Superseded by https://github.com/turbomam/biosample-xmldb-sqldb (which, despite it's name, doesn't use an XML database)

biosample-basex

Using XQueries in a BaseX database to convert NCBI's BioSample database from XML to SQLite.

Querying the native BioSample database requires familiarity with, and access to tools designed for XPath, XQuery or XSLT. The SQLite database generated by this software can be queried from Python programs or a multitude of commercial, open-source, CLI and GUI tools, using the most common query language. No citation provided :-)

BioSample metadata background

When a scientist submits data (like taxonomic identification or gene expression) to NCBI, they are also required to submit metadata about the BioSamples from which that data was collected.

NCBI uses the term attribute to describe certain named information about the BioSamples. Submitters can tag this information about their BioSamples with any attribute names they choose. The NCBI performs curation of the BioSamples, and attributes that appear to share the same meaning as controlled terms from standards like the GSC's MIxS are given harmonized names.

These BioSample attributes should not be confused with XML markup construct that is also called attribute. In fact, the BioSample attributes are actually XML elements (from the path BioSampleSet/BioSample/Attributes/Attribute) although @attribute_name and @harmonized_name are XML attributes.

There are several other paths with elements, and attributes of elements that describe the BioSamples. For example, see the Path Index section of reports/biosample_set_1_info_index.txt

BaseX database limits

As of early January, 2022, the number of XML nodes necessary to model the 22,786,924 BioSamples was 2,397,333,775. This repo uses the BaseX open-source XML database, which is feature rich and performs well. However, BaseX does have per-database limits, such as no more than 2,147,483,648 nodes per database. Therefore, NCBI's biosample_set.xml.gz is split into two chunks with sed and then loaded in two BaseX databases. Queries written in the XQuery language are then executed over the two databases, generating various TSV files.

SQLite output

After running the XQueries, the generated TSVs are loaded into a SQLite database. SQLite was selected because the entire, indexed and ready-to-use database can be shared as a single, compressible file and because of the plethora of compatible tools, which do not require setting up a persistent database server. Examples include Python's built-in sqlite module, the sqlite command line client (which can be downloaded from https://www.sqlite.org/download.html or installed with package managers that are included in many Unix-like operating systems), or with a graphical tool like DBeaver.

Tables in the SQLite database

Note: some column names like id and temp may be reserved words. These should be wrapped in double quotes when writing queries.

• all_attribs
  • All "attribute" elements from all BioSamples, whether they have a harmonized name or not, in a long format.
  • Pretty slow to query, despite indexing several columns.
  • CREATE INDEX all_attribs_idx on all_attribs("harmonized_name", "raw_id", "attribute_name");
  • Maybe the indexes weren't set up properly? (I haven't tried querying without indexes.)
  • 324,806,062 rows
  • columns (TEXT unless otherwise specified):
    • raw_id (INTEGER)
    • attribute_name
    • harmonized_name
    • value
• env_package_repair
  • Standardizes the user-submitted env_package values
  • From env_package_repair_curated.tsv
  • The Makefile has a step to create a target/env_package_repair_new.tsv. No tools are provided yet for automatically reconciling the new file with the previous curated file.
  • 99 rows
  • columns (TEXT unless otherwise specified):
    • env_package_raw
    • count (INTEGER)
    • env_package (indexed)
    • checklist_fragment
    • curation_confidence
    • notes
    • curation_complete
• harmonized_wide
  • Just the values of those attributes that have harmonized names, as TEXT, in wide format. Also includes the BioSample raw_ids, as INTEGERs.
  • 22,608,173 rows
  • 516 columns including raw_id
  • Provided with indices on raw_id_idx and env_package. More could be added for better queries at the cost of disk footprint.
• harmonized_wide_repaired
  • Exact same shape as harmonized_wide. Serves as a place to collect values that have gone though some cleanup/repair process. Currently env_package is the only column repaired by this repo, by way of the env_package_repair table.
  • Provided with an index on raw_id_idx
• non_attribute_metadata
  • Metadata about the BioSamples that doesn't come from "attribute" entities. External IDs and links have been simplified compared to previous versions. Among other things, DOIs have been omitted , since they are only available for ~ 400 out of the 20+ million BioSamples.
  • 24,580,658 rows. That's greater than the number of harmonized_wide rows as some BioSamples don't have any attributes with harmonized names.
  • columns (TEXT unless otherwise indicated):
    • id (Not native to the biosample_set.xml file. Built from the prefix "BIOSAMPLE:" and the primary_id)
    • accession
    • raw_id (INTEGER, primary key)
    • primary_id
    • sra_id
    • bp_id (while these are stored as TEXT, they are numerical BioProject IDs and should really be joined with the BioProject XML database to get BioProject accessions.)
    • model
    • package
    • package_name
    • status
    • status_date
    • taxonomy_id
    • taxonomy_name
    • title
    • paragraph
• biosample_basex_merged view
  • non_attribute_metadata + harmonized_wide tables

