Comprehensive analysis of human small RNA sequencing data provides insights into expression profiles and miRNA editing

Abstract

MicroRNAs (miRNAs) play key regulatory roles in various biological processes and diseases. A comprehensive analysis of large scale small RNA sequencing data (smRNA-seq) will be very helpful to explore tissue or disease specific miRNA markers and uncover miRNA variants. Here, we systematically analyzed 410 human smRNA-seq datasets, which samples are from 24 tissue/disease/cell lines. We tested the mapping strategies and found that it was necessary to make multiple-round mappings with different mismatch parameters. miRNA expression profiles revealed that on average ∼70% of known miRNAs were expressed at low level or not expressed (RPM < 1) in a sample and only ∼9% of known miRNAs were relatively highly expressed (RPM > 100). About 30% known miRNAs were not expressed in all of our used samples. The miRNA expression profiles were compiled into an online database (HMED, http://bioinfo.life.hust.edu.cn/smallRNA/). Dozens of tissue/disease specific miRNAs, disease/control dysregulated miRNAs and miRNAs with arm switching events were discovered. Further, we identified some highly confident editing sites including 24 A-to-I sites and 23 C-to-U sites. About half of them were widespread miRNA editing sites in different tissues. We characterized that the 2 types of editing sites have different features with regard to location, editing level and frequency. Our analyses for expression profiles, specific miRNA markers, arm switching, and editing sites, may provide valuable information for further studies of miRNA function and biomarker finding.

Introduction

MicroRNAs (miRNAs) are ∼22 nt long endogenous small non-coding regulatory RNAs, which play key regulatory roles in various developmental, physiological and pathological processes by the negative regulation of gene expression.¹ There are more than 2000 miRNAs that have been identified in the human genome as described in miRBase database (Release 20).² Previous studies have demonstrated that the expression of most miRNAs in the human genome is tissue, biological process and disease specific.^3-5 Specific expressed miRNAs and differentially expressed miRNAs have been widely studied as potential biomarker for diagnosis, subtype classification, prognosis, and therapy in human diseases, especially in various cancers.^6,7

Advances in high-throughput next-generation sequencing technology have resulted in the popular use of small RNA sequencing (smRNA-seq) as a method to detect miRNA expressions. SmRNA-seq can not only detect the entire repertoire of expressed miRNAs in a sample, but also reveal a series of miRNA variants, such as miRNA editings and additions.^8,9 Currently, there are hundreds of human smRNA-seq data deposited into public databases such as NCBI GEO and SRA databases. These data sets are of great value in investigating the expression profiles, editing and function of miRNAs. However, these datasets were obtained by different groups and each of them studied miRNAs in one or a few tissues and diseases. Except for a few databases, such as miRGator v3.0, microRNA.org and YM500, which show the miRNA expression level based on smRNA-seq data,^10–12 there are still very few integrative studies which would facilitate further researches. A comprehensive analysis of the current available smRNA-seq data will be very helpful in exploring tissue or disease specific miRNA markers and uncovering the expression of different miRNA variants.

Besides miRNA expression, smRNA-seq data contain plenty of variants of miRNAs, called miRNA isoforms (isomiRs). Through systematical analysis, Cloonan et al. found that isomiRs are biologically relevant and functionally cooperative partners of canonical miRNAs.⁹ RNA editing is a post-transcriptional event that recodes hereditary information and it occurs in both mRNAs and miRNAs.¹³ It has been reported that about 16% of human pri-miRNAs are subject to A-to-I (G) editing and it has a large impact on miRNA-mediated gene silencing.¹³ We also analyzed single nucleotide polymorphisms (SNPs) on miRNA genes and predicted their effects on miRNA biogenesis and target binding.¹⁴ Editing sites in miRNAs like SNPs, can interrupt the binding of miRNA to mRNA, and hence influence the functions of miRNA. miRNA editing is found to be an evolutionarily conserved mechanism and regulates target mRNA selection and silencing efficiency.^15,16 The YM500 database also provides information about miRNA isoforms and arm switching discovery,¹² however, there are very few comprehensive miRNA editing analysis in different tissues and diseases.

In this study, we aim to make a comprehensive analysis of miRNA expression, editing, arm switching and also discuss the usage of smRNA-seq data. We critically investigated the miRNA mapping strategies to characterize miRNA expression profiles and editings in 410 data sets. Finally, we concisely showed all the miRNA expression data in an online database, Human MicroRNA Expression Database (HMED), which is freely available at http://bioinfo.life.hust.edu.cn/smallRNA/. These results may broaden and deepen our knowledge of the miRNA specific expression and biological functions.

Results

Summary of the smRNA-seq samples and the overall miRNA expressions

Aiming to make a survey of human miRNA expression and editing, we obtained comprehensive datasets of human smRNA-seq data by downloading more than 700 human smRNA-seq data sets (each for a sample) from NCBI SRA database. After quality control and data filtering as stated in the methodology, we left 410 datasets covering more than 10 tissues, different diseases and cell lines (Table 1). Among them, 352 were from disease or disease control samples and most of them are from cancer and cancer control tissues including breast, liver, leukemia, lung and others.

