miRTarCLIP consists of six steps (see Methods for details). It automatically removes adaptor sequences from raw reads, filters low quality reads, reverts C to T, aligns reads to 3'UTRs, scans for read clusters, identifies high confidence miRNA target sites, and provides annotations from external databases (Figure 1 and Figure 2). All of the clusters and miRNA target sites and annotations from external databases are automatically presented in a web-based browser created according to a template. The browser also provides a summary table of putative miRNA target sites with scores from TargetScan, target site locations, target gene annotations, and seed region types. In addition, this system takes advantage of the multi-threading technology to enhance the performance.

As mentioned above, several databases and tools, CLIPZ, doRiNA, starBase, and PARalyzer analyze CLIP/PAR-CLIP sequencing datasets. Table 1 lists the major differences among several resources for CLIP/PAR-CLIP data analysis. CLIPZ provides a web service environment for online analysis. PARalyzer 25 is the only stand-alone tool before this work, but it only handles PAR-CLIP data and does not provide a graphical interface. Here, our miRTarCLIP is implemented as a stand-alone tool, which can analyze the new CLIP-seq/PAR-CLIP data on users' local desktops. It provides high-confidence miRNA-target sites with information in detail and presents them in a web-based interface.

Most uniquely, miRTarCLIP performs a C to T reversion in its workflow for PAR-CLIP dataset, which works along with multithreading techniques to significantly reduce the running time. After mapping reverted reads to 3' UTRs (see Methods), miRTarCLIP clusters reads to search for possible miRNA target sites and uses TargetScan to identify miRNAs that target them. If a candidate miRNA and its target sites had experimental verifications according to miRTarBase, the systems will rank these MTIs on the top of the list in a web-based browser.

To demonstrate how our system works on CLIP-seq data, it is necessary to apply a dataset for analysis. Additional file 1 shows the web interface of the miRTarCLIP analysing a CLIP-seq data from Chi et al. 15 (BrainA_130_50_fastq). In Additional file 1, Lamc1 and mmu-miR-124 were input in the "Gene Symbol" box and the "miRNA name" box respectively. Lamc1 and miR-124 were chosen because this MTI (miR-124::Lamc1) was experimentally verified by Chi et al. 15 . Figure 3 summarizes the complete annotations and visualization results. In Figure 3A, all possible miRNA-target sites in a read cluster are shown with the miRNA seeds on top. In this case, the read cluster in Lamc1 3' UTR (position 2418-2449) suggests a candidate AGO-Lamc1-miRNA terney complex. According to miRNA expression and the context scores given by TargetScan, miRTarCLIP ranks mmu-miR-124a the most possible miRNA that is involved in this MTI, which is as what we anticipated. Figure 3B gives the locations of miRNA target sites (i.e., miRNA_start and miRNA_end) and the context score from TargetScan.

We took the AGO1 PAR-CLIP sequencing dataset (SRR048973) from Hafner et al. 16 as an example. According to Hafner et al. 16 , miR-103a is a highly expressed miRNA and it targets PAG1. Hafner et al. 16 indicated a high T to C conversion at the region between 8^th -13^th nucleotide in the miRNA target sites. miRTarCLIP identified the same region (position 9 in this case) that contains the most T to C conversion (Figure 4). The system also provides multiple sequence alignments for visualizing conserved target sites among 23 species (Additional file 2). In this case, miR-103a target sites in PAG1 are clearly the conserved ones, but they are less likely targeted in rats because this region is not shown in the alignment (see Additional file 2, 10116 is a taxonomy id of rat). Figure 3 and 4 indicate that miRTarCLIP can produce similar results of the original study and provide novel insights of MTIs.

The PAR-CLIP reveals a higher efficiency in RNA co-immunoprecipitation than the regular CLIP. The PAR-CLIP incorporates 4-thiouridine (4SU) into transcripts and applies more energetic UV to enhance the crosslinking between proteins and RNAs, but it also produces artificial T to C conversion. Reads with these errors are difficult to map. Therefore, existing tool, like PARalyzer 25 , allows a read to have two mismatches against the reference. However, it dramatically increases the search space and time needed for finding a good match, and in some cases, it could lead to mistaken mappings (see Discussions).

