1 Development and validation of a novel signature to predict overall survival in “driver-gene-negative” lung adenocarcinoma (LUAD): results of a multicenter study

3 stratification, classification signature, Wnt/β-catenin-pathway pathway” identified 27 articles. However, we did not identify any previous articles that had explored the potential prognostic role of developmental genomics in “driver-gene-negative” LUAD. Therefore, there is an urgent need for precise risk stratification to this LUAD subgroup. In this study, we found that the Wnt/β-catenin pathway was a dominant component in the top ten enriched pathways in “driver-gene-negative” LUAD and we performed a retrospective study to test the ability of Wnt/β-catenin-pathway-based protein expression profiles to predict patient prognosis at the time of diagnosis. A prognosis-related CSDW signature was identified successfully. Our study indicated that patients with “driver-gene-negative” LUAD can be accurately sorted into low-risk and high-risk groups according to the CSDW signature. This signature could be helpful for clinical decision-making, guiding follow-up schedules and developing new therapy targets for “driver-gene-negative” LUAD. Abstract Purpose Examining the role of developmental signaling pathways in “driver-gene-negative” LUAD (LUAD patients negative for EGFR, KRAS, BRAF, HER2, MET, ALK, RET and ROS1 were identified as “driver-gene-negative”) may shed light on the clinical research and treatment for this LUAD subgroup. We aimed to investigate whether developmental signaling pathways activation can stratify the risk of “driver-gene-negative” LUAD. using


Introduction
Lung adenocarcinoma (LUAD) accounts for approximately 40% of lung cancer cases and remains a major cause of cancer mortality worldwide (1). Over the last decade, targeting oncogenic driver genes, primarily epidermal growth factor receptor (EGFR) and anaplastic lymphoma kinase (ALK), has significantly prolonged the survival of LUAD patients (2,3). Several genomic sequencing programs have indicated that more than half of LUAD patients present with EGFR/KRAS/ALK gene mutations (4). Recent immunotherapy, such as the use of a PD-1/PD-L1 monoclonal antibody, has shed new light on the remaining "driver-gene-negative" LUAD cases (5), which approximately account for 14%-20% in LUAD subtype (3,4). However, the large number of non-responders and the immune-related toxicities of immune check-point inhibitors present new challenges for clinical application of antibodies targeting the PD-1/PD-L1 axis in LUAD (6,7). Thus, ongoing research to develop new biomarkers that reflect tumor heterogeneity will guide individual treatment for "driver-gene-negative" LUAD patients.
on July 10, 2020. © 2018 American Association for Cancer clincancerres.aacrjournals.org Downloaded from Author manuscripts have been peer reviewed and accepted for publication but have not yet been edited.
Thus, a comprehensive understanding of development-related signaling pathways involved in cancer pathogenesis is essential for improvement of "driver-gene-negative" LUAD therapy. Herein, we performed a retrospective analysis of a large cohort of patients with "driver-gene-negative" LUAD. This study was designed to initially establish genome-wide transcriptome profiling in 52 patients and then to select candidate genes in 189 training patients, validate those genes in 437 patients from three cohorts and construct a prognostic classifier (a Wnt/β-catenin-pathway-based CSDW signature named by the four candidate genes: CTNNB1 or β-catenin, SOX9, DVL3 and Wnt2b) for those patients.

Study Design
We selected 626 formalin-fixed, paraffin-embedded (FFPE) samples from LUAD patients between September 2003 and June 2015. None of the patients underwent any anti-tumor therapy before biopsy sampling. "driver-gene-negative" status was determined in FFPE tissues according to the workflow shown in Supplementary Fig.   S1. For this study, 371 cases were from the First Affiliated Hospital of Sun Yat-sen University (SYSUFH set) and were divided randomly into training (189 cases) and internal validation (182 cases) cohorts. In the external validation cohorts, 152 cases were from Sun Yat-sen University Cancer Center (SYSUCC set), and 103 cases were from The Central Hospital of Wuhan (WUHAN set). The study was approved by the institutional review board of each hospital center, and was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent.
The entire 626 FFPE samples were confirmed by two pathologists independently ( Supplementary Fig. S2).
The discovery, training and validation phases were designed to identify and validate the prognosis-related signature of "driver-gene-negative" LUAD ( Supplementary Fig.   S3). In the discovery phase, a total of 60 pairs of fresh tumor and adjacent normal tissues were selected randomly from the 189 patients with driver-gene negative LUAD in the SYSUFH set. The selected 60 pairs samples must meet the following criteria: 15  Agilent Feature Extraction software (version 11.0.1.1) was used to analyze acquired array images. Quantile normalization and subsequent data processing were performed using the GeneSpring GX v12.1 software package (Agilent Technologies).
Genes that were expressed at a low level or close to the background level were excluded from further analyses. After quantile normalization of the raw data, genes that had at least 1 out of 3 samples flagged as detected ("All Targets Value") were chosen for further data analysis. Differentially expressed genes with statistical significance (p<0.05) between 52 paired LUAD and adjacent normal tissue samples were identified through volcano plot filtering and fold change filtering. Hierarchical clustering was performed using the R scripts. GO analysis and KEGG pathway analysis were performed using the standard enrichment computation method (13,14).
The genome-wide microarray data were available from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus with accession number GSE115002.

