Ensemble Feature Selection and Meta-Analysis of Cancer miRNA Biomarkers

Lopez-Rincon Alejandro; Martinez-Archundia Marlet; Martinez-Ruiz Gustavo Ulises; Tonda Alberto

1 Several studies have shown the properties of microRNA types (miRNAs) as oncogenes 2 and tumor suppressors [1] [2] [3] . Since then, many sophisticated techniques, such as 3 high-throughput technologies, microarray, mass spectrometry and especially the Next 4 Generation Sequencing (NGS), have been developed for their identification [4] . However, 5 it is clear that the development of computational tools is needed for the interpretation 6 of results from these high-throughput experiments [5] . Indeed, computational assisted 7 methods are used for the identification of miRNAs from different genome organisms, for 8 example in Caenorhabditis briggsae [6] and in Epstein-Barr virus (EBV or HHV4), a 9 member of the human herpesvirus (HHV) [7] . Furthermore, several computational 10 techniques can be applied to accurately predict miRNA expressions, as seen for example 11 in [8] . 12 Succeeding the earliest evidence of miRNA involvement in human cancer by Croce 13 and collaborators [9] , various studies demonstrate that miRNA expression is deregulated 14 in human cancer through diverse mechanisms [10] . Additionally, in comparison to the 15 impractical and invasive methods currently used for cancer diagnosis [11, 12] , miRNA 16 biomarkers can be detected directly from biological fluids (such as blood, urine, saliva 17 and pleural fluid [13] ), and they can also be used as biomarkers to detect tumors at an 18 early stage, which is extremely important for survival. For example, the 5-year survival 19 rate for lung cancer is 5%, but an early diagnosis can boost it to almost 50% [14] . Thus, 20 miRNA expression profiles correlate with clinical variables, highlighting their potential 21 value as prognostic and/or diagnostic tools. 22 In such a context of increasing availability of data, it is of utmost practical 23 importance to build databases of miRNA expressions data for cancer research [15] [16] [17] [18] [19] , 24 and also to extract features that could be used as cancer biomarkers [20] [21] [22] . For 25 example, miRNA hsa-mir-21 is mentioned as a marker for patients with squamous cell 26 lung carcinoma [23], with astrocytoma [24], breast cancer [25] , and gastric cancer [26] . 27 Following this idea, the scientific community is currently looking for miRNA signatures, 28 representing the minimal number of miRNAs to be measured for discriminating between 29 different stages and types of cancer. 30 Current NGS technologies such as Applied Biosystems, SOLiD3,or HiSeq from 31 Illumina are able to extract thousands of components in genome sequences [27] , and 32 traditional linear statistical analysis are not suited to manage such quantities of 33 most relevant miRNAs to use as biomarkers for cancer classification. Typically, 39 classifiers trained on a dataset will not use the whole set of available features to 40 separate classes, but just a subset which could be ordered by relative importance, with 41 a different meaning given to the list by the specific technique. The top 100 biomarkers 42 in the list are then evaluated as a potential reduced signature for classification. Finally, 43 the top 50 miRNAs are compared to a meta-analysis of the medical literature, to 44 validate the results automatically produced by the machine learning algorithms. 45 Unsurprisingly, most of the miRNAs identified by the classifiers are also considered 46 important by the specialized literature: 15 of them, however, are still understudied, and 47 they could thus repr