Table 1.

Summary of miRNA sequencing datasets

Tissue/Disease/Cell linea	# of Data sets
Muscle	1
Pancreas	1
Testis	1
Heart	2
Spleen	2
Brain	3
Breast milk exosomes	4
Kidney	4
Tonsil	5
Haematopoietic tissues
ALL	9
Control	5
Liver
HCC	4
Control	11
Skin
Psoriasis	24
Control	12
Lung
Lung cancer	25
Control	26
Breast
Breast cancer	203
Control	19
Cell Lines
NC	2
Lymphoma	2
Melanoma	8
Hela	9
HEK293T	11
Others	17
Total	410

^aNC: Nasopharyngeal Carcinoma. ALL: Acute Lymphoblastic Leukemia. HCC: hepatocellular carcinoma.

As described in the Method section, we made a test and set up a new mapping strategy to map clean reads to the human genome and miRNAs. As a result, we found that among all the smRNA-seq reads，miRNAs account for ∼78.79% of the total reads, while 3.71% were rRNAs fragments, and ∼1% were from piRNAs and snoRNAs (Fig. 1 ). Using the RPM value to normalize the miRNA expression, we got the whole mapping ratio of different small RNAs in our samples (Fig. 1). We found that on average about 70% of known mature miRNAs were expressed at a very low level or not expressed (RPM < 1) in a sample, and about 9% of known miRNAs had relatively high expression (RPM > = 100). The whole profile illustrated that most human mature miRNAs were found to be expressed at low levels. About 30% of miRNAs were not expressed in all of our analyzed samples. It doesn't mean they are not expressed in any tissues because of the limitation of our used tissue types and number of samples for some tissues. After checking these unexpressed miRNAs in the miRBase deep sequencing data, we found that most of them only had < 10 reads and only a few miRNAs had large reads in samples we didn't collect. Six miRNAs highly expressed in almost all samples (> = 400), which are hsa-miR-21-5p, hsa-let-7f-5p, hsa-let-7a-5p, hsa-miR-26a-5p, hsa-miR-103a-3p, and hsa-miR-191-5p. Their average RPM was higher than 8000 and hsa-miR-21-5p had the highest average RPM of 219,057. We also found that hsa-miR-21-5p had a higher expression in disease samples than normal samples. There were 881 miRNAs (43.14%) without expression and only 2 miRNAs (hsa-miR-21–5p and hsa-miR-26a-5p) with a high expression in all the 280 disease samples. If we summarize the data according to tissues, there were 854 miRNAs (41.82%) which were not expressed in any of these tissues.

Figure 1.

The average distribution of smRNA-seq clean reads and the ratio of miRNAs with different expressions. (a) The average distribution of smRNA-seq clean reads. (b) The average ratio of miRNAs with different expressions (RPM values) in a sample.

Tissue/disease specific miRNAs and disease/control differentially expressed miRNAs

A tissue specific miRNA was defined with specifically highly expression in a special tissue, which includes both the normal and disease of that tissue. Based on the whole expression profile, we observed that a large number of miRNAs were expressed specifically in some tissues and diseases. We identified dozens of tissue or disease specific miRNAs (Table 2) using very strict cutoffs combined Shannon entropy and Z score as described in Methods. We noticed that brain, testis, and the HEK293T cell line had relatively more specific miRNAs. Besides the single tissue/disease specific miRNAs, we also found some miRNAs which were selectively high expression in 2 sample types, such as hsa-miR-1 and hsa-miR-133a were highly expressed in both heart and muscle tissues.

Table 2.

Tissue/disease/cell line specific miRNAs

Tissue	Specific miRNAs in one tissue/disease/cell line	Tissue	Selective miRNAs in 2 tissues/diseases/cell lines
Brain	miR-1243p, miR-129-5p, miR-9-5p, miR-338-5p, miR-598, miR-219-2-3p	Heart/Muscle	miR-1, miR-133a
Breast	miR-22-3p, miR-377-3p, miR-497-5p	BMEa/Breast cancer	miR-200a-3p
BMEa	miR-200a-5p, miR-511	BMEa/Pancreas	miR-375
HEK293T	miR-18a-3p, miR-301b, miR-4521, miR-138-1-3p, miR-548b-3p, miR-20a-3p, miR-216a-5p	Brain/HEK293T	miR-9-3p miR-137
Hela	miR-582-3p	Brain/Breast	miR-370
Heart	miR-208b, miR-378a-3p	Brain/Melanoma	miR-584-5p
Liver	miR-122-5p/3p, miR-192-5p	Brain/Pancreas	miR-1224-5p
Lymphoma	miR-302c-3p, miR-3689e, miR-3923, miR-3689a/b-5p	HEK293T/Kidney	miR-218-5p, miR-615-3p
Melanoma	miR-211-5p	Lymphoma/Testis	miR-513c-5p
Muscle	miR-206	Lung/Testis	miR-34c-5p
Pancreas	miR-216b, miR-217	Melanoma/Testis	miR-508-5p miR-509-3-5p, miR-509-3p/5p
Skin	miR-203a
Testis	miR-202-3p/5p, miR-514a-3p, miR-506-3p, miR-514a/b-5p, miR-508-3p

