The known targets and new predictions share many statistically enriched functional terms, including "developmental protein", "homeobox", and "Wnt signalling pathway".
These results lend credibility to the SVM predictions since the predictions share a significant number of functional categories with the training set.
While false positives undoubtfully remain, the fact that the predictions share the properties of functional overrepresentation and bias of genomic location (properties not selected for in the filters) with known miRNAs provides independent indication that we indeed do not massively overpredict new miRNA-genes.
After this filtering, the trends in sensitivity and specificity remained, but were less pronounced: mirCoS found 181 known human miRNAs, compared to 176 for RegEx, and the total number of predictions was reduced to 1667 and 694, respectively, with 337 predictions shared between the two methods.
Predictions shared by all three gene prediction pipelines were considered candidate genes and annotated by running BLAST against the NCBI non-redundant protein databases (nr) and Buchnera aphidicola str.
The same philosophy applies to process interaction predictions: Shared ontology terms suggest such interactions in GOSynthetic but we incorporated IntAct (9) as well as organism-specific databases such as TAIR, Wormbase, Flybase, EcoCyc, MGD, SGD, ZFIN and so on.
In general, any two prediction methods share less than 12% (and as low as 1.4%) identical predicted genes in a head-to-head comparison.
Studies developing or validating a multivariable prediction model share specific challenges for researchers.
Studies developing or validating a multivariable prediction model share specific challenges for researchers (Steyerberg et al, 2013).
