Babel

Introduction

The Biomedical Data Translator integrates data across many data sources. One source of difficulty is that different data sources use different vocabularies. One source may represent water as MESH:D014867, while another may use the identifier DRUGBANK:DB09145. When integrating, we need to recognize that both of these identifiers are identifying the same concept.

Babel integrates the specific naming systems used in the Translator, creating equivalent sets across multiple semantic types, and following the conventions established by the biolink model. It checks these conventions at runtime by querying the Biolink Model service. Each semantic type (such as chemical substance) requires specialized processing, but in each case, a JSON-formatted compendium is written to disk. This compendium can be used directly, but it can also be served by the Node Normalization service or another frontend.

We anticipate that the simple approach taken here will soon be overtaken by more advanced probabilistic procedures, so caution should be taken in building strong dependencies against the Babel code.

Configuration

The ./kubernetes directory contains Kubernetes manifest files that can be used to set up a Pod to run Babel in. They'll give you an idea of the disk space and memory requirements needed to run this pipeline.

Before running, read through config.yaml and make sure that the settings look correct. You will need to update the version numbers of some databases that need to be downloaded, or change the download and output directories.

A UMLS API key is required in order to download UMLS and RxNorm databases. You will need to set the UMLS_API_KEY environmental variable to a UMLS API key, which you can obtain by creating a profile on the UMLS Terminology Services website.

Building Compendia

To run Babel, you will need to install uv. uv manages the Python environment and installs dependencies for you.

Compendia building is managed by snakemake. To build, for example, the anatomy related compendia, run

uv run snakemake --cores 1 anatomy

Currently, the following targets build compendia and synonym files:

• anatomy
• chemical
• disease
• gene
• genefamily
• protein
• macromolecular_complex
• taxon
• process
• publications

And these two build conflations:

• geneprotein
• drugchemical

Each target builds one or more compendia corresponding to a biolink model category. For instance, the anatomy target builds compendia for biolink:AnatomicalEntity, biolink:Cell, biolink:CellularComponent, and biolink:GrossAnatomicalStructure.

You can also just run:

uv run snakemake --cores 1

without a target to create all the files that are produced as part of Babel, including all reports and alternate exports.

If you have multiple CPUs available, you can increase the number of --cores to run multiple steps in parallel.

Build Process

The information contained here is not required to create the compendia, but may be useful to understand. The build process is divided into two parts:

1. Pulling data from external sources and parsing it independent of use.
2. Extracting and combining entities for specific types from these downloaded data sets.

This distinction is made because a single data set, such as MeSH or UMLS may contain entities of many different types and may be used by many downstream targets.

Pulling Data

The datacollection snakemake file coordinates pulling data from external sources into a local filesystem. Each data source has a module in src/datahandlers. Data goes into the babel_downloads directory, in subdirectories named by the curie prefix for that data set. If the directory is misnamed and does not match the prefix, then labels will not be added to the identifiers in the final compendium.

Once data is assembled, we attempt to create two extra files for each data source: labels and synonyms. labels is a two-column tab-delimited file. The first column is a CURIE identifier from the data source, and the second column is the label from that data set. Each entity should only appear once in the labels file. The labels file for a data set does not subset the data for a specific purpose, but contains all labels for any entity in that data set.

synonyms contains other lexical names for the entity and is a 3-column tab-delimited file, with the second column indicating the type of synonym (exact, related, xref, etc.)

Creating compendia

The individual details of creating a compendium vary, but all follow the same essential pattern.

First, we extract the identifiers that will be used in the compendia from each data source that will contribute, and place them into a directory. For instance, in the build of the chemical compendium, these ids are placed into /babel_downloads/chemical/ids. Each file is a two-column file containing curie identifiers in column 1, and the Biolink type for that entity in column 2.

Second, we create pairwise concords across vocabularies. These are placed in e.g. babel_downloads/chemical/concords. Each concord is a three-column file of the format:

<curie1> <relation> <curie2>

While the relation is currently unused, future versions of Babel may use the relation in building cliques.

Third, the compendia is built by bringing together the ids and concords, pulling in the categories from the id files, and the labels from the label files.

Fourth, the compendia is assessed to make sure that all the ids in the id files made into one of the possibly multiple compendia. The compendia are further assessed to locate large cliques and display the level of vocabulary merging.

Building with Docker

You can build this repository by running the following Docker command:

$ docker build .

It is also set up with a GitHub Action that will automatically generate and publish Docker images to https://github.com/TranslatorSRI/Babel/pkgs/container/babel.

Known issue: if you want to use git fetch from this Docker image, you need to manually remote the Basic authentication command from .git/config before it will work. We're tracking this at #119.

Running with Docker

You can also run Babel with Docker. There are two directories you need to bind or mount from outside the container:

$ docker run -it --rm --mount type=bind,source=...,target=/home/runner/babel/babel_downloads --entrypoint /bin/bash ggvaidya/babel

The download directory (babel/babel_downloads) is used to store data files downloaded during Babel assembly.

The script scripts/build-babel.sh can be used to run snakemake with a few useful settings (although just running uv run snakemake --cores 5 should work just fine.)

Deploying with Kubernetes

The kubernetes/ directory has example Kubernetes scripts for deploying Babel to a Kubernetes cluster. You need to create three resources:

• kubernetes/babel-downloads.k8s.yaml creates a Persistent Volume Claim (PVC) for downloading input resources from the internet.
• kubernetes/babel-outputs.k8s.yaml creates a PVC for storing the output files generated by Babel. This includes compendia, synonym files, reports and intermediate files.
• kubernetes/babel.k8s.yaml creates a pod running the latest Docker image from ggvaidya/babel. Rather than running the data generation automatically, you are expected to SSH into this pod and start the build process by:

  1. Edit the script scripts/babel-build.sh to clear the DRY_RUN property so that it doesn't , i.e.:

     export DRY_RUN=

  2. Creating a screen to run the program in. You can start a Screen by running:

     $ screen

  3. Starting the Babel build process by running:

     $ bash scripts/babel-build.sh

     Ideally, this should produce the entire Babel output in a single run. You can also add --rerun-incomplete if you need to restart a partially completed job.

     To help with debugging, the Babel image includes .git information. You can switch branches, or fetch new branches from GitHub by running git fetch origin-https.

  4. Press Ctrl+A D to "detach" the screen. You can reconnect to a detached screen by running screen -r. You can also see a list of all running screens by running screen -l.

  5. Once the generation completes, all output files should be in the babel_outputs directory.