Lung adenocarcinoma and lung squamous cell carcinoma cancer classification, biomarker identification, and gene expression analysis using overlapping feature selection methods

Lung cancer is one of the deadliest cancers in the world. Two of the most common subtypes, lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), have drastically different biological signatures, yet they are often treated similarly and classified together as non-small cell lung cancer (NSCLC). LUAD and LUSC biomarkers are scarce, and their distinct biological mechanisms have yet to be elucidated. To detect biologically relevant markers, many studies have attempted to improve traditional machine learning algorithms or develop novel algorithms for biomarker discovery. However, few have used overlapping machine learning or feature selection methods for cancer classification, biomarker identification, or gene expression analysis. This study proposes to use overlapping traditional feature selection or feature reduction techniques for cancer classification and biomarker discovery. The genes selected by the overlapping method were then verified using random forest. The classification statistics of the overlapping method were compared to those of the traditional feature selection methods. The identified biomarkers were validated in an external dataset using AUC and ROC analysis. Gene expression analysis was then performed to further investigate biological differences between LUAD and LUSC. Overall, our method achieved classification results comparable to, if not better than, the traditional algorithms. It also identified multiple known biomarkers, and five potentially novel biomarkers with high discriminating values between LUAD and LUSC. Many of the biomarkers also exhibit significant prognostic potential, particularly in LUAD. Our study also unraveled distinct biological pathways between LUAD and LUSC.

Cadherin EGF LAG seven-pass G-type receptor 2 TUBA1C Tubulin alpha 1c S100A2 S100 calcium binding protein A2 KRT5 Keratin 5 KRT14 Keratin 14 KRT6A Keratin 6A TP63 Tumor protein P63 NAPSA Napsin A aspartic peptidase MLPH Melanophilin DSC3 Desmocollin 3 Lung cancer is the most commonly diagnosed malignant tumor and is a leading cause of cancer-associated mortality. It is the second highest cause of new cancer cases in both genders in the United States and is the second leading cause of cancer deaths in females globally 1,2 . The most common subtypes of lung cancers are lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), classified together as non-small cell lung cancer (NSCLC) 3,4 . However, recent studies have suggested that LUAD and LUSC should be classified and treated as different cancers 5 .
Identifying the mechanisms underlying LUAD and LUSC is needed to develop useful biomarkers for better diagnosis and design therapeutic interventions. Multiple gene expression and immunohistochemistry studies have identified biological pathways and biomarkers that differentiate between LUAD and LUSC [6][7][8] . Other studies classified cancers using both novel and traditional machine learning or feature selection methods [9][10][11][12] . However, few have investigated cancers by applying multiple feature selection methods and selecting the overlapping features.
In this study, we downloaded LUAD and LUSC RNA-Seq datasets from The Cancer Genome Atlas (TCGA) 13 and analyzed them with five feature selection methods with ranking abilities: Differential Gene Expression Analysis (DGE), Principal Component Analysis (PCA), Least absolute shrinkage and selection operator (Lasso), minimal-Redundancy-Maximal Relevance (mRMR), and Extreme Gradient boosting (XGboost). DGE applies a normalization method and uses the negative binomial distribution to detect significant changes in gene expression across samples 14,15 . Many studies have shown that DGE, though being the most widely used algorithm to detect differentially expressed genes, often yields some false positive results; in addition, it is often sensitive to outliers [14][15][16][17] . On the other hand, XGboost is a tree-based machine learning method that is not sensitive to outliers but is prone to overfitting 17,18 . To minimize this problem, we chose to use Lasso, a linear regression technique that avoids overfitting but can be influenced by highly correlated features and potentially leading to false discoveries [17][18][19][20] . mRMR is then used to maximize the relevance between the features and the output, and minimize the relevance among the feature themselves, thus, limiting highly correlated features [21][22][23] . PCA is another well-known and widely used feature reduction technique in machine learning to reduce high dimensional data into orthogonal principal components, which also removes correlated features 17,18 . However, amidst other disadvantages, the result of PCA by itself is often not interpretable 17,18 . These algorithms were also chosen because of their ability to rank features or select a reasonable number of features. In short, overlapping these algorithms is promising because different methods select features using different criteria. Since each method has its strengths and weaknesses, focusing on the overlapping features will optimize the strengths and minimize the weaknesses of each method, thereby reducing the number of false positives and producing reliable results. This study will serve as a proof of concept for the validity of the approach to overlap feature selection methods while investigating NSCLC subtype differences and discovering novel biomarkers.

