HH-suite User Guide

Table of Contents

• Introduction
• Installation of the HHsuite and its databases
  • Supported platforms and hardware requirements
    • Support for SSE2 instruction set
  • Installation from source code
    • MPI Support
  • HH-suite databases
  • Running HHblits efficiently on a computer cluster
• Brief tutorial to HHsuite tools
  • Overview of programs
  • Searching databases of HMMs using HHsearch and HHblits
  • Generating a multiple sequence alignment using HHblits
  • Example: Comparative protein structure modeling using HHblits and MODELLER
  • Visually checking an MSA for corrupted regions
  • Building customized databases
  • Example: Building a database from the PDB
  • Modifying or extending existing databases
• Frequently asked questions
  • How do I report a bug?
  • What are HMM-HMM comparisons and why are they so powerful?
  • When can the HH-suite be useful for me?
  • What does homology mean and why is it important?
  • How can I verify if a database match is homologous?
  • What does the maximum accuracy alignment algorithm do?
  • How is the MSA diversity Neff calculated?
  • Do HHsearch and HHblits work fine with multi-domain sequences?
  • How do I best build HMMs for integral membrane proteins?
  • How do I reconcile overlapping and conflicting domain predictions?
  • How can I build a phylogenetic tree for HMMs?
  • Can I weight some residues more than others during the HMM-HMM alignment?
  • Don’t I need to calibrate my query or database HMM?
  • Should I use the -global option to build MSAs if I am intersted in global alignments to these database HMMs?
  • How many iterations of hhblits should I use to search my customized genome database?
  • How can I retrieve database A3M or HHM files from my HH-suite database?
  • How can I build my own UniProt database for HHblits?
  • Will you offer an HHsuite database that includes environmental sequences?
  • How do I read the sequence/HMM headers in the pdb70 databases?
  • Why do I get different results when reversing the roles of query and template?
  • Why do I sometimes get the same database hit twice with different probabilities?
  • Why do I get different results using HHblits and HHsearch when I search with the same query through the same database?
  • Why do I get different results when I perform 2 and 1+1 iterations with HHblits?
  • Why do I get a segmentation fault when calling hhblits on our new machine?
  • I obtain different results for the same query sequence on the web server and on my local version using HHblits and HHsearch.
  • I find an alignment of my query and a template with the exact same sequence, but they are aligned incorrectly.
• HHsearch/HHblits output: hit list and pairwise alignments
  • Summary hit list
  • HMM-HMM pairwise alignments
• File formats
  • Multiple sequence alignment formats
    • Example for aligned FASTA format:
    • The same alignment in A2M format looks like this:
    • The same alignment in A3M:
    • Secondary structure information in A3M/A2M or FASTA MSAs for HHsearch/HHblits
    • Name lines in alignments
  • HHsearch/HHblits model format (hhm-format)
• Summary of command-line parameters
  • hhblits – HMM-HMM-based lighting-fast iterative sequence search
  • hhsearch – search a database of HMMs with a query MSA or HMM
  • hhfilter – filter an MSA
  • hhmake – build an HMM from an input alignment
  • hhalign – align a query MSA/HMM to a template MSA/HMM
  • reformat.pl – reformat one or many alignments
  • addss.pl – add predicted secondary structure to an MSA or HMM
  • hhmakemodel.pl and hhmakemodel.py – generate MSAs or coarse 3D models from HHsearch results file
  • hhsuitedb.py – Build an HHsuite database
  • splitfasta.pl – Split multi-sequence FASTA file into single-sequence files
  • renumberpdb.pl – Renumber indices in PDB file to match input sequence indices
  • cif2fasta.py – Create a fasta file from cif files
  • pdbfilter.py – Filter sequences from the PDB (requires MMseqs2)
• License
• Contributors and Contact
• Acknowledgments

Summary

The HH-suite is an open-source software package for sensitive protein sequence searching based on the pairwise alignment of hidden Markov models (HMMs). It contains HHsearch and HHblits among other programs and utilities. HHsearch takes as input a multiple sequence alignment (MSA) or profile HMM and searches a database of HMMs (e.g. PDB, Pfam, or InterPro) for homologous proteins. HHsearch is often used for protein structure prediction to detect homologous templates and to build highly accurate query-template pairwise alignments for homology modeling. In the CASP9 competition (2010), a fully automated version of HHpred based on HHsearch and HHblits was ranked best out of 81 servers in template-based structure prediction. HHblits can build high-quality MSAs starting from single sequences or from MSAs. It transforms these into a query HMM and iteratively searches through Uniclust30 database by adding significantly similar sequences from the previous search to the updated query HMM for the next search iteration. Compared to PSI-BLAST, HHblits is faster, up to twice as sensitive and produces more accurate alignments. HHblits uses the same HMM-HMM alignment algorithms as HHsearch, but it employs a fast prefilter that reduces the number of database HMMs for which to perform the slow HMM-HMM comparison from tens of millions to a few thousands.

References:

• Steinegger M, Meier M, Mirdita M, Vöhringer H, Haunsberger S J, and Söding J (2019) HH-suite3 for fast remote homology detection and deep protein annotation, BMC Bioinformatics, 473. doi: 10.1186/s12859-019-3019-7

• Steinegger M, Meier M, Mirdita M, Vöhringer H, Haunsberger S J, and Söding J (2019) HH-suite3 for fast remote homology detection and deep protein annotation, bioRxiv, 560029. doi: 10.1101/560029

• Mirdita M, von den Driesch L, Galiez C, Martin M J, Söding J, and Steinegger M (2016) Uniclust databases of clustered and deeply annotated protein sequences and alignments, Nucleic Acids Res., 45, D170–D176. doi: 10.1093/nar/gkw1081

• Remmert M, Biegert A, Hauser A, and Söding J (2011) HHblits: Lightning-fast iterative protein sequence searching by HMM-HMM alignment. Nat. Methods 9, 173-175. doi: 10.1038/nmeth.1818

• Angermüller C, Biegert A, Söding J (2014) Discriminative modelling of context-specific amino acid substitution probabilities. Bioinformatics 28 (24), 3240-3247. doi: 10.1093/bioinformatics/bts622

• Söding J (2005) Protein homology detection by HMM-HMM comparison. Bioinformatics 21, 951-960. doi: 10.1093/bioinformatics/bti125

Introduction

The HH-suite is an open-source software package for highly sensitive sequence searching and sequence alignment. Its two most important programs are HHsearch and HHblits. Both are based on the pairwise comparison of profile hidden Markov models (HMMs).

Profile HMMs are a concise representation of multiple sequence alignments (MSAs). Like sequence profiles, they contain for each position in the master sequence the probabilities to observe each of the 20 amino acids in homologous proteins. The amino acid distributions for each column are extrapolated from the homologous sequences in the MSA by adding pseudocounts to the amino acid counts observed in the MSA. Unlike sequence profiles, profile HMMs also contain position-specific gap penalties. More precisely, they contain for each position in the master sequence the probability to observe an insertion or a deletion after that position (the log of which corresponds to gap-open penalties) and the probabilities to extend the insertion or deletion (the log of which corresponds to gap-extend penalties). A profile HMM is thus much better suited than a single sequence to find homologous sequences and calculate accurate alignments. By representing both the query sequence and the database sequences by profile HMMs, HHsearch and HHblits are more sensitive for detecting and aligning remotely homologous proteins than methods based on pairwise sequence comparison or profile-sequence comparison.

HHblits can build high-quality multiple sequence alignments (MSAs) starting from a single sequence or from an MSA. Compared to PSI-BLAST, HHblits is faster, finds up to two times more homologous proteins and produces more accurate alignments. It uses an iterative search strategy, adding sequences from significantly similar database HMMs from a previous search iteration to the query HMM for the next search. Because HHblits is based on the pairwise alignment of profile HMMs, it needs its own type of databases that contain multiple sequence alignments and the corresponding profile HMMs instead of single sequences. The HHsuite database Uniclust30 is generated regularly by clustering the UniProt KB into groups of similar sequences alignable over at least 80% of their length and down to ∼30% pairwise sequence identity. This database can be downloaded together with HHblits. HHblits uses the HMM-HMM alignment algorithms in HHsearch, but it employs a fast prefilter that reduces the number of database HMMs for which to perform the slow HMM-HMM comparison from tens of millions to a few thousands. At the same time, the prefilter is sensitive enough to reduce the sensitivity of HHblits only marginally in comparison to HHsearch.

By generating highly accurate and diverse MSAs, HHblits can improve almost all downstream sequence analysis methods, such as the prediction of secondary and tertiary structure, of membrane helices, functionally conserved residues, binding pockets, protein interaction interfaces, or short linear motifs. The accuracy of all these methods depends critically on the accuracy and the diversity of the underlying MSAs, as too few or too similar sequences do not add significant information for the predictions. As an example, running the popular PSIPRED secondary structure prediction program on MSAs generated by HHblits instead of PSI-BLAST improved the accuracy of PSIPRED significantly even without retraining PSIPRED on the HHblits alignments.

HHsearch takes as input an MSA (e.g. built by HHblits) or a profile HMM and searches a database of HMMs for homologous proteins. The PDB70 database, for instance, consists of profile HMMs for a set of representative sequences from the PDB database; the scop70 database has profile HMMs for representative domain sequences from the SCOP database of structural domains; the Pfam domain database is a large collection of curated MSAs and profile HMMs for conserved, functionally annotated domains. HHsearch is often used to predict the domain architectures and the functions of domains in proteins by finding similarities to domains in the pdb70, Pfam or other databases.

In addition to the command line package described here, the interactive web server runs HHsearch and HHblits. It offers extended functionality, such as Jalview applets for checking query and template alignments, histogram views of alignments, and building 3D models with MODELLER.

In the CASP9 competition (Critical Assessment of Techniques for Protein Structure Prediction) in 2010, a fully automated version of HHpred based on HHsearch and HHblits was ranked best out of the 81 servers in template-based structure prediction, the category most relevant for biological applications, while having an average response time of minutes instead of days like most other servers .

Other popular programs for sensitive, iterative protein sequence searching are PSI-BLAST and HMMER. Since they are based on profile-to-sequence and HMM-to-sequence comparison, respectively, they have the advantage over HHblits and HHsearch of being able to search raw sequence databases.

Installation of the HHsuite and its databases

The HH-suite source code for most Linux distributions and macOS, and its utility scripts written in Perl and Python can be found at:

https://github.com/soedinglab/hh-suite

In addition you have to install the Protein Data Bank parser for Python3 which can be found at:

https://github.com/soedinglab/pdbx

Databases can be downloaded at:

http://wwwuser.gwdg.de/~compbiol/data/hhsuite/databases/hhsuite_dbs/

Supported platforms and hardware requirements

HH-suite has been extensively tested on 64-bit Linux. We have done limited testing under BSD and macOS. HH-suite typically requires at least 4GB of main memory. Memory requirements can be estimated as follows:

Memory req.
  = query_length * max_db_seq_length * num_threads * (4 (SSE2) or 8 (AVX2)) byte
    + query_length * max_db_seq_length * num_threads * 8 byte
    + 1 GB

Here, num_threads is the number of threads specified with option -cpu <int>.

Due to the high number of random file accesses by HHblits, it is important for good efficiency that the databases used by HHblits are placed in RAM disk or on a local solid state drive (SSD). See section Running HHblits efficiently on a computer cluster for details.

Support for SSE2 instruction set

HHblits requires a CPU supporting at least the SSE2 (Streaming SIMD Extensions 2) instruction set. Every 64-bit CPU also supports SSE2. By default, the HH-suite binaries are compiled using the most recent supported instruction set on the computer used for compiling.

However, you may limit the used instruction set to SSE2 with the cmake compile flag -DHAVE_SSE2=1. The precompiled standard binaries are compiled with this flag.

Installation from source code

mkdir -p ~/programs/hh-suite && cd ~/programs/hh-suite
git clone https://github.com/soedinglab/hh-suite.git .
mkdir build && cd build
cmake ..
make

Set ${INSTALL_BASE_DIR} to the absolute path of the base directory where you want install HH-suite. For example, to install into /usr/local/hhsuite:

INSTALL_BASE_DIR=/usr/local/hhsuite
cmake -DCMAKE_INSTALL_PREFIX="$INSTALL_BASE_DIR" ..
make install

The HH-suite binaries will then be placed in /usr/local/hhsuite/bin.

Some old HH-suite Python and Perl scripts will try to read the HHLIB environment variable to locate HH-suite binaries and some required data files.

Set the HHLIB environment variable to ${INSTALL_BASE_DIR}, e.g, for bash use:

export HHLIB="${INSTALL_BASE_DIR}"

Put the location of your HH-suite binaries and scripts into your PATH:

export PATH="$PATH:${INSTALL_BASE_DIR}/bin:${INSTALL_BASE_DIR}/scripts"

To avoid typing these commands every time you open a new shell, you may add the following lines to the .bashrc or equivalent file in your home directory that is executed every time a shell is started:

export HHLIB="/usr/local/hhsuite"
export PATH="$PATH:$HHLIB/bin:$HHLIB/scripts"

MPI support

Since MPI is usually specific to the local cluster installation, the _mpi suffixed binaries have to be compiled from source. During the cmake invocation the build system will automatically check if MPI development files can be located and if found produce these binaries.

Generally you should refer to your local cluster admin team to find out how to load MPI development files for your local system. If you only want to use the _mpi suffixed binaries on a single node you can also try installing the development files as follows:

On Ubuntu/Debian run:

apt install mpi-default-bin mpi-default-dev

On macOS run:

brew install openmpi

HH-suite databases

The following HH-suite databases, which can be searched by HHblits and HHsearch, can be downloaded at:

http://wwwuser.gwdg.de/~compbiol/data/hhsuite/databases/hhsuite_dbs

• Uniclust30: Based on UniProt KB, clustered to 30% seq. identity :octocat:
• PDB70: Representatives from PDB (70% max. sequence identity), updated weekly :octocat:
• scop70: Representatives from SCOP (70% max. sequence identity) :octocat:
• Pfam A: Protein Family database :octocat:

Click on the :octocat: links to go to the GitHub repository that contains the scripts to generate the respective database.

The databases consist of usually to following six files, which all start with the name of the database, followed by different extensions:

<dbname>_cs219.ffdata   packed file with column-state sequences for prefiltering
<dbname>_cs219.ffindex  index file for packed column-state sequence file
<dbname>_a3m.ffdata     packed file with MSAs in A3M format
<dbname>_a3m.ffindex    index file for packed A3M file
<dbname>_hhm.ffdata     packed file with HHM-formatted HMMs
<dbname>_hhm.ffindex    index file for packed HHM file

The packed files <dbname>_cs219.ffdata, <dbname>_hhm.ffdata and <dbname>_a3m.ffdata contain simply the concatenated A3M MSAs and HHMs, respectively, with a \0 character at the end of each file. They are therefore human-readable and are parsable for specific MSAs or models using tools such as grep or search functions in text editors (which however should be able to ignore the \0 character). The .ffindex files contain indices to provide fast access to these packed files.

To get started, download the Uniclust30 database. For example:

cd ~/programs/hh  % change to HHsuite directory
mkdir databases; cd databases
wget http://wwwuer.gwdg.de/~compbiol/uniclust/2018_08/uniclust30_2018_08_hhsuite.tar.gz
tar xzvf uniclust30_2018_08_hhsuite.tar.gz

Note that, in order to generate multiple sequence alignments (MSAs) by iterative sequence searching using HHblits, you need to search the Uniclust30, since this databases covers essentially the largest part of the known sequence universe. The PDB70, PfamA, and SCOP are not appropriate to build MSAs by iterative searches.

Uniclust30 is obtained by clustering the UniProt Protein database. Clusters contain sequences that need to be almost full-length (80%) alignable and typically have pairwise sequence identities down to 30%. The clustering is done by MMseqs2, a very fast algorithm for all-against-all sequence comparison and clustering developed in our group. Sequences in each cluster are globally aligned into an MSA (using ClustalO). The clusters in Uniclust30 thus treat all member sequences equally. You need this type of database to build MSAs using iterative HHblits searches.

In the PDB70 and scop70 databases, each master sequence from the original sequence database is represented by an MSA. The MSAs are built by HHblits searches starting with the master sequence as a query. The MSAs and HMMs typically carry the name and annotation of the master sequence. In contrast to the clusters in the Uniclust30 database, sequences can in principle occur in several MSAs. These homologous sequences merely serve to contribute evolutionary information to the master sequence. As the sequences in this type of database do not cover the entire sequence space, they are not suited for iterative searches. See Building customized databases for how to build your own databases.

Note on efficiency: Always try to keep the HHsuite databases on a local SSD drive of the computer running HHblits/HHsearch, if possible. Otherwise file access times on your (shared) remote file server of local HD drive might limit the performance of HHblits/HHsearch.

Running HHblits efficiently on a computer cluster

When HHblits runs on many cores of a compute cluster and accesses the HHsuite database(s) via a central file system, hard disk random access times can quickly become limiting instead of the available compute power. To understand why, let us estimate the random file access bottleneck at an example. Suppose you want to search the Uniclust30 database with HHblits on four compute nodes with 32 CPU cores each. You let each job run on a single core, which is the most efficient way when processing many jobs. Each HHblits iteration through Uniclust30 on a modern 3GHz Intel/AMD core takes around, say, 60s on average. After each search iteration, HHblits needs to read around (3000) a3m or hhm files on disk that passed the prefilter. The total number of expected file accesses per second is therefore:

number of file accesses =
    4 nodes * 32 CPU cores * 3000 files / 60s = 6400s

A typical server hard disk has a latency of around 5ms and can therefore randomly access only 200 files per second. Hence, if all cores need to access the files from the same RAID disk array, 4 CPU cores would already saturate the RAID’s disk drives! (You can slightly decrease latency by a factor of two by using RAID configurations with striping, but this won’t help much.)

