$0 - Bulk loads gff3 files into a chado database.
% $0 [options]
% cat <gff-file> | $0 [options]
--gfffile The file containing GFF3 (optional, can read
--fastafile Fasta file to load sequence from
--organism The organism for the data
(use the value 'fromdata' to read from GFF organism=xxx)
--dbprofile Database config profile name
--dbname Database name
--dbuser Database user name
--dbpass Database password
--dbhost Database host
--dbport Database port
--analysis The GFF data is from computational analysis
--noload Create bulk load files, but don't actually load them.
--nosequence Don't load sequence even if it is in the file
--notransact Don't use a single transaction to load the database
--drop_indexes Drop indexes of affected tables before starting load
and recreate after load is finished; generally
does not help performance.
--validate Validate SOFA terms before attempting insert (can
cause script startup to be slow, off by default)
--ontology Give directions for handling misc Ontology_terms
--skip_vacuum Skip vacuuming the tables after the inserts (default)
--no_skip_vaccum Don't skip vacuuming the tables
--inserts Print INSERT statements instead of COPY FROM STDIN
--noexon Don't convert CDS features to exons (but still create
--recreate_cache Causes the uniquename cache to be recreated
--remove_lock Remove the lock to allow a new process to run
--save_tmpfiles Save the temp files used for loading the database
--random_tmp_dir Use a randomly generated tmp dir (the default is
to use the current directory)
--no_target_syn By default, the loader adds the targetId in
the synonyms list of the feature. This flag
--unique_target Trust the unicity of the target IDs. IDs are case
sensitive. By default, the uniquename of a new target
will be 'TargetId_PrimaryKey'. With this flag,
it will be 'TargetId'. Furthermore, the Name of the
created target will be its TargetId, instead of the
--dbxref Use either the first Dbxref annotation as the
primary dbxref (that goes into feature.dbxref_id),
or if an optional argument is supplied, the first
dbxref that has a database part (ie, before the ':')
that matches the supplied pattern is used.
--delete Instead of inserting features into the database,
use the GFF lines to delete features as though
the CRUD=delete-all option were set on all lines
(see 'Deletes and updates via GFF below'). The
loader will ask for confirmation before continuing.
Works like --delete except that it does not ask
--fp_cv Name of the feature property controlled vocabulary
(defaults to 'feature_property').
--noaddfpcv By default, the loader adds GFF attribute types as
new feature_property cv terms when missing. This flag
** dgg note: should rename this flag: --[no]autoupdate
for Chado tables cvterm, cv, db, organism, analysis ...
--manual Detailed manual pages
--custom_adapter Use a custom subclass adaptor for Bio::GMOD::DB::Adapter
Provide the path to the adapter as an argument
--private_schema Load the data into a non-public schema.
--use_public_cv When loading into a non-public schema, load any cv and
cvterm data into the public schema
--end_sql SQL code to execute after the data load is complete
Allow Parent tags to refer to IDs outside the current
Note that all of the arguments that begin 'db' as well as organism can be
provided by default by Bio::GMOD::Config, which was installed when 'make
install' was run. Also note the the option dbprofile and all other db* options
are mutually exclusive--if you supply dbprofile, do not supply any other db*
options, as they will not be used.
The GFF in the datafile must be version 3 due to its tighter control of the
specification and use of controlled vocabulary. Accordingly, the names of
feature types must be exactly those in the Sequence Ontology Feature
Annotation (SOFA), not the synonyms and not the accession numbers (SO
accession numbers may be supported in future versions of this script).
Note that the ##sequence-region directive is not supported as a way of declaring
a reference sequence for a GFF3 file. The ##sequence-region directive is not
expressive enough to define what type of thing the sequence is (ie, is it a
chromosome, a contig, an arm, etc?). If your GFF file uses a ##sequence-region
directive in this way, you must convert it to a full GFF3 line. For example,
if you have this line:
##sequence-region chrI 1 9999999
Then is should be converted to a GFF3 line like this:
chrI . chromosome 1 9999999 . . . ID=chrI
Here is summary of how GFF3 data is stored in chado:
- Column 1 (reference sequence)
- The reference sequence for the feature becomes the
srcfeature_id of the feature in the featureloc table for that feature.
