Stockholm Bioinformatics Center, SBC
Lecture notes: Structural biochemistry and bioinformatics 2001

Lecture 13 Nov 2001, Per Kraulis

Hidden Markov models

4. Sequence profiles

Sequence patterns using regular expressions (such as PROSITE) have a problem with large multiple alignments of divergent families: As more sequences are added, the probability that there will be even a few constant or even strongly conserved sites will diminish. There will always be an exception to the rule. In order to avoid missing a known member of a family, the regexp has to be made more general, but then the danger of including garbage increases. This is the typical sensitivity-specificity problem.

There is another approach. Sequence profiles (Gribskov et al 1987) are essentially patterns where each position in the sequence of the segment (or motif) has been assigned a probability value for each possible amino-acid residue type. Instead of requiring a yes/no response to the question "does the amino acid in the sequence fit the pattern?", we now get a response "it fits at a level of 0.9", or "it fits at level of 0.1". The idea is to make the process softer. Add together the soft responses to an overall sum and then make a decision. Don't make the decision at each comparison step.

One can use an analogy: An exam for students can be designed so that a correct answer is required for each and every question in the exam, although each question may be fairly simple. This corresponds to the regular expression approach. Another type of exam gives points for each correct answer, sums up all points at the end, and decides whether the student has passed or not based on the sum. This corresponds to profiles. A student may be unable to answer one particular question, but can make up for it by answering other questions correctly.

This approach works as if a substitution matrix had been defined for each position in the sequence. This requires that the alignment contains many sequences, which should be as varied as the family really is. Good statistics is necessary. If some parts of the family tree are missing, then the profile will not give members from that part of the tree high scores. It is therefore common to add in information from a Dayhoff-type substitution matrix (or similar); this is like mixing a pure position-dependent matrix with a pure general substitution matrix.

In the course book, two other approaches (fingerprints and blocks) are described. These are also based on numerical methods that attempt to soften the regexp idea. The use they make of the profile method is somewhat different. For more details, refer to the book.

Copyright © 2001 Per Kraulis $Date: 2001/11/09 15:32:22 $