Hatzigeorgiou replies:

Xu noticed a discrepancy between his results for the unions and intersections of different target prediction programs and the precompiled data on the provided websource TargetCombo (http://www.diana.pcbi.upenn.edu/cgi-bin/TargetCombo.cgi).

He correctly points out that a possible explanation for this discrepancy is differences in the versions of the prediction programs. For the TargetCombo website, we used the precompiled lists provided by TargetScanS, PicTar and MiRanda. The discrepancy can be largely explained by two different sets of predictions provided on the MiRanda website (http://www.microrna.org/mammalian/index.html). One set of predictions is accessible by web search (http://cbio.mskcc.org/cgi-bin/mirnaviewer/mirnaviewer.pl) and another set, used for TargetCombo, is available for direct download in the form of Excel files.

We manually investigated several putative miR-155 targets that are discrepant between the two MiRanda prediction sets. For example, the ATP2B1 gene can be found as a predicted target of miR-155 in the downloadable Excel files but not in the web-searchable set of miRanda predictions. In contrast, the RREB1 gene can be found in the latter but not in the former set of predictions. Finally, genes such as CUGBP2 can be found in both prediction sets. This would explain many of the differences in the results presented by Xu and those on the TargetCombo website.

In the first step of the published analysis, which investigates the performance of different programs on experimentally supported targets, we manually inspected the web-searchable set of miRanda predictions (performed in April 2006). When preparing the TargetCombo website, we used the predictions provided in the miRanda Excel files (July 2006 version) instead because these facilitated a straightforward procedure for automatic data extraction. Additionally, between October 23, 2006 and November 15, 2006 there was a discrepancy on a limited set of MiRanda targets listed in TargetCombo, caused by a parsing error in the web site script. This error has been fixed and in no way affects the analysis as described in the publication. We have also added a cautionary note on the TargetCombo website that multiple versions of some programs, as well as differences in gene-identification systems, can indeed yield noticeably different prediction sets.