To efficiently search with HHblits with millions of sequences in parallel on a compute cluster, we therefore recommend one of the first two options.

1. The ideal solution is to automatically load the needed HHsuite databases from the file system into the virtual RAM drive of each compute node upon boot-up. This requires enough memory in your computer. Software for RAM drives exist on all major operating systems. On Linux from kernel 2.4 upwards, a RAM drive is installed by default at /dev/shm. When memory is low, this software can intelligently swap rarely used files to a hard drive.

2. Another good option is to hold the HHsuite databases on local solid state drive (SSD) of each of the cluster’s compute nodes that can accept HHsuite jobs. The local SSDs also remove the file access bottleneck, as SSDs have typical random access times below 100ms. Reading 3000 files from a SSD after prefiltering thus should take below 0.3s.

3. Another good option might be to use a RAM disk (or SSD) on your file server to hold the HHsuite databases. Then file access time is not problematic anymore, but you need to estimate if data throughput might get limiting. The four servers from our previous example would need to read data at a rate:

transfer rate = 4 nodes * 32 CPU cores * 3000 files * 5kB (av. file size) / 60s 
              = 256MB/s

If you have 16 instead of 4 compute servers or if data transfer from the file server has to be shared with many other compute servers, the network transfer may be limiting the performance.

1. A not so good option is to keep the HHsuite databases on local hard disk drives (HDDs). In this way, you at least distribute the file accesses to as many HDDs as you have compute nodes. But as we have seen, already at around 4 cores per HDD, you will be limited by file access rather than CPU power.

2. You (or your sysadmin) should at least make sure that your compute nodes have sufficient memory reserves and proper cache settings in order to keep the prefiltering database file in cache (size: a few GB). You can check this by running HHblits on a fresh compute node twice and observing that the second run takes much less time to read the prefiltering database file. But this only solves the problem of having to read the prefiltering file every time hhblits is called. But, as discussed, the usually more serious bottleneck is reading the a3m and hhm files that passed the prefilter via the file system. These can also be cached, but since each time only a small subset of all files are read, caching is inefficient for these.

Brief tutorial to HHsuite tools

Overview of programs

• hhblits: Iteratively) search an HHsuite database with a query sequence or MSA
• hhsearch: Search an HHsuite database with a query MSA or HMM
• hhmake: Build an HMM from an input MSA
• hhfilter: Filter an MSA by max sequence identity, coverage, and other criteria
• hhalign: Calculate pairwise alignments etc. for two HMMs/MSAs
• hhconsensus: Calculate the consensus sequence for an A3M/FASTA input file
• reformat.pl: Reformat one or many MSAs
• addss.pl: Add PSIPRED predicted secondary structure to an MSA or HHM file
• hhmakemodel.pl: Generate MSAs or coarse 3D models from HHsearch or HHblits results
  
• hhmakemodel.py: Generates coarse 3D models from HHsearch or HHblits results and modifies cif files such that they are compatible with MODELLER
• hhsuitedb.py: Build HHsuite database with prefiltering, packed MSA/HMM, and index files
• splitfasta.pl: Split a multiple-sequence FASTA file into multiple single-sequence files
• renumberpdb.pl: Generate PDB file with indices renumbered to match input sequence indices
• HHPaths.pm: Configuration file with paths to the PDB, BLAST, PSIPRED etc.
• mergeali.pl: Merge MSAs in A3M format according to an MSA of their seed sequences
• pdb2fasta.pl: Generate FASTA sequence file from SEQRES records of globbed pdb files
• cif2fasta.py: Generate a FASTA sequence from the pdbx_seq_one_letter_code entry of the entity_poly of globbed cif files
• pdbfilter.pl: Generate representative set of PDB/SCOP sequences from pdb2fasta.pl output
• pdbfilter.py: Generate representative set of PDB/SCOP sequences from cif2fasta.py output

Call a program without arguments or with -h option to get more detailed explanations.

Searching databases of HMMs using HHsearch and HHblits

We will use the MSA query.a3m in the data/ subdirectory of the HHsuite as an example query. To search for sequences in the scop70_1.75 database that are homologous to the query sequence or MSA in query.a3m, type

hhsearch -cpu 4 -i data/query.a3m -d databases/scop70_1.75 -o data/query.hhr

The database scop70_1.75 can be obtained from our download server. If the input file is an MSA or a single sequence, HHsearch calculates an HMM from it and then aligns this query HMM to all HMMs in the scop70_1.75 database using the Viterbi algorithm. After the search, the most significant HMMs are realigned using the more accurate Maximum Accuracy (MAC) algorithm. After the realignment phase, the complete search results consisting of the summary hit list and the pairwise query-template alignments are written to the output file, data/query.hhr. The hhr result file format was designed to be human readable and easily parsable.

The -cpu 4 option tells HHsearch to start four OpenMP threads for searching and realignment. This will typically results in almost fourfold faster execution on computers with four or more cores. Since the management of the threads costs negligible overhead.

The HHblits tool can be used in much the same way as HHsearch. It takes the same input data and produces a results file in the same format as HHsearch. Most of the HHsearch options also work for HHblits, which has additional options associated with its extended functionality for iterative searches. Due to its fast prefilter, HHblits runs between 30 and 3000 times faster than HHsearch at the cost of only a few percent lower sensitivity.

The same search as above is performed here using HHblits instead of HHsearch:

hhblits -cpu 4 -i data/query.a3m -d databases/scop70_1.75 -o data/query.hhr -n 1

HHblits first scans the column state sequences in scop70_1.75_cs219.ffdata with its fast prefilter. HMMs whose column state sequences pass the prefilter are read from the packed file scop70_1.75_hhm.ffdata (using the index file scop70_1.75_hhm.ffindex) and are aligned to the query HMM generated from query.a3m using the slow Viterbi HMM-HMM alignment algorithm. The search results are written to the default output file query.hhr. The option -n 1 tells HHblits to perform a single search iteration. (The default is 2 iterations.)

Generating a multiple sequence alignment using HHblits

To generate an MSA for a sequence or initial MSA in query.a3m, the database to be searched should cover the entire sequence space, such as the Uniclust30. The option -oa3m <msa_file> tells HHblits to generate an output MSA from the significant hits, -n 1 specifies a single search iteration.

hhblits -cpu 4 -i data/query.seq -d databases/uniclust30_2018_08/uniclust30_2018_08 -oa3m query.a3m -n 1

At the end of the search, HHblits reads from the packed database file containing the MSAs the sequences belonging to HMMs with E-value below the threshold. The E-value threshold for inclusion into the MSA can be specified using the -e <E-value> option. After the search, query.a3m will contain the MSA in A3M format.

We could do a second search iteration, starting with the MSA from the previous search, to add more sequences. Since the MSA generated after the previous search contains more information than the single sequence in query.seq, searching with this MSA will probably result in many more homologous database matches.

hhblits -cpu 4 -i query.a3m -d databases/uniclust30_2018_08/uniclust30_2018_08 -oa3m query.a3m -n 1

Instead, we could directly perform two search iterations starting from query.seq:

hhblits -cpu 4 -i data/query.seq -d databases/uniclust30_2018_08/uniclust30_2018_08 -oa3m query.a3m -n 2 

See Why do I get different results when I perform 2 and 1+1 iterations with HHblits? for an explanation why the results can be slightly different.

In practice, it is recommended to use between 1 and 4 iterations for building MSAs, depending on the trade-off between reliability and specificity on one side and sensitivity for remotely homologous sequences on the other side. The more search iterations are done, the higher will be the risk of non-homologous sequences or sequence segments entering the MSA and recruiting more of their kind in subsequent iterations. This is particularly problematic when searching with sequences containing short repeats, regions with amino acid compositional bias and, although less dramatic, with multiple domains. Fortunately, this problem is much less pronounced in hhblits as compared to PSI-BLAST due to hhblits’s lower number of iterations, its more robust Maximum Accuracy alignment algorithm, and the higher precision of its HMM-HMM alignments.

The parameter -mact (maximum accuracy threshold) lets you choose the trade-off between sensitivity and precision. With a low mact-value (e.g. -mact 0.01) very sensitive, but not so precise alignments are generated, whereas a search with a high mact-value (e.g. -mact 0.9) results in shorter but very precise alignments. The default value of -mact in HHblits is 0.35.

To avoid unnecessarily large and diverse MSAs, HHblits stops iterating when the diversity of the query MSA – measured as number of effective sequences (N_eff) – grows past a threshold of 10.0. This threshold can be modified with the --neffmax <float> option. See HHsearch/HHblits model format for a description of how the number of effective sequences is calculated in HH-suite.

To avoid that the final MSAs grows unnecessarily large, by default, the database MSAs that are going to be merged with the query MSA into the result MSA are filtered with the active filter options (by default -id 90 and -diff 1000). The -all option turns off the filtering of the result MSA. Use this option if you want to get all sequences in the significantly similar Uniclust30 clusters:

hhblits -cpu 4 -i data/query.seq -d databases/uniclust30_2018_08/uniclust30_2018_08 -oa3m query.a3m -all

The A3M format uses small letters to mark inserts and capital letters to designate match and delete columns (see Multiple sequence alignment formats), allowing you to omit gaps aligned to insert columns. The A3M format therefore uses much less space for large alignments than FASTA but looks misaligned to the human eye. Use the reformat.pl script to reformat query.a3m to other formats, e.g. for reformatting the MSA to Clustal and FASTA format, type:

reformat.pl a3m clu query.a3m query.clu
reformat.pl a3m fas query.a3m query.fas

Next, to add secondary structure information to the MSA we call the script addss.pl. For addss.pl to work, you have to make sure that the paths to BLAST and PSIPRED in the file $HHLIB/scripts/HHPaths.pm are correctly filled in:

addss.pl query.a3m

When the sequence has a SCOP or PDB identifier as first word in its name, the script tries to add the DSSP states as well. Open the query.a3m file and check out the two lines that have been added to the MSA.

❗️ While secondary structure scoring does improve homology detection a bit, this step is extremely time consuming and currently hard to set up due to hardcoded old versions of tools. We recommend to skip this step.

Now you can generate a hidden Markov model (HMM) from this MSA:

hhmake -i query.a3m -o query.hhm

The default output file is query.hhm. By default, the option -M first will be used. This means that exactly those columns of the MSAs which contain a residue in the query sequence will be assigned to Match / Delete states, the others will be assigned to Insert states. (The query sequence is the first sequence not containing secondary structure information.) Alternatively, you may want to apply the 50%-gap rule by typing -M 50, which assigns only those columns to Insert states which contain more than 50% gaps. The -M first option makes sense if your alignment can best be viewed as a seed sequence plus aligned homologs to reinforce it with evolutionary information. This is the case in the SCOP and PDB versions of our HMM databases, since here MSAs are built around a single seed sequence (the one with known structure). On the contrary, when your alignment represents an entire family of homologs and no sequence in particular, it is best to use the 50% gap rule. This is the case for Pfam or SMART MSAs, for instance. Despite its simplicity, the 50% gap rule has been shown to perform well in practice.

When calling hhmake, you may also apply several filters, such as maximum pairwise sequence identity (-id <int>), minimum sequence identity with query sequence (-qid <int>), or minimum coverage with query (-cov <int>). But beware of reducing the diversity of your MSAs too much, as this will lower the sensitivity to detect remote homologs.

Example: Comparative protein structure modeling using HHblits and MODELLER

A three-dimensional (3D) structure greatly facilitates the functional characterization of proteins. However, for many proteins there are no experimental structures available, and thus, comparative modeling to known protein structures may provide useful insights. In this method, a 3D structure of a given protein sequence (target) is predicted based on alignments to one or more proteins of known structures (templates). In the following, we demonstrate how to create alignments for a unresolved protein with HHblits and the PDB70 database. We then convert search results from HHblits to build a comparative model using MODELLER (v9.16).

In 1999, Wu et all. reported the genomic sequence and evolutionary analysis of lactate dehydrogenase genes from Trichomonas vaginalis (TvLDH). Surprisingly, the corresponding protein sequence was most similar to the malate dehydrogenase (TvMDH) of the same organism implying TvLDH arose from TvMDH by convergent evolution. In the meantime, the structure of TvLDH has been resolved, however, for instructional purposes suppose that there is no 3D structure for the protein.

To get started we obtain the protein sequence of TvLDH from GeneBank (accession number AF060233.1). We first copy-paste the protein sequence into a new file named query.seq. The content of this file should somewhat look similar to this.

>TvLDH
MSEAAHVLITGAAGQIGYILSHWIASGELYGDRQVYLHLLDIPPAMNRLTALTMELEDCAFPHLAGFVATTDP
KAAFKDIDCAFLVASMPLKPGQVRADLISSNSVIFKNTGEYLSKWAKPSVKVLVIGNPDNTNCEIAMLHAKNL
KPENFSSLSMLDQNRAYYEVASKLGVDVKDVHDIIVWGNHGESMVADLTQATFTKEGKTQKVVDVLDHDYVFD
TFFKKIGHRAWDILEHRGFTSAASPTKAAIQHMKAWLFGTAPGEVLSMGIPVPEGNPYGIKPGVVFSFPCNVD
KEGKIHVVEGFKVNDWLREKLDFTEKDLFHEKEIALNHLAQ

The search results obtained by querying the TvLDH sequence against the PDB70 will be significantly better if we use a MSA instead of a single sequence. For this reason, we first query the protein sequence of TvLDH against the Uniclust30 database which covers the whole protein sequence space. By executing

hhblits -i query.seq -d databases/uniclust30_2018_08/uniclust30_2018_08 -oa3m query.a3m  -cpu 4 -n 1

we obtain a MSA in a3m format which contains several sequences that are similar to TvLDH. Now we can use the query.a3m and search the PDB70 database for similar protein structures.

hhblits -i query.a3m -o results.hhr -d databases/pdb70 -cpu 4 -n 1

Note that we now output a hhr file results.hhr instead of an a3m file. Before we convert the search results to a format that is readable by MODELLER, let us quickly inspect results.hhr.

Query         TvLDH
Match_columns 333
No_of_seqs    2547 out of 8557
Neff          11.6151
Searched_HMMs 1566
Date          Tue Aug 16 11:35:02 2016
Command       hhblits -i query.a3m -o results.hhr -d pdb70 -cpu 32 -n 2

 No Hit                             Prob E-value P-value  Score    SS Cols Query HMM  Template HMM
  1 7MDH_C MALATE DEHYDROGENASE; C 100.0 1.5E-39 1.3E-43  270.1   0.0  326    3-333    31-358 (375)
  2 4UUL_A L-LACTATE DEHYDROGENASE 100.0 1.8E-35 1.6E-39  244.5   0.0  332    1-332     1-332 (341)
  3 4UUP_A MALATE DEHYDROGENASE (E 100.0 2.2E-35 1.9E-39  243.8   0.0  332    1-332     1-332 (341)
  4 4UUM_B L-LACTATE DEHYDROGENASE 100.0 5.7E-35 5.1E-39  241.5   0.0  333    1-333     1-333 (341)
  5 1CIV_A NADP-MALATE DEHYDROGENA 100.0 1.6E-34 1.5E-38  241.3   0.0  326    2-332    40-367 (385)
  6 1Y7T_A Malate dehydrogenase(E. 100.0 3.4E-34 3.1E-38  235.3   0.0  324    1-331     1-325 (327)
  7 4I1I_A Malate dehydrogenase (E 100.0 3.8E-34 3.4E-38  236.4   0.0  319    2-327    22-343 (345)
  8 1BMD_A MALATE DEHYDROGENASE (E  99.9 5.9E-34 5.3E-38  233.9   0.0  324    1-331     1-325 (327)
  9 4H7P_B Malate dehydrogenase (E  99.9 9.9E-34 8.9E-38  233.9   0.0  319    2-327    22-343 (345)
 10 2EWD_B lactate dehydrogenase,   99.9 1.1E-33 9.5E-38  231.1   0.0  306    1-328     1-314 (317)
 
 (...)

We find that there are several templates that have a high similarity to our query. Interestingly, the hit with the most significant E-value score is also malate dehydrogenase. We will use this structure as a basis for our comparative model. In order to build the model we first have to obtain the template structure(s). We can get 7MDH by typing the following commands

mkdir templates
cd templates
wget http://files.rcsb.org/download/7MDH.cif 
cd ..

To convert our search results results.hhr into an alignment that is readable by MODELLER we use hhmakemodel.py.

python3 hhmakemodel.py results.hhr templates/ TvLDH.pir ./ -m 1

This script takes four positional arguments: the results file in hhr format, the path to the folder containing all templates in cif format, the output pir file, and folder where the processed cif files should be written to. The -m flag tells hhmakemodel.py to only include the first hit in the pir alignment. The pir file together with processed cifs can be used as an input for MODELLER (please refer to the MODELLER documentation for further help).

Visually checking an MSA for corrupted regions

Iterative search methods such as PSI-BLAST and HHblits may generate alignments containing non-homologous sequence stretches or even large fractions of non-homologous sequences. The cause for this is almost always the overextension of homologous alignments into non-homologous regions. This has been termed homologous overextension in . This effect occurs particularly in multidomain and repeat proteins. A single overextended alignment in the search results leads to more of such sequences to be included in the profile of the next iteration – usually all homologous to the first problematic sequence but with even longer non-homologous stretches. Thus, after three or more iterations, large sections of the resulting MSA may be non-homologous to the original query sequence. This risk of homologous overextension is greatly reduced in HHblits in comparison to PSI-BLAST, because fewer iterations are usually necessary for HHblits and because HHblits uses the Maximum Accuracy alignment algorithm (see What does the maximum accuracy alignment algorithm do?), which is much less prone to overextend alignments than the Smith-Waterman/Viterbi algorithm used by most other programs including PSI-BLAST. Still, in important cases it is worth to visually check the MSAs of the query and the matched database protein for the presence of corrupted regions containing non-homologous sequence stretches.