^aBME: Breast milk exosomes. To simplify the table, we ignored the "hsa-" in each miRNA name. miRNAs separated by "/" such as miR-122–5p/3p indicate 2 miRNAs miR-122-5p and miR-122-3p.

miRNAs which play key regulatory roles in diseases may display differential expression in the disease samples comparing to control samples. To identify disease and control differentially expressed miRNAs, we used the edgeR tool for those data sets with both disease and control samples. We identified 51, 41, 8 and 2 differentially expressed miRNAs between breast cancer, lung cancer, HCC, psoriasis and their corresponding control samples, respectively (Table S1). We noticed there were some overlapping disease/control differentially expressed miRNAs in different diseases. Five of the 8 HCC/control miRNAs overlapped with breast cancer/control differential expression miRNAs. Even more, 15 breast cancer/control miRNAs overlapped with lung cancer/control differential expression miRNAs. Notably, hsa-miR-130b-3p and hsa-miR-425-5p were found in 3 different cancer/control differentially expressed miRNA lists, namely: breast cancer, lung cancer and HCC. These two miRNAs were all highly expressed in cancer than in the control, although their absolute expression (RPM) values were not at the highest level.

miRNA editing sites identified in smRNA-seq data

For the discovery of miRNA editing sites, we adopted stricter filtration steps to reduce false positives (see Method section). Finally, 145 samples were remained for the detection of editing sites. These samples were: 36 skin samples, 52 lung samples, 12 liver samples, 12 breast samples, 10 kidney samples, 7 Hela samples, 6 blood samples, 7 melanoma samples and 3 brain samples. We identified 494 miRNA editing sites in these samples and 438 of them occurred in less than 5 samples. We considered these low-frequency editing sites as low credibility sites as they may be the consequence of random error in sample preparation, PCR amplification or sequencing. The other 56 significant modification sites occurred in 5 or more samples were considered for further analysis. These include 24 A-to-G (indicating A-to-I editing) sites, 23 C-to-U sites, 5 G-to-U sites, 1 A-to-C site, 1 G-to-A site, 1 U-to-A and 1 U-to-C site (Table S2). Remarkably, most of them were A-to-I and C-to-U modifications, which are widely discussed in RNA editing reports.^17,18 Most (70.8%) of detected A-to-I modification sites are supported by literature, highlighting the accuracy of our methods and results (Table 3 and Table S2). Although the numbers of editing sites of A-to-I and C-to-I were similar, the number of samples with A-to-I editing (936) was much more than that of C-to-U editing type (364) (Table S2 and Fig. S1).

Table 3.

Widespread A-to-I and C-to-U modifications in different tissues

miRNA	Sitea	Editing Samples	Editing levels (%)	Exp. levelsb (RPM)	Surrounding Sequence	Ref.c
A-to-I(G)
hsa-miR-99a-5p	1	93	4.66	5903	5’ UAAAC 3’	(1),(3)
hsa-miR-100-5p	1	23	0.38	3456	5’ CAAAC 3’	(3)
hsa-miR-151a-3p	3	40	0.43	666	5’ CUAGA 3’	(2),(3)
hsa-miR-200b-3p	5	95	1.16	3037	5’ AUACU 3’	(3)
hsa-miR-27a-3p	6	59	0.24	5043	5’ ACAGU 3’	(3)
hsa-miR-27a-5p	1	44	1.78	90	5’ GCAGG 3’	(2)
hsa-miR-376a-5p	3	46	4.11	10	5’ GUAGA 3’	(2),(3)
hsa-miR-376c-3p	6	54	3.18	384	5’ AUAGA 3’	(2),(3)
hsa-miR-379-5p	5	79	3.03	153	5’ GUAGA 3’	(2),(3)
hsa-miR-381-3p	4	93	12.62	67	5’ AUACA 3’	(3)
hsa-miR-411-5p	5	98	39.96	60	5’ GUAGA 3’	(3)
hsa-miR-497-5p	2	36	1.40	1194	5’ CCAGC 3’	(3)
hsa-miR-664a-5p	8	18	3.05	26	5’ CUAGG 3’
hsa-miR-944	6	34	1.57	17	5’ UUAUU 3’
C-to-U
hsa-miR-99a-5p	13	10	0.32	5903	5’ UCCGA 3’
hsa-miR-99a-5p	17	6	0.29	5903	5’ AUCUU 3’
hsa-miR-191-5p	10	12	0.33	8427	5’ AUCCC 3’
hsa-miR-23a-3p	3	46	4.77	8302	5’ AUCAC 3’
hsa-miR-23b-3p	3	43	4.83	2311	5’ AUCAC 3’
hsa-miR-30a-5p	10	14	0.28	10429	5’ AUCCU 3’
hsa-miR-30a-5p	11	31	0.11	10429	5’ UCCUC 3’
hsa-miR-30d-5p	10	33	0.22	6468	5’ AUCCC 3’
hsa-miR-30d-5p	11	31	0.30	6468	5’ UCCCC 3’
hsa-miR-30e-5p	10	12	0.53	5097	5’ AUCCU 3’
hsa-miR-30e-5p	11	15	0.71	5097	5’ UCCUU 3’

^asite position in mature miRNA.

