On Predicting lung cancer subtypes using ‘omic’ data from tumor and tumor-adjacent histologically-normal tissue

To test our hypothesis, we extracted datasets containing gene expression and DNA methylation beta values from the Cancer Genome Atlas (TCGA) data portal for lung adenocarcinoma (LUAD [24]) and lung squamous cell carcinoma (LUSC, [25]). Additionally, we also used the gene expression dataset of lung adenocarcinoma patients, described by Landi et al. [26], GEO accession number GDS3257. Table 1 describes the characteristics of the samples we used for this study. For each dataset, it provides information on the type of ‘omic’ data type, source of data, assay platform, including number of features (i.e. genes or DNA methylation sites), and the number of sample distribution – that is, tumor tissue (T and TAHN) – within each subtype, where available. The formatted TCGA dataset used in this study, along with sample IDs, are provided in Additional file 1 (TAHN_ADC vs. Tumor_ADC in gene expression), Additional file 2 (TAHN_SCC vs. Tumor_SCC in gene expression), and Additional file 3 (TAHN_ADC vs. Tumor_ADC in methylation). The annotations from TCGA to identify these samples are provided in Additional file 4 (Appendix A).

We followed a supervised classification process on 10-fold cross-validation. That is, for each fold we partitioned the dataset into training and test, where the former contains 90 % of the samples, while the latter contains the remaining 10 %. We ensured that each partition maintains the same class distribution as the whole dataset (stratified). In each fold, we analyzed the datasets using the experimental design as illustrated in Fig. 1. According to the design, there are four main components, namely, a) Feature Selection, b) Discretization, c) Model Building and d) Evaluation. We additionally perform Gene Functional Analysis, and apply Clustering methods to better understand the characteristics of the features chosen by this framework. Below, we explain each component in detail.

High-throughput platforms, such as gene expression and methylation microarrays, generate high-dimensional data that is typically very complex for analysis. Feature selection is a machine learning pre-processing step that tries to find a subset of the original variables (also called features or attributes) that are highly associated with the target class variable (i.e. phenotype, like a disease state). We used the ReliefF algorithm [27] to rank all variables and select the top scoring ones. ReliefF is a multivariate filter algorithm that estimates how well a given variable can distinguish the target class given the instances that are near to each other. The initial number of variables (17,814 in gene expression, and 27,578 in methylation) is reduced to the top 30 scoring variables. In previous studies [28], it has been reported that 30 is a sufficient number of genes to create computational classification models. With this number of genes, the classification models created would have a good trade-off between relevance and complexity of the model.

Similarly, we also selected the differentially expressed (DE) genes and differentially methylated (DM) probe sites from each dataset using Limma, which is an R-language package for the analysis of microarray data [29]. Limma uses a t-statistic to rank genes in order of evidence for differential expression. It first fits linear models for each gene (lmFit), and then it uses empirical Bayes (eBayes) moderation to adjust the standard error of the models by borrowing information from the rest of the genes (average variance across all genes). This method is very effective in finding differentially expressed (DE) genes in microarray data, however with methylation datasets it has not been equally successful [30]. The output of finding the DE genes and DM probe sites with Limma can be seen as a feature selection method (or ranked list). Similarly to the ReliefF selection, we selected the top 30 most DE genes and DM probe sites (based on log₂-fold change) to build a classifier for comparison with ReliefF. The output of the resulting classifiers was evaluated using the area under the receiver operating characteristic curve (AUC) performance metric in the test datasets.

Most ‘omic’ data such as gene expression and methylation are represented with continuous values. However, many machine learning algorithms are designed to only handle discrete (categorical) data, using nominal variables, while real-world applications, like ‘omic’ data analysis, typically involves continuous-valued variables. Discretization, the process of transforming continuous values into discrete ones, has been shown to improve the performance of machine learning classifiers [31]. To discretize the variables, we used the Fayyad and Irani’s minimum description length principle cut (MDLPC) [32]. This method, which is widely used in the machine learning community, applies a supervised greedy search strategy to recursively find the minimal number of cut-points in each variable that minimizes the entropy of the resulting subintervals.

For continuous methylation values ranging from 0 to 1, three possible strategies for discretization can occur. The first strategy is when a fixed cut-point is determined arbitrarily for all variables (for example, choosing > 0.5 methylated, while ≤ 0.5 could refer to unmethylated). The second strategy, when an expert-based discretization is made for all variables (i.e. unmethylated < 0.1, partially methylated between 0.1 and 0.8, and methylated > 0.8 [33]). The third strategy is when a supervised discretization method creates independent cut-points for each variable. For the first and second strategies, the same discretization scheme (i.e. same number of intervals or cut-points) is used for all variables. However, this approach is suboptimal for a classification task. For instance, when using MDLPC we observed that the methylation site cg19782598 was discretized into two categories: methylated (>0.86) and unmethylated (≤0.86); while methylation site cg11693019 was discretized into three categories: methylated (>0.76), partially methylated (between 0.76 and 0.47), and unmethylated (<0.47). Thus, supervised discretization could help identify appropriate cut-points for each variable, as opposed to the others, which naïvely assume the same cut-points for variables.

