Limits...
NAViGaTing the micronome--using multiple microRNA prediction databases to identify signalling pathway-associated microRNAs.

Shirdel EA, Xie W, Mak TW, Jurisica I - PLoS ONE (2011)

Bottom Line: Since microRNA binding is sequence-based but possibly structure-specific, work in this area has resulted in multiple databases storing predicted microRNA:target relationships computed using diverse algorithms.Our pathway-based analysis of mirDIP data suggests microRNAs are involved in intra-pathway signalling.We identify two distinct classes of microRNAs, suggesting a hierarchical organization of microRNAs co-targeting genes both within and between pathways, and implying differential involvement of universe and intra-pathway microRNAs at the disease level.

View Article: PubMed Central - PubMed

Affiliation: Department of Medical Biophysics, University of Toronto, Toronto, Ontario, Canada.

ABSTRACT

Background: MicroRNAs are a class of small RNAs known to regulate gene expression at the transcript level, the protein level, or both. Since microRNA binding is sequence-based but possibly structure-specific, work in this area has resulted in multiple databases storing predicted microRNA:target relationships computed using diverse algorithms. We integrate prediction databases, compare predictions to in vitro data, and use cross-database predictions to model the microRNA:transcript interactome--referred to as the micronome--to study microRNA involvement in well-known signalling pathways as well as associations with disease. We make this data freely available with a flexible user interface as our microRNA Data Integration Portal--mirDIP (http://ophid.utoronto.ca/mirDIP).

Results: mirDIP integrates prediction databases to elucidate accurate microRNA:target relationships. Using NAViGaTOR to produce interaction networks implicating microRNAs in literature-based, KEGG-based and Reactome-based pathways, we find these signalling pathway networks have significantly more microRNA involvement compared to chance (p<0.05), suggesting microRNAs co-target many genes in a given pathway. Further examination of the micronome shows two distinct classes of microRNAs; universe microRNAs, which are involved in many signalling pathways; and intra-pathway microRNAs, which target multiple genes within one signalling pathway. We find universe microRNAs to have more targets (p<0.0001), to be more studied (p<0.0002), and to have higher degree in the KEGG cancer pathway (p<0.0001), compared to intra-pathway microRNAs.

Conclusions: Our pathway-based analysis of mirDIP data suggests microRNAs are involved in intra-pathway signalling. We identify two distinct classes of microRNAs, suggesting a hierarchical organization of microRNAs co-targeting genes both within and between pathways, and implying differential involvement of universe and intra-pathway microRNAs at the disease level.

Show MeSH

Related in: MedlinePlus

MicroRNA target prediction processing and evaluation.Panel A: Identification of microRNA targets is highly dependent on the experimental setup. Filtering by cell type and microarray platform on an identical initial prediction set can cause a divergence of up to 40% in the final target lists. Panel B: MicroRNA over-expression in different experimental settings results in poor overlap of identified targets. Venn diagram of discrepancy between in vitro microRNA over-expression experiments of mir-124. Panel C: Comparison of precision and recall across microRNA prediction databases, measured by computing the average values for all microRNA predictions by a particular database compared to their matched low stringency “ground truths”. Panel D: The percentage of remaining predictions by considering overlap across 2, 3, 4, and 5 prediction databases. Panel E: Precision measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of predicted targets that were shown to be true across in vitro experiments. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods. Panel F: Recall measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of in vitro targets covered by predictions. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods.
© Copyright Policy
Related In: Results  -  Collection


getmorefigures.php?uid=PMC3045450&req=5

pone-0017429-g002: MicroRNA target prediction processing and evaluation.Panel A: Identification of microRNA targets is highly dependent on the experimental setup. Filtering by cell type and microarray platform on an identical initial prediction set can cause a divergence of up to 40% in the final target lists. Panel B: MicroRNA over-expression in different experimental settings results in poor overlap of identified targets. Venn diagram of discrepancy between in vitro microRNA over-expression experiments of mir-124. Panel C: Comparison of precision and recall across microRNA prediction databases, measured by computing the average values for all microRNA predictions by a particular database compared to their matched low stringency “ground truths”. Panel D: The percentage of remaining predictions by considering overlap across 2, 3, 4, and 5 prediction databases. Panel E: Precision measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of predicted targets that were shown to be true across in vitro experiments. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods. Panel F: Recall measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of in vitro targets covered by predictions. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods.

