@Article{Lange:Maticzka:Mohl:Globa_local_Predi:NAR2012,
  author =	 {Lange, Sita J. and Maticzka, Daniel and M{\"o}hl, Mathias and 
                  Gagnon, Joshua N. and Brown, Chris M. and Backofen, Rolf},
  title =	 {Global or local? {Predicting} secondary structure and 
                  accessibility in {mRNAs}},
  journal =	 NAR,
  year =	 {2012},
  volume =	 {40},
  number =	 {12},
  pages =	 {5215-26},
  user =	 {sita},
  pmid =	 {22373926},
  doi = 	 {10.1093/nar/gks181},
  note = 	 {SJL and DM contributed equally to this work.},
  issn = 	 {1362-4962},
  issn = 	 {0305-1048},
  abstract =	 {Determining the structural properties of mRNA is key to 
                  understanding vital post-transcriptional processes. As 
                  experimental data on mRNA structure are scarce, accurate 
                  structure prediction is required to characterize RNA 
                  regulatory mechanisms. Although various structure prediction 
                  approaches are available, it is often unclear which to 
                  choose and how to set their parameters. Furthermore, no 
                  standard measure to compare predictions of local structure 
                  exists. We assessed the performance of different methods 
                  using two types of data: transcriptome-wide enzymatic 
                  probing information and a large, curated set of 
                  cis-regulatory elements. To compare the approaches, we 
                  introduced structure accuracy, a measure that is applicable 
                  to both global and local methods. Our results showed that 
                  local folding was more accurate than the classic global 
                  approach. We investigated how the locality parameters, 
                  maximum base pair span and window size, influenced the 
                  prediction performance. A span of 150 provided a reasonable 
                  balance between maximizing the number of accurately 
                  predicted base pairs, while minimizing effects of incorrect 
                  long-range predictions. We characterized the error at 
                  artificial sequence ends, which we reduced by setting the 
                  window size sufficiently greater than the maximum span. Our 
                  method, LocalFold, diminished all border effects and 
                  produced the most robust performance.}
}