Tissue Microarrays (TMAs)
TMAs are increasingly becoming the preferred method for large-scale and epidemiological biomarker investigations (15,16). To prepare the TMA blocks, a tissue section stained with hematoxylin and eosin (HE) was prepared from each patient, and then, representative paired LUAD and adjacent normal tissues were immunoblotting. GAPDH and histone served as the cytoplasmic and nuclear control, respectively. Immunoblotting of the other 40 proteins was analyzed as described previously (19). Antibodies were diluted in a 5% milk/TBST solution containing 5% NaN3. The antibody list and dilution ratios are presented in Table S1. Signals were captured using an enhanced chemiluminescence plus kit (Millipore, Billerica, USA).

Quantitative Real-time PCR (Q-PCR)
Total RNA extraction from fresh LUAD tissues was performed using a Qiagen Research.

Statistical Analysis
In the discovery phase, volcano plot analysis was used to select potential genes based on absolute fold change in combination with t-test p-values. In the training phase, independent prognostic factors were identified using a LASSO Cox regression model. The Kaplan-Meier method was applied using GraphPad Prism 5.0 software and was used to analyze survival differences between groups. Multivariate survival analysis was performed using the Cox regression model to test the independent significance of various factors. Covariates included CSDW (high risk vs low risk), age (≥60 years vs <60 years); gender (female vs male), differentiation degree (low, median vs high), tumor size (≥5 cm vs <5 cm) and TNM stage (III/IV vs I/II). Most patients, especially those terminally ill patients with stage III/IV, received multiple treatment, which mainly included chemotherapy with different drugs, radiation therapy with different intensity and cycles and even some patients had tried Chinese medicine. It is extremely difficult for us to evaluate the distinguishing ability of CSDW risk score in stage III/IV patients with diverse therapeutic schedule. Hence, we excluded the effect of the complicated treatment factor on the prognosis in our analyses.
The raw mean optical density for each protein in each experiment was adjusted for technical normalization before biological normalization. X-tile software (version 3.6.1, Yale University School of Medicine, New Haven, CT, USA) was used to determine the optimal cutoff value for the risk score. The optimal cutoff value was Research.
on July 10, 2020. © 2018 American Association for Cancer clincancerres.aacrjournals.org Downloaded from Author manuscripts have been peer reviewed and accepted for publication but have not yet been edited.
Author Manuscript Published OnlineFirst on November 2, 2018; DOI: 10.1158/1078-0432.CCR-  defined as the value that produced the largest χ2 in the Mantel-Cox test and was used to separate patients into low-and high-risk subgroups (20,21). Due to the high dimension and correlation between some covariates, the variable selection was performed with LASSO-Cox regression, which is a well-established method for selection of the most predictive markers for time-to-event analysis (21,22). The regression coefficients penalize the size of the parameter, and thus, unimportant variables (variables whose coefficients are close to zero) are removed from the model. LASSO regression shrinks the coefficient estimates toward zero, with the degree of shrinkage dependent on an additional parameter, λ. Ten-time cross-validations were used to determine the optimal values of λ (23,24), and we chose λ via 1-s.e. criteria (25). The coefficients of the multivariable Cox regression model were used to construct a nomogram with the "rms" package of R software.
The performance of the nomogram was assessed by the concordance index (C-index) via a bootstrap method and was explored graphically by calibration curves.
Time-dependent receiver operating characteristic (ROC) analysis was applied to compare the predictive accuracy of CSDW with the clinicopathological parameters.
All other statistical tests were performed with R software version 3.0.3 (R Foundation for Statistical Computing). A 2-sided p-value< 0.05 was considered statistically significant.

