Comparison among different programming systems for organizing protein families.

Caveat: Each system is actively upgraded, often with the motive to incorporate features derived from the others.  Hence,  a capability cited as lacking in some program in the comments below may appear in a later version.  Links to pages describing the individual programs should be followed to find the latest innovations in each system.
 

Psi-Blast

• Implementation:
• At NCBI, and other major bioinformatics institutes.  Interactive interface only.
• Local version at UTHSCSA. This is an implementation of NCBI's downloadable Standalone Blast Suite.  It provides some additional flexibility of use.  Interactive interface or batch processing.
• Downloadable versions for PCs are found at NCBI.  Downloading and installing a local copy of the Blast suite is easy.  Downloading and maintaining an up-to-date copy of the databases is not so easy.
• Characteristics:
• Always does submodel vs. subsequence scoring.  This means that alignments for differing domains with different sets of homologues will develop during the same iterative search.  However, you can not automatically constrain it to only add sequences that match across all domains.  You can do that manually from the interactive interface.
• Statistical evaluation of significant matches is empirically based.
• Statistics are automatically adjusted to database size, and the database can be restricted in various ways.  This can give greatly increased sensitivity if the database restriction is based on valid outside knowledge, or a horrible artifact if the database restriction is arbitrary.  The database size for computing E values can be overridden by a user entry.
• The ability to extend to highly divergent sequences depends on the availability of intermediate sequences with which to form the PSSM.
• Major advantages:
• Speed.  Because of the simple gap penalty form, Psi-Blast is fast enough to search the entire protein database on each iteration.  Hence it is an excellent choice for the front end of an analysis wherein the found sequences will be subjected to a more intensive algorithm.
• The NCBI interface gives a particularly good graphic representation for evaluating domain structure of a sequence.
• Will often find extensively divergent homologues of 15% or less identity.
• Major disadvantages:
• Simple gap function gives relatively poor alignments.  Alignments have to be worked over retrospectively by another system.  Output of aligned sequences is not in a standard format for input to other programs, hence porting the found information can generate a lot of fiddling.  The poor alignments involving divergent sequence may be somewhat offset by manually increasing the gap penalty while reducing the gap extension penalty.
• Validity of the statistical evaluation becomes increasingly questionable for increasing numbers of sequences in the PSSM.  The user has to use subjective criterion to decide when to cut off  the search, or it may branch into another family related only by chance similarity.
• Variations:
• Reverse Psi-Blast:
• An ad hoc reverse Psi-Blast control can be performed by starting a Psi-Blast search from a proposed distant homologue and seeing if it finds its way back to the initial key sequence.  If the reverse search finds lots of variously related sequences but fails to find the initial key, that would tend to reject a true relationship.  If the reverse search can not extend far because of a lack of intermediate sequence, then it provides no information about the proposed relationship.
• RPS-Blast.  A library of PSSM models describing various protein families is maintained at NCBI.  The CD search facility searches a sequence against the library of models using the Psi-Blast algorithm.  The same capability is represented locally by the RPS blast implementation.  PSSM models for this purpose are generally developed from alignments generated by programs better at placing gaps than is Psi-Blast.  Most of NCBI's models are derived from HMMER HMMs in Pfam.
• Starting a forward search from a previously developed PSSM (or alignment).
• NCBI's version is rigged to allow entry of a PSSM derived from a prior search (or otherwise), and continue expanding it through iterative searches of the database.
• The local version can accept an alignment to perform this function.  This is an excellent way to take a protein family alignment formed by some program better at placing gaps and then continue searching for homologues with the fast Psi-Blast algorithm.
• PSI-TBLASTN (local only).  The PSSM developed by Psi-Blast is used to search a nucleotide database translated in all six frames.  This can use the sensitivity of Psi-Blast to find genes in genomes missed by the original annotators.
• Secondary structure prediction.
• The Psi-Pred system, accessed either remotely, or locally, compiles the PSSM for a key sequence and uses that additional information to improve secondary structure prediction.

