Curation pipeline overview

Software needed. Tools necessary to install to run the pipeline are listed below. You should install each of these programs on your computer.

Use TextWrangler or equivalent text editor for taking notes about the curation process, modifying scripts, and viewing files. Do NOT use Word or equivalent.


This document outlines the procedure for gathering sequences and generating an alignment and tree.

Step 0. Decide upon the scope of the group you will initially curate. Your group should be of a manageable size in terms of taxonomic breadth and total available sequences (e.g. a subset of Rhizaria or diatoms rather than the entire clade). It will be more useful for you and EukRef in general to finish curation of a smaller group. You can expand later.

Step 1. Generate a curated starting sequence set, alignment and tree for your group. This initial sequence set will form the basis of your entire curation effort. Therefore it is critical that these sequences are well-curated, and you need an alignment of these sequences and a reliable phylogenetic tree built from them. There are multiple ways to get this initial sequence set, but in each case you should ensure it meets the criteria below. Additionally, you need a file that contains diverse outgroups: 5-10 sequences from all of the neighboring clades (e.g., if you are working on ciliates you should include apicomplexa, dinoflagellates, and stramenopiles). This alignment MUST be untrimmed (e.g. contain all gaps) so that additional sequences can be added. The initial tree must be checked to make sure it covers the diversity of the group and reflects phylogenetic relationships expected based on current research/literature.  Any errant sequences need to be removed from the alignment.

Criteria for the initial set of sequences, alignment, and tree:

  • The initial set of sequences should be small, ~200 sequences or fewer, so that you can easily check the resulting alignment and tree.
  • Sequences should cover the breadth of the clade, including divergent groups.
  • The alignment and tree should have outgroups included to detect sequences that branch outside of your clade.
  • The alignment should be untrimmed
  • The tree produced from this alignment should display expected relationships and problematic sequences should be removed.

Three possible ways to get your initial alignment:

  • You may already have a curated starting alignment and tree from previous work. In this case, you are ready to go after checking your tree and alignment. Proceed to Step 5 .
  • Download sequences from GenBank based on taxonomy. This will target the sequences from cultured isolates or those that were identified in other ways (e.g. morphologically identified and picked).  Proceed to Step 2 .
  • Alternatively, you may wish to start with the SILVA or PR2 set of sequences for your group. You must ensure that these sequences meet the criteria above.  Proceed to Step 2 .
Step 2. Retrieve initial set of sequences and cluster

NOTE 1: Users can choose a different similarity threshold if desired. For example, you may wish to work at 99% identity for small clades. In this case, replace 0.97 with 0.99 throughout.

NOTE 2: In all commands, you must substitute your clade name for “NAME”.

2a) Retrieve initial sequences using GenBank, PR2 or SILVA

Recommended: SILVA or PR2 database. Use the grep command to pull out your sequences of interest. Note that clade names often differ in these different databases.

GenBankuse advanced search at NCBI, or add the text below to the NCBI query box. Replace XXXX with the Taxonomy ID. Find the Taxonomy ID by searching NCBI taxonomy:

txidXXXX[Organism:exp] (18S OR SSU) NOT (mitochondrial OR mitochondria OR complete genome)

This will target targeted cultured isolates or isolates that were morphologically identified

After you have downloaded your sequences, clean fasta headers to contain the GenBank accession number only.

2b) Cluster your sequences to produce a manageable number of sequences using usearch (two commands). Similarity threshold is set at 97% here, but can be adjusted by changing the “-id” command.  These commands choose the longest sequence as the representative sequence for each cluster. We will use also this step to remove sequences shorter than 500 bp. You will use these representative sequences for your initial alignment/tree.

Step 3. Build initial alignment.

3a) Assemble a set of close outgroup sequences. Outgroup sequences must be full-length. You should include outgroups from all of the clades that are neighbors to your group of interest, 5-10 sequences per group. These should be chosen based on your knowledge (ask for help if needed). You can get outgroup sequences from GenBank, SILVA or PR2. See step 2 for details. Add your outgroup sequences to clustered starting sequences (NAME.clustered.fasta file)

3b) Align using MAFFT. Use the default MAFFT algorithm (command below) unless there is a specific reason to use a different algorithm or approach.

3c) Trim the alignment using trimal.

Step 4. Build initial tree.

4a) Build phylogenetic tree using RAxML.

Constraint trees: You can use a constraint tree in RAxML (-g option) based on previously published relationships for your clade, for example based on morphology and/or multigene phylogeny. Constraint trees should be used with caution and should ONLY constrain robust relationships (those observed in multiple studies). All taxa in the constraint tree must be in the alignment. The constraint tree must be bifurcating (polytomies are not allowed). Constraint trees can be edited in Mesquite.

Alternatively, you can use FastTree (is a faster algorithm in case you don’t want or can’t wait for the RAxML results, it will give you a “quick and dirty” tree).