The recommended procedure to visually check an MSA is the following. We first reduce the MSA to a small set of sequences that could still fit into a single window of an alignment viewer:

hhfilter -i query.a3m -o query.fil.a3m -diff 30

The option -diff causes hhfilter to select a representative set of at least 30 sequences that best represent the full diversity of the MSA. Sequences that contain non-homologous stretches are therefore usually retained, as they tend to be the most dissimilar to the main sequence cluster.

Next, we remove all inserts (option -r) with respect to the first, master sequence in the MSA. The resulting MSA is sometimes called a master-slave alignment:

reformat.pl -r query.fil.a3m query.fil.fas

This alignment is now very easy to inspect for problematic regions in any viewer that allows to color the residues according to their physico-chemical properties. We can recommend alnedit or jalview, for example:

java -jar alnedit.jar query.fil.fas .

Building customized databases

It is possible to build custom HH-suite databases using the same tools we use to build the standard HH-suite databases (except Uniclust30). An example application is to search for homologs among all proteins of an organism. To build your own HH-suite database from a set of sequences, you first need to generate an MSA with predicted secondary structure for every sequence in the set.

🗒 The following calls assume you have compiled the HH-suite with MPI support. You can substitute hhblits_mpi with hhblits_omp, ffindex_apply_mpi with ffindex_apply and cstranslate_mpi with cstranslate -f to use the OpenMP (single compute node) parallelization instead.

1. Creating a database of HHblits compatible MSAs

Starting with sequences

First of all we convert the input FASTA file to a FFindex database with ffindex_from_fasta:

ffindex_from_fasta -s <db>_fas.ff{data,index} <db.fas> 

Now, to build an MSA with HHblits for each sequence in <db>_fas.ff{data,index}, run

mpirun -np <number_threads> \
  hhblits_mpi -i <db>_fas -d <path_to/uniclust30> -oa3m <db>_a3m_wo_ss -n 2 -cpu 1 -v 0

The MSAs are written to the ffindex <db>_a3m_wo_ss.ff{data,index}. To be sure that everything went smoothly, check that the number of lines in <db>_a3m.ffindex is the same as the number of lines in <db>_fas.ffindex.

The number of HHblits search iterations and the HMM inclusion E-value threshold for HHblits can be changed from their default values (2 and 0.01, respectively) using the -n <int> and -e <float> options. A higher number of iterations such as -n 3 will result in very high sensitivity to discover remotely homologous relationships, but the ranking among homologous proteins will often not reflect their degree of relationship to the query. The reason is that any similarities that are higher than the similarities among the sequences in the query and database MSAs cannot be resolved.

Starting with MSAs

If you already have MSAs in aligned FASTA format, place all of them first in a single folder that does not contain any other files to create a single FFindex database:

cd msa/
ffindex_build -s ../<db>_msa.ff{data,index} .
cd ..

Now you can use hhconsensus to add a consensus sequence and do the format conversion to the more space efficient A3M format:

OMP_NUM_THREADS=1 mpirun -np <number_threads> ffindex_apply_mpi <db>_msa.ff{data,index} \
  -i <db>_a3m_wo_ss.ffindex -d <db>_a3m_wo_ss.ffdata \
    -- hhconsensus -M 50 -maxres 65535 -i stdin -oa3m stdout -v 0
rm <db>_msa.ff{data,index}

Annotationg secondary structure information (optional)

Now, add PSIPRED-predicted secondary structure and if possible DSSP secondary structure annotation to all MSAs:

❗️ While secondary structure scoring does improve homology detection a bit, this step is extremely time consuming and currently hard to set up due to hardcoded old versions of tools. We recommend to skip this step. Rename _a3m_wo_ss.ff{data,index} to _a3m.ff{data,index} instead to proceed with the next step.

mpirun -np <number_threads> ffindex_apply_mpi <db>_a3m_wo_ss.ff{data,index} \
  -i <db>_a3m.ffindex -d <db>_a3m.ffdata -- addss.pl -v 0 stdin stdout
rm <db>_a3m_wo_ss.ff{data,index}

Computing hidden markov models

We also need to generate an HHM model for each MSA:

mpirun -np <number_threads> ffindex_apply_mpi <db>_a3m.ff{data,index} \\
  -i <db>_hhm.ffindex -d <db>_hhm.ffdata -- hhmake -i stdin -o stdout -v 0

Computing context states for prefiltering

In order to build the FFindex database containing the column state sequences for prefiltering for each a3m we run:

mpirun -np <number_threads> cstranslate_mpi -f -x 0.3 -c 4 -I a3m -i <db>_a3m -o <db>_cs219 

❗️ Be careful with the parameters. Leaving anything out, might result in a database that looks superficially correct, but will perform very badly.

❗️ Before version 3.2 you would need to specify the -b parameter to cstranslate to produce a HHblits3 compatible database, without it you would produce a database that could be still post-processed to be readable by HHblits2.x. We have since dropped support for HHblits2.x and the -b parameter.

Putting everything together

Next we want to reorder the entries in the hmm and a3m databases. For this purpose we sort the files in the data files according to the number of columns in the MSAs (i.e. the sequence length of the consensus sequence). The number of columns can be retrieved in the third row of <db>_cs219.ffindex.

sort -k3 -n -r <db>_cs219.ffindex | cut -f1 > sorting.dat
    
ffindex_order sorting.dat <db>_hhm.ff{data,index} <db>_hhm_ordered.ff{data,index}
mv <db>_hhm_ordered.ffindex <db>_hhm.ffindex
mv <db>_hhm_ordered.ffdata <db>_hhm.ffdata
    
ffindex_order sorting.dat <db>_a3m.ff{data,index} <db>_a3m_ordered.ff{data,index}
mv <db>_a3m_ordered.ffindex <db>_a3m.ffindex
mv <db>_a3m_ordered.ffdata <db>_a3m.ffdata

Example: Building a database from the PDB

To make efficient sequence searches in the PDB we provide a precompiled PDB70 database containing PDB sequences clustered at 70% sequence identity. However, to find a larger variety of PDB templates, larger databases with more redundancy might be required. In this tutorial, we outline the steps required to build a custom PDB database. See the PDB70 GitHub repository for the actual scripts used to build this database.

First, download the entire PDB database from RSCB in cif file format (this is the successor of the pdb file format) by executing

rsync --progress -rlpt -v -z --port=33444 rsync.wwpdb.org::ftp/data/structures/divided/mmCIF <cif_dir>

and unzip the files into a single directory <all_cifs>. Then, run cif2fasta.py by typing

python3 cif2fasta.py -i <all_cifs> -o pdb100.fas -c <num_cores> -p pdb_filter.dat

The script scans the folder <all_cifs> for files with the suffix *.cif and write each sequence and its associated chain identifier annotated in pdbx_seq_one_letter_code entry from the entity_poly table to the fasta file pdb100.fas. By specifying the optional -p flag, cif2fasta.py creates an additional file pdb_filter.dat which is required by pdbfilter.py in a later step. Note that cif2fasta.py by default removes sequences which are shorter than 30 residues and/or comprise only the residue ’X’.

If you wish to exhaustively search the PDB, skip the following steps and continue with the instructions described in section Building customized databases. However, to increase the speed of database searches, e.g. on systems with limited resources, you can reduce the number of sequences by clustering them with MMseqs2 and selecting representative sequences with pdbfilter.py. To cluster the sequences of pdb100.fas at a sequence identity of X and a coverage of Y run MMseqs2 using these options.

mmseqs createdb pdb100.fas <clu_dir>/pdb100
mmseqs cluster <clu_dir>/pdb100 <clu_dir>/pdbXX_clu /tmp/clustering -c Y --min-seq-id X
mmseqs createtsv <clu_dir>/pdb100 <clu_dir>/pdb100 <clu_dir>/pdbXX_clu <clu_dir>/pdbXX_clu.tsv

MMseqs2 yields a tab separated file pdbXX_clu.tsv which contains cluster assignments for all sequences. Representative sequences are selected by pdbfilter.py which chooses up to three sequences for each cluster by identifying the ones having either the highest resolution A, the largest R-free value or the largest "completeness".

We compute the completeness of a protein structure by dividing the number of residues that are found in the ATOM section by the total number of residues declared in the pdbx_seq_one_letter_code entry of the entity_poly table.

python3 pdbfilter.py pdb100.fas pdbXX_clu.tsv pdb_filter.dat pdbXX.fas -i pdb70_to_include.dat -r pdb70_to_remove.dat

pdbfilter.py takes the original fasta file (pdb100.fas) and the annotation file pdb_filter.dat which both were created by cif2fasta.py, and the cluster assignments from MMseqs2 (pdb70_clu.tsv) as input and outputs the final pdbXX.fas. Use this fasta file to complete the creation of your database (see section Building customized databases).

Modifying or extending existing databases

Assume you have a list of filenames in files.dat you want to remove from an HHsuite database.

ffindex_modify -s -u -f files.dat <db>_a3m.ffindex
ffindex_modify -s -u -f files.dat <db>_hhm.ffindex
ffindex_modify -s -u -f files.dat <db>_cs219.ffindex

This deletes the file entries from the ffindex files, however the files are still in the ffdata file. This way HHblits and HHsuite won’t be able to use them. If you want to get rid of them in the ffdata file you may rebuild the databases.

ffindex_build -as <db>_cs219_ordered.ff{data,index} -i <db>_cs219.ffindex -d <db>_cs219.ffdata
mv <db>_c219_ordered.ffindex <db>_cs219.ffindex
mv <db>_cs219_ordered.ffdata <db>_cs219.ffdata

sort -k3 -n <db>_cs219.ffindex | cut -f1 > sorting.dat

ffindex_order sorting.dat <db>_hhm.ff{data,index} <db>_hhm_ordered.ff{data,index}
mv <db>_hhm_ordered.ffindex <db>_hhm.ffindex
mv <db>_hhm_ordered.ffdata <db>_hhm.ffdata

ffindex_order sorting.dat <db>_a3m.ff{data,index} <db>_a3m_ordered.ff{data,index}
mv <db>_a3m_ordered.ffindex <db>_a3m.ffindex
mv <db>_a3m_ordered.ffdata <db>_a3m.ffdata

If you want to check your HHsuite database you may run:

hhsuitedb.py -o <db> --cpu 1

hhsuitedb.py supports the flag --force that tries to fix the database. This may lead to the deletion of MSAs from the database.

Additionally you may add a3m, hhm and cs219 files with globular expressions to an existing database. Please keep in mind that related a3m, hhm and cs219 files should have the same name. To make it easier you may add just a3m files and the corresponding hhm and cs219 files will be calculated by the script.

hhsuitedb.py -o <db> --cpu 1 --ia3m=<a3m_glob> --ics219=<cs219_glob> --ihhm=<hhm_glob>

Frequently asked questions

How do I report a bug?

If you found a bug, please report it to us using GitHub issues.

What are HMM-HMM comparisons and why are they so powerful?

When searching for remote homologs, it is wise to make use of as much information about the query and database proteins as possible in order to better distinguish true from false positives and to produce optimal alignments. This is the reason why sequence-sequence comparison is inferior to profile-sequence comparison. Sequence profiles contain for each column of a multiple alignment the frequencies of the 20 amino acids. They therefore contain detailed information about the conservation of each residue position, i.e., how important each position is for defining other members of the protein family, and about the preferred amino acids. Profile Hidden Markov Models (HMMs) are similar to simple sequence profiles, but in addition to the amino acid frequencies in the columns of a multiple sequence alignment they contain information about the frequency of inserts and deletions at each column. Using profile HMMs in place of simple sequence profiles should therefore further improve sensitivity. Using HMMs both on the query and the database side greatly enhances the sensitivity/selectivity and alignment quality over sequence-profile based methods such as PSI-BLAST. HHsearch is the first software to employ HMM-HMM comparison and HHblits is the first profile-profile comparison method that is fast enough to do iterative searches to build MSAs.

When can the HH-suite be useful for me?

Sequence search methods such as BLAST, FASTA, or PSI-BLAST are of prime importance for biological research because functional information of a protein or gene can be inferred from homologous proteins or genes identified in a sequence search. But quite often no significant relationship to a protein of known function can be established. This is certainly the case for the most interesting group of proteins, those for which no ortholog has yet been studied. In cases where conventional sequence search methods fail, HHblits and HHsearch quite often allow to make inferences from more remotely homologous relationships. HHblits builds better MSAs, with which more remote homologs can then be found using HHsearch or HHblits, e.g. by searching the PDB or domain databases such as Pfam. If the relationship is so remote that no common function can be assumed, one can often still derive hypotheses about possible mechanisms, active site positions and residues, or the class of substrate bound. When a homologous protein with known structure can be identified, its stucture can be used as a template to model the 3D structure of the protein of interest , since even protein domains that shared a common ancestor some 3 billion years ago mostly have similar 3D structures. The 3D model may then help to generate hypotheses to guide experiments.

What does homology mean and why is it important?

Two protein sequences are homologous to each other if they descended from a common ancestor sequence. Generally, homologous proteins (or protein fragments) have similar structure because structures diverge much more slowly than their sequences . Depending on the degree of divergence between the sequences, the proteins may also have similar cellular functions, ligands, protein interaction partners, or enzymatic mechanisms. On the contrary, proteins that have a similar structure by convergence (i.e., by chance) are said to be analogous. They don’t generally share similar functions or biochemical mechanisms and are therefore much less helpful for making inferences. HHsearch and HHblits are tools for homology detection and as such do not normally detect analogous relationships.

How can I verify if a database match is homologous?

Here is a list of things to check if a database match really is at least locally homologous.

Check probability and E-value: HHsearch and HHblits can detect homologous relationships far beyond the twilight zone, i.e., below 20% sequence identity. Sequence identity is therefore not an appropriate measure of relatedness anymore. The estimated probability of the template to be (at least partly) homologous to your query sequence is the most important criterion to decide whether a template HMM is actually homologous or just a high-scoring chance hit. When it is larger than 95%, say, the homology is nearly certain. Roughly speaking, one should give a hit serious consideration (i.e., check the other points in this list) whenever (1) the hit has (>50%) probability, or (2) it has (>30%) probability and is among the top three hits. The E-value is an alternative measure of statistical significance. It tells you how many chance hits with a score better than this would be expected if the database contained only hits unrelated to the query. At E-values below one, matches start to get marginally significant. Contrary to the probability, when calculating the E-value HHsearch and HHblits do not take into account the secondary structure similarity. Therefore, the probability is a more sensitive measure than the E-value.

Check if homology is biologically suggestive or at least reasonable: Does the database hit have a function you would expect also for your query? Does it come from an organism that is likely to contain a homolog of your query protein?

Check secondary structure similarity: If the secondary structure of query and template is very different or you can’t see how they could fit together in 3D, then this is a reason to distrust the hit. Note however that if the query alignment contains only a single sequence, the secondary structure prediction is quite unreliable and confidence values are overestimated.

Check relationship among top hits: If several of the top hits are homologous to each other, (e.g. when they are members of the same SCOP superfamily), then this will considerably reduce the chances of all of them being chance hits, especially if these related hits are themselves not very similar to each other. Searching the SCOP database is very useful precisely for this reason, since the SCOP family identifier (e.g. a.118.8.2) allows to tell immediately if two templates are likely homologs.

Check for possible conserved motifs: Most homologous pairs of alignments will have at least one (semi-)conserved motif in common. You can identify such putative (semi-)conserved motifs by the agglomeration of three or more well-matching columns (marked with a | sign between the aligned HMMs) occurring within a few residues, as well as by matching consensus sequences. Some false positive hits have decent scores due to a similar amino acid composition of the template. In these cases, the alignments tend to be long and to lack conserved motifs.

Check residues and role of conserved motifs: If you can identify possible conserved motifs, are the corresponding conserved template residues involved in binding or enzymatic function?

Check query and template alignments: A corrupted query or template alignment is the main source of high-scoring false positives. The two most common sources of corruption in an alignment are (1) non-homologous sequences, especially repetitive or low-complexity sequences in the alignment, and (2) non-homologous fragments at the ends of the aligned database sequences. Check the query and template MSAs in an alignment viewer such as Jalview.

Realign with other parameters: change the alignment parameters. Choose global instead of local mode, for instance, if you expect your query to be globally homologous to the putative homolog. Try to improve the probability by changing the values for minimum coverage or minimum sequence identity. You can also run the query HMM against other databases.

Build the query and/or database MSAs more aggressively: If your query (or template) MSA is not diverse enough, you could increase sensitivity substantially by trying to include more remotely homologous sequences into the MSA. Have non-homologous sequences or sequence segments been accidentally included? You can also try to build a more diverse MSA manually: Inspect the HHblits results after the first iteration and consider including hits above the E-value inclusion threshold of 0.001, based on biological plausibility, relatedness of the organism, a reasonable looking alignment, or just guessing. Then start the second HHblits search iteration HHblits with this manually enriched alignment.

Verify predictions experimentally: The ultimate confirmation of a homologous relationship or structural model is, of course, the experimental verification of some of its key predictions, such as validating the binding to certain ligands by binding assays, measuring biochemical activity, or comparing the knock-out phenotype with the one obtained when the putative functional residues are mutated.

What does the maximum accuracy alignment algorithm do?

HHblits and HHsearch use a better alignment algorithm than the quick and standard Viterbi method to generate the final HMM-HMM alignments. Both realign all displayed alignments in a second stage using the more accurate Maximum Accuracy (MAC) algorithm . The Viterbi algorithm is employed for searching and ranking the matches. The realignment step is parallelized (-cpu <int>) and typically takes a few seconds only.

Please note: Using different alignment algorithms for scoring and aligning has the disadvantage that the pairwise alignments that are displayed are not always very similar to those that are used to calculate the scores. This can lead to confusing results where alignments of only one or a few residues length may have obtained significant E-values. In such cases, run the search again with the -norealign option, which will skip the MAC-realignment step. This will allow you to check if the Viterbi alignments are valid at all, which they will probably not be. The length of the MAC alignments can therefore give you additional information to decide if a match is valid. In order to avoid confusion for users of our HHpred server , the -norealign option is the default there, whereas for you pros who dare to use the command line package, realigning is done by default.

