A) Workflow for circRNA enrichment with RNase-R. (B) Representation of the newly discovered circRNAs after RNase-R treatment. Bars indicate the mean of the replicas (n=2). Error bars indicate SD. (C) CircRNA/linear HK fold-change after RNase-R treatment (n=2). (D) Comparison of circRNAs/mRNA cognates in RNase-R/mock-treated samples. Bars indicate the mean of the replicas (n=2). Error bars indicates SD. (E) Correlation of the nCounter replicas (n=2) for each treatment. Pearson’s coefficient is indicated. 6Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. nCounter for circRNA detection in NSCLC samples C. Pedraz-Valdunciel et al. 0.99–1.00 (Fig. S6). As a result, 250 ng of total RNA was selected for the rest of the study. 3.3. circRNA expression in NSCLC fresh cell lines A set of seven lung cancer cell lines were selected according to their driver mutation, along with two normal epithelial cell lines (Table 2). Duplicates of equal RNA concentrations were run in all cases. Out of the 78 circRNAs included in the panel, 33 were expressed in all cell lines. Nineteen were expressed in epithelial cells and not in all lung cancer cells, while only one, circFUT8 was only expressed in all lung cancer cell lines (Fig. S7). Nineteen circRNAs included in the panel were not found in any of the assessed cell lines. Fifty-one wasthehighestnumberofcirculartranscriptsdisplayed by any cell line (AALE). The NCI-H2228 cell line showed the lowest number, with only 40 circRNAs detected (Fig. 2A). Overall, total raw counts were significantly higher in normal epithelial lung cell lines compared to cancer cell lines (Fig. S8). Hierarchical clustering led to a separation of the KRAS cell lines and normal epithelial cell Row Z-score Differential circRNA expression in lung cancer (n = 7) versus normal epithelial (n = 2) cell lines A549 HOP-62 PC9 HC-827 H1666 H3122 H2228 AALE HBEC30KT 0 20 40 60 Cell lines Nº different circRNAs detected (arbirary units) AB C Fig. 2. CircRNA analysis in lung cancer (A549, HOP-62, PC9, HCC-827, H1666, H3122, and H2228) and epithelial cells (AALE and HBEC30KT). (A) Bar plot representing total circRNAs detected in each of the cell lines. (B) Hierarchical clustering of cell lines based on circRNA expression. (C) Differential circRNA expression analysis of log 2 -normalized counts between lung cancer and normal lung cells. 7Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. C. Pedraz-Valdunciel et al.nCounter for circRNA detection in NSCLC samples lines from the rest (Fig. 2B). The two EGFR mutant cell lines positioned together, showing a distinctive group of downregulated circRNAs (circBNC2, circCLK1, circCHD2, and circNUPL2) compared to the rest of the cell lines. Finally, differential expression analysis revealed four circRNAs that allowed for differentiation between the seven-lung cancer cells and normal epithelial cells. CircPIK3R1, circFARSA, and circCHST15 were found downregulated in the cancer cell lines, while circFUT8 was upregulated (Fig. 2C). 3.4. circRNA expression in FFPE NSCLC versus nontumor tissue A total of 53-lung cancer samples and 16 control tissue samples were selected and processed with the circRNA nCounter panel. Initial analysis included normalization of counts for each circRNA as described in the methods section, followed by unsupervised hierarchical clustering of patient samples based on total circRNA expression. A partial separation between cohorts was achieved, indicating a group of circRNAs with discriminatory potential (Fig. 3A). A differential expression analysis revealed a cluster of 10 differentially expressed circRNAs. CircEPB41L2, circBNC2, circSOX13, and circFOXP1 were downregulated in lung cancer tissues, while circRUNX1, circCHD9, circACACA, circFUT8, circRHOQ, and circC1ORF116 were overexpressed (Fig. 3B). Additionally, we also investigated