That featureloc generally assigned a rank of zero if there are other
locations associated with this feature (for instance, for a match
feature), the other locations will be assigned featureloc.rank values
greater than zero.
- Column 2 (source)
- The source is stored as a dbxref. The chado instance must
of an entry in the db table named 'GFF_source'. The script will then
create a dbxref entry for the feature's source and associate it to the
feature via the feature_dbxref table.
- Column 3 (type)
- The cvterm.cvterm_id of the SOFA type is stored in
- Column 4 (start)
- The value of start minus 1 is stored in featureloc.fmin
(one is subtracted because chado uses interbase coordinates, whereas GFF
uses base coordinates).
- Column 5 (end)
- The value of end is stored in featureloc.fmax.
- Column 6 (score)
- The score is stored in one of the score columns in the
analysisfeature table. The default is analysisfeature.significance. See
the section below on analysis results for more information.
- Column 7 (strand)
- The strand is stored in featureloc.strand.
- Column 8 (phase)
- The phase is stored in featureloc.phase. Note that there is
currently a problem with the chado schema for the case of single exons
having different phases in different transcripts. If your data has just
such a case, complain to firstname.lastname@example.org to find ways to
address this problem.
- Column 9 (group)
- Here is where the magic happens.
- Assigning feature.name, feature.uniquename
- The values of feature.name and feature.uniquename are
assigned according to these simple rules:
- If there is an ID tag, that is used as
- otherwise, it is assigned a uniquename that is equal to
'auto' concatenated with the feature_id.
- If there is a Name tag, it's value is set to
- otherwise it is null.
Note that these rules are much more simple than that those that Bio::DB::GFF
uses, and may need to be revisited.
- Assigning feature_relationship entries
- All Parent tagged features are assigned
feature_relationship entries of 'part_of' to their parent features.
Derived_from tags are assigned 'derived_from' relationships. Note that
parent features must appear in the file before any features use a Parent
or Derived_from tags referring to that feature.
- Alias tags
- Alias values are stored in the synonym table, under both
synonym.name and synonym.synonym_sgml, and are linked to the feature via
the feature_synonym table.
- Dbxref tags
- Dbxref values must be of the form 'db_name:accession',
where db_name must have an entry in the db table, with a value of db.name
equal to 'DB:db_name'; several database names were preinstalled with the
database when 'make prepdb' was run. Execute 'SELECT name FROM db' to find
out what databases are already available. New dbxref entries are created
in the dbxref table, and dbxrefs are linked to features via the
- Gap tags
- Currently is mostly ignored--the value is stored as a
featureprop, but otherwise is not used yet.
- Note tags
- The values are stored as featureprop entries for the
- Any custom (ie, lowercase-first) tags
- Custom tags are supported. If the tag doesn't already exist
in the cvterm table, it will be created. The value will stored with its
associated cvterm in the featureprop table.
- When the Ontology_term tags are used, items from the Gene
Ontology and Sequence Ontology will be processed automatically when the
standard DB:accession format is used (e.g. GO:0001234). To use other
ontology terms, you must specify that mapping of the DB indentifiers in
the GFF file and the name of the ontologies in the cv table as a comma
separated tag=value pairs. For example, to use plant and cell ontology
terms, you would supply on the command line:
--ontology 'PO=plant ontology,CL=cell ontology'
where 'plant ontology' and 'cell ontology' are the names in the cv table
exactly as they appear.
- Target tags
- Proper processing of Target tags requires that there be two
source features already available in the database, the 'primary' source
feature (the chromosome or contig) and the 'subject' from the similarity
analysis, like an EST, cDNA or syntenic chromosome. If the subject feature
is not present, the loader will attempt to create a placeholder feature
object in its place. If you have a fasta file the contains the subject,
you can use the perl script, gmod_fasta2gff3.pl, that comes with this
distribution to make a GFF3 file suitable for loading into chado before
loading your analysis results.
