An Approach to Identify Individual Functional Single Nucleotide Polymorphisms and Isoform MicroRNAs

MicroRNAs (miRNAs) and single nucleotide polymorphisms (SNPs) play important roles in disease risk and development, especially cancer. Importantly, when SNPs are located in pre-miRNAs, they affect their splicing mechanism and change the function of miRNAs. To improve disease risk assessment, we propose an approach and developed a software tool, IsomiR_Find, to identify disease/phenotype-related SNPs and isomiRs in individuals. Our approach is based on the individual's samples, with SNP information extracted from the 1000 Genomes Project. SNPs were mapped to pre-miRNAs based on whole-genome coordinates and then SNP-pre-miRNA sequences were constructed. Moreover, we developed matpred2, a software tool to identify the four splicing sites of mature miRNAs. Using matpred2, we identified isomiRs and then verified them by searching within individual miRNA sequencing data. Our approach yielded biomarkers for biological experiments, mined functions of miRNAs and SNPs, improved disease risk assessment, and provided a way to achieve individualized precision medicine.


Introduction
MicroRNAs (miRNAs) are 18-24 nucleotides (nts) long, sing-stranded, noncoding RNAs. The biogenesis of miRNAs generally follows a canonical process: primary miRNAs (pri-miRNAs) are cleaved by the RNase III Drosha enzyme, generating precursor miRNAs (pre-miRNAs). These pre-miRNAs are transported to the cytoplasm and further cleaved into 22 nts long miRNA:miRNA * duplexes under the action of the RNase III Dicer enzyme [1]. Only miRNAs are loaded into the RNA-induced silencing complex for mRNA transcript target recognition [2]. Reliance of miRNA biogenesis on sequence can determine the typical secondary structure and thermodynamics for correct processing, and sequence variation around processing sites, mature miRNAs, and flanking regions of pre-miRNAs lead to changes in baseparing, structure, stability, and thermodynamics. Therefore, a single nucleotide polymorphism (SNP) in miRNAs may influence processing [3,4] and disrupt expression, biosynthesis, or activity of miRNAs, altering miRNA biogenesis and function [5,6]. On the other hand, the recent advent of next-generation sequencing showed that mature miRNAs present some sequence variants, or isoforms, with corresponding "reference" mature sequences that generate multiple variants named isoform miRNAs (isomiRs) [7,8]. Recent studies showed that SNPs may affect the miRNA maturation process [9], and isomiR generation is mainly attributed to imprecise cleavage by Drosha and Dicer, RNA editing, and SNPs, taking miR-934-T/G for example, variant occurs at Drosha sit, it leads to 5p product reduce, and isomiRs with longer sequence than 5p increased, most dramatic change happens in the 3p product, the effect of SNP in mutant samples gives rise to isomiRs with mutant type in the Drosha sit, and variant produces more isomiRs than canonical mature miRNA resulting from wild type [10].
Given that impaired miRNA processing can lead to substantial decrease in miRNA and increase in isomiRs expression levels, and isomiRs differentially regulate the targeted mRNA transcripts, complex phenotypes and diseases may result [8]. SNPs in miRNA can thus have significant phenotypic consequences and cause various diseases. Because SNPs are inherited genetic variations, they can be detected by highthroughput technology, and, therefore, SNPs and isomiRs are alternative or complementary markers to tissue-based biomarkers. Identification of SNPs and isomiRs is of great significance to explore the functions of SNPs and miRNAs, revealing different human phenotypes and disease risk.
In recent years, owing to the important roles of SNPrelated miRNA (SNP-miRNA) in disease risk and development, several databases and software tools such as MirSNP [11], PolymiRTS [12], SubmiRine [13], MicroSNiPer [14], Mirsnpscore [15], and mrSNP [16] were developed. However, these databases mainly focus on predicting the effects of SNPs on miRNAs targets. For the study SNP and isomiR function, a series of isomiR-related databases and software tools were developed. IsomiRex can identify miRNA, isomiRs, and differential expression based on next-generation sequencing data [17]. miR-isomiRExp analyzes the expression levels of miRNAs/isomiRs and track miRNA/isomiR processing mechanisms to uncover functional characteristics of these molecules [18]. The isomiR databank collected 308,919 isomiRs associated with 4,706 mature miRNAs, revealing the function of these isomiRs [19]. miRNASNP is a SNP-related miRNA database based on miRBase and dbSNP databases and contains 2257 SNPs located in 1596 pre-miRNA loci and flanking regions [20]. MSDD captured experimentally validated relationships between miRNAs, SNPs, genes, and diseases and contains 182 human miRNAs, 197 SNPs, 153 genes, and 525 interactions between them [20]. In addition, MiRVaS provides positions of mutations in miRNA (seed, maturation, stem, ring, hairpin arm, and flanking region) and predicts the changes that these variations will cause in miRNA structure [4]. MiRvar studied the effects of SNPs on the miRNA maturation mechanism, extracted 106 SNPs located in 85 miRNAs, and identified the typical miRNA and related isomiRs by analyzing the allele frequencies of 23 SNPs, suggesting that these SNPs have specific functions [21].
