Background Lung adenocarcinoma is the most common type of lung cancer, with high mortality worldwide. Its occurrence and development were thoroughly studied by high-throughput expression microarray, which produced abundant data on gene expression, DNA methylation, and miRNA quantification. However, the hub genes, which can be served as bio-markers for discriminating cancer and healthy individuals, are not well screened. Result Here we present a new method for extracting gene predictors, aiming to obtain the least predictors without losing the efficiency. We firstly analyzed three different expression microarrays and constructed multi-interaction network, since the individual expression dataset is not enough for describing biological behaviors dynamically and systematically. Then, we transformed the undirected interaction network to directed network by employing Granger causality test, followed by the predictors screened with the use of the stepwise character selection algorithm. Six predictors, including TOP2A, GRK5, SIRT7, MCM7, EGFR , and COL1A2 , were ultimately identified. All the predictors are the cancer-related, and the number is very small fascinating diagnosis. Finally, the validation of this approach was verified by robustness analyses applied to six independent datasets; the precision is up to 95.3% ∼ 100%. Conclusion Although there are complicated differences between cancer and normal cells in gene functions, cancer cells could be differentiated in case that a group of special genes expresses abnormally. Here we presented a new, robust, and effective method for extracting gene predictors. We identified as low as 6 genes which can be taken as predictors for diagnosing lung adenocarcinoma.

The progression of a disease associates with changes in genomic activity, but it remains a challenge to screen genetic biomarkers for clinical applications. The disease progression, in dynamic network methods (DNM), can be analogous to an animated film composed of discrete frames, where each frame represents a temporary state of the time-varying gene-gene interaction network. The major shortage therein is that the transition between two neighboring temporary states was beyond investigation. Here, we develop an updated computational methodology named after VD-analysis. Because single-gene biomarkers were not approved capable of representing a complex biological process, we firstly introduce V-structurea gene module composed of three genes and two interactions among themand define it as unit module. We then identify the perturbed pathways that mark the disease progression, followed with the V-structures identified which drive the pathway perturbations. Such driver V-structures can be taken as eligible biomarkers for clinical applications. To test the feasibility of this method, we apply it to a time course dataset of gene expression related to mouse type-II diabetes (T2D). Result indicates that the whole process of T2D is exactly divided into 3 stages and that the driver V-structures inferred for each stage are qualified biomarkers. In summary, our method contributes to the description of dynamic disease progression and the V-structure biomarkers facilitate the treatments of disease.

No abstract