The sequenced libraries were selectively chosen as they represented an array of different melanocytic subtypes ranging from normal pigment cells (melanoblasts and melanocytes) to non-UV exposed (acral, mucosal and uveal), intermittently UV-exposed and chronically UV-exposed melanomas. We also included a cell line derived from a giant congenital nevus. shows sample unsupervised hierarchical clustering (Pearson's correlation) based on expression of all known and predicted miRNAs. It is worthy of note that within this dendogram the group of six cutaneous melanomas cluster together.
We note that the most highly “plastic” cell type included in this study, i.e. melanoblasts, had the largest number of unique sequence tags as well as the most known and predicted miRNAs, whereas the most differentiated cell type (melanocytes) had the least (Tables S1A and S1B
, and ). It is thus tempting to speculate that the miRNA repertoire of a cell diminishes as it terminally differentiates. The various melanoma samples (and the nevocyte cell line) had intermediate numbers, consistent with the notion that these tumor cells have undergone some de-differentiation.
Our analyses of the sequence reads generated from the libraries was relatively stringent with the requirement of reads having to be present at least twice to be included. The removed single reads accounted for up to 20% of some libraries. While it is likely that a proportion of these single reads represent real miRNAs we set this cut-off to filter out potential sequence errors. Another stringency criterion was that each read had to map with no mismatches. Again, some of the unmatched reads may represent new miRNAs but the potential for clonal amplification of PCR-generated artefacts meant that it was more prudent for these reads to be eliminated. Additionally, as a function of depth of coverage, there may be more extremely rare miRNAs to discover in these libraries, albeit the likelihood of these having a significant impact on melanocyte biology is relatively low.
The miRanalyzer software 
proved to be a very useful tool especially in categorising the sequence reads into different classes. This was evident in the percent of reads that mapped to the transcriptome. Due to their size it is nearly impossible to separate mature miRNA transcripts from degraded mRNAs during library preparation.
Recently, Linsen and colleagues 
discussed the apparent sequencing bias toward certain small RNAs thereby preventing the accurate determination of their absolute numbers. Their paper used the let-7 family as an example since they account for the majority of the read counts observed. In the present study we have also noted that let-7 family members were highly expressed across all libraries and conclude that this is indeed a true reflection of their relative abundance as the mRNA expression levels of some of their known gene targets, the RAS family 
and HMGA2 
, show an inverse correlation with read proportion and relative gene expression (data not shown). A possible explanation for the very high read count of the let-7s, in particular let-7a and let-7f, may be because they are the master regulators of many genes involved in cell proliferation 
and as such they require an overall high level of expression to exert an effect on each of their target genes.
The massively parallel sequencing of a diverse set of melanoma and pigment cell libraries revealed a total of 539 known mature and mature-star sequences along with the prediction of 279 candidate novel miRNAs, of which 109 were common to 2 or more libraries, with 3 present in all libraries. This is a large increase upon the number of published miRNAs currently in the miRBase database. As a proof of principle we designed a custom Taqman assay for one of the 3 miRNAs from the high confidence list (MELmiRNA_677) that was present in all libraries. We were able to confirm expression of MELmiRNA_677, which showed strong correlation between relative sequence abundance and ΔCt value (R2
0.8). These data suggest that the prediction (miRanalyzer) and filtering tools (CID-miRNA) we have employed can yield high quality predictions.
highlights that many of the novel candidates are of quite high abundance in the libraries sequenced. This raises the question of why these novel candidates haven't been documented before? The answer may in part be due to some of these miRNAs being specific for the melanocytic lineage, or that many low abundance miRNAs are only likely to be detected through next-generation sequencing efforts. Screening of a large panel of samples from different cell types would be useful to assess these possibilities. Pigment cell specific miRNAs might be of benefit for melanoma diagnosis and early detection of distant metastases and disease recurrence by measuring the circulating levels of these potential biomarkers in blood (reviewed in 
). The primary candidates we have identified so far that may satisfy this goal are MELmiRNA_197, MELmiRNA_434, and MELmiRNA_677, since they are present in all libraries. Moreover, candidate novel miRNAs which appear unique to a particular library may also be useful as histotype-specific markers, especially for identification of metastatic tumors of unknown origin. A more extensive analysis of a larger subset of samples of each melanoma histotype is required to answer this question. It would also be appropriate to compare miRNA profiles from uncultured melanomas and melanocytic nevi with those presented here for purified pigment cells cultured in vitro, since the in vivo microenvironment might alter the relative abundance of some miRNAs, and conceivably condition whether some miRNAs are expressed. However, in order to overcome the obvious problem of tissue heterogeneity in fresh tumor samples, such comparisons would require micro dissection of the tumor cells. Indeed, even within a tumor there could be sub-regional differences in miRNA profiles. Thus, to more comprehensively characterize the melanoma miRNAome, assessment of a large number of melanomas at different stages of progression is needed.
These data have also highlighted the importance of deep sequencing to discover and quantify miRNAs rather than simply relying on off-the-shelf microarrays that contain only a set of known miRNAs. While such arrays are useful screening tools, they do not allow for the discovery of novel candidates that may be specific to a particular cell lineage or series of samples. Moreover, they do not consider the expression of isomirs, some of which are more abundant than the annotated mature miRNA. Many of the miRNA microarrays constructed to date also lack the full complement of miRNA star sequences. Taken together, deep sequencing could be considered the gold standard for comprehensive analysis of the miRNAome.