Results
Study design and overview. We obtained LUAD and LUSC RNA-Seq data from TCGA 13 and the summary of their clinical information was provided in Table 1, with more comprehensive details available on TCGA website 13 . We selected discriminatory genes by overlapping DGE, PCA, mRMR, XGboost, and lasso as depicted in Fig. 1. The genes that were overlapped by two or more algorithms were validated and used for LUAD and LUSC classification as well as gene expression analysis. The genes that were overlapped by three or more algorithms were selected as biomarker candidates, and their diagnostic values were assessed using ROC analysis and AUC value, and then further verified in an external dataset, GSE28582 24,25 , which is a microarray dataset that includes 50 LUAD and 28 LUSC samples The prognostic values of the biomarker candidates were also assessed using Kaplan Meier Plotter 26 .  (Table S1) were selected based on the ranking of the algorithm. Also, 148 genes in Xgboost (Table S1) and 68 genes in lasso (Table S1) using probability or prediction threshold of 0.5 were identified and selected. The different number of genes selected was due to the nature of the algorithm, with most of the parameters in each www.nature.com/scientificreports/ algorithm were set to default. The specifics of each metric can be found in the code at the data availability section Since each of these methods has its own selection criteria, the overlapping genes must satisfy multiple selection criteria, making them significant candidate biomarkers that differentiate LUAD and LUSC. Therefore, the five independent sets of top genes were compared with a Venn diagram to identify the overlapping genes detected by multiple algorithms. Venn diagram (Fig. 2) comparison detected 131 genes (Table S2) overlapped by two or more algorithms and 17 genes ( Table 2) overlapped by three or more algorithms.
Validation of selected genes. To evaluate how effective the selected genes are in classifying LUAD and LUSC, we used random forest to validate the top 500 genes selected from PCA, mRMR, and DGE, as well as the 148 genes from xgboost and 68 genes from lasso ( Table S1). All of the validation results for each feature selection method returned high classification accuracies of over 90% (Table 3). To compare to the previous feature  www.nature.com/scientificreports/ selection methods, the overlapping 131 genes were validated the same way as the other algorithms. The binary classification statistics (Table 3) were calculated using LUAD as 'positive' and LUSC as 'negative' . The overlapping 131 genes showed comparable, if not better, results to the other algorithms ( Table 3). The 17 proposed biomarkers also showed to be effective classifiers, having statistics comparable to the other algorithms despite only using 17 genes. Heatmaps for the top 131 and the top 17 genes were also generated ( Fig. 3A,B). Both heatmaps, in particular the heatmap with 17 genes, displayed clear borders separating LUAD from LUSC. Dot plots of the gene expression distribution between LUAD and LUSC for each of the 17 genes are displayed in Fig. 4.  as biomarker candidates 87 . The x-axis represents the samples and the y-axis represents the genes.  (Fig. 5A), KRT5 has the highest AUC of 0.9731, thereby displaying the most significant diagnostic potential in classifying LUAD and LUSC, consistent with the study reported by Jain Xiao et al. 6 in which KRT5 also had the highest diagnostic potential. All of the upregulated genes show significant discrimination potential as well (Fig. 5A,B).
To minimize the inherent RNA expression noise and to ensure that these results are reproducible, an external dataset GSE28582 was used for external validation. AUC and ROC were also used to analyze the 17 genes in GSE28582 validation dataset (Fig. 6). Largely consistent with our result, most of the genes show AUC values well above 0.9; all except one gene, ARHGEF38, have AUC values above 0.8 (Fig. 6).
Kaplan Meier plotter analysis of the 17 potential biomarkers. Of the 17 potential biomarkers (Table 2), only CELSR2 shows a significant prognostic p-value in LUSC, with its higher expression correspond- www.nature.com/scientificreports/ ing to a more favorable prognosis in LUSC (Table 4). In contrast, many genes show significant prognostic potential in LUAD. High expressions of KRT17, KRT6A, S100A2, TRIM29, REPS1, and GPC1 correspond to an unfavorable prognosis in LUAD, while high expressions of PERP, ELFN2, ARHGAP12, and QSOX1 correspond to a favorable prognosis in LUAD (Table 4).  www.nature.com/scientificreports/ GO term enrichment analysis. To further understand the biological differences between LUAD and LUSC, we performed gene expression analysis by splitting the identified 131 genes into two groups: 57 downregulated and 74 upregulated genes in LUSC compared to LUAD. Functional pathway annotation of these two groups of genes was performed using The Database for Annotation, Visualization and Integrated Discovery (DAVID) 27 analysis tool with Gene Ontology (GO) biological pathway enrichments. GO terms with P-value < 0.01 were obtained (Tables S3 and S4). The top 10 most significantly upregulated and downregulated GO terms ranked by p-value are shown in Table 5. In addition, DAVID has the functionality to group similar GO terms into clusters of the same biological pathway. To elucidate the potential biological differences between LUAD and LUSC, the top five most significantly upregulated and downregulated clusters ranked by enrichment scores were determined (Table 6 and Tables S5 and S6).