The study of SNP-miRNA focuses on several aspects: prediction of the effects of SNPs on miRNA targets, SNP-miRNA database development, the relationship of SNP-miRNA with disease or expression change, and identification of SNP-miRNA-related isomiRs using next-generation sequencing. The mechanism of mature miRNA processing should be further studied. Moreover, alternate splicing caused by variants was predicted by computational methods, and it was verified that up to 94 % of variants causing alternate splicing were correctly classified [22]. Because the mature sequence and processing mechanism can be predicted based on the biological characters of miRNAs using machine learning methods, several software tools were employed to identify the mature miRNAs (processing sites) from a given pre-miRNA, such as miRdup [23], miRmat [24], MatureByes [25], mirExplore [26], MaturePred [27], MiRPara [28], miRRim2 [29], Microprocessor SVM [30], and MiRduplexSVM [31].
In our previous study, we developed software to predict mature miRNAs [32,33] from a novel pre-miRNA, and the 1000 Genomes Project contains the genomes of 2504 individuals from 26 populations. It includes more than 88 million variants, 3.6 million short insertions/deletions, and 60,000 structural variants, and the distribution of genetic variations across the global sample plays important roles in disease [34]. However, few studies have focused on the identification of SNPs and isomiRs from individual samples. Considering the urgent needs of individualized precision medicine, in this study we provide an approach for identifying individual functional SNPs and the corresponding isomiRs, SNPs, and related pre-miRNAs from individual samples.

Materials and Methods
. . Data. Pre-miRNAs data (miRNA ID, Chromosome, Start position coordinate, End position coordinate, and strand) from 1881 human pre-miRNAs were downloaded from miRBase (release V20  The software GenomeAnalysisTK.jar was used to extract chromosome data and the VCF file of the 23 chromosomes from each sample. During this process, the files including human g1k v37.dict, human g1k v37.fasta, and human g1k v37.fasta.fai were used for SNP search within the human reference sequence. All 23 chromosomes from each example were integrated into one file presenting all SNPs and SNP-pre-miRNAs.
. . Construction of SNP-Pre-miRNA Sequences. To construct SNP-pre-miRNA sequences, SNPs must be mapped to the pre-miRNAs based on whole-genome coordinates. The coordinates of pre-miRNA can be extracted from miRNA information, and thus we downloaded pre-miRNA data from We extracted pre-miRNA and SNP data based on start and end position coordinates of pre-miRNAs and the start position of SNP.
SNPs and miRNAs in the human genome sequence may be located in the plus or trans-strand and, therefore, the coordinates of SNPs and pre-miRNAs must be converted. The SNPs positions of each pre-miRNA in the wholegenome coordinates were calculated. Next, in the pre-miRNA sequences, the SNPs position in the pre-miRNA was searched from front to back, and the trans-strand was searched from back to front to find the corresponding nucleotides.
Taking hsa-mir-9-2 and hsa-mir-1-1 as examples, SNP rs41265488 is located in the trans-strand, and we thus calculated the number of nucleotides between pre-miRNA and SNP as 76 and then located the SNP position in the pre-miRNA from back to front, and the number of spacer nucleotides is 76. SNP rs6122014 in has-mir-1-1 is located in the plus strand, and the number of nucleotides between pre-miRNA and SNP is 2, and thus we located the SNP position in the pre-miRNA from front to back, and the number of spacer nucleotides is 2. Calculation of SNPs position based on plusand trans-strands is shown in Figure 1.