- CDS and UTR features
- The way CDS features are represented in Chado is as an
intersection of a transcript's exons and the transcripts polypeptide
feature. To allow proper translation of a GFF3 file's CDS features, this
loader will convert CDS and UTR feature lines to corresponding exon
features (and add a featureprop note that the exon was inferred from a
GFF3 CDS and/or UTR line), and create a polypeptide feature that spans the
genomic region from the start of translation to the stop.
If your GFF3 file contains both exon and CDS/UTR features, then you will
want to suppress the creation of the exon features and instead will only
want a polypeptide feature to be created. To do this, use the --noexon
option. In this case, the CDS and UTR features will still be converted to
exon features as described above.
Note that in the case where your GFF file contains CDS and/or UTR features
that do not belong to 'central dogma' genes (that is, that have a gene,
transcript and CDS/exon features), none of the above will happen and the
features will be stored as is.
- Loading fasta file
- When the --fastafile is provided with an argument that is
the path to a file containing fasta sequence, the loader will attempt to
update the feature table with the sequence provided. Note that the ID
provided in the fasta description line must exactly match what is in the
feature table uniquename field. Be careful if it is possible that the
uniquename of the feature was changed to ensure uniqueness when it was
loaded from the original GFF. Also note that when loading sequence from a
fasta file, loading GFF from standard in is disabled. Sorry for any
- This script does not use sequence-region directives for
anything. If it represents a feature that needs to be inserted into the
database, it should be represented with a full GFF line. This includes the
reference sequence for the features if it is not already in the database,
like a chromosome. For example, this:
##sequence-region chr1 1 213456789
should change to this:
chr1 UCSC chromosome 1 213456789 . . . ID=chr1
- This application will, by default, try to load all of the
data at once as a single transcation. This is safer from the database's
point of view, since if anything bad happens during the load, the
transaction will be rolled back and the database will be untouched. The
problem occurs if there are many (say, greater than a 2-300,000) rows in
the GFF file. When that is the case, doing the load as a single
transcation can result in the machine running out of memory and killing
processes. If --notranscat is provided on the commandline, each table will
be loaded as a separate transaction.
- SQL INSERTs versus COPY FROM
- This bulk loader was originally designed to use the
PostgreSQL COPY FROM syntax for bulk loading of data. However, as
mentioned in the 'Transactions' section, memory issues can sometimes
interfere with such bulk loads. In another effort to circumvent this
issue, the bulk loader has been modified to optionally create INSERT
statements instead of the COPY FROM statements. INSERT statements will
load much more slowly than COPY FROM statements, but as they load and
commit individually, they are more more likely to complete successfully.
As an indication of the speed differences involved, loading yeast GFF3
annotations (about 16K rows), it takes about 5 times longer using INSERTs
versus COPY on my laptop.
- Deletes and updates via GFF
- There is rudimentary support for modifying the features in
an existing database via GFF. Currently, there is only support for
deleting. In order to delete, the GFF line must have a custom tag in the
ninth column, 'CRUD' (for Create, Replace, Update and Delete) and have a
recognized value. Currently the two recognized values are CRUD=delete and
IMPORTANT NOTE: Using the delete operations have the potential of creating
orphan features (eg, exons whose gene has been deleted). You should be
careful to make sure that doesn't happen. Included in this distribution is
a PostgreSQL trigger (written in plpgsql) that will delete all orphan
children recursively, so if a gene is deleted, all transcripts, exons and
polypeptides that belong to that gene will be deleted too. See the file
modules/sequence/functions/delete-trigger.plpgsql for more
- The delete option will delete one and only one feature for
which the name, type and organism match what is in the GFF line with what
is in the database. Note that feature.uniquename are not considered, nor
are the coordinates presented in the GFF file. This is so that updates via
GFF can be done on the coordinants. If there is more than one feature for
which the name, type and organism match, the loader will print an error
message and stop. If there are no features that match the name, type and
organism, the loader will print a warning message and continue.
- The delete-all option works similarly to the delete option,
except that it will delete all features that match the name, type and
organism in the GFF line (as opposed to allowing only one feature to be
deleted). If there are no features that match, the loader will print a
warning message and continue.