In the upregulated group, most pathways are concentrated in cell adhesion, intermediate filament organization, and cell junction assembly. In the downregulated group, the most significant cluster is platelet degranulation and cell exocytosis, as well as other pathways such as tyrosine kinase signaling pathway, homeostasis, protein translation and circulatory system. These results suggest that LUSC tends to express more genes related to cell adhesion and cytoskeleton organization, and LUAD tends to express more genes involved in platelet degranulation and exocytosis, along with other signaling pathways.
Reactome gene expression analysis. Reactome pathways 28 were also generated for both upregulated and downregulated groups. The most significantly upregulated pathway is the cornification, or the keratiniza-  www.nature.com/scientificreports/ tion pathway (Fig. 7, Table S7), along with other similar pathways related to cell adhesion, which is consistent with GO term analysis. TP53 regulation pathway, which is often implicated in cancer, is among the top enriched pathways as well (Table S7). For the downregulated group, the most significant pathway is peptide elongation synthesis (Fig. 8, Table S8), which GO term analysis also reveals to be significant.

KEGG gene expression analysis.
Only the p53 signaling pathway appeared in the upregulated group (Table 7) in Kyoto Encyclopedia of Genes and Genomes (KEGG) 29 gene expression analysis. Though it has a p-value of slightly over 0.01, this result is consistent with Reactome analysis which ranks TP53 regulation as the second most upregulated pathway after keratinization and other cell junction related pathways. Only the lysosome seems to be significant in the downregulated group ( Table 7). The lysosomal pathway is coherent with platelet degranulation and exocytosis, as reported in GO term analysis. Even though the ribosomal pathway has a p-value slightly greater than 0.05, it is most likely important as it is also shown to be significantly enriched in both GO and Reactome term analyses (Tables S3 and S8).

Discussion
Previous studies have utilized traditional feature selection and machine learning methods for cancer diagnosis, detection, and classification 10,11,22 , but few have extended them to study potential biomarkers and biological pathways to discriminate between LUAD and LUSC. To improve cancer classification accuracy, novel machine learning, and feature selection methods have been developed 12,[30][31][32] . However, few studies have used overlapping features from different methods for classification, gene expression analysis, and biomarker discovery. To provide a proof of concept for the validity of this method, we took advantage of the capabilities and the strengths of PCA, www.nature.com/scientificreports/ mRMR, XGboost, DGE, and lasso to select 131 overlapping genes for classification and gene expression analysis, and 17 genes for classification and potential biomarkers. Overall, the overlapping 131 genes showed several high-ranking metrics with lasso and PCA methods. Though the best method may vary depending on the metric, the classification result of using both the overlapping 131 and 17 genes was by many metrics comparable if not better than the other methods that use more genes. The 131 overlapped genes achieved the highest sensitivity with PCA, the second highest accuracy with lasso, and the second highest F-measure overall, indicating that overlapping feature selection methods can be used to perform cancer classification. Furthermore this method proves to be valuable in biomarker discovery. In agreement with our result, previous studies have reported levels of several genes to be greatly elevated in LUSC compared to LUAD; these genes include KRT6 6,8,33,34 ,KRT5 6,8,35 ,KRT14 8,33,34 ,KRT17 8,33 , PERP 8,33 , TRIM29 8,33 , GPC1 8 , CELSR2 8 , S100A2 8 , and TUBA1C 36 . Also, consistent with our result, levels of QSOX1 33 and MUC1 8 were reported to be lower in LUSC than in LUAD. Many current biomarkers such as Tumor Protein P63 (TP63), Napsin A Aspartic Peptidase (NAPSA), Melanophilin (MLPH), Desmocollin 3 (DSC3), and others are also part of the top 131 genes selected by our method 33,[37][38][39][40] . To our knowledge, ARHGAP12, ARHGEF38, ELFN2, NECTIN1, and REPS1 are among the top 17 genes in this study to be identified as biomarkers for the first time. All 17 candidate biomarkers, except ARHGEF38, are also validated in GSE28582 exhibiting high discriminating potential. Although the selection of ARHGEF38 may be due to bias in the TCGA dataset, it is important to note that there are many more samples in TCGA compared to GSE28582; GSE28582 as a microarray dataset is also significantly worse than RNAseq at detecting gene expression differences when the expression values are low or when the fold change is less than 2 [41][42][43] . Notably, ARHGEF38 has relatively lower fold change and expression value.