In computational genomics, heatmaps are used to graphically show the level of expression that a selected group of genes have in a cohort of patient samples. A heatmap can also be built with methylation intensity values. We build heatmaps from the genes selected by Limma and ReliefF to further validate the results obtained with these feature selection methods. The clusters are a visual representation of the class discrimination ability of the genes selected.

The order in which genes (rows) and samples (columns) are ordered in the heatmap matrix is often based on an agglomerative hierarchical clustering. We used the Minkowski measure to calculate the pairwise distances between elements, and then aggregated the closest elements in clusters using the Ward linkage calculation of distances between clusters. This combination of Minkowski distance and Ward linkage has been shown to perform well in biomedical and synthetic datasets [34].

We also performed Gene Functional Analysis using QIAGEN’s Ingenuity® Pathway Analysis tool (IPA®, QIAGEN Redwood City, www.​ingenuity.​com) to gain insight into the biological role of the genes selected by our framework. First, all gene symbols selected were used as input for the IPA platform, which will search for correlations between these genes and functions or pathways in their curated literature. A p-value is computed using Fisher’s right-tailed exact test for the gene list to a function/pathway it may be associated with. The p-values indicate the likelihood of association between the gene set (as selected by ReliefF) and a specific function (set of genes associated with a function) to have occurred due to random chance alone. A p-value of less than 0.05 is considered to be significantly better than random chance. Methylation probe sites were mapped into their corresponding gene symbols that they methylate.

In the machine learning literature, a classifier is a computational model that can differentiate between two (or more) states of disease. Bayesian networks [35] are particularly useful classifiers that are very popular in the classification of biomedical data. A Bayesian network (BN) is a probabilistic graphical representation of random variables (nodes) and probabilistic dependencies among them (arcs). Once a Bayesian network is learned, the structure and conditional probability tables can be used to calculate the posterior probabilities for a new case to be a member of a given class, i.e. the probabilities of a new case being ADC given the BN and the data. P(ADC = True|BN, data). A special case of BN is the naïve Bayesian classifier (NB), which assumes a strong conditional independence among the variables. In a NB structure, the target node (i.e. class variable) is the parent for all other features, and there are no arcs among those children nodes. The child nodes are independent given the parent, which facilitates the calculation of posterior probabilities by substituting the joint probability with the product of their probabilities. NBs have been shown to predict poorly in high-dimensional genomic datasets [36], but it is expected that the use of a feature selection method (ReliefF or Limma) will improve the NB classification performance. Moreover, its simplicity makes it a powerful tool to be considered in a biomedical classification framework, while giving us insights into the baseline performance on a given dataset.

We evaluated the NB classifiers using the area under the receiver operating characteristic (AUC), which is a measurement of the area created by plotting the performance of a classifier for the true positive rate versus the false positive rate. When presented with a test dataset, the Bayesian network calculates a posterior probability for every case, and a threshold is used to assign the class for the new cases. The curve is constructed by varying the threshold to which the probability is considered for class determination. Also, the 95 % confidence interval (C.I.) of the AUC was calculated using DeLong’s method for variance estimation [37].

AUC (equivalent to c-statistic) is a useful measurement of the ability of models to discriminate between two (or more) classes [38]. Calibration deals with agreement between observed outcomes and predictions. For this purpose, we used the Brier Skill Score (BSS) [39] creates an index between −1 and 1 that provides information as of how far away the results of any classifier are in relation to the unskilled classifier. The unskilled classifier is one that only considers the distribution of data. A classifier with a positive BSS would therefore be skilled and unbiased.

The classification performance for all models is high (\( \mathrm{A}\mathrm{U}\mathrm{C}\ge 0.81 \)), with positive calibration (\( \mathrm{B}\mathrm{S}\mathrm{S}>0 \)). This positive calibration is a good indication that the models will perform well for other cases, and that they were not biased by the distribution of the data.

In the classification task of TAHN_ADC vs. Tumor_ADC, the naïve Bayesian model created obtained high predictive performances (\( \mathrm{A}\mathrm{U}\mathrm{C}\ge 0.99\;\mathrm{with}\;\mathrm{ReliefF},\;\mathrm{and}\ge 0.81\;\mathrm{with}\;\mathrm{Limma} \)). The classification task TAHN_SCC vs. Tumor_SCC also obtained high predictive performances (\( \ge 0.99\;\mathrm{with}\;\mathrm{both}\;\mathrm{feature}\;\mathrm{selection}\;\mathrm{methods} \)). The molecular differences between TAHN and tumor tissue show distinctive signatures regardless of ‘omic’ dataset, feature selection method or lung cancer subtype. The results for these classification tasks were as expected since the tissue architecture between TAHN and Tumor is recognizable under a microscope if enough tissue samples are provided. They also could be achieved with the relatively small number of normal tissues available for analysis, since these normal tissues are very homogenous in expression and methylation features.