- The run lock
- The bulk loader is not a multiuser application. If two
separate bulk load processes try to load data into the database at the
same time, at least one and possibly all loads will fail. To keep this
from happening, the bulk loader places a lock in the database to prevent
other gmod_bulk_load_gff3.pl processes from running at the same time. When
the application exits normally, this lock will be removed, but if it
crashes for some reason, the lock will not be removed. To remove the lock
from the command line, provide the flag --remove_lock. Note that if the
loader crashed necessitating the removal of the lock, you also may need to
rebuild the uniquename cache (see the next section).
- The uniquename cache
- The loader uses the chado database to create a table that
caches feature_ids, uniquenames, type_ids, and organism_ids of the
features that exist in the database at the time the load starts and the
features that will be added when the load is complete. If it is possilbe
that new features have been added via some method that is not this loader
(eg, Apollo edits or loads with XORT) or if a previous load using this
loader was aborted, then you should supply the --recreate_cache option to
make sure the cache is fresh.
- By default, if there is sequence in the GFF file, it will
be loaded into the residues column in the feature table row that
corresponds to that feature. By supplying the --nosequence option, the
sequence will be skipped. You might want to do this if you have very large
sequences, which can be difficult to load. In this context, "very
large" means more than 200MB.
Also note that for sequences to load properly, the GFF file must have the
##FASTA directive (it is required for proper parsing by Bio::FeatureIO),
and the ID of the feature must be exactly the same as the name of the
sequence following the > in the fasta section.
- The ORGANISM table
- This script assumes that the organism table is populated
with information about your organism. If you are unsure if that is the
case, you can execute this command from the psql command-line:
select * from organism;
If you do not see your organism listed, execute this command to insert it:
insert into organism (abbreviation, genus, species, common_name)
values ('H.sapiens', 'Homo','sapiens','Human');
substituting in the appropriate values for your organism.
- Parents/children order
- Parents must come before children in the GFF file.
- If you are loading analysis results (ie, blat results, gene
predictions), you should specify the -a flag. If no arguments are supplied
with the -a, then the loader will assume that the results belong to an
analysis set with a name that is the concatenation of the source (column
2) and the method (column 3) with an underscore in between. Otherwise, the
argument provided with -a will be taken as the name of the analysis set.
Either way, the analysis set must already be in the analysis table. The
easist way to do this is to insert it directly in the psql shell:
INSERT INTO analysis (name, program, programversion)
VALUES ('genscan 2005-2-28','genscan','5.4');
There are other columns in the analysis table that are optional; see the
schema documentation and '\d analysis' in psql for more information.
Chado has four possible columns for storing the score in the GFF score
column; please use whichever is most appropriate and identifiy it with
--score_col flag (significance is the default). Note that the name of the
column can be shortened to one letter. If you have more than one score
associated with each feature, you can put the other scores in the ninth
column as a tag=value pair, like 'identity=99', and the bulk loader will
put it in the featureprop table (provided there is a cvterm for identity;
see the section above concerning custom tags). Available options are:
- significance (default)
A planned addtion to the functionality of handling analysis results is to allow
"mixed" GFF files, where some lines are analysis results and some
are not. Additionally, one will be able to supply lists of types (optionally
with sources) and their associated entry in the analysis table. The format
will probably be tag value pairs:
--analysis match:Rice_est=rice_est_blast, \
- Grouping features by ID
- The GFF3 specification allows features like CDSes and
match_parts to be grouped together by sharing the same ID. This loader
does not support this method of grouping. Instead the parent feature must
be explicitly created before the parts and the parts must refer to the
parent with the Parent tag.
- External Parent IDs
- The GFF3 specification states that IDs are only valid
within a single GFF file, so you can't have Parent tags that refer to IDs
in another file. By specificifying the "allow_external_parent"
flag, you can relax this restriction. A word of warning however: if the
parent feature's uniquename/ID was modified during loading (to make it
unique), this functionality won't work, as it won't beable to find the
original feature correctly. Actually, it may be worse than not working, it
may attach child features to the wrong parent. This is why it is a bad
idea to use this functionality! Please use with caution.
Allen Day <email@example.com>, Scott Cain <firstname.lastname@example.org>
Copyright (c) 2011
This library is free software; you can redistribute it and/or modify it under
the same terms as Perl itself.