Quantitative measure for regularizers

 

The traditional way in computational biology to demonstrate that a technique or set of parameters is better is to pick a biologically interesting problem and compare methods for solving it. Many of the regularizers in Section 3 have been validated in this way [HH92, BHK 93, TAK94].

This sort of anecdotal evidence is very valuable for establishing that techniques are useful in real biological problems, but is very difficult to quantify. It is difficult to determine how much improvement is expected on different problems, and whether the improved technique is better in general, or just on the specific problem it was applied to.

In this paper, the regularizers are compared quantitatively on a rather generic problem--independently encoding the columns of multiple alignments. This generic problem has some attractive features:

• There are large data sets of multiple alignments available for training, making it easy to optimize the parameters of methods.
• Many of the search and alignment techniques that will use the regularizers are attempting to produce multiple alignments, and so finding a good regularizer for the encoding problem should produce a good regularizer for the search and alignment algorithms.
• By using trusted alignments, we have fairly high confidence that each amino acid distribution we see is for amino acids from a single biological context, and not just an artifact of a particular search or alignment algorithm.

Throughout this paper, the trusted alignments used are the BLOCKS database [HH91] with the sequence weighting scheme mentioned in Section 5.

• Encoding cost
• More efficient computation of encoding cost