1. Pfam: multiple sequence alignments and HMM-profiles of protein domains
Xianhui Li

2. Outline
What is Pfam?
What is a Hidden Markove model (the methodology underlying Pfam)?
How to use Pfam and sample output

3. pfam
Pfam is a database of multiple alignments of protein domains or conserved protein regions.
The alignments represent some evolutionary conserved structure which has implications for the protein's function.
Profile hidden Markov models (profile HMMs) built from the Pfam alignments can be very useful for automatically recognizing that a new protein belongs to an existing protein family, even if the homology is weak.

4. Overview of Pfam Database
Pfam A contains curated families each with an associated profile HMM that can be used for alignment and database searching
Annotation --contains several compulsory fields
Seed alignment– a manually verified multiple alignment of a representative set of sequences
HMM –profile— turned a multiple sequence alignment into a position-specific scoring system.
Full alignment– generated automatically from the seed HMM-profile by searching Swisssprot for all detectable members and aligning them to the HMM profile
Pfam B are clustered automatically, allowing Pfam to be comprehensive

5. Pfam Sequence Database Coverage
Figure 2. Pfam 2.0 contains domains from nearly half of all Swissprot 34 proteins.
The automatic clusters in Pfam-B 2.0 contain domains from 33% of the Swissprot
proteins that do not contain Pfam domains. When counting residue-by-residue,
roughly a third of Swissprot is covered by Pfam and Pfam-B each. Pfam-B does
not include proteins known to be fragments or segments shorter than 30 residues;
the figures for unique sequences are therefore overestimated.
residue
Sequence
Data shown is from Pfam v2.0 as of 1998 with 527 families. Current version is Pfam 12.0 (January 2004) contains alignments and models for 7316 protein families, based on the Swissprot 42.5 and SP-TrEMBL 25.6 protein sequence databases

6. Markov Model
Simplest example: Each state emits (or, equivalently, recognizes) a particular element with probability 1.
Begin
Emit 1
Emit 2
Emit 4
Emit 3
End
Example sequences:

7. Probabilistic Emission
If we let the states define a set of emission probabilities for elements, we can no longer be sure which state we are in given a particular element of a sequence BCCD or BCCD ?
0.5
0.25
0.75
0.9
0.1
0.2
0.8
1.0
Begin
A (0.8) B(0.2)
B (0.7) C(0.3)
C (0.1) D (0.9)
C (0.6) A(0.4)
End

8. Hidden Markov Models (HMM)
Emission uncertainty means the sequence doesn't identify a unique path. The states are “hidden”
Probability of a sequence is sum of all paths that can produce it:
0.5
0.25
0.75
0.9
0.1
0.2
0.8
1.0
Begin
A (0.8) B(0.2)
B (0.7) C(0.3)
C (0.1) D (0.9)
C (0.6) A(0.4)
End
p(bccd) = 0.5 * 0.2 * 0.1 * 0.3 * 0.75 * 0.6 * 0.8 * * 0.7 * 0.75 * 0.6 * 0.2 * 0.6 * 0.8 * = =

9. HMMs for homology
Homology model: ancestral residue (match) states, insertion states, deletion states.
start
insert
match
delete
end

10. Profile HMM

11. Searching Pfam
Web site: provide users the ability to search query protein sequences against one, all, or a few PfamHMM.
_http://
_http://genome.wustl.edu/Pfam --
. Software: Users can use Pfam HMM-profile to search locally using the freely available HMMERsoftware package at:

12. Sample Pfam Query Results
Score
Query Start
Query End
Hmm Start
Hmm end
Pfam Family
Description
97.57
104
153 1 50
DAG_PE bind
Phorbol Estser/ diacylglycerol binding domain
92.44
169
216
DAG_PE-bind
137.88
240
328 92 C2
C2 domain
276.16
413
674
247
pkinase
Eukaryotic protein kinase domain
84.44
675
741 69
pkinase_C
Protein kinase C terminal domain
70.99
807
857 17

13. Acknowledgements
Some slides adapted from lectures by Larry Hunter at University of Colorado Health Sciences Center
Altmann Lab for critical comments