Non-small cell lung cancer (NSCLC) is one leading cause of cancer deaths in many countries (1). It can be categorized into three major histological subtypes among which adenocarcinoma (AC) and squamous cell carcinoma (SCC) account roughly for 40 and 30% of the lung cancer (LC) cases, respectively (2). Increasing evidence supports that AC and SCC differ in the composition of genes and molecular characteristics. For instance, Hou et al (3) found that in contrast to the AC-associated genes are highly enriched to tight junction and cell adhesion molecules, the SCC-associated genes are more correlated to cell communication. Therefore, they are currently regarded as two distinct diseases.

Currently, treatment choices for the NSCLC patients mainly depend on the stage at which cancer was diagnosed regardless of the histological subtype. For example, patients at the stage IA usually undergo surgical resection and rarely prescribe to adjuvant chemotherapy. But the recurrence rates of patients at the same stage of cancer are heterogeneous, making such homogeneous treatment choices implausible. It is becoming critical to evaluate the risk profiles of patients using a reliable molecular/gene signature. Nevertheless, due to the fundamental differences between AC and SCC of NSCLC patients, it is hypothesized that specific genes are related to recurrence/survival rates for each histology subtype (4–6).

To deal with the issue of high dimensionality commonly existing in gene expression profiles, downsizing from thousands of genes to a minimal gene signature with maximal predictive ability is of the essence. In statistics, this process is referred to as feature selection (7). Efforts have been devoted to distinguish AC from SCC using gene expression profiles and various feature selection algorithms (8–12), and more recently to identify prognostic markers for each specific subtype (4–6).

Genes are highly correlated and can be grouped into many gene sets correspondingly. Depending on if these group structures are taken into account, a feature selection algorithm may be classified into either a pathway-based or a gene-based method. Studies have demonstrated that compared to its gene-based counterpart, a pathway-based feature selection algorithm in which pathway information is utilized to assist the selection process has a better predictive performance, stability or biological interpretation (13–18). Specifically for the NSCLC applications, several pathway-based feature selection algorithms have been applied to distinguish its major subtypes and/or histological stages (8,11,14).

As a data visualization method, the Radial Coordinate Visualization (RadViz) method (19) can display more than two variables in a 2-dimensional projection. It can also be used to search for biologically interesting patterns and select relevant genes highly associated with the phenotype of interest (9,20). In a Radviz projection, features such as genes are presented as anchor points spaced around the perimeter of a circle while samples are as points inside the circle. Each point (i.e., a sample) is held in place by springs that are attached at the other end to the feature anchors (i.e., genes). The stiffness of each spring is proportional to the sample's corresponding gene expression value and the point ends up at the position where the spring forces for these anchors are in equilibrium. When used for the purpose of feature selection, RadViz may be roughly regarded as a gene-based method since it does not account for any pathway information.

In this article, we first ordered genes using a novel ranking method in bioinformatics-the GeneRank method (21) which ranks genes according to not only its expression level but also its connectivity with other genes in the gene-to-gene interaction network, and then we restricted the genes under consideration to those ranked on the top by the GeneRank method and used RadViz to select relevant genes in the restricted search space. The proposed procedure is a combination of the pre-filtering and RadViz, in which the connectivity information is also incorporated. We applied the proposed procedure to a set of NSCLC data to establish diagnostic gene signatures for the classification between AC and SCC and prognostic signatures for the survival prediction of NSCLC patients.

We applied the proposed procedure to the NSCLC application and obtained two sets of gene signatures-one for the AC/SCC segmentation and the other for high/low risk segmentation, being herein referred to as the diagnostic signature and the prognostic signature, respectively. We consider four scenarios: the genes under consideration were restricted to the first 200, or 500, or 1,000 highest-ranked genes and then were all 1,952 genes in the last scenario. The corresponding RadViz projections with optimal gene subsets are presented in Figs. 2 and 3, from which we observed that the final diagnostic and prognostic gene signatures were barely overlapped in all scenarios. As discussed previously (5), it is unsurprising to observe no or only limited overlaps between the diagnostic signatures and the prognostic signatures since the outcomes under investigation for these two sets of signatures differ in nature.