The posterior probability threshold is controlled with the -mact option. This parameter controls the alignment algorithm’s greediness. More precisely, the MAC algorithm finds the alignment that maximizes the sum of posterior probabilities minus mact for each aligned pair. Global alignments are generated with -mact 0, whereas -mact 0.5 will produce quite conservative local alignments.

The -global and -local options now refer to both the Viterbi search stage as well as the MAC realignment stage. With -global (-local), the posterior probability matrix will be calculated for global (local) alignment. When -global is used in conjunction with -realign, the mact parameter is automatically set to 0 in order to produce global alignments. In other words, both following two commands will give global alignments:

hhsearch -i <query> -d <db> -realign -mact 0
hhsearch -i <query> -d <db> -realign -global

The first version uses local Viterbi to search and then uses MAC to realign the proteins globally (since mact is 0) on a local posterior probability matrix. The second version uses global Viterbi to search and then realigns globally (since mact is automatically set to 0) on a global posterior matrix. To detect and align remote homologs, for which sometimes only parts of the sequence are conserved, the first version is clearly better. It is also more robust. If you expect to find globally alignable sequence homologs, the second option might be preferable. In that case, it is recommended to run both versions and compare the results.

How is the MSA diversity Neff calculated?

The number of effective sequences of the full alignment, which appears as NEFF in the header of each hhm file, is the average of local values Neff_M(i) over all alignment positions i. The values Neff_M(i) are given in the main model section of the hhm model files (subsection HHsearch/HHblits model format). They quantify the local diversity of the alignment in a region around position i. More precisely, Neff_M(i) measures the diversity of subalignment Ali_M(i) that contains all sequences that have a residue at column i of the full alignment. The subalignment contains all columns for which at least 90% of these sequences have no end gap. End gaps are gaps to the left of the first residue or to the right of the last residue. The latter condition ensures that the sequences in the subalignment Ali_M(i) cover most of the columns in it. The number of effective sequences in the subalignment Ali_M(i) is exp of the average sequence entropy over all columns of the subalignment. Hence, Neff_M is bounded by 0 from below and 20 from above. In practice, it is bounded by the entropy of a column with background amino acid distribution f_a:

N_eff < sum_(a=1)^20 f_a*log(f_a) ≈ 16

Similarly, Neff_I(i) gives the diversity of the subalignment Ali_I(i) of all sequences that have an insert at position i, and Neff_D(i) refers to the diversity of subaligment Ali_D(i) of all sequences that have a Delete (a gap) at position i of the full alignment.

Do HHsearch and HHblits work fine with multi-domain sequences?

HHblits and HHsearch have been designed to work with multi-domain queries. However, the chances for false positives entering the query alginment during the HHblits iterations is greater for multi-domain proteins. For long sequences, it may therefore be of advantage to first search the PDB or the SCOP domain database and then to cut the query sequence into smaller parts on the basis of the identified structural domains. Pfam or CDD are - in our opinion - less suitable to determine domain boundaries.

How do I best build HMMs for integral membrane proteins?

Despite the biased amino acid composition in their integral membrane helices, membrane proteins behave very well in HMM-HMM comparisons. In many benchmark tests and evolutionary studies we have not noticed any higher rates of false positives than with cytosolic proteins.

How do I reconcile overlapping and conflicting domain predictions?

For example when domain A is predicted from residues 2-50 with 98% probability and domain B from 2-200 with 95% probability? The probability that a pair of residues is correctly aligned is the product of the probability for the database match to be homologous (given by the values in the Probab column of the hit list) times the posterior probability of the residue pair to be correctly aligned given the database match is correct in the first place. The posterior probabilities are specified by the confidence numbers in the last line of the alignment blocks (0 corresponds approximately to 0-10%, 9 to 90-100%). Therefore, an obvious solution would be to prune the alignments in the overlapping region such that the sum of total probabilities is maximized. There is no script yet that does this automatically.

How can I build a phylogenetic tree for HMMs?

I would use a similarity measure like the raw score per alignment length. You might also add the secondary structure score to the raw score with some weight. Whereas probabilities, E-values, and P-values are useful for deciding whether a match is a reliable homolog or not, they are not suitable for measuring similarities because they strongly depend on the length of the alignment, roughly like:

P-value ∝ exp^(λ * average_similarity * length),

with some constant lambda of order 1. The probability has an even more complex dependence on length. Also, the “Similarity” given above the alignment blocks is a simple substitution matrix score per column between the query and template master sequences and does not capture the evolutionary information contained in the MSAs (see HMM-HMM pairwise alignments). One note of caution: Large, diverse MSAs are usually more sensitive to find homologs than narrower ones. Therefore, I would limit the diversity of all HMMs to some reasonable number (perhaps around 5 or 7, depending on how far diverged your HMMs are). This filtering can be done using:

hhfilter -i <MSA.a3m> -o <MSA_filt.a3m> -neff 5

Can I weight some residues more than others during the HMM-HMM alignment?

For example to select templates for homology modeling, it could make sense to weight residues around the known functional site more strongly. This might also positively impact the query-template alignment itself. Well, directly weighting query or template match states is not possible. What you can do with a bit of file fiddling, however, may even be better: You can modify the amount of pseudocounts that hhblits, hhsearch, or hhalign will automatically add to the various match state columns in the query or template profile HMM. This is equivalent to assuming a higher conservation for these positions, since pseudocounts are basically mutations that are added to the observed counts in the training sequences to generalize to more remotely related sequences. In HHsuite, the more diverse the sequence neighborhood around a match state column i is, as measured by the number of effective sequences Neff_M(i) (see How is the MSA diversity Neff calculated?), the fewer pseudocounts are added. By manually increasing Neff_M(i) for a match state column you increase the assumed conservation of this position in the protein. The Neff_M(i) values are found in the hhm-formatted HMM model files in the eighth column of the second line of each match state block in units of 0.001 (see HHsearch/HHblits model format). You could, for example, simply increase Neff_M(i) values by a factor of two for important, conserved positions. Setting Neff_M(i) to 99999 will reduce pseudocounts to effectively zero. A word of caution is advised: HMMs built from a diverse set of training sequences already contain quantitative information about the degree of conservation for each position.

Don’t I need to calibrate my query or database HMM?

No. If you don’t specify otherwise, the two parameters of the extreme-value distribution for the query are estimated by a neural network from the lengths and diversities N_eff of query and database HMMs that was trained on a large set of example queries-template pairs, in an approach similar to the one used in . However, the old calibration is still available as an option in HHsearch.

Should I use the -global option to build MSAs if I am intersted in global alignments to these database HMMs?

Never use -global when building MSAs by iterative searches with HHblits. Remember that global alignments are very greedy alignments, where alignments stretch to the ends of either query of database HMM, no matter what. Global alignments will therefore often lead to non-homologous segments getting included in the MSA, which is catastrophic, as these false segments will lead to many false positive matches when searching with this MSA. But you may use -global when searching (with a single iteration) for global matches of your query HMMs through your customized HHsuite database, for example.

How many iterations of hhblits should I use to search my customized genome database?

One iteration! Doing more than one search iterations in general only makes sense in order to build up a diverse MSA of homologous sequences by searching through the Uniclust30 or a clustered NR databases.

How can I retrieve database A3M or HHM files from my HH-suite database?

For reasons of efficiency, the MSA A3M and HHM model files are packed into two files, <db>_a3m.ffdata and <db>_hhm.ffdata, each with its own index file. The contained files are easy to extract from the packed files. For example, to dump d1aa7a_.a3m in scop70 to standard output, type

ffindex_get scop70_a3m.ffdata scop70_a3m.ffindex d1aa7a_.a3m

You may write the extracted file to a separate file by appending '> d1aa7a_.a3m' to the above command. If you want to extract all files in the ffindex data structures to directory tmp/scop70, type

ffindex_unpack scop70_a3m.ffdata scop70_a3m.ffindex tmp/scop70/ .

How can I build my own UniProt database for HHblits?

The procedure to cluster the UniProt databases is more complicated than building a database for a genome or for the sequences in the PDB. As its first step it involves clustering these huge databases down to 20%-30% sequence identity. The clustering is done in our lab using MMseqs2. We provide this database - the Uniclust30 - including all scripts to build it. These scripts generate A3M files, HHM files, and consensus sequences. Because of the large number of files to generate, these scripts need to be run on a computer cluster and this would require considerable computer savviness.

Will you offer an HHsuite database that includes environmental sequences?

Currently we do not provide environmental databases for the HH-suite. Please use MMseqs2 to search through such datasets. MMseqs2 can easily search through the largest data such, such as assembled proteins from environmental metagenomes. We offer a Soil Reference Catalog (SRC) and a Marine Eukaryotic Reference Catalog, with billions of environmental proteins.

How do I read the sequence/HMM headers in the PDB70 databases?

The pdb70 database is clustered to 70% maximum pairwise sequence identity to reduce redundancy. The PDB identifiers at the end of the description line of sequence X list the sequences with lower experimental resolution that were removed for having higher than 70% sequence identity with PDB sequence X. The asterisks indicate that a multi-atom ligand is bound in the structure.

Why do I get different results when reversing the roles of query and template?

When taking A as the query, the pairwise alignment and significance scores with template B in the database can be different than when B is the query and A the template. By default hhblits, hhsearch, and hhalign add context-specific pseudocounts to the query MSA or HMM whereas the much faster substitution matrix pseudocounts are added to the database HMMs/MSAs for reasons of speed. If you want to combine the significance estimates from the forward and reverse comparisons, we recommend to take the geometric mean of Probability, P-value and E-value. Since the less diverse HMM or MSA will profit more from the context-specific pseudocounts than the more diverse HMM or MSA, it might be even better to take the estimate from the comparison in which the HMM/MSA with the smaller diversity (measured as number of effective sequences, NEFF) is used as the query.

Why do I sometimes get the same database hit twice with different probabilities?

Each line in the summary hit list refers to an alignment with an HMM in the database, not to the database HMM itself. Sometimes, alternative (suboptimal) alignments covering a different part of either the query or the database HMM may appear in the hit list. Usually, both the optimal and the alternative alignments are correct, in particular when the query or the database HMMs represent repeat proteins.

Why do I get different results using HHblits and HHsearch when I search with the same query through the same database?

There are two reasons. First, some hits that hhsearch shows might not have passed the prefilter in hhblits. The option -prepre_smax_thresh <bits> lets you modify the minimum score threshold of the first, gapless alignment prefilter (default is 10 bits). Option -pre_evalue_thresh <E-value> sets the maximum E-value threshold for the second, gapped alignment prefilter (default is 1000). Second, while the probabilities and P-values of hhblits and hhsearch should be the identical for the same matches, the E-values are only similar. The reason is that hhblits heuristically combines the E-values of the second prefilter with the E-values from the full HMM-HMM Viterbi alignment into total E-values . These E-values are slightly better in distinguishing true from false hits, because they combine the partly independent information from two comparisons.

Why do I get different results when I perform 2 and 1+1 iterations with HHblits?

When you do a single iteration and start a second iteration with the A3M output file from the first iteration, the alignments should be nearly the same as if you do two iterations (-n 2) Why not exactly the same? Because when doing multiple iterations in a single hhblits run, alignments for each hit are only calculated once. When the same database HMM is found in later search iterations, the “frozen”, first alignment is used instead of recomputing it. This increases robustness with respect to homologous overextension. However, the significance values can be slightly different between the cases with 2 and 1+1 iterations. The reason is that the HHblits P-value is calculated by a weighted combination of the P-value of the Viterbi pairwise HMM alignment and the prefilter E-value. The prefilter E-values differ in the two cases described, since in the first case the E-value is calculated for the first prefilter search (before the first full search) and in the second case it is calculated for the second prefilter search, performed with the query MSA obtained after the first search iteration.

Why do I get a segmentation fault when calling hhblits on our new machine?

Often, segmentation faults can be caused by too little available memory. HHblits needs memory to read the entire column state file into memory for prefiltering. For Viterbi HMM-HMM alignment it needs about:

Memory Viterbi = Query_length * max_db_seq_length * (num_threads+1) * 5 byte 

Here, num_threads is the number of threads specified with option -cpu <int>. For realignment, HHblits needs in addition:

Memory Realign = Query_length * max_db_seq_length * (num_threads+1) * 8 byte 

But the amount of memory for realignment can be limited using the -maxmem <GB> option, which will cause the realignment to skip the longest sequences.

Make sure to enable the memory overcommitment on your machine. On Linux either add vm.overcommit_memory=1 to /etc/sysctl.conf and then reboot or run the command sysctl vm.overcommit_memory=1.

Under Linux, you can check your memory limit settings using ulimit -a. Make sure you have at least 4GB per thread. Try

ulimit -m 4000000 -v 4000000 -s 8192 -f 4000000

Also try using -cpu 1 -maxmem 1 to reduce memory requirements. If it still does not work, follow the tipps in subsection How do I report a bug? to report your problems to us.

I obtain different results for the same query sequence on the web server and on my local version using HHblits and HHsearch.

• Check the exact command line calls for HHblits and HHsearch from the log file accessible on the results page of the server. If any of the commands differ, rerun your command line version with the same options.

• Check if the query multiple sequence alignments on the server and in your versions contain the same number of sequences and the same diversity. If not, try to find out why (see previous point, for example.)

• For some of the alignments with differing scores, check if the db aligment on the server is the same as the one in your local database. You can download the db MSA by clicking the alignment logo above the query-template pairwise profile-profile alignment.

I find an alignment of my query and a template with the exact same sequence, but they are aligned incorrectly.

I find a gap inserted at residue position 92 in query and position 99 in

template:

Q ss_pred             CcEEEeeCCCC-CCCeEEEEEECCeeEEEEEEECCCCcEEEEEEeeCCHHHHHHHHhhCCCccceeEEEecccCCCCCCe
Q Pr_1             81 GAFLIRESESA-PGDFSLSVKFGNDVQHFKVLRDGAGKYFLWVVKFNSLNELVDYHRSTSVSRNQQIFLRDIEQVPQQPT  159 (217)
Q Consensus        81 ~~~~~~~~~~~-~~~~~~s~~~~~~~~~~~i~~~~~~~~~~~~~~f~s~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~  159 (217)
                    |.+++|.+... ++.+.+ +..++.++|+++.....+.|+.+...|.++..++.++...+....................
T Consensus        81 G~flvR~s~~~~~~~~~l-v~~~~~v~h~~i~~~~~g~~~~~~~~f~sl~~Lv~~y~~~~~~~~~~~~l~~~~~~~~~~~  159 (217)
T 1gri_A           81 GAFLIRESESAPGDFSLS-VKFGNDVQHFKVLRDGAGKYFLWVVKFNSLNELVDYHRSTSVSRNQQIFLRDIEQVPQQPT  159 (217)
T ss_dssp             TCEEEEECSSSTTCEEEE-EEETTEEEEEEEEECSSSCEESSSCEESSHHHHHHHHHHSCSSSSTTCCCCBCCCCCCCCC
T ss_pred             CeEEEEECCCCCCCcEEE-EECCCceeEEEEEECCCCcEEEeeEecCCHHHHHHHhhcCCCccccceeccccccccCCce

The seemingly incorrect alignment is explained by the fact that HHsearch / HHblits aligns profile HMMs, not sequences directly. In this example, the query and template multiple sequence alignments (MSAs) from which the profile HMMs were built are not identical, as one can see for instance from the differing Consensus annotation strings and different predicted secondary structure strings. The underying MSAs probably contain the exact same alignment ambiguity in this region, with some member sequences being shifted by one positions between positions 92 and 99 relative to the master sequence of the MSA.