Each pre-miRNA and one or more SNPs can then be mapped based on position coordinates. Because a pre-miRNA may include one or more SNPs, each SNP was used to substitute canonical nucleotides in the pre-miRNA. The combination method was used to construct the SNP-pre-miRNAs of each pre-miRNA: where is the number of SNP-pre-miRNA of the i th pre-miRNA and n is the number of SNPs of the i th pre-miRNA.
Using this method, all possible SNP-pre-miRNA sequences can be obtained without losing important information.
. . IsomiR Identification in SNP-Pre-miRNAs. In our previous study, we developed matpred [33] to identify mature miRNAs from pre-miRNAs using biological characteristics of the 5 arm start sites. As is known, isomiRs are always generated from 3 arm heterogeneity [8,10]. Therefore, we developed an extended version of matpred, named mat-pred2, a software tool capable of identifying mature miRNA based on all processing site characteristics. The underlying reasoning of matpred2 is to consider the four processing sites as identified objects. Firstly, we extracted the training dataset based on the four processing sites. Next, the biological features from each site were extracted, and at last the classifier model was trained. With the exception of dataset extraction training, the other steps of the training model, including feature extraction, feature selection, and model training, are those of matpred.
Structural specificity feature is the nucleotide and structural composition at each site. Taken M and N represent paired nucleotide and no-paired nucleotide, then structural specificity feature is one of the following combinations: In addition, we extracted the number of "-" in sequence. In particular, from 3nt to 8nt on the right of Drosha, from 9nt to 12nt on the right of Drosha, and from 2nt on the left of Drosha to 2nt on the right of Drosha were extracted. And the distance from 5 arm Drosha sit to terminal loop was adopted.
Owing to the models used for each site, matpred2 lacks the low recognition accuracy problem caused by the uncertainty in the 5 and 3 ends of the mature miRNA length and migration distance that affects the recognition of mature miRNA in double-stranded sequence. In addition, based on the biological feature differences between Drosha and Dicer sites, the two sites were separated in the prediction model to make the classifier more representative.
For each processing site of SNP-pre-miRNA, matpred identifies five mature miRNA candidates. To improve the accuracy of our prediction results and remove false positive data, the five candidates are all considered as canonical miRNA. In consequence, miRNAs that differed from the candidates were identified as isomiRs. Related SNPs were identified as functional SNPs.
. . So ware Development. To make our approach convenient for application, we provide IsomiR Find, which is designed to identify individual SNPs and SNP-pre-miRNAs. All dicer sites of mature miRNAs were identified using matpred2 [32]. Moreover, mature miRNA candidates were predicted to provide the normal miRNA and isomiRs, and experimentally validated isomiRs were screened based on miRNA sequencing data.

Results and Discussion
. . Approach Design and Implementation. A computational pipeline was established to identify individual functional SNPs and corresponding isomiRs as shown in Figure 2.
Firstly, we extracted the SNPs of each chromosome from the VCF files. Secondly, we combined all the SNPs of each chromosome by alignment of their positions with miRBase, obtained the SNPs and related pre-miRNAs of each sample, and then constructed the SNP-pre-miRNA sequences of this sample. Thirdly, we identified the four cleavage sites of these SNP-pre-miRNAs using matpred2, and then the normal miRNA and isomiR candidates were extracted. Finally, the individual miRNA sequencing data were used to validate and mine the isomiRs.
. . SNP Extraction from Individual Samples. To test our approach, samples from a British population in England and Scotland (GBR) were selected. GBR data from the 1000 Genome Project includes DNA polymorphism and miRNA sequencing data. DNA polymorphism data of each sample were obtained based on karyotype, and, using our approach, we extracted the SNPs of each chromosome and integrated the SNPs of all chromosomes.
Taking one sample from GBR, HG00096, as an example, our approach mapped SNPs on genomic locations of pre-miRNAs, and 96 SNPs were identified in 90 pre-miRNAs (Supplementary Table 1). SNP location on pre-miRNAs from the mapped data of sample HG00096 is shown in Figure 3.
A total of 30 SNPs were identified in the mature miRNA region, 10 in the seed regions, 10 in the terminal loop, and 25 in the stem region. The results show that the mature miRNA region of pre-miRNA has the highest, and the terminal loop the lowest, SNP density.
. . Construction of SNP-Pre-miRNAs. SNPs were mapped to pre-miRNAs based on their positions in the genome. For example, if miRNA end and start positions are 33484789 and 33484781, respectively, and the position of the SNP rs114964240 is 33484783, then the SNP is located in has-mir-187.