^bThe average expression level of miRNAs calculated from miRNA expression pipeline in all samples.

^c(1). Discovered by Blow ⁶¹; (2). Discovered in brain by Kawahara ^62–64; (3). Discovered in brain by Alon.¹⁹

Twenty-five miRNA editing sites are widespread in 4 tissues

Alon et al. identified many statistically significant A-to-I modifications from pooled human brain and also found some frontal lobe specific A-to-I modifications.¹⁹ Thus, the editing sites for each tissue may be composed of 2 parts: the widespread editing sites and the tissue-specific editing sites. As described in Methods, we characterized the “widespread miRNA editing sites", which occur in more than 3 tissues and in more than 50% of samples in each tissue. Finally, we identified 25 widespread miRNA editing sites, including 14 A-to-I sites and 11 C-to-U sites (Fig. 2, Table 3). Twelve of the 14 A-to-I editing have been reported previously in the brain. According to the editing frequency (number of edited samples/number of total samples) of each site in different tissues, the 25 widespread editing sites were clearly classified into 4 clusters using unsupervised hierarchical clustering (Fig. 2). The C-to-U and A-to-I editing sites were clustered into different clades representative their different tissue distributions. The two clusters of the C-to-U sites were mainly found in liver and normal breast tissues, while the 2 clusters of the A-to-I modifications were mainly edited in skin, lung tissues and breast tissues (Fig. 2).

Figure 2.

Clustering of 25 widespread miRNA editing sites in different tissues. (A) The editing frequency (edited sample number/total sample number) of each editing site in different samples using unsupervised hierarchical clustering with average linkage and Spearman Rank correlation. (B) The average editing level (edited read number/total miRNA read number) of each editing site in different samples. ID for each editing site is composed of mature miRNA name, editing site on mature miRNA and editing type. Numbers in the brackets after the tissue names are the numbers of samples. For detailed editing frequency and editing level data, please refer to Table S2.

Besides the different tissue distribution, we also observed other distinguishing characteristics between A-to-I and C-to-U editings. First, all the 14 widespread A-to-I modifications were located at the position 1–8 of mature miRNA, which may change ‘seed region’ or cause ‘seed shifts’ ^20,21 (Table 3), thus may cause to change their target sets. On the contrary, only 2 out of 11 widespread C-to-U modifications were located at position 1–8 of mature miRNA. Several miRNAs had 2 C-to-U editing sites, especially members of the miR-30 family (miR-30a/d/e). Second, the C-to-U edited miRNAs shared a consistently high expression level, while the expressions of miRNAs with A-to-I editing were much lower and non-uniform. Third，the editing expression ratios (editing reads /total reads of the miRNA) and edited samples of most (11/14) of the widespread A-to-I modifications are higher than those of C-to-U sites. Another evidence for the distinguishing functional mechanism by A-to-I and C-to-U modifications was their difference on motif preference. We noticed the A-to-I and C-to-U modifications had different sequence patterns, which are a UAG motif in A-to-I editing sites and AUC or UCC motifs in C-to-U editing sites (Fig. S2). Recent studies found non-seed interactions between miRNA and mRNA were also statistically significant but had only modest functional effects compared with seed interaction.^22,23 These observations signify that the A-to-I and C-to-U editing may have a distinguishing mechanism to exert the functions and need more studies to explore functional mechanism.

Disease specific A-to-I miRNA editing sites

We further analyzed those modifications showing a great difference between normal and disease tissues. To make a systematic analysis on their functional changes, we only reported those editing sites occurred in the seed region, as most of target prediction algorithm only using seed region to search targets (Table 4). We did not find any significant difference on the miRNA editing expression ratio and sample ratio between the normal and psoriasis skin tissues. Thus, we focused on data in breast, liver and lung tissues. In these 3 tissues, we counted 11 A-to-I modifications which showed a very significant difference in their normal and tumor tissues (Table 4). Among them, the expression of 4 miRNAs with modifications showed an up-regulation trend in tumor tissues, while the other 7 showed a down-regulation trend. Three (hsa-let-7d-3p, hsa-miR-24-2-5p, and hsa-miR-589-3p) were only found in lung tissues and others were widespread edited miRNAs in all the 3 tissues. We observed that hsa-miR-381-3p and hsa-miR-589-3p showed high editing levels in both normal lung and lung cancer tissues. These two modifications were previously reported to be over represented in neuronal functions.¹⁹ Target prediction show that only a small number of targets (averagely ∼20% of canonical miRNA targets) overlap between canonical miRNAs and editing miRNAs, suggesting editing in seed regions will greatly change the function of miRNAs

Table 4.