HHsearch/HHblits output: hit list and pairwise alignments

Summary hit list

Let’s do a search with the human PIP49/FAM69B protein, for which we generated an MSA in query.a3m with two iterations of HHblits in Generating a multiple sequence alignment using HHblits:

Query         sp|Q5VUD6|FA69B_HUMAN Protein FAM69B OS=Homo sapiens GN=FAM69B PE=2 SV=3
Match_columns 431
No_of_seqs    149 out of 272
Neff          5.2 
Searched_HMMs 13730
Date          Wed Jan  4 17:44:24 2012
Command       hhsearch -i query.a3m -d /cluster/user/soeding/databases/scop -cpu 18 

 No Hit                             Prob E-value P-value  Score    SS Cols Query HMM  Template HMM
  1 d1qpca_ d.144.1.7 (A:) Lymphoc  99.7 4.5E-17 3.2E-21  154.3  10.2   99  203-320    56-157 (272)
  2 d1jpaa_ d.144.1.7 (A:) ephb2 r  99.7 4.3E-17 3.1E-21  156.8   8.8   99  203-321    75-177 (299)
  3 d1uwha_ d.144.1.7 (A:) B-Raf k  99.7 5.1E-17 3.7E-21  154.8   7.7  100  203-322    52-154 (276)
  4 d1opja_ d.144.1.7 (A:) Abelson  99.7 6.2E-17 4.5E-21  154.8   8.3  100  203-321    61-164 (287)
  5 d1mp8a_ d.144.1.7 (A:) Focal a  99.6 9.9E-17 7.2E-21  151.3   8.6  100  203-322    56-158 (273)
  6 d1sm2a_ d.144.1.7 (A:) Tyrosin  99.6 1.2E-16 8.8E-21  150.3   8.8   99  203-321    48-150 (263)
  7 d1u59a_ d.144.1.7 (A:) Tyrosin  99.6 2.4E-16 1.7E-20  150.9   9.5   99  203-321    57-158 (285)
  8 d1xbba_ d.144.1.7 (A:) Tyrosin  99.6 2.2E-16 1.6E-20  150.2   8.6   97  203-320    56-155 (277)
  9 d1vjya_ d.144.1.7 (A:) Type I   99.6 2.6E-16 1.9E-20  151.3   8.8   98  204-320    46-156 (303)
 10 d1mqba_ d.144.1.7 (A:) epha2 r  99.6 4.4E-16 3.2E-20  148.0   8.7  193  203-422    57-272 (283)
...
 64 d1j7la_ d.144.1.6 (A:) Type II  97.3 0.00014   1E-08   65.0   6.3   33  292-324   184-216 (263)
 65 d1nd4a_ d.144.1.6 (A:) Aminogl  96.7  0.0012 8.5E-08   58.5   6.6   31  292-322   176-206 (255)
 66 d1nw1a_ d.144.1.8 (A:) Choline  96.6  0.0011 7.8E-08   63.9   5.8   37  203-239    92-128 (395)
 67 d2pula1 d.144.1.6 (A:5-396) Me  95.6  0.0071 5.2E-07   58.3   6.4   32  290-322   222-253 (392)
 68 d1a4pa_ a.39.1.2 (A:) Calcycli  91.7    0.12 8.9E-06   40.0   5.4   62  140-202    18-80  (92)
 69 d1ksoa_ a.39.1.2 (A:) Calcycli  91.2    0.17 1.2E-05   39.5   5.8   56  147-203    28-83  (93)
 70 d1e8aa_ a.39.1.2 (A:) Calcycli  90.5    0.23 1.7E-05   38.3   6.0   56  147-203    27-82  (87)
...
175 d1qxpa2 a.39.1.8 (A:515-702) C  23.7      29  0.0021   28.8   3.8   49  137-197    69-118 (188)
176 d1tuza_ a.39.1.7 (A:) Diacylgl  23.5      55   0.004   25.3   5.3   55  143-201    44-106 (118)
177 d1ggwa_ a.39.1.5 (A:) Cdc4p {F  23.1      26  0.0019   27.0   3.2   66  129-197    35-101 (140)
178 d1topa_ a.39.1.5 (A:) Troponin  22.8      72  0.0052   24.5   6.0   58  140-199    65-123 (162)
179 d1otfa_ d.80.1.1 (A:) 4-oxaloc  22.5      66  0.0048   21.5   5.0   40  267-306    12-53  (59)
180 d1oqpa_ a.39.1.5 (A:) Caltract  22.2      32  0.0023   24.0   3.2   32  165-197     3-34  (77)
181 d1df0a1 a.39.1.8 (A:515-700) C  21.7      43  0.0032   27.0   4.5   51  137-199    67-118 (186)
182 d1zfsa1 a.39.1.2 (A:1-93) Calc  21.1      41   0.003   24.6   3.8   30  170-199     8-38  (93)
183 d1snla_ a.39.1.7 (A:) Nucleobi  20.9      23  0.0016   26.2   2.2   24  174-197    18-41  (99)

The summary hit list that is written to the screen shows the best hits from the database, ordered by the probability of being a true positive (column 4: ’Prob’). The meaning of the columns is the following:

No: the index of the datbase match.

Hit: the first 30 characters of the name line.

Prob: the Probability of template to be a true positive. For the probability of being a true positive, the secondary structure score in column SS is taken into account, together with the raw score in column Score. True positives are defined to be either globally homologous or they are at least homologous in parts, and thereby locally similar in structure. More precisely, the latter criterion demands that the MAXSUB score between query and hit is at least 0.1. In almost all cases the structural similarity will we be due to a global OR LOCAL homology between query and template.

E-value: The E-value gives the average number of false positives (’wrong hits’) with a score better than the one for the template when scanning the database. It is a measure of reliability: E-values near to 0 signify a very reliable hit, an E-value of 10 means about 10 wrong hits are expected to be found in the database with a score at least this good. Note that E-value and P-value are calculated without taking the secondary structure into account!

P-value: The P-value is the E-value divided by the number of sequences in the database. It is the probability that in a pairwise comparison a wrong hit will score at least this good.

Score: the raw score is computed by the Viterbi HMM-HMM alignment excluding the secondary structure score. It is the sum of similarities of the aligned profile columns minus the position-specific gap penalties in bits. The column similarity score is the log-sum-of-odds score (base 2) as described in the original HHsearch paper (Soding, Bioinformatics 2005). The gap penalties are the log2 of the state transition probabilities, e.g. from match state to insert or delete to match state.

SS: the secondary structure score. This score tells you how well the PSIPRED-predicted (3-state) or actual DSSP-determined (8-state) secondary structure sequences agree with each other. PSIPRED confidence values are used in the scoring, low confidences getting less statistical weight.

Cols: the number of aligned Match columns in the HMM-HMM alignment.

Query HMM: the range of aligned match states from the query HMM.

Template HMM: the range of aligned match states from the database/template HMM and, in parenthesis, the number of match states in the database HMM.

HMM-HMM pairwise alignments

The output file d1bpya1.hhr contains the same hit list plus the pairwise HMM alignments. One example is give here:

No 68 
>d1a4pa_ a.39.1.2 (A:) Calcyclin (S100) {Human (Homo sapiens), P11 s100a10, calpactin [TaxId: 9606]}
Probab=91.65  E-value=0.12  Score=40.00  Aligned_cols=62  Identities=16%  Similarity=0.149  Sum_probs=42.0

Q ss_pred             ccCCCCCCcHHHHHHHHHHHHHhhcccCccHHHHHHHHHhhhhccCCCCcCHHHHHHH-HHHHH
Q sp|Q5VUD6|FA69  140 FDKPTRGTSIKEFREMTLSFLKANLGDLPSLPALVGQVLLMADFNKDNRVSLAEAKSV-WALLQ  202 (431)
Q Consensus       140 ~d~p~~g~s~~eF~emv~~~i~~~lg~~~~l~~L~~~~~~~~d~nk~g~vs~~e~~sl-waLlq  202 (431)
                      ||+..-..|.+||.+++.......++.+.+ ...+..++..+|.|+||+|++.|...+ ..|..
T Consensus        18 yd~ddG~is~~El~~~l~~~~~~~~~~~~~-~~~v~~~~~~~D~n~DG~I~F~EF~~li~~l~~   80 (92)
T d1a4pa_          18 FAGDKGYLTKEDLRVLMEKEFPGFLENQKD-PLAVDKIMKDLDQCRDGKVGFQSFFSLIAGLTI   80 (92)
T ss_dssp             HHGGGCSBCHHHHHHHHHHHCHHHHHHSCC-TTHHHHHHHHHCTTSSSCBCHHHHHHHHHHHHH
T ss_pred             HcCCCCEEcHHHHHHHHHHhccccccccCC-HHHHHHHHHHHhCCCCCCCcHHHHHHHHHHHHH
Confidence            444433449999999998876655554332 234566677899999999999997544 44443

This alignment shows an EF hand embedded in a kinase domain in PIP49/FAM69B. The first line, which begins with with a >, contains the name and description line of the template/database HMM. (We use "template HMM" and "matched database HMM" synonymously.) The next line summarizes the main statistics for the alignment: The probability for the query and template HMMs to be homologous (Probab), the E-value, the raw Score, and the number of aligned columns are repeated from the summary hit list. The Identities give the percentage of aligned residue pairs of the query and the template master sequences that are identical. The Similarity is the arithmetic mean of the substitution scores between the aligned residue pairs from the query and template master sequences. The substitution matrix is the same as the one used to calculate the pseudocounts for the database HMMs, by default the Gonnet matrix. (The matrix can be changed with the -Blosom<XX> option.)

The Sum_probs value is the sum over the posterior probabilities of all aligned pairs of match states. These probabilities are calculated by the Forward-Backward algorithm. (They are used by the maximum accuracy algorithm which computes the final alignments.) When the template HMM has secondary structure annotation from DSSP, the sum_probs value runs only over aligned pairs for which the template has a valid DSSP state, not a - sign. A - would indicate that the structural coordinates of that residue are missing in the template. For homology modelling, this special treatment of templates with known structure makes sum_probs a useful feature to use for ranking templates.

The pairwise alignment consists of one or more blocks with the following lines:

Q ss_dssp:      the query secondary structure as determined by DSSP (when available)
Q ss_pred:      the query secondary structure as predicted by PSIPRED (when available)
Q <Q_name>:     the query master sequence
Q Consensus:    the query alignment consensus sequence

The predicted secondary structure states are shown in capital letters if the PSIPRED confidence value is between 0.7 and 1.0, for lower confidence values they are given in lower-case letters. With the option -ssconf, ss_conf lines can be added to the alignments which report the PSIPRED confidence values by numbers between 0 and 9 (as in versions up to 1.5).

The consensus sequence uses capital letters for well conserved columns and lower case for partially conserved columns. Unconserved columns are marked by a tilde ~. Roughly speaking, amino acids that occur with (>+ 60%) probability (before adding pseudocounts) are written as capital letters and amino acids that have (>= 40%) probability are written as lower case letters, where gaps are included in the fraction counts. More precisely, when the gap-corrected amino acid fraction

p_i(a) * N_eff(i)/(N_eff + 1)

is above 0.6 (0.4) an upper (lower) case letter is used for amino acid a. Here, p_i(a) is the emission probability for a in column i, N_eff is the effective number of sequences in the entire multiple alignment (between 1 and 20) and N_eff(i) is the effective number of sequences in the subalignment consisting of those sequences that do not have a gap in column i. These percentages increase approximately inversely proportionally with the fraction of gaps in the column, hence a column with only cysteines and 50% gaps gets a lower case letter.

The line in the middle shows the column score between the query and template amino acid distributions. It gives a valuable indication for the alignment quality.

= : column score below -1.5
- : column score between -1.5 and -0.5
. : column score between -0.5 and +0.5
+ : column score between +0.5 and +1.5
| : column score above   +1.5

A unit of column score corresponds approximately to 0.6 bits. From the column score line the excellent alignment around the conserved D.n.DG.i...E motif in the turn between two helices is evident. The alignment around the gap by contrast scores only a bit better than zero per residue and is therefore not very reliable.

After the template block, which consists of the following lines,

T Consensus:    the template alignment consensus sequence
T <T_name>:     the template domain sequence
T ss_dssp:      the template secondary structure as determined by DSSP (when available)
T ss_pred:      the template secondary structure as predicted by PSIPRED (when available)

The last line in the block (Confidence) reports the reliability of the pairwise query-template alignment. The confidence values are obtained from the posterior probabilities calculated in the Forward-Backward algorithm. A value of 8 indicates a probability that this pair of HMM columns is correctly aligned between 0.8 and 0.9. The Confidence line is only displayed when the -realign option is active.

File formats

Multiple sequence alignment formats

Multiple alignments can be read in A2M, A3M, or aligned FASTA format. (Check the -M option for using an input format different from the default A3M). You can transform MSAs from Clustal or Stockholm format to A3M or aligned FASTA with the reformat.pl utility supplied in this package.

To reformat from Clustal format to A3M:

reformat.pl test.aln test.a3m

or explicitly, if the formats can not be recognized from the extensions:

reformat.pl clu a3m test.clustal test.a3m

To reformat from Stockholm to aligned FASTA:

reformat.pl test.sto test.fas

Example for aligned FASTA format:

>d1a1x__ b.63.1.1 (-) p13-MTCP1 {Human (Homo sapiens)}
PPDHLWVHQEGIYRDEYQRTWVAVVEE--E--T--SF---------LR----------ARVQQIQVPLG-------DAARPSHLLTS-----QL
>gi|6678257|ref|NP_033363.1|:(7-103) T-cell lymphoma breakpoint 1 [Mus musculus]
HPNRLWIWEKHVYLDEFRRSWLPVVIK--S--N--EK---------FQ----------VILRQEDVTLG-------EAMSPSQLVPY-----EL
>gi|7305557|ref|NP_038800.1|:(8-103) T-cell leukemia/lymphoma 1B, 3 [Mus musculus]
PPRFLVCTRDDIYEDENGRQWVVAKVE--T--S--RSpygsrietcIT----------VHLQHMTTIPQ-------EPTPQQPINNN-----SL
>gi|11415028|ref|NP_068801.1|:(2-106) T-cell lymphoma-1; T-cell lymphoma-1A [Homo sapiens]
HPDRLWAWEKFVYLDEKQHAWLPLTIEikD--R--LQ---------LR----------VLLRREDVVLG-------RPMTPTQIGPS-----LL
>gi|7305561|ref|NP_038804.1|:(7-103) T-cell leukemia/lymphoma 1B, 5 [Mus musculus]
----------GIYEDEHHRVWIAVNVE--T--S--HS---------SHgnrietcvt-VHLQHMTTLPQ-------EPTPQQPINNN-----SL
>gi|7305553|ref|NP_038801.1|:(5-103) T-cell leukemia/lymphoma 1B, 1 [Mus musculus]
LPVYLVSVRLGIYEDEHHRVWIVANVE--TshS--SH---------GN----------RRRTHVTVHLW-------KLIPQQVIPFNplnydFL
>gi|27668591|ref|XP_234504.1|:(7-103) similar to Chain A, Crystal Structure Of Murine Tcl1
-PDRLWLWEKHVYLDEFRRSWLPIVIK--S--N--GK---------FQ----------VIMRQKDVILG-------DSMTPSQLVPY-----EL
>gi|27668589|ref|XP_234503.1|:(9-91) similar to T-cell leukemia/lymphoma 1B, 5;
-PHILTLRTHGIYEDEHHRLWVVLDLQ--A--ShlSF---------SN----------RLLIYLTVYLQqgvafplESTPPSPMNLN-----GL
>gi|7305559|ref|NP_038802.1|:(8-102) T-cell leukemia/lymphoma 1B, 4 [Mus musculus] 
PPCFLVCTRDDIYEDEHGRQWVAAKVE--T--S--SH---------SPycskietcvtVHLWQMTTLFQ-------EPSPDSLKTFN-----FL
>gi|7305555|ref|NP_038803.1|:(9-102) T-cell leukemia/lymphoma 1B, 2 [Mus musculus]
---------PGFYEDEHHRLWMVAKLE--T--C--SH---------SPycnkietcvtVHLWQMTRYPQ-------EPAPYNPMNYN-----FL

The sequence name and its description must be contained in a single name line beginning with the > symbol and followed directly by the sequence name. The residue data is contained in one or more lines of arbitrary length following the name line. No empty lines should be used. In aligned FASTA the gaps are written with - and the n'th letter of each sequence (except newlines) is understood to build the n'th column of the multiple alignment.

The same alignment in A2M format looks like this:

>d1a1x__ b.63.1.1 (-) p13-MTCP1 {Human (Homo sapiens)}
PPDHLWVHQEGIYRDEYQRTWVAVVEE..E..T..SF.........LR..........ARVQQIQVPLG.......DAARPSHLLTS.....QL
>gi|6678257|ref|NP_033363.1|:(7-103) T-cell lymphoma breakpoint 1 [Mus musculus]
HPNRLWIWEKHVYLDEFRRSWLPVVIK..S..N..EK.........FQ..........VILRQEDVTLG.......EAMSPSQLVPY.....EL
>gi|7305557|ref|NP_038800.1|:(8-103) T-cell leukemia/lymphoma 1B, 3 [Mus musculus]
PPRFLVCTRDDIYEDENGRQWVVAKVE..T..S..RSpygsrietcIT..........VHLQHMTTIPQ.......EPTPQQPINNN.....SL
>gi|11415028|ref|NP_068801.1|:(2-106) T-cell lymphoma-1; T-cell lymphoma-1A [Homo sapiens]
HPDRLWAWEKFVYLDEKQHAWLPLTIEikD..R..LQ.........LR..........VLLRREDVVLG.......RPMTPTQIGPS.....LL
>gi|7305561|ref|NP_038804.1|:(7-103) T-cell leukemia/lymphoma 1B, 5 [Mus musculus]
----------GIYEDEHHRVWIAVNVE..T..S..HS.........SHgnrietcvt.VHLQHMTTLPQ.......EPTPQQPINNN.....SL
>gi|7305553|ref|NP_038801.1|:(5-103) T-cell leukemia/lymphoma 1B, 1 [Mus musculus]
LPVYLVSVRLGIYEDEHHRVWIVANVE..TshS..SH.........GN..........RRRTHVTVHLW.......KLIPQQVIPFNplnydFL
>gi|27668591|ref|XP_234504.1|:(7-103) similar to Chain A, Crystal Structure Of Murine Tcl1
-PDRLWLWEKHVYLDEFRRSWLPIVIK..S..N..GK.........FQ..........VIMRQKDVILG.......DSMTPSQLVPY.....EL
>gi|27668589|ref|XP_234503.1|:(9-91) similar to T-cell leukemia/lymphoma 1B, 5;
-PHILTLRTHGIYEDEHHRLWVVLDLQ..A..ShlSF.........SN..........RLLIYLTVYLQqgvafplESTPPSPMNLN.....GL
>gi|7305559|ref|NP_038802.1|:(8-102) T-cell leukemia/lymphoma 1B, 4 [Mus musculus] 
PPCFLVCTRDDIYEDEHGRQWVAAKVE..T..S..SH.........SPycskietcvtVHLWQMTTLFQ.......EPSPDSLKTFN.....FL
>gi|7305555|ref|NP_038803.1|:(9-102) T-cell leukemia/lymphoma 1B, 2 [Mus musculus]
---------PGFYEDEHHRLWMVAKLE..T..C..SH.........SPycnkietcvtVHLWQMTRYPQ.......EPAPYNPMNYN.....FL

A2M format is derived from aligned FASTA format. It looks very similar, but it distinguishes between match/delete columns and insert columns. This information is important to uniquely specify how an alignment is transformed into an HMM. The match/delete columns use upper case letters for residues and the - symbol for deletions (gaps). The insert columns use lower case letters for the inserted residues. Gaps aligned to inserted residues are written as . Lines beginning with a hash # symbol will be treated as commentary lines in HHsearch/HHblits (see below).

