Recognizing entities in text is the first step towards machines that can extract medical insights out of enormous document repositories like pubmed.  

Imagine asking your computer "which therapies are most effective for my disease?" To answer this kind of question machines can read millions of documents, but first they must know which words are therapies and diseases.  Models that identify entities in text are called Named Entity Recognition (NER) models.

Let's build an NER model for genes from 2000 abstracts reviewed in sysrev's Gene Hunter project. In ~5 minutes your model will identify genes in text like the below:

Automatically identified genes from a model trained on sysrev.com/p/3144 annotations.asdf

This is called annotated text and we can represent it in python with a simple list.

["Objectives: We investigated whether polymorphisms (SNPs) in the promoter region of TNFA, or in the autoinflammatory TNFRSF1A", 
{entities: [(84,88,'GENE'),(117,125,'GENE'),...]}]
Each annotated text block is a python list with raw text and a dictionary of character offsets for each annotation. Here two gene annotations are shown at offsets 84 and 117 representing TNFA and TNFRSF1A.

Start by installing the PySysrev package: pip install PySysrev.  Then just execute the next 13 lines of code to have your very own gene NER model.  

import PySysrev, spacy, random
TRAIN_DATA = PySysrev.processAnnotations(project_id=3144,label='GENE')
Getting spacy.io ready annotations from gene hunter is a one liner. TRAIN_DATA is a list of annotated paragraphs. Get a pandas dataframe with PySysrev.getAnnotations(project_id=3144)
nlp = spacy.blank('en')          # create a spacy model
nlp.meta['name'] = 'gene'        # name the model 'gene'

ner = nlp.create_pipe('ner')     # create an NER stage
ner.add_label('GENE')            # add the label 'GENE' to the stage

nlp.add_pipe(ner)                # put the pipe together
optimizer = nlp.begin_training() # get an optimizer for training the model
Set up a spacy NER model optimizer in just a few lines.
for itn in range(30):
    random.shuffle(TRAIN_DATA)                     #shuffle examples 
    text = [item[0] for item in TRAIN_DATA]        #get training text items
    annotations = [item[1] for item in TRAIN_DATA] #get training annotations
    nlp.update(text, annotations, sgd=optimizer, drop=0.6)
Train the model! This can take a while. In larger scale examples you probably want to use minibatching like spaCy's util.minibatch to update the model in batches.

Thats it! Your model is done. Lets give it a test:

doc = nlp("""Epigenetic Silencing of the mutL homolog 1 (MLH1) Promoter in
Relation to the Development of Gastric Cancer (GC) and its use as a
Biomarker for Patients with Microsatellite Instability.""")

from spacy import displacy
displacy.serve(doc,style="ent")
Epigenetic Silencing of the mutL homolog 1 ( MLH1 GENE ) Promoter in Relation to the Development of Gastric Cancer (GC) and its use as a Biomarker for Patients with Microsatellite Instability.
Our nlp model successfully identifies genes in text. You can generate visualizations from your NER results like the above by using spaCy visualizers.

It looks like the model did a decent job on this snippet. It matched MLH1 but missed the full name. It also knew GC wasn't a gene.  

With more training this model could be further improved.  To get an alert when the next gene hunter sysrev starts just Subscribe to this blog.

You can get started doing annotations in a public sysrev today for free. Learn all about it at "Creating Annotations".

@misc{Luechtefeld2019simplener,
  title={How to extract genes from text with Sysrev and spaCy},
  author={Luechtefeld,Thomas},
  year={2019 (accessed <your date>)},
  url="https://blog.sysrev.com/simple-ner/"
}