Hidden Markov Models (TIGR)

Michelle Gwinn, TIGR curators; 2003

A Hidden Markov Model (HMM) is a statistical representation of patterns found in a data set. When using HMMs with proteins, the HMM is a statistical model of the patterns of the amino acids found in a multiple alignment of a set of proteins called the “seed”. Seed proteins are chosen based on sequence similarity to each other. Seed members can be chosen with different levels of relationship to each other. They can be members of a superfamily (ex. ABC transporter, ATP-binding proteins), they can all share the same exact specific function (ex. biotin synthase) or they could share another type of relationship of intermediate specificity (ex. subfamily, domain). New proteins can be scored against the model generated from the seed according to how closely the patterns of amino acids in the new proteins match those in the seed. There are two scores assigned to the HMM which allow annotators to judge how well any new protein scores to the model. Proteins scoring above the “trusted cutoff” score can be assumed to be part of the group defined by the seed. Proteins scoring below the “noise cutoff” score can be assumed to NOT be a part of the group. Proteins scoring between the trusted and noise cutoffs may be part of the group but may not. One of the important features of HMMs is that they are built from a multiple alignment of protein sequences, not a pairwise alignment. This is significant, since shared similarity between many proteins is much more likely to indicate shared functional relationship than sequence similarity between just two proteins. The usefulness of an HMM is directly related to the amount of care that is taken in chosing the seed members, building a good multiple alignment of the seed members, assessing the level of specificity of the model, and choosing the cutoff scores correctly. In order to properly assess what functional relevance an above-trusted scoring HMM match has to a query, one must carefully determine what the functional scope of the HMM is. If the HMM models proteins that all share the same function then it is likely possible to assign a specific function to high-scoring match proteins based on the HMM. If the HMM models proteins that have a wide variety of functions, then it will not be possible to assign a specific function to the query based on the HMM match, however, depending on the nature of the HMM in question, it may be possible to assign a more general (family or subfamily level) function. In order to determine the functional scope of an HMM, one must carefully read the documentation associated with the HMM. The annotator must also consider whether the function attributed to the proteins in the HMM makes sense for the query based on what is known about the organism in which the query protein resides and in light of any other information that might be available about the query protein. After carefully considering all of these issues the annotator makes an annotation.