The same alignment in A3M:

>d1a1x__ b.63.1.1 (-) p13-MTCP1 {Human (Homo sapiens)}
PPDHLWVHQEGIYRDEYQRTWVAVVEEETSFLRARVQQIQVPLGDAARPSHLLTSQL
>gi|6678257|ref|NP_033363.1|:(7-103) T-cell lymphoma breakpoint 1 [Mus musculus]
HPNRLWIWEKHVYLDEFRRSWLPVVIKSNEKFQVILRQEDVTLGEAMSPSQLVPYEL
>gi|7305557|ref|NP_038800.1|:(8-103) T-cell leukemia/lymphoma 1B, 3 [Mus musculus]
PPRFLVCTRDDIYEDENGRQWVVAKVETSRSpygsrietcITVHLQHMTTIPQEPTPQQPINNNSL
>gi|11415028|ref|NP_068801.1|:(2-106) T-cell lymphoma-1; T-cell lymphoma-1A [Homo sapiens]
HPDRLWAWEKFVYLDEKQHAWLPLTIEikDRLQLRVLLRREDVVLGRPMTPTQIGPSLL
>gi|7305561|ref|NP_038804.1|:(7-103) T-cell leukemia/lymphoma 1B, 5 [Mus musculus]
----------GIYEDEHHRVWIAVNVETSHSSHgnrietcvtVHLQHMTTLPQEPTPQQPINNNSL
>gi|7305553|ref|NP_038801.1|:(5-103) T-cell leukemia/lymphoma 1B, 1 [Mus musculus]
LPVYLVSVRLGIYEDEHHRVWIVANVETshSSHGNRRRTHVTVHLWKLIPQQVIPFNplnydFL
>gi|27668591|ref|XP_234504.1|:(7-103) similar to Chain A, Crystal Structure Of Murine Tcl1
-PDRLWLWEKHVYLDEFRRSWLPIVIKSNGKFQVIMRQKDVILGDSMTPSQLVPYEL
>gi|27668589|ref|XP_234503.1|:(9-91) similar to T-cell leukemia/lymphoma 1B, 5;
-PHILTLRTHGIYEDEHHRLWVVLDLQAShlSFSNRLLIYLTVYLQqgvafplESTPPSPMNLNGL
>gi|7305559|ref|NP_038802.1|:(8-102) T-cell leukemia/lymphoma 1B, 4 [Mus musculus] 
PPCFLVCTRDDIYEDEHGRQWVAAKVETSSHSPycskietcvtVHLWQMTTLFQEPSPDSLKTFNFL
>gi|7305555|ref|NP_038803.1|:(9-102) T-cell leukemia/lymphoma 1B, 2 [Mus musculus]
---------PGFYEDEHHRLWMVAKLETCSHSPycnkietcvtVHLWQMTRYPQEPAPYNPMNYNFL

The A3M format is a condensed version of A2M format. It is obtained by omitting all . symbols from A2M format. Hence residues emitted by Match states of the HMM are in upper case, residues emitted by Insert states are in lower case and deletions are written -. A3M-formatted alignments can be reformatted to other formats like FASTA or A2M with the reformat.pl utility:

reformat.pl test.a3m test.a2m

Lines beginning with a hash # symbol will be treated as commentary lines in HHsearch/HHblits (see below). Please note that A3M, though very practical and space-efficient, is not a standard format, and the name A3M is our personal invention.

Secondary structure information in A3M/A2M or FASTA MSAs for HHsearch/HHblits

The alignments read in by HHblits, HHsearch or HHmake can also contain secondary structure information. This information can be included in sequences with special names, like in this A3M file:

>ss_dssp
CCSEEEEEETTEEEETTSCEEEEEEEECSSCEEEEEECCCCCCCSCCCHHHHTTCSSCSEEEEETTTEEEETTSC
>aa_dssp
PPDHLWVHQEGIYRDEYQRTWVAVVEEETSFLRARVQQIQVPLGDAARPSHLLTSQLPLMWQLYPEERYMDNNSR
>aa_pred 
PPDHLWVHQEGIYRDEYQRTWVAVVEEETSFLRARVQQIQVPLGDAARPSHLLTSQLPLMWQLYPEERYMDNNSR
>ss_pred 
CCCEEEEECCCEECCCCCEEEEEEEEECCCCCCEEEEEEECCCCCCCCCCCCCCCCCCCEEEECCCCCEECCCCC
>ss_conf 
987689961870104587078999970578640132153103788788777774424614787217702035631
>d1a1x__ b.63.1.1 (-) p13-MTCP1 {Human (Homo sapiens)}
PPDHLWVHQEGIYRDEYQRTWVAVVEEETSFLRARVQQIQVPLGDAARPSHLLTSQLPLMWQLYPEERYMDNNSR
>gi|6678257|ref|NP_033363.1|:(7-103) T-cell lymphoma breakpoint 1 [Mus musculus]
HPNRLWIWEKHVYLDEFRRSWLPVVIKSNEKFQVILRQEDVTLGEAMSPSQLVPYELPLMWQLYPKDRYRSCDSM
>gi|7305557|ref|NP_038800.1|:(8-103) T-cell leukemia/lymphoma 1B, 3 [Mus musculus]
PPRFLVCTRDDIYEDENGRQWVVAKVETSRSpygsrietcITVHLQHMTTIPQEPTPQQPINNNSLPTMWRLESMNTYTGTDGT
>gi|11415028|ref|NP_068801.1|:(2-106) T-cell lymphoma-1; T-cell lymphoma-1A [Homo sapiens]
HPDRLWAWEKFVYLDEKQHAWLPLTIEikDRLQLRVLLRREDVVLGRPMTPTQIGPSLLPIMWQLYPDGRYRSSDSS

The sequence with name >ss_dssp contains the 8-state DSSP-determined secondary structure. >aa_dssp and >aa_pred contain the same residues as the query sequence (>d1a1x__ in this case). They are optional and used merely to check whether the secondary structure states have correctly been assigned to the alignment. >ss_pred contains the 3-state secondary structure predicted by PSIPRED, and >ss_conf contains the corresponding confidence values. The query sequence is the first sequence that does not start with a special name. It is not marked explicitly.

Name lines in alignments

If you would like to create HMMs from alignments with a specified name which differ from the name of the first sequence, you can do so by adding name lines to your FASTA, A2M, or A3M alignment:

#PF02043 Bac_chlorC:  Bacteriochlorophyll C binding protein
>ss_pred
CCCCHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHHCCCCCCCCCCCCCCCCCCCCCCCCCCCCHHHHHHHCC
>ss_conf
9863234658887677777750019999999998878886403445886666775655576678777660667633039
>CSMA_CHLAU/1-79
ATRGWFSESSAQVAQIGDIMFQGHWQWVSNALQATAAAVDNINRNAYPGVSRSGSGEGAFSSSPSNGFRPKRIRSRFNR
>CSMA_CHLPH/2-51
NGGGVFTDILAASGRIFEVMVEGHWATVGYLFDSLGKGVSRINQNAYGNM-----------------------------
...

When creating an HMM from an A3M file with hhmake, the first word of the name line is used as the name and file name of the HMM (PF02043 in this case). The following is an optional description. The descriptions will appear in the hit list and alignment section of the search results. The name lines can be arbitrarily long and there can be any number of name/description lines included, marked by a ’#’ as the first character in the line. Note that name lines are read by HHmake but are not a part of the standard definition of the FASTA or A2M format.

HHsearch/HHblits model format (hhm-format)

HMMs can be read by HHsearch/HHblits in HHsuite’s own hhm format. Performance is severely reduced when using HMMER-format, because HMMER adds more pseudocounts than HHsearch/HHblits that we cannot remove anymore. We therefore strongly advise against using models from HMMER with HHsuite. We recommend to get the original MSA files and generate hhm-formatted models using hhmake, as described in subsection Generating a multiple sequence alignment using HHblits.

If you absolutely need to use HMMER format, you can convert it to hhm format with hhmake:

hhmake -i test.hmm -o test.hhm

This works only for a single HMM per file, not for concatenated HMMs. You may add predicted secondary structure to the hmm file with addss.pl as for hhm format.

HHsearch/HHblits uses a format HMM that is unchanged since HHsearch version 1.5. This is the example of an HHM model file produced by HHmake:

HHsearch 1.5
NAME  d1mvfd_ b.129.1.1 (D:) MazE {Escherichia coli}
FAM   b.129.1.1
FILE  d1mvfd_
COM   hhmake1 -i d1mvfd_.a3m -o test.hhm 
DATE  Wed May 14 10:41:06 2011
LENG  44 match states, 44 columns in multiple alignment
FILT  32 out of 35 sequences passed filter (-id 90 -cov 0 -qid 0 -qsc -20.00 -diff 100)
NEFF  4.0 
SEQ
>ss_dssp
CBCEEETTEEEEECCHHHHHHTTCCTTCBEEEEEETTEEEEEEC
>ss_pred
CCCCCCCCCCCCCCHHHHHHHHCCCCCCEEEEEEECCEEEEEEC
>ss_conf
93233467666600578899808998986889874993798739
>Consensus
sxIxKWGNSxAvRlPaxlxxxlxlxxgdxixxxxxxxxivlxPv
>d1mvfd_ b.129.1.1 (D:) MazE {Escherichia coli}
SSVKRWGNSPAVRIPATLMQALNLNIDDEVKIDLVDGKLIIEPV
>gi|10176344|dbj|BAB07439.1|:(1-43) suppressor of ppGpp-regulated growth inhibitor [Bacillus halodurans]
TTIQKWGNSLAVRIPNHYAKHINVTQGSEIELSLgSDQTIILKP-
>gi|50120611|ref|YP_049778.1|:(3-43) suppressor of growth inhibitory protein ChpA [Erwinia carotovora]
-TVKKWGNSPAIRLSSSVMQAFDMTFNDSFDMEIRETEIALIP-
>gi|44064461|gb|EAG93225.1|:(2-42) unknown [environmental sequence]
-SVVKWGSYLAVRLPAELVLELGLKEGDEIDLVKDDGPVRVR--
>gi|31442758|gb|AAP55635.1|:(1-44) PemI-like protein [Pediococcus acidilactici]
TRLAKWGNSKAARIPSQIIKQLKLDDNQDMTITIENGSIVLTPI
>gi|44419085|gb|EAJ13619.1|:(3-43) unknown [environmental sequence]
SAIQKWGNSAAVRLPAVLLEQIDASVGSSLNADVRPDGVLLSP-
>gi|24376549|gb|AAN57947.1|:(3-44) putative cell growth regulatory protein [Streptococcus mutans UA159]
SAINKWGNSSAIRLPKQLVQELQLQTNDVLDYKVSGNKIILEKV
>gi|11344928|gb|AAG34554.1|:(1-44) MazE [Photobacterium profundum]
TQIRKIGNSLGSIIPATFIRQLELAEGAEIDVKTVDGKIVIEPI
>gi|45681193|ref|ZP_00192636.1|:(2-44) COG2336: Growth regulator [Mesorhizobium sp. BNC1]
-TIRKIGNSEGVILPKELLDRHNLKTGDALAIVEEGSDLVLKPV
#
NULL  3706 5728 4211 4064 4839 3729 4763 4308 4069  3323  5509 4640 4464 4937 4285 4423 3815 3783 6325 4665
HMM   A    C    D    E    F    G    H    I    K     L     M    N    P    Q    R    S    T    V    W    Y
      M->M M->I M->D I->M I->I D->M D->D Neff NeffI NeffD
      0    *    *    0    *    0    *    *    *     *
S 1   *    *    *    *    *    *    *    *    *     *     *    *    *    *    *    1012 988  *    *    *   1
      0    *    *    *    *    *    *    2817 0     0

S 2   2307 *    *    *    *    *    *    *    *     *     *    *    *    3178 3009 2179 1546 *    *    *   2
      0    *    *    *    *    *    *    3447 0     0

V 3   *    *    *    *    *    *    *    917  *     3009  *    *    *    *    *    *    *    1530 *    *   3
      0    *    *    *    *    *    *    3447 0     0
  .
  .
  .
V 44  *    *    *    *    *    *    *    1309 *     *     *    *    *    *    *    *    *    745  *    *   44
      0    *    *    0    *    *    *    2533 0     0

//

The first line (HHsearch 1.5) gives the format version, which corresponds to the HHsearch version for which this format was first introduced. Newer versions of HHsearch/HHblits may use previous format versions. The NAME line gives the name of the HMM and an optional description. The first 30 characters of this field are used in the summary hit list of the search results in hhr format, the full name line is given above the query-template alignments of the search results. The FAM line contains the family if the sequence is from SCOP of PFAM (used for calibration). COM is the command that was used to generate the file. NEFF is the diversity of the alignment, calculated as exp of the negative entropy averaged over all columns of the alignment.

The SEQ section contains a number of aligned, representative (pseudo) sequences in A3M format and is terminated with a line containing only a #. The first sequence represents the DSSP secondary structure (if available, i.e., if contained in the A3M or FASTA alignment from which the HMM model was built), the second and third sequences contain the predicted secondary structure and the corresponding confidence values in the range 0–9 (if available). The fourth sequence is the consensus annotation sequence that is shown in the pairwise query-template alignments in the hhsearch output. The first real sequence after the pseudo sequences is the seed or master sequence from which the alignment was built (>d1mvfd_, in our example). If the alignment does not represent a single master sequence but an entire family, as in the case of PFAM alignments for example, the first real sequence may be a consensus sequence calculated for the entire alignment. This master sequence is shown in the pairwise query-template alignments in the hhsearch output.

The next line specifies the null model frequencies, which are extracted from the selected substitution matrix used to add pseudocounts. Each of the positive integers is equal to 1000 times the negative logarithm of the amino acid frequency (which is between 0 and 1):

-1000 * log_2(frequency)

After the two annotation lines that specify the order of columns for the emission and transition probabilities that follow, there is a line which is not currently read by HHsearch and that lists the transition frequencies from the begin state to the first Match state, Insert state and Delete state.

The last block contains two lines for each column of the HMM. The first line starts with the amino acid in the master sequence at that column in the HMM and the column number. Following are 20 positive integers representing the match state amino acid emission frequencies (see equation above). Asterisks * stand for a frequency of 0 (which would otherwise be represented by 99999). Please note that, unlike in HMMER format, the emission frequencies do not contain pseudo-counts in the HHsearch model format. The second line contains the seven transition frequencies coded as above. The three local diversities, Neff_M, Neff_I, and Neff_D are given in units of 0.001 (see next paragraph). The end of the model is indicated by a line containing only //.

Summary of command-line parameters

This is just a brief summary of command line parameters for the various binaries and perl scripts as they are displayed by the programs when calling them without command line parameters. On the help pages of the HHpred/HHblits web server you can find more detailed explanations about some of the input parameters (Parameters section) and about how to interpret the output (Results section). The FAQ section contains valuable practical hints on topics such as how to validate marginally significant database matches or how to avoid high-scoring false positives.

hhblits – HMM-HMM-based lighting-fast iterative sequence search

HHblits is a sensitive, general-purpose, iterative sequence search tool that represents both query and database sequences by HMMs. You can search HHsuite databases starting with a single query sequence, a multiple sequence alignment (MSA), or an HMM. HHblits prints out a ranked list of database HMMs/MSAs and can also generate an MSA by merging the significant database HMMs/MSAs onto the query MSA.

The binary hhblits_omp supports the parallelization over several queries with OpenMP. In this case the input needs to be the basename of an ffindex with MSA’s. The outputs are treated as the basenames for output ffindices.

Assume the ffindex database /home/user/databases/scop_a3m.ff{data,index}, the corresponding basename is /home/user/databases/scop_a3m

Usage: hhblits -i query [options] 
 -i <file>      input/query: single sequence or multiple sequence alignment (MSA)
                in a3m, a2m, or FASTA format, or HMM in hhm format

<file> may be 'stdin' or 'stdout' throughout.