Moreover, studies have shown that biomarkers for diagnosis and prognosis are most reliable when they are biologically related to the disease in addition to being statistically significant 44,45 . Although this study is primarily data-driven, the results reveal biomarkers that would corroborate with a knowledge-based approach. For instance, the most significant candidate biomarkers between LUAD and LUSC are all cytokeratins and cadherins, which is reasonable because they are markers of squamous epithelial cells. In particular, NECTIN1, as a novel cadherin biomarker, consistently demonstrates high discriminating potential both in the TCGA and the external validation dataset; it also directly binds and signals fibroblast growth factor receptor 46 , a pathological signaling pathway that is more prominent in LUSC 47,48 . NECTIN1 also serves a key role in herpes simplex virus type 1 (HSV-1) viral entry and is important in oncolytic therapy in squamous cell carcinomas 49,50 . Similarly, it is logical that MUC1 can be used to identify LUAD, as it is a marker for columnar cells from which LUAD arise. In addition to  www.nature.com/scientificreports/ satisfying the aims of both data-driven and knowledge-based approach, many of the 17 genes identified through this method show significant prognostic importance, particularly in LUAD ( Table 4).
The other candidate biomarkers also show strong association with cancers. ARHGEF38 and ARHGAP12 are both part of the Rho family GTPase regulators. Rho GTPases are essential to cell cytoskeletal structure, motility, and morphogenesis, and they have been implicated in many cancer proliferation and metastases [51][52][53][54] . The other upregulated genes ELFN2, QSOX1, and MUC1 have been shown to directly promote metastasis in various cancers [55][56][57][58][59] , including lung cancer. Furthermore, the loss of certain genes upregulated in LUSC such as TRIM29 and KRT6A is associated with more cellular invasion 60,61 . Clinical differences between LUAD and LUSC are well known. In particular, LUAD has a higher metastatic rate than LUSC 62 . Studying these potential biomarkers may provide insight into tumor progression, metastatic, and therapeutic differences between LUAD and LUSC. Overall, these results not only align with known literature, but also provide reasonable and promising biomarkers, suggesting that using overlapping feature selection methods can be used to reliably detect new biomarkers. With the validity of this overlapping method shown both in cancer classification and biomarker identification, we performed gene expression analysis for further investigation.
Aside from cell adhesion or cytoskeleton organization, LUSC demonstrates higher regulation of p53 signaling in both KEGG and Reactome analyses. It is known that TP53 mutation is more common in LUSC than in LUAD [63][64][65] , and that such mutation may predominantly be a non-truncated mutation in LUSC leading to higher expression levels of genes involved in the p53 regulation pathway 66 . Moreover, P53 mutations often lose their tumor suppression function while gaining oncogenic abilities, leading to increased cell growth and proliferation compared to LUAD 67 .
The most prominent pathway associated with LUAD, compared to LUSC, is platelet degranulation and exocytosis (Tables 5, 6). Interestingly, lung cancer is the most common malignancy to coexist with venous thromboembolism, especially pulmonary embolism 68 . LUAD, in particular, has been shown to be an independent risk factor for pulmonary embolism even among lung cancers 69,70 . Because platelet granulation directly causes thrombus formation, the differential enrichment of platelet granulation pathway can therefore help explain a more active and a more common hypercoagulation and thrombotic process in LUAD compared to LUSC 71 . In addition, platelet degranulation can modulate innate immunity via the release of cytokines, and platelet-leukocyte interactions can lead to leukocyte recruitment and activation in cancer 72 . In fact, CD63, one of the genes in the platelet degranulation pathway (Tables S3 and S6), is directly involved in leukocyte recruitment through endothelial P-selectin 73 . LUSC has recently been associated with a relatively more suppressed immune response, implying a more active immune response in LUAD, which supports our result 67,74 .
There are several limitations of this study. One of them is that this study does not prioritize the RNA expression fold changes, which some groups have used to rank differentially expressed genes 75,76 . Also, although this study aims to minimize the discovery of false positive biomarkers by overlapping different feature selection methods, the proposed biomarker candidates in this study still lack experimental verification. Nevertheless, these results may shed light into the biological differences between LUAD and LUSC, as well as aid the discovery of better diagnosis and treatment for each 77,78 .