Disease-specific A-to-I modifications in different cancers

Tissue	miRNAa	Edited Samples	Editing levelsb (%)	Exp. levelsb (RPM)	Predicted Targets
					Before editing	After editing	Overlap(%)d
Breast	hsa-miR-376c-3p	1/2	1.38/5.09	3534/373	1390	2411	341(25)
	hsa-miR-411–5p	2/0	20.15/0	33/12	1029	1158	111(11)
Liver	hsa-miR-200b-3p	1/1	0.83/3.01	125/96	2241	2596	617(28)
	hsa-miR-411–5p	5/1	43.85/1.25	90/651	1029	1158	111(11)
Lung	hsa-let-7d-3pc	10/0	0.76/0	92/81	117	133	0(0)
	hsa-miR-151a-3p	22/10	0.66/0.69	621/836	1206	1783	214(18)
	hsa-miR-24–2–5pc	14/2	1.12/1.09	75/66	861	2222	198(23)
	hsa-miR-27a-3p	21/9	0.22/0.24	3811/3292	2991	514	154(5)
	hsa-miR-381–3p	25/26	9.60/15.78	125/96	2054	1852	411(20)
	hsa-miR-497–5p	24/11	1.31/1.46	214/99	4052	3412	1333(33)
	hsa-miR-589–3pc	1/7	10.91/12.80	1/2	3353	3521	1058(32)
	hsa-miR-664a-5p	8/1	2.84/3.45	16/10	2235	3758	838(37)

^aIn this table, we only present those editing sites occurred in seed region.

^bThis data format is the Editing (Expression) level in normal / cancer tissue. The expression level is the average RPM in normal / cancer tissue calculated from miRNA expression pipeline.

^cIt's lung tissue-specific editing sites, while others are widespread sites.

^dOverlap ratio is calculated by overlapped targets/targets before editing.

Arm switching of miRNA expression

Usually, a pre-miRNA has the ability to produce 2 mature miRNAs (named miR-5p and miR-3p) from its 2 stems, which may have different target profiles and biological functions. However, the expression preference between the 5p and 3p mature sequences can switch in different tissues or developmental times, which is termed as arm switching.²⁴ To study the arm switching events of miRNA expression, we further analyzed the expression of the 3p and 5p mature sequence of 266 pre-miRNAs, of which the average expression RPM of their 5p or 3p arms was higher than 10 in at least one of the 24 tissue or disease categories. Using cutoff 2 as the expression fold change of 5p/3p or 3p/5p, we found that 130 (49%) pre-miRNAs was mainly expressed by 5p and 88 (33%) pre-miRNAs was mainly expressed by 3p. While 46 (17%) pre-miRNAs existed arm switching, which means in some samples the expression of the 5p sequence is higher and in other samples the 3p expression is higher. For example, hsa-miR-142-5p was highly expressed in acute leukemia and tonsil samples, while hsa-miR-142-3p was highly expressed in lung. There were 2 (1%) pre-miRNAs (hsa-mir-545 and hsa-mir-624), of which the expression of their 3p and 5p mature miRNA was close in all the 24 tissue/disease categories.

HMED, an online Human MicroRNA Expression Database

To provide a useful resource for these human miRNA expression data to the research community, we compiled all the expression profile data into a MySQL database and developed a simple and user-friendly online website named HMED. It's freely available at http://bioinfo.life.hust.edu.cn/smallRNA/. The HMED database contains 3 major modules (Fig. 3): (1) Average miRNA expression level in different tissue and disease categories; (2) Basic information for each sequenced sample including the distribution of reads length; (3) miRNA expression level in each sample. A JavaScript based search window is also available for miRNA ID searching. Users can search the miRNA ID or SRA ID of the dataset, as well as browse the data sets by tissue/disease or by miRNA. With the quantitative information of miRNAs in 410 datasets from various diseases and tissues, the HMED enables users to conduct deeper studies into miRNA specific expression and biological functions in different tissues and diseases.

Figure 3.

Snapshot of Human MicroRNA Expression Database (HMED). (A) The homepage of HMED database. It provides miRNA profiles in different ways, including tissues/diseases, miRNAs and experiments. (B) and (E). The heatmap and detailed RPM for all miRNAs expression in a sample. (C) and (D). The average expression of a miRNA in different tissues and diseases in histogram and box plot.

Discussion

To better understand the entire repertoire of miRNA sequences in different tissues and diseases, we have conducted a comprehensive analysis of miRNA expression profiles and variants from 410 smRNA-seq data sets. Based on the analysis, we observed many tissue/disease specific miRNAs and identified 25 widespread editing sites as well as some different features between A-to-I and C-to-U editings. It is noteworthy that only a small part of miRNAs were highly expressed and most of miRNAs were lowly expressed or even not expressed in a sample. This result highlights the necessity to check the expression of candidate miRNAs, when considering the miRNA-related functions.