Patient Characteristics
The detailed clinical characteristics of 626 LUAD patients from three medical

Activation of the Wnt/β-catenin Signaling Pathway in "driver-gene-negative" LUAD
To identify the dominant genes or signaling cascades involved in oncogenesis and development of "driver-gene-negative" LUAD, a genome-wide microarray was employed to look for differential mRNA expression between 52 paired LUAD and adjacent normal tissues Supplementary Table S3). We found 960 differentially expressed genes (DE genes; |Log2FC|≥2, Tumor vs Normal; p<0.01), as shown in the volcano plot in Fig. 1A. Given the dispersed distribution of the DE genes, we applied KEGG (Kyoto Encyclopedia of Genes and Genomes) and GO (Gene Ontology) analyses to explore the role of development-related signaling pathways in LUAD pathogenesis and found that the Wnt/β-catenin pathway was a dominant component in the top ten enriched pathways ( Fig. 1B; Supplementary Fig. S4). Thus, we hypothesized that the Wnt/β-catenin pathway might be abnormally activated in "driver-gene-negative" LUAD. To confirm this hypothesis, we detected the location  Fig. S5). These findings combined with β-catenin non-specific expression characteristics in EGFR mt , KRAS mt and ALK ft LUAD ( Supplementary   Fig. S6) substantially supported the hypothesis that the Wnt/β-catenin pathway is specifically activated in "driver-gene-negative" LUAD.

Establishing the Wnt/β-catenin-pathway-based Prognostic Prediction Model
To establish a Wnt/β-catenin-pathway-based prognostic signature based on the genome-wide microarray, we focused on the differential mRNA expression profiles   Fig. S7).
We then used immunohistochemistry (IHC) to semi-quantify the protein expression of these 41 candidate molecules in 189 FFPE tissues (SYSUFH set) (Fig. 1E CDH10 and Wnt2b) were selected from 41 candidate genes by LASSO regression (Fig. 1G). Then the five candidate variables underwent multivariate Cox regression analysis of overall survival in the training cohort. And we found that there was no statistical significance for CDH10. Finally, β-catenin, SOX9, DVL3 and Wnt2b remained as the prognosis-related molecules, and the coefficients were obtained from the Cox proportional hazard model. In a word, multivariate LASSO Cox regression identified a prognosis-related CSDW signature comprising β-catenin (CTNNB1), SOX9, DVL3 and Wnt2b. The following prognosis risk score of the CSDW signature was then constructed using X-tile software based on the following four molecules: Risk score = (0.1481 × expression of β-catenin) + (0.2424 × expression of SOX9) + (0.2770 × expression of DVL3) -(0.2536 × expression of Wnt2b). The optimal cutoff value of the risk score was finally set to 3.34, which separated patients into the low-risk and high-risk groups in the SYSUFH training cohort ( Fig. 2A). The distribution of risk scores for CSDW showed that patients with lower risk scores generally had a better survival outcome than those with higher risk scores ( Fig. 2A, left panel). Compared with patients with low-risk scores, patients with high-risk scores also had shorter overall survival (HR=10.42; 95% CI: 6.46 to 16.79; p<0.001; Fig. 2A, right panel).

Validating the Wnt/β-catenin-pathway-based Prognostic Prediction Model at Three Independent Centers
To evaluate the reproducibility and stability of the CSDW signature, we performed one internal validation using the SYSUFH group (182 cases) and two external  Fig. 2D). In all independent cohorts, patients with low-risk scores had better progression-free survival (PFS) (p<0.001; Supplementary Fig. S9). For the entire cohort, 626 patients were separated into CSDW-defined low-risk (331; 52.9%) or high-risk subgroups (295; 47.1%) (p<0.001; Supplementary Fig. S10). After multivariable adjustment for clinical prognostic factors (e.g., age and TNM stage), the CSDW signature remained an independent prognostic factor in the training and validation cohorts (Table 2, all p<0.001).
In addition, the optimum cutoff scores of the four genes in the CSDW signature were generated based on the entire cohort using X-tile statistics as follows: 8.  Supplementary Fig. S11). For the entire cohort of 626 patients, protein expression levels of β-catenin, SOX9, DVL3 and Wnt2b were strongly associated with lymphatic metastasis and distant organ metastasis (Supplementary Fig. S12).

Stratification Analysis of the Wnt/β-catenin-pathway-based Prognostic Prediction Model
When the entire cohort was stratified by clinical variables (sex, age, tumor size, differentiation degree and TNM stage), the CSDW signature was still a statistically and clinically significant prognostic model ( Fig. 3; Table 2; Supplementary Fig. S13).
Based on multivariate analysis of OS (Table 2), we developed a clinically applicable nomogram, which integrated both CSDW and clinicopathological covariates to  Table S7). Statistical analysis showed that the inter-sample standard deviation (SD) was significantly larger than the intra-sample SD ( Supplementary Fig. S16), indicating that the CSDW signature is a reliable prediction tool.