The location of SNPs in pre-miRNA is calculated based on coordinates and the trans/plus strand. For example, rs114964240 is located in the trans-strand, and the distance of re114964240 from the miRNA start position of has-mir-187 is 2. The SNP position in the pre-miRNA is then calculated from back to front of the pre-miRNA sequence, and, therefore, in this SNP, A substitutes G. We then constructed the SNP-pre-miRNA sequences of HG0096 as shown in Figure 4.
Taking has-mir-187 as an example, three SNPs were mapped into has-mir-187 based on the combination method, and we constructed seven SNP-pre-miRNA sequences. The SNP-pre-miRNA sequences of HG00096 are shown in Supplementary Table 2.
. . IsomiR Identification. Matpred2 was trained to identify the four sites of mature miRNAs of pre-miRNA. The four classifiers were trained as follows: training dataset construction, feature set extraction, feature set selection, and classifier training. We defined the start and end sites of the 5 arm of pre-miRNAs as P5 5 and P5 3, respectively, and the start and end sites of the 3 arm of pre-miRNAs as P3 5 and P3 3, respectively. The main difference between the training of classifiers P5 3, P3 5, and P3 3 lies in the construction of the training data sets: the P3 5 classifier used the 5 for train dataset, and the P5 3 and P3 3 classifiers used the 3 for train  dataset. Taking hsa-mir-19a as an example, the construction of the training sets of the P5 3, P3 5, and P3 3 classifiers is shown in Figure 5.
To construct the training dataset of the P5 3 classifier, the 22 nd nucleotide in front of P5 3 was defined as P5 5, and the sequence between P5 5 and P5 3 was shifted to the left of the 3 end. Two corresponding sites of the 2 nt sequence are P3 3 and P3 5. The sequence between P5 5 and P5 3 is positive example data, whereas all sequences offset by 1 nt distance are negative example data. Similarly, the training datasets of P3 5 and P3 3 were constructed. Feature set extraction, feature set selection, and classifier training were performed as described for matpred [33].
To investigate the potential functional effects of SNPs in pre-miRNAs, we used matpred2 to identify mature miRNAs of SNP-pre-miRNAs of HG00096 (Supplementary Table 3). We predicted 101 pre-miRNAs having 115 SNPs in a different guide strand with the incorporation of variations in its sequence. The pre-miRNAs, SNPs, and potential functional effects are summarized in Table 2. All SNPs within pre-miRNAs could have a potential impact in biogenesis and function.        For the four dicer sites, we identified 11 SNPs within 7 pre-miRNAs, 39 SNPs within 47 pre-miRNAs, 11 SNPs within 12 pre-miRNAs, and 44 SNPs within 45 pre-miRNAs with the potential function to alter P5 5, P5 3, P3 5, and P3 3 sites, respectively.

u u A G U U U U G C A U A G U U G C A C U A C A a g a N c c g g u g g uAG U C A A A A C G U A U C U A A A C G U G U u g a u
. . IsomiR Validation. Although SNPs in pre-miRNAs were predicted to alter miRNA splicing sites and generate isomiRs, to validate these effects in experiments, we downloaded the miRNA sequencing data of HG00096 and searched isomiR sequences with reads from the miRNA sequencing data. We identified isomiRs using matpred2 and validated them in miRNA sequencing data as shown in Table 3.
Mature miRNAs were identified in all SNP-pre-miRNAs using the four classifiers. For P5 5, we identified isomiRs of hsa-mir2173h, and then we searched these isomiRs in the miRNA sequencing data of HG00096. We found four isomiRs with the following sequences: CCTGGGAGGTCAAGG-CTGTAGT, GCCTGGGAGGTCAAGGCTGTAG, ATT-GCTTGAGCCTGGGAGGTCA, and TTGAGCCTGGGA-GGTCAAGGCT. The corresponding number of occurrences was 1, 1, 11, and 24 reads, respectively, in the miRNA sequencing data.
For HG00096, we searched 36 isomiRs within 11 pre-miRNAs in the miRNA sequencing data.
