BLAST vs PSI-BLAST vs HMMER: Finding Remote Homologs by Sequence

Start with BLAST, and move to a profile method when BLAST stops returning significant hits. Pairwise protein BLAST (Altschul et al., J Mol Biol 1990) settles most homology questions: protein alignments routinely detect common ancestry more than 2.5 billion years back, and an E-value below 0.001 can reliably be used to infer homology (Pearson, Curr Protoc Bioinformatics 2013). It runs out in the twilight zone of roughly 20–35% identity. There, profile methods earn their cost: iterated search with PSI-BLAST (Altschul et al., Nucleic Acids Res 1997) and profile hidden Markov models, the approach HMMER implements (Eddy, PLoS Comput Biol 2011), detected three times as many relationships below 30% identity as pairwise search on a structure-graded benchmark (Park et al., J Mol Biol 1998). The price is not speed any more but specificity: a profile can learn the wrong family, and every hit it then reports looks significant.

The scope of this page: the .bloom* platform runs protein BLAST as a job, and the section on running it describes exactly what that job does. PSI-BLAST and HMMER are covered here as methods: how they score, what they detect, and how they fail.

At a Glance#

The remote-homolog row comes from one benchmark: sequences of known structure derived from SCOP with 40% identity or less between any two, at a false-positive rate of about 1 in 50,000 (Park et al. 1998). It compares the methods under identical conditions. It is not a current figure for any specific release, and the profile HMM it tested is SAM-T98 rather than HMMER.

Why Pairwise Search Runs Out#

A homology search does not measure ancestry directly. It measures excess similarity — an alignment score higher than chance would produce — and infers common ancestry from it (Pearson 2013). Two consequences follow, and both shape the choice between methods.

Percent identity is the wrong yardstick. The common rule of thumb that two proteins are homologous above 30% identity is, in Pearson's words, too conservative: statistically significant protein homologs can share less than 20% identity. For average-length proteins a bit score of 50 is almost always significant, and the E-value or the bit score, not identity, is the number to read. Identity does become meaningful once homology is established, as a rough proxy for evolutionary distance.

Below about 30% identity, single alignments stop being trustworthy. A survey of more than a million alignments between proteins of known structure put numbers on the twilight zone (Rost, Protein Eng 1999): above a cut-off of roughly 30% identity, 90% of aligned pairs were homologous; below 25%, fewer than 10% were. Alignment length matters too — 10 similar residues in an alignment of 16 says nothing about shared structure. Real homologs still exist down there. A pairwise score is simply too noisy to separate them from chance.

The database is part of the statistics. An E-value scales with the number of sequences searched. Pearson's worked example: the same alignment score is 100-fold less significant in a 10,000,000-entry database than in a 100,000-entry one. A larger database is not automatically a better search — it can push a real but weak hit out of significance. And protein searches beat DNA searches outright: protein and translated searches are 5–10-fold more sensitive than DNA:DNA comparison, which rarely detects homology beyond 200–400 million years of divergence.

BLAST: Fast, Well-Calibrated, and Pairwise#

BLAST seeds alignments from short word matches, extends them, and scores the result against a fixed substitution matrix. Its statistics come from the maximal segment pair score, and at publication it was an order of magnitude faster than existing sequence comparison tools of comparable sensitivity (Altschul et al. 1990). Gapped BLAST then roughly tripled the speed of the original while improving sensitivity to weak similarities (Altschul et al. 1997).

The strength is calibration. Protein BLAST E-values are accurate enough that a hit below 0.001 can be used to infer homology, and a hit with an E-value near 1 to a protein with a different domain architecture is what unrelated sequences should look like (Pearson 2013). The weakness is that every position in the query is scored the same way. A conserved catalytic residue and a variable surface loop get the same matrix, so a distant family member that keeps only the conserved positions can score no better than chance.

PSI-BLAST: Let the Hits Teach the Matrix#

PSI-BLAST keeps BLAST's engine and replaces the fixed matrix. After a first search it combines the statistically significant alignments into a position-specific score matrix (PSSM) and searches the database again with it, iterating as new hits refine the matrix. Each iteration runs at about the speed of gapped BLAST, and the method is in many cases much more sensitive to weak but biologically relevant similarity (Altschul et al. 1997).

One refinement matters in practice. Composition-based statistics — a scoring system tuned to each database sequence's amino acid composition — account for most of an improvement in ROC score from 0.758 to 0.895 on 103 expert-annotated yeast queries (Schäffer et al., Nucleic Acids Res 2001). The other thing to know is that the iteration that gives PSI-BLAST its reach is also its main source of error, covered under limitations below.