the possible differences in circRNA expression based on the smoking habits of the lung cancer cohort. As a result, four circRNAs (circCSPP1, circNEDD4L, circSOX13, and circCORO1C) negatively correlated with smoking status with P=0.015, P=0.043, P=0.017, and P=0.045 respectively (Student’s t-test). Next, a ML approach was used to develop a circRNA signature predictive of lung cancer. Due to the low number of samples to be analyzed (n=59), we decided to use LOOCV as a validation model, which considers only one sample for testing in each interaction reducing the bias to the minimum when compared to other techniques such as stratified cross validation. As a result, a RFE algorithm selected an 8-circRNA signature (including circSOX13, circEPB41L2, circFOXP1, circBNC2, circCORO1C, circCHD9, circSNX25, and circPIK3R1) as the final model, providing a ROC AUC of 0.965, 0.953, 0.983 with RF, KNN, and GBM classifiers, respectively (Fig. 3C). A PPV of 98.1% and NPV of 81.2% were achieved with the final model. The accuracy, sensitivity, and specificity of the signature were of 97.1%, 94.5%, and 92.8%, respectively. 3.4.1. CircRNA expression in early-stage NSCLC tissues Next, the 27 early-stage NSCLC samples (stages I– IIIA) of our cohort were compared to the 26 late-stage specimens (stages IIIB and IV) (Table 1) to assess those differentially expressed circRNAs emerging early in the disease. From the 41 circRNAs expressed in early-stage samples, 39 were shared with late-stage samples (Fig. 4A). Only 6 out of these 39 transcripts were differentially expressed when compared with the control specimens (Fig. 4B). Interestingly, one of these circRNAs (circFUT8) was found upregulated in both lung cancer tissues and lung cancer cell lines. To shed some light on the potential targets of these six circRNAs, a circRNA-miRNA network was built based on sequence-pairing prediction (Fig. 5). Using circinteractome database, 64 miRNAs were found to potentially bind to differentially expressed circRNAs, with 29 of them showing more than 1 binding site (Fig. S9). Additional ML analysis was performed in early-stage lung cancer and control samples. RFE algorithm provided a signature that included four circRNAs (circEPB41L2, circSOX13, circBNC2, circCORO1C) and provided a ROC AUC of 0.981, 0.918 with RF and KNN, respectively (Fig. 6A). PPV and NPV were of 92.6% and 87.5%, whereas accuracy, sensitivity, and specificity were of 90.6%, 92.6%, and 87.5%, respectively with the selected model. Hierarchical clustering based on the four circRNA included in the signature allowed a clear differentiation between both cohorts (Fig. 6B). 3.5. Univariate analysis related to lung cancer risk We then explored if certain patient characteristics could provide risk factors for lung cancer by performing a univariate analysis (Fig. 7). Several characteristics that could be associated with higher risk of lung cancer such as age, gender, and smoking status were evaluated. No significant association could be found between lung cancer and any of the characteristics previously mentioned. However, presented signatures for lung cancer and early-lung cancer classification were found to be significant predictive factors for lung cancer, with an odds ratio of 371 and 91, respectively. 