Our systematic analysis of multiple samples not only validated some previously reported specific miRNAs or differentially expressed miRNAs, but also found new biomarkers. For examples, the miR-122 is a widely reported liver and HCC specific miRNA,^25-27 miR-1/miR-133a are previous reported as heart and muscle specific miRNAs,^28,29 miR-124/miR-9 are key miRNAs in brain development, miR-210/miR-30a-3p are miRNAs differentially expressed in lung cancer/control³⁰ and miR-182/miR-96 are different in breast cancer/normal tissue,³¹ our results also confirmed these miRNAs.^32,33 For the testis, we identified both known specific miRNAs (miR-202-5p/3p) and novel miRNAs (miR-514/508 etc.). However, we may have missed some specific miRNAs which were reported by previous studies. This may be due to 2 reasons: (1) we used a strict method in combining the Z score, Shannon entropy and expression consistence in different samples of the same tissue/disease, (2) the more datasets used for a category, the less common miRNAs in the category. Notably, we found many common miRNAs among the differentially expressed miRNAs in different diseases. miRNAs miR-130b-3p and miR-425-5p highly expressed in 3 kinds of cancers (breast cancer, HCC and lung cancer) comparing to control samples. Their upregulations in different solid cancers have been reported by several independent and combined studies.^34-37 These 2 miRNAs were also reported to be regulated by BCR-ABL, an oncoprotein in chronic myeloid leukemia.³⁸ These evidences suggest these common miRNAs may play an important role in cancer progress across several cancer types.

Editing sites in miRNA sequences may influence miRNA functions by changing their target bindings, which confirmed by our target prediction and enrichment analysis. Thus, when a editing event frequently appears in a number of samples with a certain expression level, we should treat the edited miRNA as a novel miRNA. In this study, we identified 25 widespread miRNA editing sites in 4 tissues (Table 3). Six out of 11 identified A-to-I editing sites were also reported in glioblastoma and normal brain samples,³⁹ which partly supports the reliability of our miRNA editing identification results. Members in the hsa-miR-376 family with several other miRNAs are often found to be associated with A-to-I editing in several publications and in our results.^39,40 Recently, Choudhury et al. found that the overall miRNA editing frequencies of the miR-376 cluster were reduced in human gliomas and proved that the unedited miR-376a* promoted glioma cell migration and invasion, while the edited miR-376a* suppressed these features by targeted different targets, respectively.⁴⁰ This supports the assumption that miRNA editing has specific functions and it is very important to diseases.

In our study, although most of edited miRNA isoforms with a relatively low expression, we still found some edited miRNA isoforms with high editing level and their edited isoforms had a relatively high expression (RPM > 10). For example, editing sites in seed regions of hsa-miR-381-3p and hsa-miR-411-5p had high editing levels (12.62% and 39.96%) and edited isoform expression (RPM 11.6 and 84.7). Target prediction showed only a small part of overlapping target genes between the 2 reference miRNAs and editing isoforms, which suggest the 2 editing isoforms could function as “novel miRNAs”. Hsa-miR-381 was reported to be dysregulated in muscular disorders and ovarian cancer.^41,42 It is an "oncomir" in glioma progression and targets LRRC4 to increase the proliferation of glioma cells.⁴³ Hsa-miR-411-5p was reported to be up-regulated in myoblasts of muscle disorder and it may play a role in regulating myogenesis by suppressing myogenic factors.⁴⁴ We also observed significantly different editing levels of hsa-miR-411-5p in liver cancer and its control, although the edited samples are a few (Table 4). GO enrichment analysis revealed an increase in neuron related genes (“GO:0007399 nervous system development” and “GO:0022008 neurogenesis”) targeted by edited hsa-miR-411-5p as compared to the original miRNA. And, we observed a loss of KEGG “map04612: Antigen processing and presentation” pathway after its editing. Thus, it is worth to study the function of these editing isoforms in disease process like miR-376a in glioma.

In conclusion, our comprehensive analysis of huge smRNA-seq data sets provided a whole and tissue-specific miRNA profiles, including tissue or disease specific miRNAs and disease/control differentially expressed miRNAs, which may be the potential biomarkers. The changes of the target profiles by the editing sites in miRNAs will provide potential function clues for those miRNAs. In the expression analysis, our study unveils the existence of a relatively diverse and complex human miRNA repertoire and the editing events. We hope our study would help both basic research and biomarker applications.