HMMER: Profile Hidden Markov Models#

A profile hidden Markov model turns a multiple sequence alignment into a position-specific scoring system suited to searching databases for remotely homologous sequences (Eddy, Bioinformatics 1998). HMMER implements profile HMMs and takes either a single sequence or a multiple sequence alignment as the query. Three of its programs map onto the jobs in this comparison (descriptions from the HMMER manual pages, EddyRivasLab/hmmer):

• phmmer searches protein sequences against a protein sequence database — the profile-HMM counterpart of a BLAST search.
• jackhmmer searches iteratively — the counterpart of PSI-BLAST. Its paper reports that it detects 14% more remote homologs than SAM's iterative T2K method on its benchmark, and significantly more than PSI-BLAST (Johnson, Eddy & Portugaly, BMC Bioinformatics 2010).
• hmmsearch searches a profile HMM — built from your own alignment, or taken from a domain library — against a sequence database.

For years the objection to profile HMMs was cost. HMMER3 removed it with a filter pipeline — a fast "multiple segment Viterbi" pass that hands high-scoring hits to the full Forward/Backward algorithm — which sacrifices negligible sensitivity, is 100- to 1000-fold faster than HMMER2, and is about as fast as BLAST for protein searches (Eddy 2011). Hosted versions exist too: EMBL-EBI runs a free HMMER web server with an API (Potter et al., Nucleic Acids Res 2018).

Head to Head: How Much More Do Profiles Detect?#

The cleanest comparison is still the one that graded all three families against relationships established from structure and function rather than from sequence (Park et al. 1998). On sequences of known structure sharing 40% identity or less, at nine false positives out of 432,680 possible pairs:

Below 30% identity the gap widens: the profile methods detected three times as many relationships as the pairwise ones. Two readings matter. Profiles roughly double what a pairwise search can see in this regime. And even the best method reached about a third of the relationships the structures reveal — which is the honest case for the next section.

When Sequence Search Runs Out: Structure Search#

When profile searches stop returning significant hits, the next lever is to compare structures instead of sequences — provided you have a structure for the query, experimental or predicted. Foldseek makes that practical at database scale by describing each residue's tertiary interactions as a letter in a structural alphabet and searching those strings, cutting computation times by four to five orders of magnitude while keeping 86% and 88% of the sensitivity of DALI and TM-align (van Kempen et al., Nat Biotechnol 2023; code at steineggerlab/foldseek). A structure-only match is evidence of shared fold, which often but not always means shared ancestry, and it says less about shared function. If your query has no structure yet, the ESMFold vs AlphaFold comparison covers getting one.

How to Choose#

Plain BLAST is enough when#

• You are annotating a sequence and expect close relatives in a curated database. Most queries have them, and a significant BLAST hit is a reliable homology call.
• The hits you get are well above the noise. If the top hits score far above 50 bits with E-values far below 0.001, a profile method will mostly add more members of a family you already have.
• You need orthologs or the same protein in other species. That is high-identity territory, where pairwise statistics are at their best.

Move to PSI-BLAST when#

• BLAST returns only a few weak hits in the twilight zone and you want to extend the family one or two iterations further.
• You want a sequence profile — a PSSM — as an input to something else, and BLAST-family statistics suit the rest of your pipeline.

Move to HMMER when#

• You are working with protein domains and want to score against a family model rather than one representative sequence: build a profile from your own alignment, or take one from a domain library, and run hmmsearch.
• You need the most sensitive sequence-only search and are going to iterate anyway; jackhmmer is reported to detect more remote homologs than PSI-BLAST.
• You were avoiding HMMs on speed grounds. HMMER3 is reported to run about as fast as BLAST for protein searches.

Move to structure when#

• Every sequence method has run dry and you have a structure, or can predict one. Read the result as a fold-level relationship first.

Limitations to Keep in View#

• Iteration can corrupt a profile. PSI-BLAST's characteristic error is homologous over-extension: an alignment begins in a homologous region, runs past it into a neighbouring unrelated domain, and that sequence enters the PSSM, which then scores non-homologous proteins as significant. In searches against complete protein sequences, 5–9% of alignments at iteration 5 were non-homologous, and over-extension often began in a partial domain; not extending previously found sequences improved specificity 4–8-fold with little loss in sensitivity (Gonzalez & Pearson, Nucleic Acids Res 2010). The mechanism is a property of building a profile from your own hits, so treat any iterative search — jackhmmer included — with the same caution: inspect what joins the profile at each round, and stop early on multidomain queries.
• The benchmarks are old. The cleanest three-way comparison dates from 1998 and tested SAM-T98, not HMMER3. Composition-based statistics changed PSI-BLAST after it, and HMMER3 changed the speed argument entirely. Use the relative ordering, not the percentages, as a guide to current releases.
• E-values depend on the database you chose. The same score is more significant in a smaller database, so an E-value cannot be compared across searches of different databases without accounting for their size.
• Homology is not function. Homologous sequences have similar structures and often, but not always, similar functions (Pearson 2013). A remote homolog tells you where to look for a function, not what the function is.
• Identity cut-offs fail in both directions. Below the twilight zone they miss real homologs; on short alignments a high identity can still occur by chance (Rost 1999).

