The use of Hidden Markov Models in the Analysis of Biological 
Sequences 

by Steve Hardies. Ph.D. The University of Texas Health 
Science Center at San Antonio

Date: 10/15 M      
Time: 4:00 pm - 5:15 pm   
SB 3.02.02

In the genomic era, sequences of genes from organisms of all kinds are being 
acquired much faster than the functions of their encoded proteins can be 
experimentally characterized.  The functional description of genomes is becoming 
more and more dependent on inferring similar function through the finding that 
the encoded protein sequence is similar to that of some characterized protein.  
Detection of sequence similarity has been dominated by the Blast search algorithm 
and a related profile method called Psi-Blast. Profile methods, by definition, 
create a survey of which amino acids are observed at each position among an alignment 
of the more obviously related sequences, and then search for more distant relatives 
making use of that extra training in what residues should be considered similar at 
each position.  Blast and Psi-Blast are designed for speed, and are served from 
International Bioinformatics Centers, thus relieving individual investigators 
from having to invest in computational facilities or know-how.  However, it has 
long been known that for proteins with established 3D structures, there are many 
examples where two proteins are structurally related, but the presumptive underlying 
sequence similarity is too weak for Blast or Psi-Blast to confidently detect.  
The fundamental problem is caused by the observation that proteins not only 
experience residue substitutions during evolution, but also gain and lose residues.  
To detect sequence similarity therefore requires alignment of the compared sequences 
by adding gaps to account for these insertions and deletions.  Once the comparison 
algorithm is given the ability to add gaps to optimize the number of residue matches 
between two sequences, its flexibility to make unrelated sequences appear similar 
increases and detection of faint similarities is overpowered by this noise.

The Hidden Markov Model (HMM) provides an algorithmic and statistical basis to limit 
the placement of gaps in a sequence alignment based on where insertions and deletion 
are observed among the more closely related (and hence more easily aligned) sequences 
of the set.  Hence comparative algorithms based on HMMs experience reduced noise and 
can detect fainter traces of similarity.  The HMM method is inherently Bayesian; 
meaning that it can accommodate a mixture of information learned from the related 
sequences already aligned with "prior" information derived from theory about how 
proteins usually evolve.  The prior information can reflect facts learned by 
structural or functional investigation of one of the family members.  Hence the HMM 
is a very powerful method for incorporating a variety of insights into the algorthmic 
part of the search.  The lecture will focus on the use of HMM methods to address a 
number of difficult similarity detection problems encountered by Dr. Hardies, and 
will mainly feature the Sequence Alignment and Modeling system (SAM) obtained from 
Richard Hughey at UCSC.


All about Blast and PsiBlast:
•	http://www.ncbi.nlm.nih.gov/BLAST/Blast.cgi?CMD=Web&PAGE_TYPE=BlastDocs
All about SAM
•	http://www.soe.ucsc.edu/research/compbio/sam.html
A paper showing a number of methods of inferring similarity at a distance
•	http://www.ncbi.nlm.nih.gov/sites/entrez?Db=pubmed&TermToSearch=17673272