Hafner et al. 16 and PARalyzer's authors 25 indicated in their works that the ratio of T to C conversion is high in position 8 to 13 of the target sites. The high ratio is considered an evident sign of real miRNA target sites in PAR-CLIP data. To confirm this, we compared the T to C conversion rate within position 1-7 to that within position 8-14 of miRNA target sites, the results indicate that the T to C conversion is significantly different in these two regions (p value = 0.02, by one- tailed Student's T test, see Figure 5A, Additional file 3, Table 2). To further understand the association between T to C conversion levels and high-confidence MTIs, we looked for only highly expressed miRNAs and their target sites. The results show that the T to C conversion rates differ in an even higher degree between these regions (p value = 0.01. See Figure 5B, Table 2 and Additional file 3). These two results suggest that by incorporating miRNA expression, it is possible to reduce the false positives in finding miRNA targets. The rules of miRNA target prediction usually put constraints on the sequence conservation and miRNA seed regions 4 . Therefore, we also tested whether the conservation and seed regions play a role here. Analytical results indicate that all of the nonconserved seed regions (i.e., N78, N8, N7, see Figure 5, Table 2) and total miRNA/CN7 miRNA-target do not exhibit significant difference (p value > 0.05) (Figure 5A, Table 2 and Additional file 3). The results suggest the importance of seeds and conservation. Above results consent with the finding that T to C conversion is located in non-complementary regions of the ternary AGO complex 16 26 .

Chi et al. 15 recently analyzed MTIs in the mouse brain tissue by high throughput sequencing and CLIP. Hafner et al. 16 modified the original CLIP methods by incorporating 4-thiouridine (4SU) into transcripts to increase the efficiency of crosslinking and provide high resolution in protein-RNA binding sites. The raw data of AGO1-AGO4 can be obtained from Gene Expression Omnibus (GEO: GSM545212, GSM545213, GSM545214, GSM545215). The sequencing raw data of these two studies are used in this proposed miRTarCLIP system.

The miRNA related information, including the accessions and miRNA sequences were obtained from miRBase release 18 27 28 . microRNA indexes are created to replace miRNA names because of the inconsistency of miRNA naming among different versions of miRBase. The miRNA target prediction and 3' UTR data are obtained from TargetScan release 6.2 29 30 . The experimentally confirmed MTIs were collected from miRTarBase release 2.5 31 , which was developed previously by our group.

Figure 2 illustrates the analysis flow of miRTarCLIP pipeline. The FASTX-Toolkit 32 , SRA-Toolkit 33 , and bowtie 34 was incorporated into the miRTarCLIP analysis pipeline. The pipeline has six steps: (1) adapter trimming, (2) quality control, (3) C to T reversion, (4) read alignment, (5) cluster analysis, (6) MTI identification analysis. We also take advantage of multi-threading to enhance the performance of the algorithms.

This step removes the adapter sequence, if any, at the 3' end of each read. If a trimmed read is shorter than 15 nucleotides or contains any ambiguous nucleotides, the reads are discarded.

Following the adapter trimming step, we scan the quality at the tail of each read. The elimination rules are based on the phred quality score. Notably, the nucleotides at the 3' end are removed when their phred scores are lower than 20. Similarly, a reads is discarded if its length less than 15 nucleotides after the tail trimming. Reads with the same sequences are collapsed into one to save the time for mapping duplicates.

PAR-CLIP technology is implemented by incorporating 4-thiouridone (4SU) to cause thymidine to cytidine transition in the RNA binding protein sites on transcripts. For each cytidine in a read, this step will create a new read with that C converted to T. For instance, a read sequence, AATG

C

miRNAs target mRNAs at 3' UTRs, so instead of aligning reads to the entire genome, we use exclusively the 3' UTR sequences from TargetScan. The reads are mapped with at most one mismatch. Other tools, like PARalyzer 25 uses two mismatches to address the T to C conversion issue in PAR-CLIP dataset. However, allowing two mismatches in mapping (e.g., using bowtie) is very time consuming and error-prone. To resolve this problem, a better strategy is to revert C back to T in reads (as described in Step 3), and align them to the references with at most one mismatch, in which reduces the computational costs. We have tested our results with published PAR-CLIP data from Hafner et al. 16 (

SRX020783

These reads are clustered based on their minimum overlap between each other, at least 20% of the reads in a cluster should have the T to C conversion; the minimum number of reads in a cluster is five reads. In the PAR-CLIP dataset, a cluster reads should contain at least 20% of the T to C conversion. Whether the cluster sequence is a possible target site is confirmed using the miRNA seed region sequences extracted from miRBase.

The clustering results are used to search for possible miRNA target sites by TargetScan. If a candidate target site is experimentally validated according to miRTarBase, the system will display it on the top. Other candidates will be ranked according to the context scores assigned by TargetScan.