Background
The accumulation of genetic alteration drives cancer development and progression [
1]. The Cancer Genome Atlas (TCGA) consortium integrated comprehensive clinicopathologic annotation data together with molecular profiles of over 11,000 human tumors across 30 different human tumor types [
2]. Analyzing these large datasets can provide more exciting opportunities to better understand the tumor characteristics and discover novel and effective predictive and prognostic tumor biomarkers and therapeutic targets. While most of the previous studies on tumor classifications have focused primarily on the gene expression data, including RNA-seq and microarray data [
3‐
5]. miRNAs and isoforms of human miRNAs (isomiRs) also play essential roles and may serve as potential biomarkers for tumor classification [
6‐
11]. The isomiRs are predominantly generated from the alternative cleavage of Drosha or Dicer and 3′addition events, which produce mature miRNA different from the canonical miRNA by a few nucleotides at the 5′ or 3′ end and designated as 5’isomiR or 3’isomiR [
12,
13]. Both computational and experimental analyses revealed that the isomiRs are involved in regulating distinctive target genes and could play a crucial biological role in miRNA-mediated gene regulation [
14‐
18]. In a recently conducted study, the presence or absence of 7466 isomiRs could be effectively discriminated amongst 32 TCGA cancer types with 90% sensitivity [
19]. Moreover, an average sensitivity of 82% was achieved by using 456 most significant isomiRs. In the present study, we aimed to evaluate the effective reduction of discriminant isomiR features for multiclass TCGA tumor discrimination classification.
The filter, wrapper and embedded methods are typically utilized for feature selection, though all of them are not good at dealing with data which contain a large number of collinear variables. [
20,
21]. The isomiRs may belong to the same miRNA family, the same miRNA cluster, or some of them even have same seed region, leading to similar or related function and highly correlated expression. In the previous studies, the wrapper method could outperform embedded methods by the combined machine learning algorithm for classification [
21]. The genetic algorithms combined with the machine learning algorithm, which employs a GA as the search engine for feature subset selection and the machine learning algorithm as the classification tool, was efficiently used for classification of gene expression data [
3,
4]. This algorithm could identify and classify more than 90% of samples from 31 tumor types with a set of 20 genes [
3]. Beside the machine learning algorithms -- such as support vector machines (SVM), sparse representation (SR), sparse representation classifier (SRC), random forest (RF), and k-Nearest Neighbors (KNN) -- has been extensively applied in cancer prognosis and prediction analysis [
3,
22‐
25]. Nevertheless, GA has proven ability to detect the optimal classifier effectively for multiclass cancer discrimination [
4,
26]. GA is based on Darwin’s theory of natural evolution, and it is typically implemented using computer simulations in which an optimization problem is specified. GA is frequently used to generate high-quality solutions to optimization problems using genetic operators: selection, crossover, and mutation [
27]. In this study, we constructed a combination of the GA with random forest algorithms to detect reliable sets of cancer-associated 5’isomiRs from TCGA isomiR expression data.
Furthermore, 5’isomiR may target very different transcripts as compared with their canonical miRNAs attributed to shifting in the seed region (typically 2–7 nt of the miRNA), which is recognized to be very critical in determining miRNA target specificity [
28‐
31]. Various 5′ isomiRs play an important role in suppression and progression of cancer [
32,
33]. Using the combined GA/RF algorithms, reliable sets of candidate tumor biomarkers for multiclass tumor discrimination was detected by combining all the miRNA isoforms with same loci of 5′ end together in TCGA isomiR expression data. In this step, the miRNA isoforms with same loci of 5′ end will be left with only one in the reliable sets, which will dramatically reduce the type of isomiRs. The findings of the present study revealed that the 5’isomiRs might be utilized for effective tumor classification and classifier can achieve an average sensitivity of 91.5% with only 50–5’isomiRs.
Discussion
We report a novel GA/RF analytical model for multiclass tumor classification using the miRNA expression data that may reveal effective predictive and prognostic biomarkers and therapeutic targets for drug development. With an average sensitivity of 92%, we were able to accurately classify the tumor samples using 100 different 50–5’isomiR sets, though some 5’isomiRs appeared repetitively in the sets. These predictive 5’isomiRs sets could achieve similar prediction accuracies with slight overlap; suggesting that even less sensitive 5’isomiRs could be detected for the tumor classification. Notably, most of the tumor types could be easily distinguished with high sensitivity. However, there were also some cancers that exhibited low prediction accuracy due to similar histology and anatomical location [
19,
34]. In this study, we used GA algorithm to obtain the optimal isomiR set for maximizing the prediction accuracy in all TCGA cancer types but not for some individual cancer. Therefore, the samples from some cancers that cannot be classified by one set may be successfully classified by another set. In addition, we calculated the frequency with which each 5’isomiRs appeared in these sets. More than half of 41 highly frequent 5′ isomiRs showed different 5′ loci than the canonical miRNA, supporting that the isomiRs play a significant role in the multiclass tumor classification. It is noted that our analysis only included tumor samples, and we cannot distinguish cancer-specific isomiRs from tissue-specific biomarkers. Actually, a group from Saarland university had utilized a tissue specificity index to define the distribution of miRNA across 61 tissue biopsies of two individuals, and people can check whether the detected isomiRs correspond to the tissue-specific miRNA expression in their web-based repository [
35].
In a recent study, the RNA-seq expression data analysis revealed that many development-related genes are essential for the analysis of TCGA cancer classification [
3]. Similar clues were also revealed in the present study. 5’isomiR-233, one of the most frequently appearing 5’isomiRs in 100 generated predictor sets, derived from the shift in the seed region of canonical hsa-mir-196b-5p, which usually appears to be expressed from the intragenic regions of HOX gene clusters that are major regulators of animal development [
36]. Increasing studies have suggested that 5’isomiR-313, combined from the isomiRs with identical 5′ loci of the canonical hsa-miR-205-5p, play an important role in normal cellular development as well as in cancer development [
37,
38]. Moreover, TBX5, one of the most important genes for tumor classification from the previous study [
3], could be regulated by one of the 5 most frequently appearing 5’isomiRs in our sets (miR-10b-5p/5’isomiR-39) as derived from the TargetScanHuman prediction.
Using only 50–5’isomiRs, the present GA/RF model could achieve comparable prediction performances consistent with previous report, with an average accuracy of 90% for all isomiRs [
19]. We also detected the similar discriminatory isomiRs as their finding. For example, the isomiRs of has-miR-205-5p and has-miR-944, two of the most important miRNAs detected by the method using the presence or absence of isomiRs amongst 32 TCGA cancer types, are also listed in the ten highly frequent isomiRs from 100 generated predictor sets. The isomiR of hsa-mir-196b-5p, the most frequently appearing 5’isomiRs with a shift in seed regions found in our study, showed a high VI score in previous report [
19]. Further, we reduced the number of features by employing two strategies. In the first approach, we combined the isomiR with same 5′ loci to reduce the type of isomiRs. While in the second approach, the GA-based isomiR selection reduced the feature selection significantly. We also found that the 9 most frequently appearing 5’isomiRs could achieve an average sensitivity of 73.7%, suggesting that a reasonable accurate performance could be obtained with less number of features. The features can be further reduced by additional approaches, including hybrid GA-based machine learning method [
39]. The highly expressed 5’isomiRs (rpm > 10 in all samples) and slightly expressed 5’isomiRs (rpm < 10 nearly in all samples), demonstrated that the expression level of isomiRs could also be beneficial for the tumor classification.