1. Introduction
Lung cancer is the leading cause of cancer-related death worldwide, responsible for around 1.59 million deaths in 2012 [1]. This disease is characterized by high heterogeneity in terms of pathological features and is commonly classified into two major groups, small-cell lung carcinoma (SCLC), accounting for approximately 15% of lung cancer cases, and non-small cell lung carcinoma (NSCLC). NSCLC includes mostly squamous cell carcinoma (SqLC), adenocarcinoma (AdCa), and large cell carcinoma (LCLC) [2,3]. Most lung cancers are detected at advanced stages, reducing the likelihood of cure. Five-year survival rates for all stages are about 15% [4]. One reason for late lung cancer detection is that usually lung cancer symptoms, such as cough, chest pain, and weight loss, do not appear until the disease is already in an advanced stage. Moreover, the lung is a visceral organ with a complex branching structure, which makes an entire examination impossible. Additionally, traditional diagnostic methods like computer tomography (CT) or X-ray often yield low positive predictive values (PPV), indicating that only few patients with a positive test result were confirmed with lung cancer after additional biopsy. This circumstance results in unsatisfactory performance of many diagnostic tests, high downstream costs and unpleasant procedures for the patients [5]. Therefore, it is important to find novel markers for early detection of lung cancer by means of minimal invasive biomarker screenings using material such as blood, sputum, or exhaled breath [6]. 
Today, it is known that proteins, expressed by tumors, often evoke a humoral immune response in cancer patients [7]. This response occurs because these tumor-associated antigens (TAAs) are mostly altered in some way. They may be unique to cancer and germ cells, found only in specific tumors, or may be mutated, misfolded, overexpressed, aberrantly degraded, or aberrantly glycosylated [8,9,10]. The response of the human immune system to such TAAs becomes apparent with the production of autoantibodies targeting these antigens [11]. This suggests that the immune system is able to detect aberrant structure, distribution, and function of certain proteins involved in tumorigenesis [12]. The finding that cancer patients produce autoantibodies against tumor antigens encourages the idea that these autoantibodies could facilitate cancer diagnosis and prognosis [10,11,13,14]. Moreover, the relevance of autoantibody-based biomarkers is emphasized by the fact that autoantibodies against TAAs can be found up to five years before tumors become symptomatic or clinically detectable with conventional methods [7,15,16,17,18,19]. 
Autoantibodies have been detected in several human cancers, but only a few antigen-autoantibody reactivities have made their way into clinical practice. One reason may be that most autoantibodies occur only in a subset of cancer patients, resulting in low sensitivity of single antigens as tumor biomarkers [11,20]. This lack of sensitivity indicates that a single tumor autoantibody will not be sufficient as biomarker [21]. Therefore, biomarker panels, including cancer-specific autoantibodies are considered as promising tools for early diagnosis and prognosis of cancer. For this purpose, autoantibody signatures, the molecular finger print of antibodies in a certain disease, have to be identified [22]. 
For identification of potential TAAs, large screening sets are required [23]. A highly multiplexed approach, like protein microarrays, offers a suitable platform for autoantibody marker discovery while using only small amounts of serum or plasma samples [24,25]. This technology enables screening of humoral response against thousands of potential TAAs in order to identify novel biomarker panels for cancer diagnosis [23]. For this purpose, antigens, which are either derived from recombinant protein expression or are isolated from tumor cell lines, are immobilized and their reactivity with sera from patients is investigated [26].