3.6. Validation by RT-qPCR and Sanger sequencing of circRNA junction sites CircEPB41L2, circSOX13 and circBNC2 not only were significantly downregulated both in early and late 8Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. nCounter for circRNA detection in NSCLC samples C. Pedraz-Valdunciel et al. stages showing the highest fold-change, but also were selected by the RFE algorithm as part of the two predictive signatures. As a result, these 3 targets were selected to validate the nCounter performance using RT-qPCR. Divergent primers and probes spanning the junction sites were designed for the specific amplification of cited circular transcripts (Fig. 8A) in 10 NSCLC and 10 control FPPE tissue samples previously assessed with the nCounter circRNA panel. RT-qPCR results correlated with the data previously obtained from nCounter, indicating downregulation of cited circRNAs in NSCLC samples (Fig. 8B). A gel electrophoresis of the PCR products revealed three FFPE tissues Lung cancer (n = 53) versus non cancer (n = 16). ROC curve (RFE, 8 circRNA signature) PREDICTION REFERENCE Healthy Cancer Class.Err. Healthy Cancer 13 3 0.1875 1 52 0.0188 AC B Fig. 3. (A) Heatmap showing the circRNA expression in lung cancer and control specimens. Unsupervised clustering was performed based on total circRNA expression. (B) Volcano plot showing the circRNA log 2 fold-change in FFPE lung cancer (n=53) versus control (n=16) FFPE tissues. (C) Area under the ROC curve for the classification of lung cancer and control samples. Confusion matrix was generated based on the RF classification scores. Classification error scores are indicated. 9Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. C. Pedraz-Valdunciel et al.nCounter for circRNA detection in NSCLC samples cancer. Cancer Lett. 2018;414:301–9. https://doi.org/10. 1016/j.canlet.2017.11.022 8 Zhou W-Y, Cai Z-R, Liu J, Wang D-S, Ju H-Q, Xu RH. Circular RNA: metabolism, functions and interactions with proteins. Mol Cancer. 2020;19:172. https://doi.org/10.1186/s12943-020-01286-3 9 Ashwal-Fluss R, Meyer M, Pamudurti Nagarjuna R, Ivanov A, Bartok O, Hanan M, et al. circRNA biogenesis competes with pre-mRNA splicing. Mol Cell. 2014;56:55–66. https://doi.org/10.1016/j.molcel.2014.08. 019 10 Starke S, Jost I, Rossbach O, Schneider T, Schreiner S, Hung L-H, et al. Exon circularization requires canonical splice signals. Cell Rep. 2015;10:103–11. https://doi.org/10.1016/j.celrep.2014.12.002 11 Yu C-Y, Kuo H-C. The emerging roles and functions of circular RNAs and their generation. J Biomed Sci. 2019;26:29. https://doi.org/10.1186/s12929-019-0523-z 12 Zhou R, Wu Y, Wang W, Su W, Liu Y, Wang Y, et al. Circular RNAs (circRNAs) in cancer. Cancer Lett. 2018;425:134–42. https://doi.org/10.1016/j.canlet.2018. 03.035 13 Pedraz-Valdunciel C, Rosell R. Defining the landscape of circRNAs in non-small cell lung cancer and their potential as liquid biopsy biomarkers: a complete review including current methods. Extracell Vesicles Circ Nucl Acids. 2021;2:179–201. 14 Prokopec SD, Watson JD, Waggott DM, Smith AB, Wu AH, Okey AB, et al. Systematic evaluation of medium-throughput mRNA abundance platforms. RNA. 2013;19:51–62. https://doi.org/10.1261/rna. 034710.112 15 San Segundo-Val I, Sanz-Lozano CS. Introduction to the gene expression analysis. In: IsidoroGarc ıa M, editor. Molecular genetics of asthma. New York, NY: Springer New York; 2016. p. 29-43. 16 Kulkarni MM. Digital multiplexed gene expression analysis using the NanoString nCounter system. Curr Protoc Mol Biol. 2011;94:25B.10.21–25B.10.17. https:// doi.org/10.1002/0471142727.mb25b10s94 17 Philippeos C, Hughes RD, Dhawan A, Mitry RR. Introduction to cell culture. In: Mitry RR, Hughes RD, editors. Human cell culture protocols. Totowa, NJ: Humana Press; 2012. p. 1–13. 18 Aguado C, Gim enez-Capit an A, Rom an R, Rodr ıguez S, Jordana-Ariza N, Aguilar A, et al. RNA-based multiplexing assay for routine testing of fusion and splicing variants in cytological samples of NSCLC patients. Diagnostics (Basel). 2020;11:15. https://doi. org/10.3390/diagnostics11010015 19 Dahl M, Daugaard I, Andersen MS, Hansen TB, Grønbæk K, Kjems J, et al. Enzyme-free digital counting of endogenous circular RNA molecules in Bcell malignancies. Lab Invest. 