Tumor classification and biomarker discovery based on the 5’isomiR expression level

All the TCGA isomiR expression data for 10,999 TCGA datasets representing 33 tumor types was downloaded (April 2018) from the TCGA data portal (https://​tcga-data.​nci.​nih.​gov). Only primary solid tumor samples with infix ‘01’ in the TCGA sample barcode, with the exception of the blood samples derived from acute myeloid leukemia (LAML; sample infix ‘03’) were included in the study. After excluding all samples that were annotated as ‘potentially problematic’ datasets (file_annotations.txt files), 9085 eligible datasets (isoform.quantification.txt files) were finally included for further analysis (Table 1). For each dataset, we generated 5’isomiR profiles by combing all the sequences together with same loci at 5′ end, and the expression level of each 5’isomiR was calculated by the sum of all corresponding RPM (read per million) values. In order to avoid noise generated due to poorly expressed isomiRs, we only included the 5’isomiR profiles with read depths of ≥10 in more than 10 samples. Next, we log2-transformed the combined normalized read depths for each 5’isomiR; however, as the read depths ≤1 RPM were considered noise, we filtered them by assigning all values less than 1 the value 1 before log transformation.

In this study, we used the GA/RF based model for tumor classification. GA/RF utilizes a GA to select a set of salient features from input and classification module using RF [27]. The selected features were used as inputs to RF [22]. In a genetic algorithm, a population of strings (designated as chromosomes), which encode candidate solutions (the 5’isomiR signature in this case) to an optimization problem, evolves toward better solutions. The evolution typically starts from a population of randomly generated 5’isomiR sets and occurs in generations. In the present study, the parameters including the “chromosome” length, the “population” size and the maximum number of “generations” were set to 50 (including 50–5’isomiR set), 50, and 300, respectively. For RF classification, the randomForest was used (https://​cran.​r-project.​org/​web/​packages/​randomForest/​randomForest.​pdf). Besides, we use SVM classification for comparison, and the e1071 package in R with linear kernel function was run (https://​cran.​r-project.​org/​web/​packages/​e1071/​index.​html).

For tumor, classification prediction may vary based on different samples assigned to the training set. We repeated the above the GA/RF procedure100 times. During each of the 100 runs, the training and testing were carried out, each time using one distinct subset of randomly selected for training and the remaining subsets for testing. In a given run, the training sets were generated by randomly selecting 75% of each cancer’s available tumor datasets, and the testing sets were generated by the remaining 25% datasets. Finally, we achieved optimal 5’isomiR sets after 300 generations of GA/RF steps.

Using the TargetScanHuman (http://www.targetscan.org) and the TargetScanHuman Custom (http://​www.​targetscan.​org/​vert_​50/​seedmatch.​html) prediction of the target genes of 5’isomiRs with original seed region along with different seed region of canonical miRNA, respectively, were performed [40]. Then, the predicted target genes were submitted to the functional annotation tools of DAVID for the functional enrichment analysis [41, 42]. For functional annotation, the 3 Gene Ontology items (GOTERM_BP_FA GOTERM_CC_FAT, and GOTERM_MF_FAT) were selected with the Enrichment Thresholds or EASE set as 0.001.

Here, we have constructed a combination of the genetic algorithms (GA) with Random Forest (RF) algorithms to detect reliable sets of cancer-associated 5’isomiRs from TCGA isomiR expression data for multiclass tumor classification (Fig. 1). After 100 independent runs, the prediction accuracies of each classifier for each cancer could be obtained with 300 generations of GA. Based on the initial pre-selected set population size, we obtained 100 sets of the optimal predictive features, each of which is comprised of 50–5’isomiRs. The GA/RF and GA/SVM achieved quite similar results (the average sensitivities were 92 and 91.5%, respectively), and then our following analysis only used the result from GA/RF classifier. The 100 generated predictor sets required relatively similar classification accuracies(Fig. 2a, Fig. 2b), which indicated that our selected 5’isomiR sets were remarkably accurate for multiclass tumor classifications. Besides, the prediction accuracies for cholangiocarcinoma (CHOL), rectum adenocarcinoma (READ) and esophageal carcinoma (ESCA), were recorded to be relatively low, indicating that these tumors were often classified as other types (Fig. 2c). Interestingly, the samples of these cancers could be effectively classified in some runs by altering the training and test set, with different isomiR sets, except for READ. Further, in order to investigate which tumor types could be hardly distinguished from all others, we calculated the mean prediction sensitivity for all runs. Notably, similar tumor classification was obtained as reported previously (Fig. 2d). Moreover, the majority of samples from READ tumor were misclassified as colon adenocarcinoma (COAD), which could be attributed to similar molecular expression, histology, and anatomical location [19, 34]. These findings suggested that the GA/RF model could perform effective tumor classification by a series of largely independent optimal predictor 5′ isomiR sets.

For all 100 runs, we acquired 100 sets of GA/RF classifiers and each set comprised of 50 items. Then, the frequency of selected 5′ isomiR across all the optimal predictive set was calculated. The relative significance of each 5′ isomiR for tumor classification was assessed by counting how often it appeared in these predicted optimal feature sets. For the functional enrichment analysis, we included all 5′ isomiRs with the frequency of ≥11 in these sets. It is difficult to randomly select a gene for more than 11 times in a 100 subset from 100 runs for a given dataset of 2231 samples, and the significance of adjusted p-value was calculated as less than 0.01 followed by Bonferroni correlation for multiple testing. Finally, 41 highly frequent 5’isomiRs were selected (Table 2). Notably, of which 28 had different 5′ loci than that of the canonical miRNA from miRBase 21, and some 5’isomiRs even originated from the same pre-miRNA.

Next, we examined the expression level of these highly frequent isomiRs (Fig. 3). Many of them showed higher than 10 rpm, which was the threshold value as derived from a previous study for presence or absence of isomiRs [19]. Noticeably, some of them showed extremely lower expression, nearly all of 5’isomiRs showed less than 10 rpm. It is worth noting that we combined 5′ isomiRs with same 5′ loci together in this study.

Next, the 9 most frequently appearing 5’isomiRs from all 100 runs were selected to determine whether a rational classification could be obtained using a reduced set. The training set and the corresponding test set were randomly produced 1000 times. Then, the RF algorithm was performed to determine the prediction accuracy of multiclass tumor classification (Fig. 4a). We could still achieve a reasonable prediction accuracy (the average sensitivity was 73.7%). In the list of 9 most frequently appearing 5’isomiRs, two 5’isomiRs were from hsa-miR-10b-5p and three 5’isomiRs belonged to hsa-miR-375. In order to investigate whether these 5’isomiRs with different 5′ loci could provide an additional contribution to the tumor classification, we removed the three 5’isomiRs which had different 5′ loci from the canonical miRNA, and randomly rebuilt the training set and the corresponding test set 1000 times again. Finally, we could only obtain an average sensitivity of 72.3% by the RF algorithm, which suggested that the isomiR shown different contribution in multiclass tumor classification than the canonical miRNA.

Furthermore, to investigate the candidate target genes of the 9 most frequently appearing 5’isomiRs, the TargetScanHuman was performed for the 5–5’isomiRs with canonical seed region and the TargetScan Custom was used for other 4–5’isomiRs with a shift in seed regions. The target genes predicted were subjected to the functional enrichment analysis. Finally, 2345 genes were recognized by the DAVID web tools to analyze the functional annotation and the detection of enriched functional categories. The gene ontology enrichment analysis suggested that these target genes were highly enriched in genes implicated in the activity of transcription activator and protein kinase and cell-cell adhesion (Fig. 4b).