next up previous
Next: 5 Results for training Up: Evaluating Regularizers for Estimating Previous: 3 Posterior Counts

4 Experimental method

 

The regularizers are compared by computing the average entropy of a multiple alignment given a regularizer. The multiple alignments chosen are from the BLOCKS database [8]. The sequences are weighted using a slight variant of the Henikoffs' position-specific weighting scheme [10], as implemented by Kimmen Sjölander. Her weighting scheme is proportional to the Henikoffs' position-specific weights, but instead of having the weights sum to 1.0 for each block, they sum to the number of sequences in the block, so that blocks containing more sequences have more influence than blocks with only a few sequences. Other weighting schemes have been proposed (for example, tree distances [18] or weighting for pairs of alignments [2]); I chose the position-specific one rather arbitrarily for its ease of computation.



karplus@cse.ucsc.edu