The PheneBank Project

PheneBank aims at automatic extraction and validation of a database of human phenotype-disease associations in the scientific literature. This package provides code, data, and models for the following three purposes:

1. Named Entity Recognition (Tagging)

The model is trained to support 9 categories of entities:

• Phenotype
• Disease
• Anatomy
• Cell
• Cell_line
• GPR
• Gene_variant
• Molecule
• Pathway

2. Harmonisation (Grounding)

Map an entity to its corresponding concept in any of the following 5 ontologies:

• SNOMED (Phenotype, Disease, GPR, Anatomy, Molecule, Cell, Cell_line, Gene_variant)
• HPO (Phenotype, Disease)
• MESH (Phenotype, Disease)
• FMA (Anatomy)
• PRO (GPR)

3. Tagging and Grounding

Given an input text, extract its entities and map each to its corresponding concept in the ontologies (a pipeline containing both previous stages).

Online demo

• To be activated soon!

Data

Download the followings:

• embeddings.zip (around 1.1GB)
• data.zip (around 200MB)

Getting Started:

To get started with the pipeline, first obtain the required data and decompress them in the project directory. Then, import pipeline into your project:

from pipeline import pipe
pp = pipe()
input_text = "Risk factors for recurrent respiratory infections in preschool children in China."

Find entities in an input text:

pp.tag(input_text)

The output will look like the following (formatted for clarity). Lists of tuples, one tuple per sentence. Each tuple contains two lists: words and their corresponding tags.

[
(['Risk', 'factors', 'for', 'recurrent', 'respiratory', 'infections', 'in', 'China.'],
['O', 'O', 'O', 'B-Phenotype', 'I-Phenotype', 'I-Phenotype', 'O', 'O'])
]

Find entities in the text and harmonise (map) them to their corresponding ontologies:

pp.tag_harmonise(input_text)

The output will have each sentence as a list of tuples. Each tuple has three parts: word, tag (Null if not an entity), (the list of) corresponding concept IDs ([] if no mapping was found).

[
[
('Risk', 'Null', []),
('factors', 'Null', []),
('for', 'Null', []),
('recurrent respiratory infections', 'Phenotype', [('HP:0002205', 1.0)]),
('in', 'Null', []),
('China', 'Null', [])
]
]

Updaing ontology embeddings:

1. Place the new ontology file (eg, hp.obo) under the data directory.
2. Fix the corresponding path in utils/project_config.py.
3. Use the ontology_embedding.py script under grounding to create a new semantic embedding.

You can use the following command in the "embeddings" directory to binarise the ontology embedding:

$ ./convertvec txt2bin [embedding.txt] [embedding.bin]

(convertvec script from https://github.com/marekrei/convertvec)

Dependencies

The tagging stage relies on Anago, a Bidirectional LSTM-CRF for Sequence Labeling: https://github.com/Hironsan/anago

Citation

M.T. Pilehvar, D. Smedley, A. Bernard, and N. Collier: PheneBank: a literature-based database of phenotypes. Bioinformatics, Volume 38, Issue 4, 2022.