Materials and Methods

Data collection and mapping strategy selection

We retrieved 741 human smRNA-seq datasets sequenced by Illumina platforms from NCBI SRA database and also added 5 data sets from our own smRNA-seq data. After removing polyA degradation fragments, low quality reads with more than half of low quality base (Q < 20), trimming 3’ adaptors and applying a 18-30 nt length filter, we obtained the remain reads as clean reads. Using a cutoff which clean reads accounted for > 2/3 of the raw reads, we left 410 high quality samples. Then an in-house small RNA pipeline was used to process these data (Fig. 4). Since reads in smRNA-seq data are short, a key step in the analysis is to align clean reads to the reference sequences. The mismatch parameter used in the alignment step has a great impact on the results. So we performed a test for different mismatch parameters. When set at no mismatch, an average of 68.9% total reads and only 36.8% unique reads were mapped to the reference genome. When allowed one mismatch in 20 bases, these 2 mapping rates increased to 88.1% and 69.0%, respectively. No big increase of the mapping rates was observed when set at 2 mismatches. Further, we selected 5 smRNA-seq datasets with significantly different number of reads (from 9 K to 32 M reads) to do pre-analysis. We found that reads mapped with 1 mismatch will greatly increase the mapping rate and most of the reads with 1 mismatch were mapped to miRNAs (Table 5). Then, we constructed a trade-off mapping strategy (Fig. 4) with the following steps. Step (1); mapped the smRNA-seq clean reads (Referred as set R) to the reference genome without mismatch, removed other types of small RNAs and we were left with the candidate miRNA reads set R1. Step (2); masked all miRNA sequences on the genome with “X." Reads in set R1 mapped to the masked genome perfectly matched were removed and the remnant as set R2. Step (3); reads in set R2 were aligned to known miRNAs without mismatch. The mapped reads were considered as miRNA reads (Referred as set M1) and unmapped reads as set R3. Step (4); reads in set R3 were further aligned to the mask genome with one mismatch. The unmatched reads were aligned to known miRNAs with one mismatch and then the matched reads were regarded as miRNA reads (set M2). Step (5); we merged the matched reads in set M1 and M2 to calculate the miRNA expression. The BWA tool was used to align reads to the genome as it is fast and accurate in aligning short reads.⁴⁵ The miRExpress tool was used to align reads to known miRNAs because of its good result presentation.⁴⁶

Figure 4.

The pipeline for smRNA-seq data analysis. The parallelogram represents data processing. The shaded rectangles show the main analysis contents. The shaded parallelograms show our improved mapping strategy.

Table 5.

miRNA mapping rates of 5 datasets in testing the mapping strategy.

Strategy	SRR191529 (9322)b	SRR191522 (65428)	SRR191552 (374094)	SRR039615 (8726686)	SRR372670 (32506452)
0 mismatch	51.29%	68.52%	58.16%	72.85%	51.26%
1 mismatch after 0 mismatcha	67.12%	84.54%	69.46%	85.47%	69.35%
1 mismatch	68.79%	85.63%	70.17%	87.17%	69.62%
2 mismatch after 1 mismatcha	73.58%	89.36%	72.03%	87.69%	74.61%

^aOne(2) mismatch after 0(1) mismatch: first 0(1) mismatch mapping then 1(2) mismatch mapping the remanent reads and combine both to calculate the mapping rate. Details refer to the methods.

^bNumber of clean reads in sample from NCBI SRA.

miRNA expression profiles and differential expression

We divided the 410 samples into 24 categories according to the tissues, diseases or cell lines (Table 1). To make these data comparable, we normalized the expression level by miRNA reads per million reads (RPM), which was calculated as 1,000,000 x read counts / total clean reads.⁴⁷ We calculated the average RPM levels of each miRNA in each sample category.

We combined Shannon entropy and the Z score methods to identify miRNAs specifically expressed in one tissue/disease/cell line and selectively expressed in 2 tissues/diseases/cell lines.^48,49 Shannon entropy is used to measure the concentration ratio of the expression levels in different samples and Z score is used for outlier detection. Consider one gene's expression vector x = (x₁, x₂, … , x_N) for N tissues and an observation x_i for tissue i. The entropy of the gene is calculate as H = − ${\displaystyle {\sum }_{i=1}^N\text{p}}$_ilog₂p_i, p_i is the relative expression of miRNA x for tissue i defined as p_i = x_i/${\displaystyle {\sum }_{i=1}^N\text{x}}$_i. A modH is calculated after using a one-step Tukey's biweight to improve robustness of the expression data as the ROKU paper used.⁵⁰ The Z score is calculated as Z_i =（x_i-μ)/σ, μ is the average expression of miRNA x in all tissues and σ is the standard deviation. The quantity of the Z score represents a distance between the raw expression and average expression. For specific expressed miRNAs, we limited the Shannon entropy modH < 1.5 and the max Z score value of 24 categories Zmax > 4.66. In order to identify these selective expressed miRNAs in 2 tissues, we calculated another Z score based on 23 categories (the max Z score is Z${\displaystyle \begin{array}{l},\\ \text{max}\end{array}}$) after removing the highest expressed category. Thus, for selectively expressed miRNA in 2 categories, we restricted modH < 2.0, Zmax < 4.67 and Z${\displaystyle \begin{array}{l},\\ \text{max}\end{array}}$ > 4.1. Finally, we checked the actual RPM expression of the specific or selectively expressed miRNAs in all samples, and removed the false positive results caused by a few very large outliers. The outliers were shown in box plot on the database website. We also tried to detect the disease/control specific miRNAs for the 5 categories with both disease and control samples using the edgeR tool.⁵¹ The Benjamini-Hochberg procedure (BH) is used to control the false discovery rate.⁵² We selected those miRNAs with P value < 0.001, BH adjust P value < 0.01, and the higher RPM > = 100.