Options:                                                                        
 -d <name>      database name (e.g. uniclust30_2018_08)                        
                Multiple databases may be specified with '-d <db1> -d <db2> ...'
 -n     [1,8]   number of iterations (default=2)                               
 -e     [0,1]   E-value cutoff for inclusion in result alignment (def=0.001)       

Input alignment format:                                                       
 -M a2m         use A2M/A3M (default): upper case = Match; lower case = Insert;
               ' -' = Delete; '.' = gaps aligned to inserts (may be omitted)   
 -M first       use FASTA: columns with residue in 1st sequence are match states
 -M [0,100]     use FASTA: columns with fewer than X% gaps are match states   
 -tags/-notags  do NOT / do neutralize His-, C-myc-, FLAG-tags, and trypsin 
                recognition sequence to background distribution (def=-notags)  

Output options: 
 -o <file>      write results in standard format to file (default=<infile.hhr>)
 -oa3m <file>   write result MSA with significant matches in a3m format
 -opsi <file>   write result MSA of significant matches in PSI-BLAST format
 -ohhm <file>   write HHM file for result MSA of significant matches
 -oalis <name>  write MSAs in A3M format after each iteration
 -add_cons      generate consensus sequence as master sequence of query MSA (default=don't)
 -hide_cons     don't show consensus sequence in alignments (default=show)     
 -hide_pred     don't show predicted 2ndary structure in alignments (default=show)
 -hide_dssp     don't show DSSP 2ndary structure in alignments (default=show)  
 -show_ssconf   show confidences for predicted 2ndary structure in alignments
 -Ofas <file>   write pairwise alignments in FASTA xor A2M (-Oa2m) xor A3M (-Oa3m) format   
 -seq <int>     max. number of query/template sequences displayed (default=1)  
 -aliw <int>    number of columns per line in alignment list (default=80)       
 -p [0,100]     minimum probability in summary and alignment list (default=20)  
 -E [0,inf[     maximum E-value in summary and alignment list (default=1E+06)      
 -Z <int>       maximum number of lines in summary hit list (default=500)        
 -z <int>       minimum number of lines in summary hit list (default=10)        
 -B <int>       maximum number of alignments in alignment list (default=500)     
 -b <int>       minimum number of alignments in alignment list (default=10)     

Prefilter options                                                               
 -noprefilt                disable all filter steps                                        
 -noaddfilter              disable all filter steps (except for fast prefiltering)         
 -maxfilt                  max number of hits allowed to pass 2nd prefilter (default=20000)   
 -min_prefilter_hits       min number of hits to pass prefilter (default=100)               
 -prepre_smax_thresh       min score threshold of ungapped prefilter (default=10)               
 -pre_evalue_thresh        max E-value threshold of Smith-Waterman prefilter score (default=1000.0)
 -pre_bitfactor            prefilter scores are in units of 1 bit / pre_bitfactor (default=4)
 -pre_gap_open             gap open penalty in prefilter Smith-Waterman alignment (default=20)
 -pre_gap_extend           gap extend penalty in prefilter Smith-Waterman alignment (default=4)
 -pre_score_offset         offset on sequence profile scores in prefilter S-W alignment (default=50)

Filter options applied to query MSA, database MSAs, and result MSA              
 -all           show all sequences in result MSA; do not filter result MSA      
 -id   [0,100]  maximum pairwise sequence identity (def=90)
 -diff [0,inf[  filter MSAs by selecting most diverse set of sequences, keeping 
                at least this many seqs in each MSA block of length 50 
                Zero and non-numerical values turn off the filtering. (def=1000) 
 -cov  [0,100]  minimum coverage with master sequence (%) (def=0)             
 -qid  [0,100]  minimum sequence identity with master sequence (%) (def=0)    
 -qsc  [0,100]  minimum score per column with master sequence (default=-20.0)    
 -neff [1,inf]  target diversity of multiple sequence alignment (default=off)   
 -mark          do not filter out sequences marked by ">@"in their name line  

HMM-HMM alignment options:                                                       
 -norealign           do NOT realign displayed hits with MAC algorithm (def=realign)   
 -realign_old_hits    realign hits from previous iterations                          
 -mact [0,1[          posterior prob threshold for MAC realignment controlling greedi- 
                      ness at alignment ends: 0:global >0.1:local (default=0.35)       
 -glob/-loc           use global/local alignment mode for searching/ranking (def=local)
 -realign             realign displayed hits with max. accuracy (MAC) algorithm 
 -realign_max <int>   realign max. <int> hits (default=500)                        
 -ovlp <int>          banded alignment: forbid <ovlp> largest diagonals |i-j| of DP matrix (def=0)
 -alt <int>           show up to this many alternative alignments with raw score > smin(def=4)  
 -smin <float>        minimum raw score for alternative alignments (def=20.0)  
 -shift [-1,1]        profile-profile score offset (def=-0.03)                         
 -corr [0,1]          weight of term for pair correlations (def=0.10)                
 -sc   <int>          amino acid score         (tja: template HMM at column j) (def=1)
              0       = log2 Sum(tja*qia/pa)   (pa: aa background frequencies)    
              1       = log2 Sum(tja*qia/pqa)  (pqa = 1/2*(pa+ta) )               
              2       = log2 Sum(tja*qia/ta)   (ta: av. aa freqs in template)     
              3       = log2 Sum(tja*qia/qa)   (qa: av. aa freqs in query)        
              5       local amino acid composition correction                     
 -ssm {0,..,4}        0:   no ss scoring                                             
                      1,2: ss scoring after or during alignment  [default=2]         
                      3,4: ss scoring after or during alignment, predicted vs. predicted
 -ssw [0,1]           weight of ss score  (def=0.11)                                  
 -ssa [0,1]           ss confusion matrix = (1-ssa)*I + ssa*psipred-confusion-matrix [def=1.00)
 -wg                  use global sequence weighting for realignment!                   

Gap cost options:                                                                
 -gapb [0,inf[  Transition pseudocount admixture (def=1.00)                     
 -gapd [0,inf[  Transition pseudocount admixture for open gap (default=0.15)    
 -gape [0,1.5]  Transition pseudocount admixture for extend gap (def=1.00)      
 -gapf ]0,inf]  factor to increase/reduce gap open penalty for deletes (def=0.60) 
 -gapg ]0,inf]  factor to increase/reduce gap open penalty for inserts (def=0.60) 
 -gaph ]0,inf]  factor to increase/reduce gap extend penalty for deletes(def=0.60)
 -gapi ]0,inf]  factor to increase/reduce gap extend penalty for inserts(def=0.60)
 -egq  [0,inf[  penalty (bits) for end gaps aligned to query residues (def=0.00) 
 -egt  [0,inf[  penalty (bits) for end gaps aligned to template residues (def=0.00)

Pseudocount (pc) options:                                                        
 Context specific hhm pseudocounts:
  -pc_hhm_contxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=2) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               3: CSBlast admixture:   tau = a(1+b)/(Neff[i]+b)                 
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_hhm_contxt_a  [0,1]        overall pseudocount admixture (def=0.9)                        
  -pc_hhm_contxt_b  [1,inf[      Neff threshold value for mode 2 (def=4.0)                      
  -pc_hhm_contxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context independent hhm pseudocounts (used for templates; used for query if contxt file is not available):
  -pc_hhm_nocontxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=2) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_hhm_nocontxt_a  [0,1]        overall pseudocount admixture (def=1.0)                        
  -pc_hhm_nocontxt_b  [1,inf[      Neff threshold value for mode 2 (def=1.5)                      
  -pc_hhm_nocontxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context specific prefilter pseudocounts:
  -pc_prefilter_contxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=3) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               3: CSBlast admixture:   tau = a(1+b)/(Neff[i]+b)                 
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_prefilter_contxt_a  [0,1]        overall pseudocount admixture (def=0.8)                        
  -pc_prefilter_contxt_b  [1,inf[      Neff threshold value for mode 2 (def=2.0)                      
  -pc_prefilter_contxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context independent prefilter pseudocounts (used if context file is not available):
  -pc_prefilter_nocontxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=2) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_prefilter_nocontxt_a  [0,1]        overall pseudocount admixture (def=1.0)                        
  -pc_prefilter_nocontxt_b  [1,inf[      Neff threshold value for mode 2 (def=1.5)                      
  -pc_prefilter_nocontxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context-specific pseudo-counts:                                                  
  -nocontxt      use substitution-matrix instead of context-specific pseudocounts 
  -contxt <file> context file for computing context-specific pseudocounts (default=${HHLIB}/data/context_data.crf)
  -csw  [0,inf]  weight of central position in cs pseudocount mode (def=1.6)
  -csb  [0,1]    weight decay parameter for positions in cs pc mode (def=0.9)

Other options:                                                                   
 -v <int>       verbose mode: 0:no screen output  1:only warings  2: verbose (def=2)
 -neffmax ]1,20] skip further search iterations when diversity Neff of query MSA 
                becomes larger than neffmax (default=10.0)
 -cpu <int>     number of CPUs to use (for shared memory SMPs) (default=2)      
 -scores <file> write scores for all pairwise comparisons to file               
 -atab   <file> write all alignments in tabular layout to file                   
 -maxres <int>  max number of HMM columns (def=20001)             
 -maxmem [1,inf[ limit memory for realignment (in GB) (def=3.0)          

Example: hhblits -i query.fas -oa3m query.a3m -n 1  

hhsearch – search a database of HMMs with a query MSA or HMM

Usage: hhsearch -i query -d database [options]                       
 -i <file>      input/query multiple sequence alignment (a2m, a3m, FASTA) or HMM

<file> may be 'stdin' or 'stdout' throughout.
Options:                                                                        
 -d <name>      database name (e.g. pdb70)                        
                Multiple databases may be specified with '-d <db1> -d <db2> ...'
 -e     [0,1]   E-value cutoff for inclusion in result alignment (def=0.001)       

Input alignment format:                                                       
 -M a2m         use A2M/A3M (default): upper case = Match; lower case = Insert;
               '-' = Delete; '.' = gaps aligned to inserts (may be omitted)   
 -M first       use FASTA: columns with residue in 1st sequence are match states
 -M [0,100]     use FASTA: columns with fewer than X% gaps are match states   
 -tags/-notags  do NOT / do neutralize His-, C-myc-, FLAG-tags, and trypsin 
                recognition sequence to background distribution (def=-notags)  

Output options: 
 -o <file>      write results in standard format to file (default=<infile.hhr>)
 -oa3m <file>   write result MSA with significant matches in a3m format
 -opsi <file>   write result MSA of significant matches in PSI-BLAST format
 -ohhm <file>   write HHM file for result MSA of significant matches
 -add_cons      generate consensus sequence as master sequence of query MSA (default=don't)
 -hide_cons     don't show consensus sequence in alignments (default=show)     
 -hide_pred     don't show predicted 2ndary structure in alignments (default=show)
 -hide_dssp     don't show DSSP 2ndary structure in alignments (default=show)  
 -show_ssconf   show confidences for predicted 2ndary structure in alignments
 -Ofas <file>   write pairwise alignments in FASTA xor A2M (-Oa2m) xor A3M (-Oa3m) format   
 -seq <int>     max. number of query/template sequences displayed (default=1)  
 -aliw <int>    number of columns per line in alignment list (default=80)       
 -p [0,100]     minimum probability in summary and alignment list (default=20)  
 -E [0,inf[     maximum E-value in summary and alignment list (default=1E+06)      
 -Z <int>       maximum number of lines in summary hit list (default=500)        
 -z <int>       minimum number of lines in summary hit list (default=10)        
 -B <int>       maximum number of alignments in alignment list (default=500)     
 -b <int>       minimum number of alignments in alignment list (default=10)     

Filter options applied to query MSA, database MSAs, and result MSA              
 -all           show all sequences in result MSA; do not filter result MSA      
 -id   [0,100]  maximum pairwise sequence identity (def=90)
 -diff [0,inf[  filter MSAs by selecting most diverse set of sequences, keeping 
                at least this many seqs in each MSA block of length 50 
                Zero and non-numerical values turn off the filtering. (def=100) 
 -cov  [0,100]  minimum coverage with master sequence (%) (def=0)             
 -qid  [0,100]  minimum sequence identity with master sequence (%) (def=0)    
 -qsc  [0,100]  minimum score per column with master sequence (default=-20.0)    
 -neff [1,inf]  target diversity of multiple sequence alignment (default=off)   
 -mark          do not filter out sequences marked by ">@"in their name line  

HMM-HMM alignment options:                                                       
 -norealign          do NOT realign displayed hits with MAC algorithm (def=realign)   
 -ovlp <int>         banded alignment: forbid <ovlp> largest diagonals |i-j| of DP matrix (def=0)
 -mact [0,1[         posterior prob threshold for MAC realignment controlling greedi- 
                     ness at alignment ends: 0:global >0.1:local (default=0.35)       
 -glob/-loc          use global/local alignment mode for searching/ranking (def=local)
 -realign            realign displayed hits with max. accuracy (MAC) algorithm 
 -excl <range>       exclude query positions from the alignment, e.g. '1-33,97-168' 
 -realign_max <int>  realign max. <int> hits (default=500)                        
 -alt <int>          show up to this many alternative alignments with raw score > smin(def=4)  
 -smin <float>       minimum raw score for alternative alignments (def=20.0)  
 -shift [-1,1]       profile-profile score offset (def=-0.03)                         
 -corr [0,1]         weight of term for pair correlations (def=0.10)                
 -sc   <int>         amino acid score         (tja: template HMM at column j) (def=1)
        0       = log2 Sum(tja*qia/pa)   (pa: aa background frequencies)    
        1       = log2 Sum(tja*qia/pqa)  (pqa = 1/2*(pa+ta) )               
        2       = log2 Sum(tja*qia/ta)   (ta: av. aa freqs in template)     
        3       = log2 Sum(tja*qia/qa)   (qa: av. aa freqs in query)        
        5       local amino acid composition correction                     
 -ssm {0,..,4}    0:   no ss scoring                                             
                1,2: ss scoring after or during alignment  [default=2]         
                3,4: ss scoring after or during alignment, predicted vs. predicted
 -ssw [0,1]          weight of ss score  (def=0.11)                                  
 -ssa [0,1]          SS substitution matrix = (1-ssa)*I + ssa*full-SS-substition-matrix [def=1.00)
 -wg                 use global sequence weighting for realignment!                   

Gap cost options:                                                                
 -gapb [0,inf[  Transition pseudocount admixture (def=1.00)                     
 -gapd [0,inf[  Transition pseudocount admixture for open gap (default=0.15)    
 -gape [0,1.5]  Transition pseudocount admixture for extend gap (def=1.00)      
 -gapf ]0,inf]  factor to increase/reduce gap open penalty for deletes (def=0.60) 
 -gapg ]0,inf]  factor to increase/reduce gap open penalty for inserts (def=0.60) 
 -gaph ]0,inf]  factor to increase/reduce gap extend penalty for deletes(def=0.60)
 -gapi ]0,inf]  factor to increase/reduce gap extend penalty for inserts(def=0.60)
 -egq  [0,inf[  penalty (bits) for end gaps aligned to query residues (def=0.00) 
 -egt  [0,inf[  penalty (bits) for end gaps aligned to template residues (def=0.00)

Pseudocount (pc) options:                                                        
 Context specific hhm pseudocounts:
  -pc_hhm_contxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=2) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               3: CSBlast admixture:   tau = a(1+b)/(Neff[i]+b)                 
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_hhm_contxt_a  [0,1]        overall pseudocount admixture (def=0.9)                        
  -pc_hhm_contxt_b  [1,inf[      Neff threshold value for mode 2 (def=4.0)                      
  -pc_hhm_contxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context independent hhm pseudocounts (used for templates; used for query if contxt file is not available):
  -pc_hhm_nocontxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=2) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_hhm_nocontxt_a  [0,1]        overall pseudocount admixture (def=1.0)                        
  -pc_hhm_nocontxt_b  [1,inf[      Neff threshold value for mode 2 (def=1.5)                      
  -pc_hhm_nocontxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context-specific pseudo-counts:                                                  
  -nocontxt      use substitution-matrix instead of context-specific pseudocounts 
  -contxt <file> context file for computing context-specific pseudocounts (default=${HHLIB}/data/context_data.crf)
  -csw  [0,inf]  weight of central position in cs pseudocount mode (def=1.6)
  -csb  [0,1]    weight decay parameter for positions in cs pc mode (def=0.9)

Other options:                                                                   
 -v <int>       verbose mode: 0:no screen output  1:only warings  2: verbose (def=2)
 -cpu <int>     number of CPUs to use (for shared memory SMPs) (default=2)      
 -scores <file> write scores for all pairwise comparisons to file               
 -atab   <file> write all alignments in tabular layout to file                   
 -maxres <int>  max number of HMM columns (def=20001)             
 -maxmem [1,inf[ limit memory for realignment (in GB) (def=3.0)          

Example: hhsearch -i a.1.1.1.a3m -d scop70_1.71

hhfilter – filter an MSA

Filter an alignment by maximum pairwise sequence identity, minimum coverage, minimum sequence identity, or score per column to the first (seed) sequence etc.

Usage: hhfilter -i infile -o outfile [options]                  
 -i <file>      read input file in A3M/A2M or FASTA format                 
 -o <file>      write to output file in A3M format                         
 -a <file>      append to output file in A3M format                        

Options:                                                                  
 -v <int>       verbose mode: 0:no screen output  1:only warings  2: verbose
 -id   [0,100]  maximum pairwise sequence identity (%) (def=90)   
 -diff [0,inf[  filter MSA by selecting most diverse set of sequences, keeping 
                at least this many seqs in each MSA block of length 50 (def=0) 
 -cov  [0,100]  minimum coverage with query (%) (def=0) 
 -qid  [0,100]  minimum sequence identity with query (%) (def=0) 
 -qsc  [0,100]  minimum score per column with query  (def=-20.0)
 -neff [1,inf]  target diversity of alignment (default=off)

Input alignment format:                                                    
 -M a2m         use A2M/A3M (default): upper case = Match; lower case = Insert;
                '-' = Delete; '.' = gaps aligned to inserts (may be omitted)   
 -M first       use FASTA: columns with residue in 1st sequence are match states
 -M [0,100]     use FASTA: columns with fewer than X% gaps are match states   
                                                                          
Example: hhfilter -id 50 -i d1mvfd_.a2m -o d1mvfd_.fil.a2m          

hhmake – build an HMM from an input alignment

Build a profile hidden Markov models from an input alignment, formatted in A2M, A3M, or FASTA, or convert between HMMER format (.hmm) and HHsearch format

(.hhm).

