Gilda: Grounding Integrating Learned Disambiguation

Gilda is a Python package and REST service that grounds (i.e., finds appropriate identifiers in namespaces for) named entities in biomedical text.

Gyori BM, Hoyt CT, Steppi A (2022). Gilda: biomedical entity text normalization with machine-learned disambiguation as a service. Bioinformatics Advances, 2022; vbac034 https://doi.org/10.1093/bioadv/vbac034.

Installation

Gilda is deployed as a web service at http://grounding.indra.bio/ (see Usage instructions below), however, it can also be used locally as a Python package.

The recommended method to install Gilda is through PyPI as

pip install gilda

Note that Gilda uses a single large resource file for grounding, which is automatically downloaded into the ~/.data/gilda/<version> folder during runtime (see pystow for options to configure the location of this folder).

Given some additional dependencies, the grounding resource file can also be regenerated locally by running python -m gilda.generate_terms.

Documentation and notebooks

Documentation for Gilda is available here. We also provide several interactive Jupyter notebooks to help use and customize Gilda:

Gilda Introduction provides an interactive tutorial for using Gilda.
Custom Grounders shows several examples of how Gilda can be instantiated with custom grounding resources.
Model Training provides interactive sample code for training new disambiguation models.

Usage

Gilda can either be used as a REST web service or used programmatically via its Python API. An introduction Jupyter notebook for using Gilda is available at https://github.com/indralab/gilda/blob/master/notebooks/gilda_introduction.ipynb

Use as a Python package

For using Gilda as a Python package, the documentation at http://gilda.readthedocs.org provides detailed descriptions of each module of Gilda and their usage. A basic usage example for named entity normalization (NEN), or grounding is as follows:

import gilda
scored_matches = gilda.ground('ER', context='Calcium is released from the ER.')

Gilda also implements a simple dictionary-based named entity recognition (NER) algorithm that can be used as follows:

import gilda
results = gilda.annotate('Calcium is released from the ER.')

Use as a web service

The REST service accepts POST requests with a JSON header on the /ground endpoint. There is a public REST service running at http://grounding.indra.bio but the service can also be run locally as

python -m gilda.app

which, by default, launches the server at localhost:8001 (for local usage replace the URL in the examples below with this address).

Below is an example request using curl:

curl -X POST -H "Content-Type: application/json" -d '{"text": "kras"}' http://grounding.indra.bio/ground

The same request using Python’s request package would be as follows:

import requests
requests.post('http://grounding.indra.bio/ground', json={'text': 'kras'})

The web service also supports multiple inputs in a single request on the ground_multi endpoint, for instance

import requests
requests.post('http://grounding.indra.bio/ground_multi',
              json=[
                  {'text': 'braf'},
                  {'text': 'ER', 'context': 'endoplasmic reticulum (ER) is a cellular component'}
              ]
          )

Resource usage

Gilda loads grounding terms into memory when first used. If memory usage is an issue, the following options are recommended.

Run a single instance of Gilda as a local web service that one or more other processes send requests to.
Create a custom Grounder instance that only loads a subset of terms appropriate for a narrow use case.
Gilda also offers an optional sqlite back-end which significantly decreases memory usage and results in minor drop in the number of strings grounder per unit time. The sqlite back-end database can be built as follows with an optional [db_path] argument, which if used, should use the .db extension. If not specified, the .db file is generated in Gilda’s default resource folder.

python -m gilda.resources.sqlite_adapter [db_path]

A Grounder instance can then be instantiated as follows:

from gilda.grounder import Grounder
gr = Grounder(db_path)
matches = gr.ground('kras')

Run web service with Docker

After cloning the repository locally, you can build and run a Docker image of Gilda using the following commands:

$ docker build -t gilda:latest .
$ docker run -d -p 8001:8001 gilda:latest

Alternatively, you can use docker-compose to do both the initial build and run the container based on the docker-compose.yml configuration:

$ docker-compose up

Default grounding resources

Gilda is customizable with terms coming from different vocabularies. However, Gilda comes with a default set of resources from which terms are collected (almost 2 million entries as of v1.1.0), without any additional configuration needed. These resources include:

HGNC (human genes)
UniProt (human and model organism proteins)
FamPlex (human protein families and complexes)
CHeBI (small molecules, metabolites, etc.)
GO (biological processes, molecular functions, complexes)
DOID (diseases)
EFO (experimental factors: cell lines, cell types, anatomical entities, etc.)
HP (human phenotypes)
MeSH (general: diseases, proteins, small molecules, cell types, etc.)
Adeft (misc. terms corresponding to ambiguous acronyms)

Citation

@article{gyori2022gilda,
    author = {Gyori, Benjamin M and Hoyt, Charles Tapley and Steppi, Albert},
    title = "{{Gilda: biomedical entity text normalization with machine-learned disambiguation as a service}}",
    journal = {Bioinformatics Advances},
    year = {2022},
    month = {05},
    issn = {2635-0041},
    doi = {10.1093/bioadv/vbac034},
    url = {https://doi.org/10.1093/bioadv/vbac034},
    note = {vbac034}
}

Funding

The development of Gilda was funded under the DARPA Communicating with Computers program (ARO grant W911NF-15-1-0544) and the DARPA Young Faculty Award (ARO grant W911NF-20-1-0255).

Software documentation