MiRNA editing sites identification

It's generally thought that there may be several potential types of false-positive calls of RNA editing sites based on the smRNA-seq data. These are mainly due to: 3’ end addition of mature miRNA,⁵³ sequencing error, cross-mapping of reads⁵⁴ and known SNP sites.¹⁴ These factors will have larger effects in editing identification than in expression calculation. So, to minimize the influence of these factors in editing site identification, we used a more rigorous quality control and a more precise alignment rule as compared to our expression analysis.

We initially discarded those reads with more than 3 low-quality bases (Q < 20) or lack of 3’ adaptor. After trimming the 3’ adaptor, reads longer than 28 nt or shorter than 15 nt were also removed. Finally, after these filtering processes, we kept those samples of which total clean reads were > 2/3 of the raw data. To eliminate the influence of cross-mapping and 3’ modification, we used bowtie to map our clean data to human genome reference sequences with parameters “bowtie -t -v 1 -m 1 –trim3 2 –best –strata."⁵⁵ The –trim3 2 parameter is to trim 2 bases from the 3’ end of each read. And the –m 1 parameter allows those reads with only one best reportable alignment to be reported and suppress those multi-mapped reads. Using the strict mapping rule, lots of reads were discarded for their multiple mismatches and multi-mapped alignments. Then, we discarded those data sets with low mapped ratio (mapped reads/clean reads <0.2) or high mismatch ratio (1 mismatch mapped reads/all mapped reads >0.2).

The following steps for editing sites detection were conducted according to the method described by Alon, et al.¹⁹ Briefly, each read was mapped to the pre-miRNA locations on human reference genome. Then, a binomial statistics analysis was conducted to remove the sequencing error in each mismatch position. Multiple tests (Bonferroni and Benjamini-Hochberg corrections) were performed to get an accurate result. Finally, we downloaded the miRNA SNPs information from miRNASNP v2.0 database (http://bioinfo.life.hust.edu.cn/miRNASNP2/), and removed those SNPs from our editing results.¹⁴ Most of the identified editing sites only exist in one sample. We chose the 56 editing sites occurred in 5 or more samples for further analysis. We tested different parameters and found that using 5 samples as a cutoff would reduce both false positives and false negatives (Fig. S3). We defined a set of “widespread miRNA editing sites," which occur in more than 3 tissues and in more than 50% of samples in each tissue. If an editing site shows an obvious increase in editing level or a clear occurrence frequency between normal and disease sample, we called this site as a “disease-specific editing site” and the corresponding miRNA as a “disease-specific edited miRNA."

Functional analysis of miRNA targets

miRNA target profiles, before and after editing, show a distinct diversity with regards to whether the editing occurred on the miRNA “seed region”. We used our web online service (http://bioinfo.life.hust.edu.cn/miRNASNP2/online.php) to predict the effect impacted by the editing sites. The tools provide online server for predicting the loss and gain of miRNA:mRNA binding sites disturbed by a SNP in the miRNA seed region, by combining the predicted results of TargetScan and miRanda.^56,57 We used this tool to obtain targets of the miRNAs, before and after editing. To better understand potential functions of the miRNA targets, DAVID database was used to do functional enrichment analysis for targets.⁵⁸ The enrichment of gene ontology (GO) categories (the level 4 of ‘Biological Process’) and KEGG pathways were performed. Default DAVID human gene background was used in this analysis, and those terms with P value <0 .05 and False Discovery Rate (FDR) <0.3 were treated as statistically significant ones. The sequence logo was generated by the WebLogo tool.⁵⁹ The heatmap figure of miRNAs were drawn by MultiExperiment Viewer (MeV version 4.7).⁶⁰

miRNA arm switching

To study arm switching, we gathered all pre-miRNAs which comprised of both the 3p and 5p mature miRNAs in all samples. We calculated the relative ratio of the expression of their 3p and 5p mature miRNA, when one of their expression >10 RPM. If the 3p/5p expression ratio >2 or <0.5, we considered the miRNA to have an expression preference for 3p or 5p mature miRNA. If there was either a 5p expression preference or a 3p expression preference in different samples, we regarded the miRNA as an arm switching miRNA.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Acknowledgments

We thank Hong-Mei Zhang, Hu Chen and Hui Liu for helpful discussion.

Funding

The work was supported by National Natural Science Foundation of China (NSFC) (31171271, 31270885, 31471247, 11101300 and 81402744), Program for New Century Excellent Talents in University (NCET), Ministry of Education of China and General Financial Grant from the China Postdoctoral Science Foundation (2014M552049).

Supplemental Material

Supplemental data for this article can be accessed on the publisher's website.