In conclusion, we designed and implemented a workflow of overlapping five different feature selection methods to perform cancer classification, identify novel biomarkers, and study biological differences in NSCLC. This overlapping method proves to be reliable in both cancer classification and biomarker identification, yielding statistically promising genes that also support our current knowledge. We identified ARHGAP12, ARHGEF38, ELFN2, NECTIN1, and REPS1 as novel biomarkers, along with 12 other strong biomarker candidates. We also provided insight into potential explanations for different clinical findings and biological characteristics between LUSC and LUAD through gene expression analysis. Further validation studies of these biomarkers and biological mechanisms are therefore warranted.

Method
RNA-Seq data processing. The LUAD and LUSC HTSeq read counts data were downloaded from TCGA 13 using TCGAbiolinks from R 79,80 . As of June 2020, there were 529 LUAD and 498 LUSC samples. The samples were normalized using TMM method and standardized using the CPM (read counts per million) function in R. Genes < 1 CPM in over 600 samples were considered noise and discarded to obtain 14,010 genes. The filtered genes were analyzed with different gene selection methods to further narrow down potential gene candidates for biomarkers and pathway analyses.
Gene selection and cancer classification. Gene selection analysis was performed using five different selection methods to generate five independent sets of top genes (Fig. 1). The 5 independent sets were compared, and the resulting overlapped genes were used for cancer classification, biomarker identification, and gene expression analysis. The selection methods used were DGE, PCA, xgboost, lasso, and mRMR. DGE between LUAD and LUSC was performed using the edgeR package 81 . Though there are other options to perform differential gene expression analysis, edgeR was chosen mostly because of its speed and efficiency in analysis. Also, one of the other popular algorithm, DESeq, has also been shown to yield similar result as edgeR 16 . After using edgeR analysis and filtering for genes that have FDR < 5E−2 and log(Fold Change) > 0.5, 4702 genes were identified as differentially expressed. Top 500 of the 4702 differentially expressed genes (Table S1) were selected as top features based on their lowest p-values; validation of these genes was performed using random forest with the ranger package 82 . The top 500 genes from the first principle component in PCA and the top 500 genes ranked from mRMR 83 algorithm were selected and validated the same way as the differentially expressed genes. Genes with probability or prediction threshold over 0.5 were selected from Xgboost 84 and lasso 85 (Table S1), and validated in a similar manner as the other algorithms. For each validation, the data were randomly split into a train- www.nature.com/scientificreports/ ing set and a testing set in a 7:3 ratio, where the training set was used to construct the model while the testing set was used to evaluate the model's performance. To compare each selection method more effectively, we split the training sets and testing sets the same way for all validations. We applied fivefold cross validation to decide the optimal parameters for each training model and estimated its accuracy by applying the best determined parameters to the test set. The detailed parameters can be found in the data availability section. For classification and gene expression analysis, we selected genes that were detected by at least two methods, and they were validated using ranger 82 . We also used bootstrapping 86 with 10,000 replicates to calculate the confidence interval for the accuracy of each method, including the proposed method of classification. The genes that were detected by at least 3 methods were considered candidate biomarkers. Their diagnostic potential was determined and assessed using receiving operating characteristics (ROC) curve analysis. GSE28582 24,25 , was used as an external dataset to validate the chosen 17-gene classifier.
Prognostic value analysis using Kaplan-Meier plotter. Kaplan-Meier Plotter is an online database that contains comprehensive clinical and microarray data for various cancers, including lung cancer 26 . Prognostic values of the identified biomarkers in LUAD and LUSC were evaluated using Kaplan-Meier Plotter with each gene used as an univariate analysis. The parameters were set such that the only restricted subtypes were LUAD and LUSC, and the median was used as the cutoff. The rest of the parameters were in the default settings.
Gene expression analysis of selected genes. To further investigate and understand the biological difference between LUAD and LUSC, we performed pathway enrichment analysis using KEGG 29 , Gene Ontology (GO), and Reactome 28 . Modified Fisher's exact tests were performed using DAVID v6.8 27 . Pathways with false discovery rate (FDR) < 5% or p-value less than 0.01 were considered significant. These databases were all accessed in November 2020.

Data availability
All data generated and/or analyzed during the current study are included in this published article (and its supplementary information files). The custom code used for data analysis can be accessed at https:// github. com/ chenj oe569/ NSCLC-Resea rch.