4b) Visualize tree using FigTree. Identify errant sequences to be removed.  Go back to step 3  to refine alignment, remove errant sequences and repeat as necessary until you have a high quality starting initial alignment and tree. If your tree does not reflect known relationships you may wish to use a constraint tree in step 4a.   Proceed to step 5  once you are happy with your tree/alignment. You will need the unaligned set of sequences (without outgroups) as input for step 5 (initial_DS.fas).

Step 5. Download databases

5a) Download and unpack databases.tar.gz:

Inside the unpacked “databases” directory you will find three files: Reference_DB.fas and gb203_pr2_all_10_28_97p_noorg.fasta, and tax_d.bin.  These files are used for filtering results within All three files must be in the current working directory where you run the script in Step 6.

5b) Locate a copy of GenBank nt database in use at your institution or on your server.  [at the EukRef workshop there is a copy on the Amazon server]. GenBank NT should ideally be downloaded to a server. There will be more than 40 files totaling more than 30 GB. Downloading can take several hours. You may wish to install and use a tool like wget.  In a workshop setting download one time and allow access for all participants.

Make a new folder called DATABASE_FOLDER. From download taxdb.tar.gz and all nt.tar.gz files. Also, download all nt.tar.gz.md5 files.

Unpack with the following command:

Run md5 checksum to ensure all files are fully downloaded.

5c) Before running the pipeline run the following commands:

(you can also add BLASTDB to your PATH permanently)

Step 6. Retrieve all sequences that belong to your clade.

6a)Run the script. This script will run in a loop until no new sequences are retrieved. This may take several hours for large groups.


-i Unaligned fasta file of your clustered starting set of sequences (the output of  Step 2 , NAME.clustered.fasta). Should be refined version, with any errant sequences detected in tree inspection during  Step 4  removed. Should NOT include outgroups. Fasta headers must either be in standard GenBank format (>gi|ginumber|gb|accession| ), or have the accession number followed by a space (>accession )
-dbnt (/path/to/DATABASE_FOLDER/nt)GenBank NT file from  Step 5c  .
-dbsi (Reference_DB.udb)PR2 SSU reference database plus representative bacteria, used for filtering results. Must be in current working directory.
-n Number of sequences retrieved from GenBank per blasted sequence. recommended: 100
-p Number of CPUs.
-g Name of the most inclusive group you are working with from PR2 taxonomy. Find in the PR2 taxonomy file (ReferenceDB.fas) by grep.
-m Blast method. recommended: megablast
-idsi PR2 Blast cut-off (nothing less than 70% similar will be retrieved).
-idnt GenBank Blast cut-off (average similarity to everything on the original database).

Example command, update with your information.


Reference set of sequences (current_DB.fas),
Reference sequences with only accessions in header for downstream use (current_DB_final.fas),
List of accession numbers (accessions_current_DB.txt),
A fasta file of short reads (<500 bp), and of chimeras will also be generated.

6b) Format sequences and metadata.Run the script to pull taxonomy and environmental information from GenBank records to 1) generate your initial reference database, and 2) reformat the fasta headers for easier annotation in a tree.

First, download the GenBank format for all sequences retrieved in step 6a from NCBI batch entrez. Upload accessions ( accessions_current_DB.txt).  Click the retrieve records link.  Click “send to” then download as GenBank(full) file (.gb extension).  This will download the gb file for all of your accessions. Save as


-gb Genbank flat file
-i fasta file output from step 5 (current_DB_final.fasta)outgroup
–outgroup outgroup fasta file
-t reference taxonomy file (SILVA_128_full_taxa_map.euks.txt).


metadata file (tab delimited format) that includes taxonomy from GenBank, reference taxonomy, environmental data available in the GenBank record, and the publication associated with an accession.
fasta file with headers labeled for easier curation.

Step 7. Build an alignment with the reference sequences

7a) Cluster sequences using usearch

If the FASTA file to be clustered contains long (e.g. genomic) sequences, it is likely that the clustering step will run out of memory. Check the screen output of this “usearch -sortbylength” step, which indicates the length of the longest sequence found within the file; this should not exceed 5,000 (the hard limit imposed in Step 5).

NOTE: you should choose the -id (similarity threshold for clustering) that is appropriate for your group.

7b) Align using MAFFT. If there is not a specific reason to use a different approach use the default mafft algorithm.  See instructions below for alignments with groups known to be difficult.

It is critical to open and check your alignment in Aliview software (or similar) to see if there are misaligned blocks or sequences. Make sure you look at the entire length of the alignment. If the alignment is not good you will not produce a good tree.  If you have a poor alignment you can 1) align according to a template (seed) alignment, command below, or 2) modify your alignment by hand.

If you already have a curated alignment for your group you can use it as a template (seed) to align your new sequences with the following command.