How to Run It on .bloom*#

.bloom* runs protein BLAST as a job: blastp against UniProtKB/Swiss-Prot. You supply one protein query — a FASTA record, raw residues (10 to 10,000), or a UniProt accession — and the job returns a ranked hit list with E-values, percent identity and query coverage, plus the pairwise alignments, which you leave with as a hits JSON and an alignment text file. Details are on the BLAST page.

Why Swiss-Prot is the database to search. Two reasons, both from the statistics above. Swiss-Prot is a curated subset of UniProt, and its annotation of modified residues, active sites, variants and mutation studies supports more accurate functional inference — for a well-populated family, the accuracy of the annotation often matters more than finding the closest homolog (Pearson 2013). And a smaller, curated database keeps E-values sharp: the same alignment score is more significant against it than against a comprehensive archive many times its size.

Chaining. BLAST is a search step that feeds the rest of a structure workflow. Ask Ivy to BLAST a sequence and fold the top hit, and Ivy runs the search, then chains the hit into ESMFold — the folding trade-offs are in the ESMFold vs AlphaFold comparison. For matching other goals to the right job, see which tool?.

Access. The BLAST job runs on the free tier and needs a signed-in account; plans and tiers lists what each plan includes.

Key Takeaways#

• BLAST first. A significant protein BLAST hit (E-value below 0.001, or about 50 bits for an average-length protein) is a reliable homology call.
• Read E-values and bits, not identity. Real homologs can sit below 20% identity, and the 30% rule misses them.
• Profiles roughly double remote-homolog detection among sequences sharing 40% identity or less, and triple it below 30%, on a structure-graded benchmark.
• HMMER is the more sensitive profile method and no longer the slow one.
• Iteration is where false positives come from. Watch what enters the profile, especially on multidomain queries.
• When sequence runs out, compare structures — and read the match as a fold, not a function.

Start with the search that settles most questions: run blastp against Swiss-Prot on your own sequence, with the hits and alignments kept together, on the .bloom* platform.

References#

1. Altschul, S.F., Gish, W., Miller, W., Myers, E.W. & Lipman, D.J. "Basic local alignment search tool." Journal of Molecular Biology 215, 403–410 (1990). doi:10.1016/S0022-2836(05)80360-2
2. Altschul, S.F., Madden, T.L., Schäffer, A.A. et al. "Gapped BLAST and PSI-BLAST: a new generation of protein database search programs." Nucleic Acids Research 25, 3389–3402 (1997). doi:10.1093/nar/25.17.3389
3. Schäffer, A.A., Aravind, L., Madden, T.L. et al. "Improving the accuracy of PSI-BLAST protein database searches with composition-based statistics and other refinements." Nucleic Acids Research 29, 2994–3005 (2001). doi:10.1093/nar/29.14.2994
4. Eddy, S.R. "Profile hidden Markov models." Bioinformatics 14, 755–763 (1998). doi:10.1093/bioinformatics/14.9.755
5. Eddy, S.R. "Accelerated Profile HMM Searches." PLoS Computational Biology 7, e1002195 (2011). doi:10.1371/journal.pcbi.1002195 · code: EddyRivasLab/hmmer
6. Johnson, L.S., Eddy, S.R. & Portugaly, E. "Hidden Markov model speed heuristic and iterative HMM search procedure." BMC Bioinformatics 11, 431 (2010). doi:10.1186/1471-2105-11-431
7. Potter, S.C., Luciani, A., Eddy, S.R. et al. "HMMER web server: 2018 update." Nucleic Acids Research 46, W200–W204 (2018). doi:10.1093/nar/gky448
8. Park, J., Karplus, K., Barrett, C. et al. "Sequence comparisons using multiple sequences detect three times as many remote homologues as pairwise methods." Journal of Molecular Biology 284, 1201–1210 (1998). doi:10.1006/jmbi.1998.2221
9. Rost, B. "Twilight zone of protein sequence alignments." Protein Engineering 12, 85–94 (1999). doi:10.1093/protein/12.2.85
10. Pearson, W.R. "An Introduction to Sequence Similarity ('Homology') Searching." Current Protocols in Bioinformatics 42 (2013). doi:10.1002/0471250953.bi0301s42
11. Gonzalez, M.W. & Pearson, W.R. "Homologous over-extension: a challenge for iterative similarity searches." Nucleic Acids Research 38, 2177–2189 (2010). doi:10.1093/nar/gkp1219
12. van Kempen, M., Kim, S.S., Tumescheit, C. et al. "Fast and accurate protein structure search with Foldseek." Nature Biotechnology 42, 243–246 (2024; published online 2023). doi:10.1038/s41587-023-01773-0 · code: steineggerlab/foldseek