Mentions: To examine whether a combination of microRNA prediction databases would outperform any one source, data from 15 publicly available microRNA over-expression/knockdown experiments followed by microarray [53], [54], [55], [56], [57], [58], [59], [60], [61] was assembled (Table 3). As discussed in the Methods section, when comparing microRNA target predictions to actual microRNA targets (as determined by microarray experiments) two filtering steps were performed to increase the suitability of the target predictions for the data – filtering by both microarray and by cell type. Filtering by microarray (Table 3 column 3) eliminates targets not present on the particular chip in the experiment, and thus having no chance of appearing in the final target set. Filtering by cell type (Table 3 column 4) eliminates genes expressed at only low levels in the cell line (which would reduce their chances of showing a knock-down effect). This two-step filtering drastically changes the predictions. As illustrated in Figure 2A, beginning with an identical set of mir-1 predicted targets across all databases and filtering by cell type and chip type to make the target predictions suitable for comparison to 2 different experiments results in significantly different final prediction sets – with overlapping targets numbering only 60% of the sets – clearly demonstrating the need to tailor predictions to the setting in which the experiment was done before any comparisons are undertaken. This filtering exercise shows how critical it is to consider tissue specificity when examining microRNAs of interest. Clearly, with the availability of more in vitro and in vivo data, it will become crucial to ensure that data is organized in a tissue-specific manner to enable more accurate modelling of the interactions present in particular settings.


NAViGaTing the micronome--using multiple microRNA prediction databases to identify signalling pathway-associated microRNAs.

Shirdel EA, Xie W, Mak TW, Jurisica I - PLoS ONE (2011)

MicroRNA target prediction processing and evaluation.Panel A: Identification of microRNA targets is highly dependent on the experimental setup. Filtering by cell type and microarray platform on an identical initial prediction set can cause a divergence of up to 40% in the final target lists. Panel B: MicroRNA over-expression in different experimental settings results in poor overlap of identified targets. Venn diagram of discrepancy between in vitro microRNA over-expression experiments of mir-124. Panel C: Comparison of precision and recall across microRNA prediction databases, measured by computing the average values for all microRNA predictions by a particular database compared to their matched low stringency “ground truths”. Panel D: The percentage of remaining predictions by considering overlap across 2, 3, 4, and 5 prediction databases. Panel E: Precision measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of predicted targets that were shown to be true across in vitro experiments. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods. Panel F: Recall measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of in vitro targets covered by predictions. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods.
© Copyright Policy
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC3045450&req=5

