Profile HMM

A profile hidden Markov model (profile HMM) is a probabilistic representation of a multiple sequence alignment that captures both the sequence conservation and the positional variation — including insertions and deletions — across a protein family or domain. Developed by Sean Eddy and others in the 1990s, profile HMMs generalize the position-specific scoring matrix (PSSM) by modeling alignment as a path through a probabilistic state machine rather than as a column-by-column scoring table.

The standard profile HMM architecture, introduced in the SAM and HMMER software packages, consists of three states per alignment column: match states (emitting residues with position-specific probabilities), insertion states (emitting residues not present in the consensus), and deletion states (silent transitions that skip a column). This architecture allows the model to represent both conserved motifs and variable-length regions, making it far more expressive than a fixed-length PSSM.

Profile HMMs are the dominant method for remote homology detection in database search: HMMER and HHblits use profile-profile comparison to detect evolutionary relationships that have diverged beyond the reach of sequence-sequence methods like BLAST. The key insight is that a protein family's statistical signature — the pattern of conservation, the tolerated substitutions, the gap probabilities — persists long after individual sequences have become unrecognizably divergent.

The profile HMM is computational biology's most successful formal model of evolutionary conservation. It treats a protein family not as a set of sequences but as a probabilistic object — a distribution over sequences from which the observed members are draws. This shift from set to distribution is not merely mathematical sophistication; it is a conceptual revolution that enables inference about sequences that have never been observed. But the model makes strong assumptions — that evolution operates independently at each position, that gap probabilities are position-specific but not context-dependent, that the Markov property holds — and these assumptions are biologically false in ways that matter. The profile HMM is a beautiful approximation that works better than it should, and the field has not yet figured out why.