2018;98:1657–69. https:// doi.org/10.1038/s41374-018-0108-6 20 Olivier M, Asmis R, Hawkins GA, Howard TD, Cox LA. The need for multi-omics biomarker signatures in precision medicine. Int J Mol Sci. 2019;20:4781. https:// doi.org/10.3390/ijms20194781 21 Zhang H-D, Jiang L-H, Sun D-W, J-c H, Ji Z-L. CircRNA: a novel type of biomarker for cancer. Breast Cancer. 2018;25:1–7. https://doi.org/10.1007/s12282-0170793-9 22 Zhang J, Zhang X, Li C, Yue L, Ding N, Riordan T, et al. Circular RNA profiling provides insights into their subcellular distribution and molecular characteristics in HepG2 cells. RNA Biol. 2019;16:220– 32. https://doi.org/10.1080/15476286.2019.1565284 23 Das Mahapatra K, Pasquali L, Søndergaard JN, Lapins J, Nemeth IB, Balt as E, et al. A comprehensive analysis of coding and non-coding transcriptomic changes in cutaneous squamous cell carcinoma. Sci Rep.2020;10:3637. https://doi.org/10.1038/s41598-020-59660-6 24 Moldovan L-I, Hansen TB, Venø MT, Okholm TLH, Andersen TL, Hager H, et al. High-throughput RNA sequencing from paired lesionaland non-lesional skin reveals major alterations in the psoriasis circRNAome. BMC Med Genomics. 2019;12:174. https://doi.org/10. 1186/s12920-019-0616-2 25 Ahmadov U, Bendikas MM, Ebbesen KK, Sehested AM, Kjems J, Broholm H, et al. Distinct circular RNA expression profiles in pediatric ependymomas. Brain Pathol. 2021;31:387–92. https://doi.org/10.1111/bpa.12922 26 Li S, Sun X, Miao S, Lu T, Wang Y, Liu J, et al. hsa_circ_0000729, a potential prognostic biomarker in lung adenocarcinoma. Thoracic Cancer. 2018;9:924–30. https://doi.org/10.1111/1759-7714.12761 27 Liu X-X, Yang Y-E, Liu X, Zhang M-Y, Li R, Yin YH, et al. A two-circular RNA signature as a noninvasive diagnostic biomarker for lung adenocarcinoma. J Transl Med. 2019;17:50. https://doi. org/10.1186/s12967-019-1800-z 28 Yan Y, Zhang R, Zhang X, Zhang A, Zhang Y, Bu X. RNA-Seq profiling of circular RNAs and potential function of hsa_circ_0002360 in human lung adenocarcinom. Am J Transl Res. 2019;11:160–75. 29 Zhang F, Zhao X, Dong H, Xu J. circRNA expression analysis in lung adenocarcinoma: comparison of paired fresh frozen and formalin-fixed paraffin-embedded specimens. Biochem Biophys Res Comm. 2018;500:738– 43. https://doi.org/10.1016/j.bbrc.2018.04.145 30 Zhao J, Li L, Wang Q, Han H, Zhan Q, Xu M. CircRNA expression profile in early-stage lung adenocarcinoma patients. Cell Physiol Biochem. 2017;44:2138–46. https://doi.org/10.1159/000485953 31 Hossain ST, Malhotra A, Deutscher MP. How RNase R degrades structured RNA: role of the helicase activity and the S1 domain. J Biol Chem. 2016;291:7877–87. https://doi.org/10.1074/jbc.M116. 717991 16 Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. nCounter for circRNA detection in NSCLC samples C. Pedraz-Valdunciel et al. 32 Westholm Jakub O, Miura P, Olson S, Shenker S, Joseph B, Sanfilippo P, et al. Genome-wide analysis of Drosophila circular RNAs reveals their structural and sequence properties and age-dependent neural accumulation. Cell Rep. 2014;9:1966–80. https://doi.org/ 10.1016/j.celrep.2014.10.062 33 Guo JU, Agarwal V, Guo H, Bartel DP. Expanded identification and characterization of mammalian circular RNAs. Genome Biol. 2014;15:409. https://doi. org/10.1186/s13059-014-0409-z 34 Salzman J, Chen RE, Olsen MN, Wang PL, Brown PO. Cell-type specific features of circular RNA expression. PLoS Genet. 2013;9:e1003777. https://doi.org/10.1371/ journal.pgen.1003777 35 Szabo L, Salzman J. Detecting circular RNAs: bioinformatic and experimental challenges. Nat Rev Genet. 2016;17:679–92. https://doi.org/10.1038/nrg.2016.114 36 Zhang J, Chen S, Yang J, Zhao F. Accurate quantification of circular RNAs identifies extensive circular isoform switching events. Nat Commun. 