Designing HMMs: Motif discovery and modeling

Position Specific Scoring Matrices capture the distribution of residues observed in each position in a conserved motif, but are not a good model for variable length motifs, recognition of new instances with insertions and deletions, and positional dependencies. Moreover, PSSMs can be used to search for instances of an ungapped motif in an unlabeled sequence, but do not lend themselves to precise boundary detection. We turned to Hidden Markov models to address these limitations. HMMs provide a flexible and expressive formalism for modeling conserved sequence motifs. In addition to modeling precise conserved motifs, like the WEIRD motif, HMMs can also be used to model biologically distinct regions that are characterized by a change in underlying sequence composition, rather than a precise pattern. Examples of these include transmembrane regions, which are enriched for hydrophobic residues, and CpG islands, which have higher GC content.