The performance statistics of the resulting diagnostic signatures for both the training set and the test set are presented in Table I. Similarly, the performance statistics of the resulting prognostic signatures for both the training set and the test set are presented in Table II. There are no significant differences in terms of predictive performance for either diagnostic signatures or prognostic signatures under these four scenarios, indicating a prescreening step to downsize the genes under consideration to those that are important in terms of both pathway connectivity and expression differences shall not deteriorate the predictive performance of resulting final signatures. Even though no huge differences among those signatures exist, the signatures constructed with the first 1,000 genes outperform slightly to the signatures under the other scenarios, suggesting 1,000 is the optimal cutoff for the number of genes under consideration in this study. The heatmaps of the 8-gene diagnostic signature and the 8-gene prognostic signature under the first 1,000-gene scenario are shown in Fig. 4. In consistent to the previous observations, there existed a clear separation between AC and SCC samples but not so between the high-risk and the low-risk patients. Instead, using hierarchical clustering (as shown in Fig. 4), the samples may be classified into three clusters-the high-risk patients, the low-risk patients, and those with ambiguous labels.

Furthermore, since it is demonstrated that several genes are adequate to discriminate AC and SCC apart (9,11), we set the maximum number of genes in the RadViz projections as 3 and redid the selection of relevant genes and the final model fitting. In contrast, previous studies (5,27) have shown that compared to the diagnostic gene signatures, the identification of prognostic gene signatures is much difficult and thus less than 10 genes might be incapable of separating patients with good prognosis from those with bad prognosis. As a fix to this, we resort to the strategy of using genes with the highest frequencies in the RadViz projections (9,20). Here the final size is set at 40. The performance statistics for the 40-gene prognostic signature are tabulated in Table II as well.

The gene symbols of these 3-gene diagnostic signatures are presented in Fig. 5. We found the 3-gene signatures are very stable. While there is one gene (33.3%) existing in all these signatures, 3 of these four signatures (75%) share 2 common genes (66.7%), providing further evidence to support that several gene biomarkers are sufficient to distinguish AC and SCC. For the prognosis analysis, when we increased the size of final models to 40, a better separation between patients with good prognosis and those with bad prognosis has been achieved compared to the 8-gene signatures. But the performance of the 40-gene signatures is still below satisfactory, which may be explained by the following reasons.

First, the patients were stratified into two categories as the high-risk one and the low-risk one on the basis of their survival time. The risk status served as the outcome when training the prognostic signatures. Such an over-simplified stratification might lead to the predictive inferiority of a prognostic signature, as pointed out by (28). Considering the RadViz method is incapable of dealing with the time-to-event outcomes, we will definitely replace it with a more novel feature selection algorithm e.g., LASSO and reanalyze this NSCLC dataset in our future research.

Second, in this study we constructed the overall prognostic signature for NSCLC patients without considering their histological subtypes. As mentioned in the Introduction section, there may exist subtype-specific prognostic genes for AC and SCC. Since one major goal of this study is to illustrate the point that the diagnostic and prognostic gene signatures differ dramatically, a homogenous prognostic signature for both AC and SCC is required. Construction of subtype-specific prognostic signatures using either separate survival analysis for each specific subtype or a suitable statistical method such as (4,5) is warranted, in order to make better prediction and thus to facilitate personalized treatment strategies for NSCLC patients.

Lastly, the gene expression profile alone might not convey all information about the prognosis of NSCLC patients. If this is true, other omics data such as copy number alternation and DNA methylation data need to be integrated in order to provide a better survival prediction for the NSCLC early-stage patients.

In the present study, we trained on the same data to construct the diagnostic and prognostic gene signatures with the aids of RadViz and SVM. The gene expression profiles may contain valuable information on AC/SCC segmentation, and also valuable information on prognosis. Nevertheless, those informative genes for diagnosis might not be valuable for prognosis, and vice verse. It is unsurprising that the diagnostic signatures and the prognostic signatures share no or limited overlaps, even they are all trained on the same dataset.

With regard to that no significant prognostic gene signatures have been achieved in this study, in the Results section, we listed three reasons to explain why this happened. Given the fact we obtained substantially better C-statistics using the same datasets and Cox-models (unpublished work), we believe the stratification of patients into different risk categories on the basis of their survival time may result in huge information loss. Thus, it is emphasized that such an over-simplification shall be avoided in practice.

Depending on if the membership/label information is taken into account, a machine learning method is classified into either an unsupervised method or a supervised method. Without considering the labels/dependent variables, the information captured by an unsupervised learning method might not be meaningful for both diagnosis and prognosis, let alone there are so many irrelevant and redundant genes in gene expression profiles to blur the signals from those relevant ones, thus the process of variable selection becomes imperative where the outcome/label is always taken into consideration. Therefore, we prefer to a supervised method over an unsupervised learning method.