CLUSTAL W

Implementation:

• Remote at EBI.
• Local at UTHSCSA (local implementations have some additional flexibility).
• One can download copies for PCs from EBI.  The size of the alignment than can be dealt with is limited by the memory capacity of the machine. Clustal X is a version with a color graphics interface.

• Clustal W is not a database searching system.  It is used to align a set of sequences proposed to be homologues by another system.
• Gap penalties are more sophisticated than Psi-Blast, with decreased penalties where sequences in alignment already have a gap, and an automatic capacity to avoid gaps in hydrophobic runs.  However, they are still empirically adjusted, and the user may expect to do a lot of subjective fiddling if the alignment has very divergent sequences.
• Similar to Psi-Blast, Clustal W can accept a subset of the alignment prealigned by another system, and then carry on.
• Can accept a secondary structure map, and adjust gap penalties to avoid gapping within secondary structure elements.
• Can accept an arbitrary set of position specific gap penalties from the user.

• Often implemented together with an Neighbor Joining phylogenetic analysis package, including Bootstrap analysis.  This provide an easy interface to a means to make a statistically defensible statement about the tree relating the sequences.
• Implemented within Seaview multiple alignment editor, such that local regions of a larger alignment can be subjected to revision by clustalw.
• Default parameters widely accepted to produce an acceptable "objective" alignment on sequences >= 25% identity.
• May do better than HMMER or SAM at relating divergent but well populated families.
• ClustalW will allow preserving prior alignments of two or more subsets as the global alignment is formed.  The other methods will at most preserve the prior alignment of only one group.

• Is easily confused by long stretches of unalignable sequences within otherwise well related sequences.
• Often produces a kind of false objectivity, where the user has fiddled with the program parameters to achieve a subjectively pleasing result, rather than just manually editing the alignment.
• Has no statistical evaluatory properties.  It will produce an alignment whether the provided sequences are related or not.
• Is less good than HMMER or SAM at adding a single divergent sequence to a family.

• Pileup in the GCG system is related to ClustalW, but performs substantially less well on even moderately divergent sequences.
• T-Coffee is a related system, which is considered by some to perform slightly better (but more slowly) than ClustalW.

HMMER

• The degree of divergence to be considered belonging to a particular protein family is preconceived and a seed alignment is prepared by another method, typically ClustalW.
• The HMMER build program is used to make the HMM.
• The HMM is used to search the global protein database, and additional homologues are aligned with the original seed sequences.
• Cutoffs are recommended to exclude sequences that exceed the original expert's preconception about what should belong to the family.
• Pfam, and some other databases provide a library of such models, the original alignments created from the build process (called seed alignments), and the more global alignments.  There is also usually a facility to search a novel sequence against the library and suggest which families it may join.

Implementation.

• For searching a novel sequence against a premade library, the best place to do that is at the site hosting the library.  For Pfam, that is Sanger Center.  Since the HMMER library concept is tied to decisions made by "experts" during assembly of the family, you will only be able to retrieve the seed alignments representing the judgments of those experts from a separate database at the hosting site.
• Numbers of programming packages allow importation of the pfam library and will run the search against it.  But usually information about the seed alignments, is lost in these implementations.  Also Pfam updates monthly, so it is easy to be tricked into searching an out-of-date version.
• There is a local implementation at UTHSCSA that permits additional exploration of the families.  It should be used in conjunction with the remotes sites; not instead of them.
• The HMMER programs implemented within GCG are essentially interchangeable with the free standing ones, except for file format issues.  The Pfam database version internal to the GCG distribution suffers from falling out of date.

Characteristics.

• The program doesn't deal well with large separation between alignable regions.  Hence, the families analyzed end up corresponding to protein domains.
• The nature of the program requires separate models to be built to cover an intent to detect homology to a portion of a domain in a sequence versus the intent to evaluate homology across the entire domain.  Therefore, Pfam has two divisions to cover these two purposes.
• The model building process involves a mixture of prior expectations and the observed distribution of gaps and residues.  There is some flexibility for adjusting the prior expectations (called "priors" in the documentation).

Advantages and Disadvantages.