Usage: hhmake -i file [options]                                       
 -i <file>     query alignment (A2M, A3M, or FASTA), or query HMM         

Output options:                                                           
 -o <file>     HMM file to be written to  (default=<infile.hhm>)          
 -a <file>     HMM file to be appended to                                 
 -v <int>      verbose mode: 0:no screen output  1:only warings  2: verbose
 -seq <int>    max. number of query/template sequences displayed (def=10)  
               Beware of overflows! All these sequences are stored in memory.
 -cons         make consensus sequence master sequence of query MSA 
 -name <name>  use this name for HMM (default: use name of first sequence)   

Filter query multiple sequence alignment                                     
 -id   [0,100]  maximum pairwise sequence identity (%) (def=90)   
 -diff [0,inf[  filter MSA by selecting most diverse set of sequences, keeping 
                at least this many seqs in each MSA block of length 50 (def=100) 
 -cov  [0,100]  minimum coverage with query (%) (def=0) 
 -qid  [0,100]  minimum sequence identity with query (%) (def=0) 
 -qsc  [0,100]  minimum score per column with query  (def=-20.0)
 -neff [1,inf]  target diversity of alignment (default=off)

Input alignment format:                                                    
 -M a2m        use A2M/A3M (default): upper case = Match; lower case = Insert;
               '-' = Delete; '.' = gaps aligned to inserts (may be omitted)   
 -M first      use FASTA: columns with residue in 1st sequence are match states
 -M [0,100]    use FASTA: columns with fewer than X% gaps are match states   

Pseudocount (pc) options:                                                        
 Context specific hhm pseudocounts:
  -pc_hhm_contxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=0) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               3: CSBlast admixture:   tau = a(1+b)/(Neff[i]+b)                 
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_hhm_contxt_a  [0,1]        overall pseudocount admixture (def=0.9)                        
  -pc_hhm_contxt_b  [1,inf[      Neff threshold value for mode 2 (def=4.0)                      
  -pc_hhm_contxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context independent hhm pseudocounts (used for templates; used for query if contxt file is not available):
  -pc_hhm_nocontxt_mode {0,..,3}   position dependence of pc admixture 'tau' (pc mode, default=2) 
               0: no pseudo counts:    tau = 0                                  
               1: constant             tau = a                                  
               2: diversity-dependent: tau = a/(1+((Neff[i]-1)/b)^c)            
               (Neff[i]: number of effective seqs in local MSA around column i) 
  -pc_hhm_nocontxt_a  [0,1]        overall pseudocount admixture (def=1.0)                        
  -pc_hhm_nocontxt_b  [1,inf[      Neff threshold value for mode 2 (def=1.5)                      
  -pc_hhm_nocontxt_c  [0,3]        extinction exponent c for mode 2 (def=1.0)                     

 Context-specific pseudo-counts:                                                  
  -nocontxt      use substitution-matrix instead of context-specific pseudocounts 
  -contxt <file> context file for computing context-specific pseudocounts (default=${HHLIB}/data/context_data.crf)

Example: hhmake -i test.a3m 

hhalign – align a query MSA/HMM to a template MSA/HMM

Align a query alignment/HMM to a template alignment/HMM by HMM-HMM alignment. If only one alignment/HMM is given it is compared to itself and the best off-diagonal alignment plus all further non-overlapping alignments above significance threshold are shown. The command also allows to sample alignments randomly or to print out a list of indices of aligned residue pairs.

Usage: hhalign -i query -t template [options]  
 -i <file>      input/query: single sequence or multiple sequence alignment (MSA)
                in a3m, a2m, or FASTA format, or HMM in hhm format
 -t <file>      input/template: single sequence or multiple sequence alignment (MSA)
                in a3m, a2m, or FASTA format, or HMM in hhm format

Input alignment format:                                                       
 -M a2m         use A2M/A3M (default): upper case = Match; lower case = Insert;
               '-' = Delete; '.' = gaps aligned to inserts (may be omitted)   
 -M first       use FASTA: columns with residue in 1st sequence are match states
 -M [0,100]     use FASTA: columns with fewer than X% gaps are match states   
 -tags/-notags  do NOT / do neutralize His-, C-myc-, FLAG-tags, and trypsin 
                recognition sequence to background distribution (def=-notags)  

Output options: 
 -o <file>      write results in standard format to file (default=<infile.hhr>)
 -oa3m <file>   write query alignment in a3m or PSI-BLAST format (-opsi) to file (default=none)
 -aa3m <file>   append query alignment in a3m (-aa3m) or PSI-BLAST format (-apsi )to file (default=none)
 -Ofas <file>   write pairwise alignments in FASTA xor A2M (-Oa2m) xor A3M (-Oa3m) format   
 -add_cons      generate consensus sequence as master sequence of query MSA (default=don't)
 -hide_cons     don't show consensus sequence in alignments (default=show)     
 -hide_pred     don't show predicted 2ndary structure in alignments (default=show)
 -hide_dssp     don't show DSSP 2ndary structure in alignments (default=show)  
 -show_ssconf   show confidences for predicted 2ndary structure in alignments

Filter options applied to query MSA, template MSA, and result MSA              
 -id   [0,100]  maximum pairwise sequence identity (def=90)
 -diff [0,inf[  filter MSAs by selecting most diverse set of sequences, keeping 
                at least this many seqs in each MSA block of length 50 
                Zero and non-numerical values turn off the filtering. (def=100) 
 -cov  [0,100]  minimum coverage with master sequence (%) (def=0)             
 -qid  [0,100]  minimum sequence identity with master sequence (%) (def=0)    
 -qsc  [0,100]  minimum score per column with master sequence (default=-20.0)    
 -mark          do not filter out sequences marked by ">@"in their name line  

HMM-HMM alignment options:                                                       
 -norealign     do NOT realign displayed hits with MAC algorithm (def=realign)   
 -mact [0,1[    posterior prob threshold for MAC realignment controlling greedi- 
                ness at alignment ends: 0:global >0.1:local (default=0.35)       
 -glob/-loc     use global/local alignment mode for searching/ranking (def=local)

Other options:                                                                   
 -v <int>       verbose mode: 0:no screen output  1:only warings  2: verbose (def=2)

Example: hhalign -i T0187.a3m -t d1hz4a_.hhm -o result.hhr

reformat.pl – reformat one or many alignments

Read one or many multiple alignments in one format and write them in another format

Usage: reformat.pl [informat] [outformat] infile outfile [options] 
  or   reformat.pl [informat] [outformat] 'fileglob' .ext [options] 

Available input formats:
   fas:     aligned fasta; lower and upper case equivalent, '.' and '-' equivalent
   a2m:     aligned fasta; inserts: lower case, matches: upper case, deletes: '-',
            gaps aligned to inserts: '.'
   a3m:     like a2m, but gaps aligned to inserts MAY be omitted
   sto:     Stockholm format; sequences in several blocks with sequence name at 
            beginning of line (HMMER output)
   psi:     format as read by PSI-BLAST using the -B option (like sto with -M first -r)
   clu:     Clustal format; sequences in several blocks with sequence name at beginning 
            of line
Available output formats:
   fas:     aligned fasta; all gaps '-'
   a2m:     aligned fasta; inserts: lower case, matches: upper case, deletes: '-', 
            gaps aligned to inserts: '.'
   a3m:     like a2m, but gaps aligned to inserts are omitted
   sto:     Stockholm format; sequences in just one block, one line per sequence
   psi:     format as read by PSI-BLAST using the -B option 
   clu:     Clustal format
If no input or output format is given the file extension is interpreted as format 
specification ('aln' as 'clu')

Options:
  -v int    verbose mode (0:off, 1:on)
  -num      add number prefix to sequence names: 'name', '1:name' '2:name' etc
  -noss     remove secondary structure sequences (beginning with >ss_)
  -sa       do not remove solvent accessibility sequences (beginning with >sa_)
  -M first  make all columns with residue in first sequence match columns 
            (default for output format a2m or a3m)
  -M int    make all columns with less than X% gaps match columns 
            (for output format a2m or a3m)
  -r        remove all lower case residues (insert states) 
            (AFTER -M option has been processed)
  -r int    remove all lower case columns with more than X% gaps
  -g ''     suppress all gaps
  -g '-'    write all gaps as '-'
  -uc       write all residues in upper case (AFTER other options have been processed)
  -lc       write all residues in lower case (AFTER other options have been processed)
  -l        number of residues per line (for Clustal, FASTA, A2M, A3M formats) 
            (default=100)
  -d        maximum number of characers in nameline (default=1000)

Examples: reformat.pl 1hjra.a3m 1hjra.a2m  
          (same as reformat.pl a3m a2m 1hjra.a3m 1hjra.a2m)
          reformat.pl test.a3m test.fas -num -r 90
          reformat.pl fas sto '*.fasta' .stockholm

addss.pl – add predicted secondary structure to an MSA or HMM

Add PSIPRED secondary structure prediction (and DSSP annotation) to a multiple sequence alignment (MSA) or HMMER (multi-)model file.

If the input file is an MSA, the predicted secondary structure and confidence values are added as special annotation sequences with names >ss_pred, >ss_conf, and >ss_dssp to the top of the output A3M alignment. If no output file is given, the output file will have the same name as the input file, except for the extension being replaced by '.a3m'. Allowed input formats are A2M/FASTA (default), A3M (-a3m), CLUSTAL (-clu), STOCKHOLM (-sto), HMMER (-hmm). Note that in order to add PSIPRED and DSSP annotations corresponding paths have to be set in HHPaths.pm. To add DSSP annotations addss.pl first checks a folder of precomputed dssp files which can be obtained from here. If the dssp file for a particular structure is not available, addss.pl tries to execute the DSSP binary mkdssp. Please keep in mind that only the current DSSP version (2.2.1) supports structures in cif format, and that compiling the software may cause problems on several systems. If you have troubles to compile DSSP you may try to delete all occurrences of "static" in its Makefile.

If the input file contains HMMER models, records SSPRD and SSCON containing predicted secondary structure and confidence values are added to each model. In this case the output file name is obligatory and must be different from the input file name.

Usage: perl addss.pl <ali file> [<outfile>] [-fas|-a3m|-clu|-sto]  
  or   perl addss.pl <hmm file> <outfile> -hmm  

hhmakemodel.pl and hhmakemodel.py – generate MSAs or coarse 3D models from HHsearch results file

From the top hits in an hhsearch output file (hhr), you can

• generate a MSA (multiple sequence alignment) containing all representative template sequences from all selected alignments (options -fas, -a2m, -a3m, -pir)

• generate several concatenated pairwise alignments in AL format (option -al)

• generate several concatenated coarse 3D models in PDB format (option -ts)

In PIR, PDB and AL format, the pdb files are required in order to read the pdb residue numbers and ATOM records. The PIR formatted file can be used directly as input to the MODELLER homology modeling package.

 Usage: hhmakemodel.pl [-i] file.hhr [options]

 Options:
  -i   <file.hhr>        results file from hhsearch with hit list and alignments
  -fas <file.fas>        write a FASTA-formatted multiple alignment to file.fas
  -a2m <file.a2m>        write an A2M-formatted multiple alignment to file.a2m
  -a3m <file.a3m>        write an A3M-formatted multiple alignment to file.a3m
  -m   <int> [<int> ...] pick hits with specified indices  (default='-m 1')
  -p   <probability>     minimum probability threshold     
  -e   <E-value>         maximum E-value threshold      
  -q   <query_ali>       use the full-length query sequence in the alignment 
                         (not only the aligned part);
                         the query alignment file must be in HHM, FASTA, A2M,  
                         or A3M format.
  -N                     use query name from hhr filename (default: use same  
                         name as in hhr file)
  -first                 include only first Q or T sequence of each hit in MSA
  -v                     verbose mode

 Options when database matches in hhr file are PDB or SCOP sequences
  -pir <file.pir>        write a PIR-formatted multiple alignment to file.pir 
  -ts  <file.pdb>        write the PDB-formatted models based on *pairwise*  
                         alignments into file.pdb
  -al  <file.al>         write the AL-formatted *pairwise* alignments into file.al
  -d   <pdbdirs>         directories containing the pdb files (for PDB, SCOP, or DALI  
                         sequences)
  -s   <int>             shift the residue indices up/down by an integer           
  -CASP                  formatting for CASP (for -ts, -al options) 
                         (default: LIVEBENCH formatting)

Analogously, hhmakemodel.py generates a MSA containing all representative template sequences from all selected alignments in PIR format. Note that in hhmakemodel.py you have to specify the folder containing the *.cif files which make up the alignment. The script will modify the ATOM section of those cif files according to the residue numbering in the alignment. Use these renumbered cifs file together with the PIR alignment output in

MODELLER.

usage: hhmakemodel.py [-h] [-v] [-m INT [INT ...]] [-e FLOAT] [-r FLOAT] [-c]
                      FILE DIR FILE DIR

Creates a MODELLER alignment (*.pir) from a HHSearch results file (*.hhr).

positional arguments:
  FILE              results file from HHsearch with hit list and alignment
  DIR               path to the folder containing cif files
  FILE              output file (PIR-formatted multiple alignment)
  DIR               path to the folder where modified cif files should be
                    written to

optional arguments:
  -h, --help        show this help message and exit
  -v, --verbose     verbose mode
  -m INT [INT ...]  pick hits with specified indices (e.g. -m 2 5)
  -e FLOAT          maximum E-Value threshold (e.g. -e 0.001)
  -r FLOAT          residue ratio (filter alignments that have contribute at
                    least residues according to the specified ratio).
  -c                convert non-canonical residues (default = True)

hhsuitedb.py – Build an HHsuite database

Builds HHsuite database from a3m formatted MSAs and/or from HMMs (-o). MSAs and HMMs can also be added (-a) to or removed (-r) from an existing database.

Usage: hhsuitedb.py [options]

Options:
  -h, --help       show this help message and exit
  --ia3m=<GLOB>    Glob for a3m files
  --ics219=<GLOB>  Glob for cs219 files
  --ihhm=<GLOB>    Glob for hhm files
  -o FILE          Output hhsuite database basename
  --cpu=INT        Number of threads that shall used
  --force          Try to fix problems with the database

splitfasta.pl – Split multi-sequence FASTA file into single-sequence files

Write files into current directory and name each file by the first word after ">" in the name line. 

Usage: splitfasta.pl infile [option]
Option:
-fam       : use family-based name (for SCOP/ASTRAL sequences
-name      : use sequence name as file name (default)
-ext <ext> : extension for sequence files (default=seq)

renumberpdb.pl – Renumber indices in PDB file to match input sequence indices

 The program reads an input sequence in FASTA/A3M which must have a SCOP domain, PDB chain, or DALI 
 domain identifier (e.g. d1hz4a_, 1hz4_A, or 1hz4A_1). It reads the corresponding PDB file from the 
 given pdb directory and  generates a new PDB file by aligning the input sequence to the sequence 
 extracted from the ATOM records of the corresponding pdb chain and renumbering the residues in 
 columns 23-26 according to the position in the input sequence.
 (HETATM records for MSE (selenomethionine) will be interpreted as ATOM records for MET in the 
 alignment. MSE will be changed to MET in the output file.)
 
 Usage:   renumberpdb.pl [options] infile [outfile] 
 Example: renumberpdb.pl d1hz4a_.a3m /cluster/tmp/d1hz4a_.pdb 

 Options: 
   -o <file>    output file (default: <infile_wo_extension>.pdb
   -d <pdbdir>  give directory of pdb files (default=)
   -a [A|B]     filter alternative location identifier (e.g. A or B)

cif2fasta.py – Create a fasta file from cif files

An example of the usage of cif2fasta.py is provided in Building customized databases.

Usage: cif2fasta.py -i cif_folder -o *.fasta -c num_cores -v

cif2fasta.py takes a folder that contains cif files as input and outputs their
sequences into fasta file.

Options:
  -h, --help  show this help message and exit
  -i DIR      input cif folder.
  -o FILE     output fasta file.
  -p FILE     output PDB filter file (optional).
  -s FILE     SCOP annotation.
  -c INT      number of cores (default = 1).
  -l INT      Remove chains with a length < X (default = 30).
  -v INT      Verbose Mode (quiet = 0, full verbosity = 2).

pdbfilter.py – Filter sequences from the PDB (requires MMseqs2)

An example of the usage of pdbfilter.py is provided in Building customized databases. Note that the annotations file which is required to select the proper PDB is created by cif2fasta.py using the -p

flag.

usage: pdbfilter.py [-h] [-i FILE] [-r FILE] [-v] FILE FILE FILE FILE

pdbfilter.py selects from sequence clusters (determined by MMseqs2) the
sequences which have the best resolution, R-free factor and/or completness and
writes them to a fasta file.

positional arguments:
  FILE                  input fasta file (created by cif2fasta.py)
  FILE                  sequence clusters (MMseqs)
  FILE                  annotations file (created by cif2fasta using the -p
                        flag, contains information about resolution, R-free
                        and completness of sequences).
  FILE                  output fasta file

optional arguments:
  -h, --help            show this help message and exit
  -i FILE, --include FILE
                        include PDB chains
  -r FILE, --remove FILE
                        exclude PDB chains
  -v, --verbose         verbose mode

License

The HHsuite software package is distributed under Gnu Public License, Version 3.

This means that the HHsuite is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program in the file LICENSE. If not, see http://www.gnu.org/licenses/.

Contributors and Contact

• Johannes Söding: main code base incl. perl scripts up to version 2.0, debugging, documentation. Contact at [email protected]
• Milot Mirdita: Uniclust30 database and maintenance
• Markus Meier: uniprot-1it database and associated hhblits code, code refactoring, tests and benchmarks
  
• Martin Steinegger: Fast SIMD HMM-HMM Viterbi code and various speed-ups
  
• Michael Remmert: early HHblits version
  
• Andy Hauser: ffindex, speed/memory optimization
  
• Christof Angermüller: context-specific pseudocounts, speed optimizations
  
• Armin Meier: HHblits column state alphabets, tests and benchmarks
  
• Andreas Biegert: HHblits column state alphabet cs219, context-specific pseudocounts
  

Acknowledgments

We thank the many users of HHsuite who sent bug reports or gave critical feedback that led to improvements in the software. We are grateful to Lazlo Kajan, Stefan Möller and the Debian team for making the HHsuite software available as a Debian package. J.S. thanks Andrei Lupas for his invaluable support, encouragement, and feedback during the first five years of the work on HHsuite in his department.

For many years, the HH-suite contained code adapted from Michael Farrar for its Smith-Waterman implementation. We want to express our posthumous gratitude to Michael Farrar for his code. The current Smith-Waterman implementation is based on the SSW library from Mengyao Zhao, for which we are very grateful.

Reference:

• Farrar M (2007). Striped Smith-Waterman speeds database searches six times over other SIMD implementations. Bioinformatics 23, 156-161. doi: 10.1093/bioinformatics/btl582

• Zhao M, Lee W P, Garrison E P, Marth G T (2013). SSW Library: An SIMD Smith-Waterman C/C++ Library for Use in Genomic Applications. PLOS One 8, e82138. doi: 10.1371/journal.pone.0082138

Good luck with your work!