pone-0017429-g002: MicroRNA target prediction processing and evaluation.Panel A: Identification of microRNA targets is highly dependent on the experimental setup. Filtering by cell type and microarray platform on an identical initial prediction set can cause a divergence of up to 40% in the final target lists. Panel B: MicroRNA over-expression in different experimental settings results in poor overlap of identified targets. Venn diagram of discrepancy between in vitro microRNA over-expression experiments of mir-124. Panel C: Comparison of precision and recall across microRNA prediction databases, measured by computing the average values for all microRNA predictions by a particular database compared to their matched low stringency “ground truths”. Panel D: The percentage of remaining predictions by considering overlap across 2, 3, 4, and 5 prediction databases. Panel E: Precision measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of predicted targets that were shown to be true across in vitro experiments. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods. Panel F: Recall measurements for microRNA:target predictions by number of prediction databases, indicating the percentage of in vitro targets covered by predictions. Stringency levels refer to confidence in the microarray data and were determined by either p-value or percentage knockdown as discussed in the methods.
Mentions: To examine whether a combination of microRNA prediction databases would outperform any one source, data from 15 publicly available microRNA over-expression/knockdown experiments followed by microarray [53], [54], [55], [56], [57], [58], [59], [60], [61] was assembled (Table 3). As discussed in the Methods section, when comparing microRNA target predictions to actual microRNA targets (as determined by microarray experiments) two filtering steps were performed to increase the suitability of the target predictions for the data – filtering by both microarray and by cell type. Filtering by microarray (Table 3 column 3) eliminates targets not present on the particular chip in the experiment, and thus having no chance of appearing in the final target set. Filtering by cell type (Table 3 column 4) eliminates genes expressed at only low levels in the cell line (which would reduce their chances of showing a knock-down effect). This two-step filtering drastically changes the predictions. As illustrated in Figure 2A, beginning with an identical set of mir-1 predicted targets across all databases and filtering by cell type and chip type to make the target predictions suitable for comparison to 2 different experiments results in significantly different final prediction sets – with overlapping targets numbering only 60% of the sets – clearly demonstrating the need to tailor predictions to the setting in which the experiment was done before any comparisons are undertaken. This filtering exercise shows how critical it is to consider tissue specificity when examining microRNAs of interest. Clearly, with the availability of more in vitro and in vivo data, it will become crucial to ensure that data is organized in a tissue-specific manner to enable more accurate modelling of the interactions present in particular settings.

Bottom Line: Since microRNA binding is sequence-based but possibly structure-specific, work in this area has resulted in multiple databases storing predicted microRNA:target relationships computed using diverse algorithms.Our pathway-based analysis of mirDIP data suggests microRNAs are involved in intra-pathway signalling.We identify two distinct classes of microRNAs, suggesting a hierarchical organization of microRNAs co-targeting genes both within and between pathways, and implying differential involvement of universe and intra-pathway microRNAs at the disease level.

View Article: PubMed Central - PubMed

Affiliation: Department of Medical Biophysics, University of Toronto, Toronto, Ontario, Canada.

ABSTRACT

Background: MicroRNAs are a class of small RNAs known to regulate gene expression at the transcript level, the protein level, or both. Since microRNA binding is sequence-based but possibly structure-specific, work in this area has resulted in multiple databases storing predicted microRNA:target relationships computed using diverse algorithms. We integrate prediction databases, compare predictions to in vitro data, and use cross-database predictions to model the microRNA:transcript interactome--referred to as the micronome--to study microRNA involvement in well-known signalling pathways as well as associations with disease. We make this data freely available with a flexible user interface as our microRNA Data Integration Portal--mirDIP (http://ophid.utoronto.ca/mirDIP).

Results: mirDIP integrates prediction databases to elucidate accurate microRNA:target relationships. Using NAViGaTOR to produce interaction networks implicating microRNAs in literature-based, KEGG-based and Reactome-based pathways, we find these signalling pathway networks have significantly more microRNA involvement compared to chance (p<0.05), suggesting microRNAs co-target many genes in a given pathway. Further examination of the micronome shows two distinct classes of microRNAs; universe microRNAs, which are involved in many signalling pathways; and intra-pathway microRNAs, which target multiple genes within one signalling pathway. We find universe microRNAs to have more targets (p<0.0001), to be more studied (p<0.0002), and to have higher degree in the KEGG cancer pathway (p<0.0001), compared to intra-pathway microRNAs.

Conclusions: Our pathway-based analysis of mirDIP data suggests microRNAs are involved in intra-pathway signalling. We identify two distinct classes of microRNAs, suggesting a hierarchical organization of microRNAs co-targeting genes both within and between pathways, and implying differential involvement of universe and intra-pathway microRNAs at the disease level.

Show MeSH
Related in: MedlinePlus