• Versus ClustalW as an alignment tool, HMMER provides a more objective placement of gaps.  However, ClustalW provides the user with more capability to introduce secondary structural considerations.
• HMMER can be used as a search tool for additional homologues, whereas ClustalW can not.
• HMMER alignments can be converted to PSSMs and hence used by Psi-Blast. This gives a much faster database search method, with the poor alignment performance of Psi-Blast somewhat offset by the intelligent placement of gaps by HMMER in the original alignment.  This is the basis of NCBI's RPS-Blast system (aka CD search).
• In terms of the ease for distributing models of new families over the web that others might use: PSSMs or the associated alignments are the easiest for most others to grab and use, since if they are formatted right, they can go straight into NCBI's web page for further searching.  Many packages have HMMER implementations, so many could user HMMER HMM models directly.  SAM implementations are relatively rare, so SAM models would usually be converted to HMMER models for distribution.
• Compared to SAM, HMMER could be used cyclically to expand to larger and more even divergent families, but it is really not designed to do so.  SAM is.  Compared to SAM, HMMER's selection of priors is limited, and its interaction with secondary structure is rudimentary.
• HMMER is relatively dependent on sequences being divided up into domains in advance of model building.
• HMMER results are difficult to divorce from "expert" opinion, since both defining domains and defining the degree of divergence to be included in a protein family are based on preconceptions rather than computation.

SAM (Sequence Alignment and Modeling Tool).

Implementation

• The SAM web site at UCSC will provide assembly of a family from a single sequence, roughly like conducting four cycles of Psi-Blast.  It will provide consensus secondary structure prediction, and also scores reflecting matching to a HMM family representing different protein folds.
• The local implementation of SAM provides the target99 script which can be used to expand a family through an additional arbitrary number of cycles, define it based on arbitrary seed alignments, and perform other tricks to achieve high sensitivity in incorporating divergent members.  Secondary structure prediction can be incorporated, but not in the automatic sense it is at the UCSC site.  The local implementation does not include the fold library.

• The SAM search process from a single sequence does four cycles of searching and HMM building.  As more divergent sequences are included, it automatically switches the prior information about how gaps should be defined appropriately.
• It's final alignment is made with some automatic sensitivity to avoiding gaps in secondary structure elements.
• If there is 3D information for the sequence, this information can be strongly incorporated into the process.
• SAM has a facility to mask portions of the starting sequence, thus focusing the search and alignment process on one or more known conserved domains.
• SAM has some facility to incorporate knowledge about secondary structure in the searched sequences (as opposed to the key sequences), but this is not well developed in the package at this time.  It would require conducting a prior secondary structure prediction over all sequences in the searched database.

Advantages/ disadvantages

• Compared to HMMER, SAM can more easily incorporate multiple domains within the same family alignment.  However, the developers still advise pairing down to individual domains if possible.
• Compared to HMMER, SAM can more easily extend the family to more divergent homologues.  However, like Psi-Blast, the user has to be more on guard against jumping over into an unrelated family.
• Compared to HMMER, SAM is much better set up to start from a prior alignment and carry on extending the family.  It has flexibility about how or whether to adjust the prior alignment in the process.
• When adding sequences a prior alignment, SAM is capable of using that information to extend alignment beyond the domains that were originally aligned.  Psi-Blast does this to a limited extent. HMMER is bound to its domain models.
• There seems to be little option for distributing a SAM HMM other than converting it to a HMMER HMM, which is supported in the package.  The HMMER HMM should retain the degree of divergence recognized, and the structural sensitivity, but may not be able to handle multiple domains in the same HMM.
• SAM is very slow at searching compared to HMMER (which is already slow compared to Psi-Blast), to the point that SAM can not realistically search the entire nr database.  So it uses blast as a prefilter to pick out the set of sequences within E<300 from each sequence in the seed alignment.  The burden of Blast searching can become prohibitive if the seed alignment is very large to start with.  And further, a given divergent target might never get searched for matching to the HMM.  A reasonable work-around is to first run many iterations of Psi-Blast at some low stringency.  Then compile a separate database using fastacmd of all of those hits.  Then make SAM search that database without the Blast prefilter.