Discussion
With advances in molecular biology, multigene-based profiles have been used for risk stratification to guide treatment for various cancer types (27,28). A number of traditional mRNA expression-based prognostic techniques, combined with IHC, genome-wide microarrays, CpG microarrays and TMAs, have been demonstrated to more precisely evaluate the prognosis of patients with lung cancer (28,29). In this multicenter study, we found that the Wnt/β-catenin pathway was involved in the development of "driver-gene-negative" LUAD, and we established a CSDW signature comprising β-catenin, SOX9, DVL3 and Wnt2b, which was found to have a more powerful prognosis-prediction ability. Furthermore, this novel prognostic signature was successfully validated in an internal validation cohort and in two external validation cohorts. To the best of our knowledge, this study is the most comprehensive study to date demonstrating the prognostic significance of Wnt/β-catenin pathway activation in "driver-gene-negative" LUAD patients. This novel CSDW signature stratified "driver-gene-negative" LUAD patients into two distinct high-and low-risk subgroups. In general, high-risk patients should receive high-frequency surveillance and corresponding treatment to prevent disease progression (30). In the present study, most of the disease progression occurred within 20 months after resection for high-risk patients, which may reflect a true In the stratified analysis of tumor differentiation, tumor size, sex and age, the CSDW signature subdivided patients into significant high-and low-risk subgroups. The CSDW signature subdivided patients with stage I or II LUAD into two subgroups with tremendously different prognoses. Conversely, the distinguishing ability was weak for patients with stage III and IV disease. This difference may be primarily due to the high individual variation between patients with stage III and IV LUAD because terminally ill patients receive multimodal treatment, which may affect follow-up results (31).
The Wnt/β-catenin signaling pathway, a developmental signaling pathway, plays an important role in carcinogenesis and progression. (8,9) Consistent with this, we found that Wnt/β-catenin pathway activation is related to the prognosis of "driver-gene-negative" LUAD patients. β-Catenin, SOX9, and DVL3 were upregulated and Wnt2b was downregulated in "driver-gene-negative" LUAD tissues compared with adjacent normal tissues. Among these prognosis-related molecules, β-catenin and SOX9 have been reported to be involved in development and stem cell differentiation in many previous studies (32,33). Several reports have revealed that DVL3 acts as a recurrence predictor in prostate adenocarcinoma and that Wnt2b is a Research.
on July 10, 2020. © 2018 American Association for Cancer clincancerres.aacrjournals.org Downloaded from Author manuscripts have been peer reviewed and accepted for publication but have not yet been edited.
Prior to this report, several studies had analyzed the clinical practice of using gene expression profiles or CpG signatures as markers in NSCLC (36)(37)(38)(39). However, the lack of an independent validation cohort (37,38), an insufficient number of patients (39) and unintegrated high-throughput biomarkers (38) have restricted their widespread clinical application. In this study, we focused on "driver-gene-negative" LUAD, a rare subtype of NSCLC, and developed a feasible prognostic model based on FFPE tissue using a practical IHC-based assay. More importantly, we used a whole-genome microarray that included Notch, Hedgehog and TGF-β/smad pathways (40)(41)(42) to screen development-specific Wnt/β-catenin activation, thus providing a more specific range of options for future multitarget drug development for "driver-gene-negative" LUAD. Because large intra-sample variability might distort verification of results (29,43) and to reduce heterogeneity among different patients, ITH was taken into consideration, and the CSDW signature was validated in multicenter samples. Additionally, unlike previous signatures that utilize many genes detected by microarray or reverse transcription-PCR (37,38), we used a LASSO algorithm to incorporate Wnt/β-catenin-based immunomarkers to identify four highly effective prognostic markers. Given the smaller number of immunomarkers, our signature is more clinically feasible and will pave the way for Fourth, although we validated the CSDW signature using independent internal and external cohorts, its prognostic utility should be extended further in prospective cohorts.
In summary, we demonstrated that a Wnt-pathway-based prognosis-prediction panel is an appropriate tool for predicting OS and PFS of "driver-gene-negative" LUAD patients. Moreover, this prognostic marker panel may help clinicians develop individualized treatment programs for these patients. The results suggest that the Wnt signaling pathway is correlated with lung cancer progression and is a potential target for clinical treatment.        Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) . http://clincancerres.aacrjournals.org/content/early/2018/11/02/1078-0432.CCR-  To request permission to re-use all or part of this article, use this link Research.
on July 10, 2020. © 2018 American Association for Cancer clincancerres.aacrjournals.org Downloaded from Author manuscripts have been peer reviewed and accepted for publication but have not yet been edited.