High-dimensional microarray data require careful statistical handling. In general, the first step is log2 transformation of the data in order to achieve symmetric ratios of fold-changes, independent from increase or decrease in intensity values [27]. The second step requires adjustment of the data for systematic biases. Many microarray studies are affected by batch effects, which can often not be avoided, because not all samples can be processed in one single batch [28]. This introduces non-biological differences, which impede comparability of the processed samples. Combination of microarray data derived from different experimental batches enables increasing statistical power of microarray studies. Therefore, adjustment for batch effects has to be performed [28,29,30,31]. One method is quantile normalization, which aims to adjust the distribution of probe intensities for each array in the data set to the same level [30]. To minimize batch effects in data sets derived from multiple experiments, algorithms, such as the empirical Bayes method referred to as “combating batch effects when combining batches of gene expression microarray data” (ComBat) [31] or “distance-weighted discrimination” (DWD) [32], which have been proposed to be effective at adjustment for systematic biases, can be applied. Batch effects as well as batch effect removal methods are often investigated visually by means of principal component analysis (PCA) plots [33]. A novel method combining PCA and variance component analysis (VCA) is able to evaluate the contribution of certain sources to the variance in the microarray data set. This approach is referred to as principal variance components analysis (PVCA) [28,34]. The third step after data pre-processing is comprehensive analysis of the microarray data set. This can be done by means of the software Biometric Research Branch BRB-ArrayTools, developed by Richard Simon. This software includes a broad variety of methods for predictive classifier development. Supervised machine learning methods, such as class comparison and class prediction, are available with complete cross-validation [35]. The aim of class comparison analysis of DNA microarrays is the univariate identification of differentially expressed genes in two different specimen phenotypes [36]. The same statistical analysis methods which are usually used for gene expression studies were applied to our protein microarray data. When using protein microarrays for the identification of tumor autoantibody signatures, class comparison analysis enables determination of differentially reactive antigens with autoantibodies of cancer patients compared to controls [37]. Another method included in this software is class prediction. This method aims at developing a multivariate statistical model to predict the class of a sample based on its signal pattern on the microarray. For the development of such models, features that enable discriminating the predefined classes have to be identified and accuracy of the built predictor is estimated [38]. For this purpose, it is advisable to use complete cross-validation which includes feature selection and model development. By doing so, overfitting of the model to the used data set is avoided in contrast to incomplete cross-validation, which does not include feature selection in the process of cross-validation, or no cross-validation. Another advantage of cross-validation, especially leave-one-out cross validation (LOOCV), is that small data sets are used efficiently. Each sample is withheld once during feature selection and model development and is then classified based on the developed model. This process is iterated for each sample [39]. Analysis of microarray experiments enables the detection of features which are correlated to a phenotype or to find a predictor or classifier to predict the phenotype of a new sample [40]. These predictors could be applied in clinical diagnostic testing, for assessment of prognosis, or treatment decision [39].
In the present study, we used a high-density protein microarray for the detection of autoantibody signatures in lung cancer. This protein microarray was processed with purified immunoglobulin G (IgG) from 100 lung cancer cases and 100 matched lung cancer-free controls. The lung cancer group comprised 25 samples from each of the four main histological lung cancer types. Since the sample cohort of this study included 200 different samples, not all samples could be processed at once. Therefore, samples belonging to one histological lung cancer entity and their matched controls (results in a total of n = 50) had to be processed in two different runs. As already mentioned, adjustment for batch effects has to be done when combining data from different microarray experiments [41]. For this purpose, different normalization and batch effect adjustment methods were used, including quantile normalization, DWD, and ComBat. Currently a great number of normalization algorithms and data transformation methods are available and there are several publications comparing the performance of some of these [28,30,41,42]. However, it is still difficult to decide which method performs best for a certain data set. Moreover, it should be kept in mind that analysis results are strongly influenced by the applied normalization method [27,42]. 
We aimed at comparing different normalization and data adjustment methods in terms of effectiveness and their influence on subsequent data analysis. For this purpose, principal component analysis (PCA) was used to investigate the variance in the data set before and after applying different data pre-processing methods. Furthermore, overlaps of significant antigens derived from class comparison were compared between different sample cohorts with all normalization methods. Since the study was focusing on the identification of a tumor-autoantibody-based biomarker panel for early diagnosis of lung cancer, the development of multivariate classifiers for each histological type of lung cancer and for the combination of all lung cancer types was done by means of class prediction. This was performed with quantile normalization and ComBat adjustment. Classifier panels identified during this study are already patented [43].