Ensemble Feature Selection and Meta-Analysis of Cancer miRNA Biomarkers

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

The most important result of this work is that several miRNAs types highly ranked 151 by the classifiers appear to be still understudied in literature. Table 3 reports a 152 summary of the literature meta-analysis, from which it clearly emerges that 15 of the 153 miRNA types appear in less than 50 references connected to cancer from the PubMed 154 query: strikingly, hsa-135-a-1 appears only once, and hsa-103-1 only twice, while both 155 are considered important by more than 60% of the classifiers. These two miRNAs could 156 thus represent promising leads for future research. Interestingly, only 25 of the miRNA 157 types found by the proposed machine learning approach also appear among the most 158 expressed in the baseline analysis summarized in Figure 1 , and only 4 (hsa-mir-9-1, 159 hsa-mir-28, hsa-mir-103-1, and hsa-mir-199b) are both overexpressed and understudied, 160 suggesting that simple statistical approaches might not be enough to extract meaningful 161 information from such complex data. hsa-mir-10b, the feature with the highest S f score, while not appearing in the surveys, 165 is currently under clinical trials as oncomarker for Glioma [45] , and it is mentioned in a 166 recent article as possible oncomarker [46] . hsa-mir-135a-1 and hsa-mir-135a-2, located 167 inside chromosomes 3 and 12, respectively, generate the same mature active 168 sequence [47] . hsa-mir-9-1, hsa-mir9-2, and hsa-mir9-3, generate the oncogenic 169 hsa-mir-9 [48] . hsa-mir-944 at the moment is known to decrease malignant features in 170 gastric [49] , colorectal [50] and endometrial [51] cancers. 171 We envision that understudied miRNAs could play a key role in the biology of 172 cancer. On the other hand, the presence of single-nucleotide polymorphism (SNP) could 173 affect miRNA biology due to alterations in the miRNA maturation process, their structure, and expression levels. From these, creation and loss of miRNA targeted sites 175 by SNPs is the most inspected area [52, 53] . However, the miRNA-mediated oncogenic 176 transcriptional landscape could be a consequence of the SNPs presence in the seed 177 region of mature miRNAs [54] , that is involved in the molecular recognition with its 178 targeted mRNAs. Although the presence of SNPs in seed regions seems to be negatively 179 selected [55] , in a pathological condition as cancer, it would be interesting to perform 199 Furthermore, personal-directed anti-tumoral therapy could be achieved by measurement 200 of a specific, minimal miRNA signature, as proposed in this work. List of the top 100 most relevant features identified by the proposed methodology, in 209 order of importance: hsa-mir-10b, hsa-mir-126, hsa-mir-10a, hsa-mir-9-2, hsa-mir-30a, 210 hsa-mir-9-1, hsa-mir-375, hsa-mir-21, hsa-mir-125a, hsa-mir-143, hsa-mir-122, hsa-let-7i, 211 hsa-mir-200c, hsa-mir-196b, hsa-mir-22, hsa-mir-145, hsa-mir-205, hsa-mir-30d, 212 hsa-mir-210, hsa-mir-148a, hsa-mir-193a, hsa-mir-190b, hsa-mir-9-3, hsa-let-7c, 213 hsa-mir-1247, hsa-mir-490, hsa-mir-141, hsa-mir-19a, hsa-mir-503, hsa-mir-135a-1, 214 hsa-mir-944, hsa-mir-203, hsa-let-7b, hsa-mir-103-1, hsa-mir-584, hsa-mir-152, 215 hsa-mir-30e, hsa-mir-106a, hsa-mir-183, hsa-let-7f-1, hsa-mir-202, hsa-mir-199b, 216 hsa-mir-200b, hsa-mir-194-2, hsa-mir-29c, hsa-mir-30b, hsa-mir-101-2, hsa-mir-192, 217 hsa-mir-135a-2, hsa-mir-28, hsa-mir-211, hsa-mir-200a, hsa-mir-598, hsa-mir-1-2, 218 hsa-mir-95, hsa-mir-99a, hsa-mir-378, hsa-mir-194-1, hsa-mir-199a-1, hsa-mir