2020;11:90. https://doi.org/10.1038/s41467-019-13840-9 37 Bachmayr-Heyda A, Reiner AT, Auer K, Sukhbaatar N, Aust S, Bachleitner-Hofmann T, et al. Correlation of circular RNA abundance with proliferation – exemplified with colorectal and ovarian cancer, idiopathic lung fibrosis and normal human tissues. Sci Rep. 2015;5:8057. https://doi.org/10.1038/srep08057 38 Yang J, Jia Y, Wang B, Yang S, Du K, Luo Y, et al. Circular RNA CHST15 Sponges miR-155-5p and miR194-5p to promote the immune escape of lung cancer cells mediated by PD-L1. Front Oncol. 2021;11:595609. https://doi.org/10.3389/fonc.2021.595609 39 Hang D, Zhou J, Qin N, Zhou W, Ma H, Jin G, et al. A novel plasma circular RNA circFARSA is a potential biomarker for non-small cell lung cancer. Cancer Med. 2018;7:2783–91. https://doi.org/10.1002/cam4.1514 40 Huang T, She K, Peng G, Wang W, Huang J, Li J, et al. MicroRNA-186 suppresses cell proliferation and metastasis through targeting MAP3K2 in non-small cell lung cancer. Int J Oncol. 2016;49:1437–44. https://doi. org/10.3892/ijo.2016.3637 41 Ruan L, Chen J, Ruan L, Yang T, Wang P. MicroRNA-186 suppresses lung cancer progression by targeting SIRT6. Cancer Biomark. 2018;21:415–23. https://doi.org/10.3233/CBM-170650 42 Zhu H, Lu Q, Lu Q, Shen X, Yu L. Matrine regulates proliferation, apoptosis, cell cycle, migration, and invasion of non-small cell lung cancer cells through the circFUT8/ miR-944/YES1 axis. Cancer Manag Res. 2021;13:3429–42. https://doi.org/10.2147/CMAR.S290966 43 Tan Z, Zheng H, Liu X, Zhang W, Zhu J, Wu G, et al. MicroRNA-1229 overexpression promotes cell proliferation and tumorigenicity and activates Wnt/bcatenin signaling in breast cancer. Oncotarget. 2016;7:24076–87. https://doi.org/10.18632/oncotarget.8119 44 Liu J, Wang D, Long Z, Li W. CircRNA8924 promotes cervical cancer cell proliferation, migration and invasion by competitively binding to MiR-518d-5p /519-5p family and modulating the expression of CBX8. Cell Physiol Biochem. 2018;48:173–84. https://doi.org/ 10.1159/000491716 45 Fasihi A, Soltani BM, Ranjbaran ZS, Bahonar S, Norouzi R, Nasiri S. Hsa-miR-942 fingerprint in colorectal cancer through Wnt signaling pathway. Gene. 2019;712:143958. https://doi.org/10.1016/j.gene.2019. 143958 46 Ge C, Wu S, Wang W, Liu Z, Zhang J, Wang Z, et al. miR-942 promotes cancer stem cell-like traits in esophageal squamous cell carcinoma through activation of Wnt/b-catenin signalling pathway. Oncotarget. 2015;6:10964–77. Supporting information Additional supporting information may be found online in the Supporting Information section at the end of the article. Fig. S1. nCouter probe design allows specific recogition of the circRNAs included in the panel. Fig. S2. nCounter workflow for circRNA expression studies in FFPE lun tissues. Fig. S3. Reproducibility experiment comparing the log2 of normalized counts by nCounter from three independent RNA samples derived from the PC9 cell line. Fig. S4. Bioanalyzer profiles of faired fresh (left) and FFPE (right) PC9 cell line-derived RNA. Fig. S5. nCounter analysis of FFPE PC9 cell line. Fig. S6. Total RNA concentration assessment for circRNA analysis using the nCounter platform. Fig. S7. Venn diagram showing circRNAs identified in all healthy cells (19) versus those only expressed in all lung cancer cell lines (1). Fig. S8. Overall total number of raw counts in lung cells. Fig. S9. Different miRNA binding sites of dysregulated circRNAs in early-stage lung cancer tissues. Fig. S10. Diagram showing the tracking of those circRNAs of the circRNA nCounter panel not detected in assessed FFPE tissues. Table S1. Diagnosis and associated pathologies of the control cohort. Table S2. Characteristics of FFPE samples included in the study. Tumor and lymphocyte infiltration is indicated. Table S3. circRNA and mRNA candidates included in the nCounter panel. 17Molecular Oncology (2022) ª2022 The Authors. Molecular Oncology published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies. C. Pedraz-Valdunciel et al.nCounter for circRNA detection in NSCLC samples