Remember to open your alignment file and check before proceeding.

7c) Trim the alignment using trimal

Step 8. Build RaxML trees and clean up.

8a) Build phylogenetic tree using RAxML. NOTE: do not use more threads (-T option) than available processors. Typically no more than 4 on a laptop and potentially many more on a cluster. NOTE: you can add a backbone constraint tree with the option -g.  This should be done if clades within your group are always monophyletic in the literature (e.g. in phylogenomic or multigene analyses), but are not monophyletic in 18S trees. The sequences (tips) in this constraint tree must be in your alignment (they should be a subset of your sequences), and the names must match exactly.

Alternatively, you can use FastTree (is a faster algorithm in case you don’t want or can’t wait for the RAxML results, it will give you a “quick and dirty” tree).

Step 9. Visualize your tree, identify errant sequences that should be removed.

Download your tree e.g. “RAxML_bestTree.clade”. Open tree in FigTree and annotate taxa with either “remove” for taxa you wish to remove, or you can modify/add to the existing name. You can only annotate taxa, NOT nodes or clades.  You can select entire clades to annotate all taxa at once. Annotated nodes cannot be collapsed (annotations within collapsed node will not be reported).
“Save as” the tree with annotations for example “RAxML_bestTree.clade_annotated.tre”, which is a nexus file.

The “metadata_out.txt” will be your tab delimited metadata file with “remove” in the last column for the sequences you want to remove.  You can sort your metadata_out.txt file in excel to get all of the accessions of sequences that should be removed from your tree. Save these accessions to be removed in a file called accessions_remove.txt. This is the input for the next step:

Go back to Step 7  to align and build a new tree, repeat until you are happy with your current set of sequences.

Step 10. Build Reference Tree.

Once you are happy with the set of sequences in your tree and your alignment you are ready to build your final reference tree. Your final, trimmed alignment is in input into RAxML. NOTE: you can add a backbone constraint tree with the option -g.  The sequences (tips) in this constraint tree must be in your alignment (they should be a subset of your sequences), and the names must match exactly.

Step 11. Moving from tree to annotations.  Expanding annotations to all sequences.

11a) Use the script Trim your metadata file to include only representative sequences of clusters — these are the sequences in your tree.


-i is the fasta file with your clustered sequences containing only the sequences used in your reference tree.  -m is your metadata.txt file from


-o is metadata_ref.txt, which only has the accessions in your reference tree. This will make annotating easier.

11b) You need to annotate classification your metadata_ref.txt. Note that you should have GenBank taxonomy and SILVA taxonomy in your metadata_ref.txt file so you know the starting point. In many cases, it will be easiest to annotate clades directly on your reference tree. You can annotate the tree and then export annotations to your metadata_ref.txt file using instructions below.

Download your reference tree e.g. “RAxML_bestTree.clade”. Open tree in FigTree and annotate taxa with the name of the taxon. You can only annotate the taxa, NOT nodes or clades.  You can select entire clades to annotate all taxa at once. Annotated nodes cannot be collapsed (annotations within collapsed node will not be reported).
“Save as” the tree with annotations for example “RAxML_bestTree.clade_annotated.tre”, which is a nexus file.

Run script to add annotations in your tree to your metadata_ref.txt file. The output is metadata_ref_out.txt

The “metadata_ref_out.txt” will be your tab delimited metadata file that includes the Accession Number of representative for your clusters and the annotations that you made on the tree in the last column.

11c) Use the script to expand your curation to all sequences within the clusters.


-i is the reference metadata you just added annotations to: metadata_ref_out.txt

-c is the cluster file (.uc) from usearch. Use the most recent version if you clustered multiple times.


-o is the expanded annotated metadata file: metadata_ref_expanded.txt

Step 12. Create reference database.

Now the taxonomy and metadata curation fun begins! See Building your database page for details.

If you have any question or suggestion please feel free to add it to the comment section below, visit our forum or send us an e-mail
RAxML tips

  • MPI vs. PTHREADS: If you’re running RAxML on your laptop, use PTHREADS. If you’re running RAxML on a server/cluster with a lot of cores (>4), please consider using MPI (or HYBRID) version as it allows for better parallelization. In both cases, make absolutely sure that the number of threads/MPI processes you use is equal to the number of CPU cores you have.
  • GTRCAT vs. GTRGAMMA: GTRCAT model is faster to compute compared to the standard GAMMA model, and might better
    account for among-site heterogeneity as it has more rate categories (default: 25). However, it shouldn’t be used on alignments with few taxa (<50-100) because of the risk of overparametrization.
  • +I models: we generally do not recommend using GTRCATI/GTRGAMMAI, since the use two orthogonal approaches to model rate heterogeneity, which might lead to problems in numerical optimization.
  • Please refer to RAxML manual for further details:

Leave a Reply