As of 2022-01-03

• biosample_basex.db.gz: 7.7 GB
• biosample_basex.db: 67.9 GB

Usage

BaseX and SQLite can be installed with homebrew on Macs or apt-get on Ubuntu Linux machines. They're both open-source software. SQLite3 3.32 or greater is required.

Downloading the BaseX .zip archive makes it a little more straightforward to increase the amount of memory allocated by the launch scripts. For example, on a 32 GB machine, basex/bin/basex might look like this:

#!/usr/bin/env bash

# Path to this script
FILE="${BASH_SOURCE[0]}"
while [ -h "$FILE" ] ; do
  SRC="$(readlink "$FILE")"
  FILE="$( cd -P "$(dirname "$FILE")" && \
           cd -P "$(dirname "$SRC")" && pwd )/$(basename "$SRC")"
done
MAIN="$( cd -P "$(dirname "$FILE")/.." && pwd )"

# Core and library classes
CP=$MAIN/BaseX.jar:$MAIN/lib/custom/*:$MAIN/lib/*:$CLASSPATH

# Options for virtual machine (can be extended by global options)
BASEX_JVM="-Xmx24g $BASEX_JVM"

# Run code
exec java -cp "$CP" $BASEX_JVM org.basex.BaseX "$@"

https://docs.basex.org/wiki/Main_Page

https://basex.org/download/

NERSC cori specific notes

make all takes roughly 8 hours.

• cd /global/cfs/cdirs/m3513/endurable/biosample
• git clone [email protected]:turbomam/biosample-basex.git
• wget the latest BaseX zip archive
  • 2022-01-02: wget https://files.basex.org/releases/9.6.4/BaseX964.zip
• unzip BaseX964.zip
• increase the Java memory allocation in basex/bin/basex
  • BASEX_JVM="-Xmx96g $BASEX_JVM"
  • That may be more than necessary. Seems to run OK on an Intel MacBook Pro with 24 GB allocated.
• cp biosample-basex/template.env biosample-basex/.env
• edit biosample-basex/.env
  • BASEXCMD = ../basex/bin/basex
  • The value for del_from should be roughly half the number of BioSamples. The suggested value of 12500001 should be reasonable but could be revised after downloading and unpacking the BioSample XML file. Run something like tail -n 50 target at the shell prompt, note the id attribute of the last <BioSample>, and set del_from to an integer close to one half of that last id value.
• Start a session manager like screen. If you get disconnected, it may be tricky to log back into the same cori node because of load balancing, but scripts started before the disconnection should run to completion. It doesn't hurt to force your client computer to stay awake with something like caffeine.
• module load python
  • On cori, this makes common Python packages like pandas available. On other systems, users should create a venv virtual environment, enter it, and run pip install -r requirments.txt
• cd biosample-basex/
• make all

After running make final_sqlite_gz_dest, the SQLite will be available at https://portal.nersc.gov/project/m3513/biosample/