. . So ware Implementation. IsomiR Find was developed in perl and JavaScript. To use IsomiR Find, the files human g1k v37.dict, human g1k v37.fasta, and human g1k v37.fasta.fai were needed to extract individual SNPs, as well as GenomeAnalysisTK.jar, libsvm-3.20, and ViennaRNA-1.8.5. The user selects the sample for study from the 1000 Genomes Project database, and IsomiR Find identifies its SNPs and associated isomiRs. IsomiR Find is freely Table 3: Identification of isomiRs using matpred2 and validation using miRNA sequencing data. For the convenience of users, we provide a virtual machine image of the installed IsomiR Find tool to use IsomiR Find in windows system, it can be downloaded from https://pan.baidu.com/s/1YCQQYy-RMT0O5hJY8UvYOA." IsomiR Find is a freely accessible tool for identify candidate SNPs and related isoform miRNAs in individuals from DNA sequencing data; especially, it can also predict isomiRs for novel miRNA sequences with novel SNPs. Owing to MatPred2 including four models for predicting four dicer sits of 5 and 3 arm, our tool can predict the isomiR from both the arms. Because MatPred2 is trained based on conservation features of pre-miRNA, it can be used to identify mature miRNAs of novel pre-miRNAs. Therefore, isomiR can predict various types isomiR such as 5 addition, 5 trimming, 3 addition, 3 trimming, 3 nontemplate addition, and 5 nontemplate addition no matter with seed SNP or tail SNP.

Conclusions
MiRNAs and SNPs play important roles in diseases. miRNA function is closely related to its generation mechanism, and SNPs located in the pre-miRNA may affect its function. To improve personal disease risk assessment, we developed an approach to identify heterogeneous isomiRs generated by an individual's SNPs affecting the pre-miRNA maturation mechanism.
For this approach, we developed a freely accessible software tool, IsomiR Find, to identify disease-causing candidate SNPs and associated miRNAs in individuals, and screen candidate SNPs and related isomiRs. The presented algorithm is, to our knowledge, the first approach that aims at isomiR identification in individuals. It will provide a deeper understanding of transcriptome mutation, cell mechanism discovery, and SNP functional exploration in the human genome. Furthermore, identification of relevant isomiRs can provide a valuable reference for biological experiments, and it can be used to investigate the relationship between SNPs in individuals and diseases. Our approach and the software IsomiR Find are freely accessible for identification of individual SNP-pre-microRNAs and related isomiRs. The software facilitates the identification of candidate diseasecausing SNPs and associated miRNAs in individuals and the screening of candidate SNPs and related isomiRs for experimental validation.
Our approach is applied to the 1000 genomes project which detect most variants with frequencies as low as 1% for 2,504 samples. Our approach mined single nucleotide variation (SNV) functions from the perspective of SNV affecting the miRNA maturation mechanism, providing a new idea for the study of SNV functions in noncoding regions. Because our approach is based on DNA sequence, in clinical applications it can be used to analyze the effects of each SNV on miRNA processing in the individual's DNA sequence with one drop of blood and then predict the function of SNV on phenotype and disease. In addition, our approach solves the technical bottlenecks arising from a small number of heterogeneous miRNAs, low SNV allele frequency, and undesirable miRNA sequencing samples and is expected to make breakthrough discoveries in the function of heterogeneous miRNAs and SNVs in diseases and phenotypes.
Our tool focus on identifying isomiRs which derived from SNP effects on miRNAs mature mechanism. In addition, the origins of isomiRs also include imprecise cleavage of Drosha and Dicer, RNA editing, 3 addition events, and TRBP regulation; it should be studied in further work.
As we known, 5 isomiR will have a large influence on target specificity as it has seed sequences, so an effective tool that provides a target can significantly understand their molecular mechanism; we shall make efforts in our future work to provide a tool to predict the target gene for isomiRs.

Data Availability
The pre-miRNAs data used to support the findings of this study have been deposited in the miRBase (release V20). The data are available from ftp://mirbase.org/pub/ mirbase/20/hairpin.fa.zip. SNPs data of individual samples in 1000 Genome Project including 23 Chromosome data used to support the findings of this study have been deposited in the 1000 Genomes Project. The data are available from ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/. The human GRCh37 reference sequence data (human g1k v37.dict, human g1k v37.fasta, and human g1k v37.fasta.fai) used to support the findings of this study have been deposited in ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/technical/reference/. The miRNA sequencing data of individual sample in 1000 Genome Project used to support the findings of this study have been deposited in https://www.ebi.ac.uk/ arrayexpress/experiments/E-GEUV-2/samples/.