In the classification task of Tumor_ADC vs. Tumor_SCC the predictive performance was very high (\( \mathrm{A}\mathrm{U}\mathrm{C}\ge 0.89,\;\mathrm{f}\mathrm{o}\mathrm{r}\;\mathrm{gene}\;\mathrm{expression},\;\mathrm{and}\ge 0.89\;\mathrm{with}\;\mathrm{methylation} \)). Previous studies for the same classification task also show a similar classification performance. For example, Ben-Hamo et al. [40] correctly classified 85 %, using linear models. Meanwhile, Cai et al. [10] obtained an accuracy of 86 % using ensemble methods; Li et al. [41] achieved an AUC of 0.98 using Support Vector Machines; and Zhang et al. [42] achieved AUCs of 0.89 using naïve Bayesian models. Similarly, the study by Chang and Ramoni [22] achieved an accuracy of 0.95, using naïve Bayesian models. It is worth noting that none of these studies used methylation datasets and they fail to clearly recognize the importance of TAHN tissue for classification.

The classification task of TAHN_ADC vs. TAHN_SCC also had very high evaluation performances (\( \mathrm{A}\mathrm{U}\mathrm{C}=1 \)). This high performance means that all samples were correctly classified. We hypothesize that an explanation of this excellent result can be attributed to the distinctive epigenetic differences between lung tissues. We did not evaluate the gene expression in this classification task due to the lack of an available dataset. To the best of our knowledge reporting of TAHN tissue in public repositories is still an open challenge that should be addressed to improve experimental designs of other studies. A study by Haaland et al. [43], showed that there are differentially expressed genes between TAHN tissues in prostate cancer. In our study, we investigate DNA methylation data to indicate that the same differences could also be found in lung cancer TAHN tissues, and we hypothesize that the use of TAHN tissues might also help in the classification performance of other cancer types.

The classification task of TAHN-Tumor_ADC vs. TAHN-Tumor_SCC is a novel approach, where a mix of tissue types are used to classify between lung cancer subtypes. The noise introduced by mixing tissue types is overcome by our experimental design, which was able to obtain a very good classification performance (\( \mathrm{A}\mathrm{U}\mathrm{C}\ge 0.92 \)). Despite, the ‘noisy’ tissue samples, a simple naïve Bayesian classifier can accurately classify between lung cancer subtypes. This classification performance is confirmed by the heatmap analysis in Fig. 2c, where the tumor tissue of ADC creates a distinct cluster, while the remaining samples cluster together in three distinct subclusters. Furthermore, our Gene Functional Analysis using IPA® shows strong associations to cancer pathways, with 19 genes found to be associated with adenocarcinoma, squamous-cell carcinoma and carcinoma of the lung. Out of these 19 genes we found 4 genes associated specifically with lung cancer subtypes: AKR1B10, AQP10, CXCR2, TP73.

Lung cancer patients could benefit with a potentially novel approach for subtyping. The diagnosis of adenocarcinoma vs. squamous cell carcinoma is routinely accomplished using histology supplemented by immunohistochemistry (TTF-1 and p63/p40). It is therefore not likely that our approach would change this practice, which is well established, quick and inexpensive. Rather, we suggest that the use of epigenomic changes could help in the small number of tumors which remain difficult to classify. However, the primary importance of our work may be in providing additional understanding of the origins of squamous cell and adenocarcinomas, which suggest that these phenotypes are associated with, or perhaps even derived from, different epigenomic phenotypes. Epigenomic alterations, in the form of DNA methylation, prevent the binding of transcription machinery, resulting in gene silencing [44]. Moreover, DNA methylation signatures are different between tissue types and between tumors and normal surrounding tissue [20]. In our study, tumor-adjacent histologically normal tissue samples were used to classify lung cancer subtypes with excellent results. This classification performance was achieved when no tumor samples were involved (TAHN_ADC vs. TAHN_SCC), and when a mix of tissue was used (TAHN-Tumor_ADC vs. TAHN-Tumor_SCC). The high AUC results are an indication of the diagnostic potential of this technology.

Our study had some limitations, which include the following: 1) There were a limited number of tumor-adjacent histologically normal tissue samples used. However, the homogeneity of these normal tissues we observed suggests that additional normal tissues would not improve the classifier. 2) The resulting classifiers were not validated in another dataset outside of TCGA lung samples. 3) Each ‘omic’ classifier is independent of one another. In the future, we would like to explore data integration models in a multi-omic approach. 4) The classification problem of discriminating cancer subtypes of adenocarcinoma and squamous cell carcinoma could also be explored in a pan-cancer analysis, to validate the same finding seen in our study of lung cancer subtypes. 5) Due to the challenge of data availability, in this study we did not analyze biopsies with varying percentages of tumor and TAHN tissue (mixed biopsies). Instead, we took relatively ‘pure’ biopsies of either tumor or TAHN to classify between lung cancer subtypes. A future study could consider the molecular classification or discovery of cancer given a mixture of tumor and TAHN tissue. For example, an analysis of ‘omic’ data from cancerous and non-cancerous tumor tissues, as well as TAHN tissue for both types of tumors, might be performed in the same way as presented in this manuscript.