DEEP learning in bio NLP

Background

Named entity recognition (NER) is an important technique that promises to improve information classification and retrieval in biomedical natural language processing (NLP). However, existing approaches primarily rely on either laborious manual curation or feature engineering. Here we adopt deep learning techniques in NLP and repurpose the vast amount of entity-freetext pairs available in the BioSample to train a scalable NER model.

Jupyter notebooks

Key notebooks

Code	Usage
mergeEntities	Merge all the highly similar BioSample entities using cosine similarities
deep_sra_train_and_test	Train an entity recognition model using SRA meta data with entity groupings from mergeEntities
deep_sra_predict	Classify text entity using the trained NER model
Parsing and merging of BioSample data	---

Independent validation

Code	Usage
validationDataGenration.ipynb	validation data generation for comparison against curation and Metamap
NER in batch	predict NER based on all possible sentence segments
scoreAgainstManualCuration_entity_membership	score against manual curation
Parse metamap data
Score against metamap
Auxilary notebooks that probably not used or not critical towards understanding of manuscripts

Code	Usage
downloadFromPMC	download the pubmed text
train_pmc_word2vec.ipynb	Train a word2vec model based on pubmed text, used the pretrained one in the manuscript at the end
uploadToSynapse

Data location

Please download the data from the following websites:

File name	Usage
allSRS.pickle.gz	all BioSample SRS annotations
word vectors	Spacy word vector model
meta data	bioSample to Study mapping table
pretrained models	pretrained LSTM and Spacy word models

Unused data location	Usage
https://www.synapse.org/#!Synapse:syn15661258	all SRX annotations
https://www.synapse.org/#!Synapse:syn16805240	PUBMED ID conversions

Manuscript auxilary data	Description
Machine annotated validation data	Example output of deep NER annotations from NER in batch
Data curated using dataturk

Manuscript

Preprint titled "Creating a scalable deep learning based Named Entity Recognition Model for biomedical textual data by repurposing BioSample free-text annotations " is available at the bioAriv.

Dependency

If u have anaconda, install relevant packages using following command lines:

• conda env create -f environment.yml
• source activate deep_nlp_cpu

License

This work is under MIT license.

#!head -n 20 ./pubmed/PMC0019XXXXX/PMC1913286.txt