This repo contains the source code and dataset for the PhenoTagger.
PhenoTagger is a hybrid method that combines dictionary and deep learning-based methods to recognize Human Phenotype Ontology (HPO) concepts in unstructured biomedical text. It is an ontology-driven method without requiring any manually labeled training data, as that is expensive and annotating a large-scale training dataset covering all classes of HPO concepts is highly challenging and unrealistic. Please refer to our paper for more details:
- Use Transformers instead of kears-bert to load deep learning models.
- Use Tensorflow.keras instead of Keras.
- Add a PubMedBERT model.
- Re-train phenotype models using the newest version of HPO (hp/releases/2023-10-09)
- add negation and uncertainty detection function using NegBio
(2022-11-24)
- build a app demo for PhenoTagger
(2022-05-10):
- Fix some bugs to speed up the processing time.
- Add a Bioformer model (a light weight BERT in biomedical domain).
- Re-train phenotype models using the newest version of HPO (hp/releases/2022-04-14)
- Dependency package
- Data and model preparation
- Instructions for tagging text with PhenoTagger
- Instructions for training PhenoTagger
- Performance on HPO GSC+
- Citing PhenoTagger
PhenoTagger has been tested using Python3.9/3.10 on CentOS and uses the following dependencies on a CPU and GPU:
To install all dependencies automatically using the command:
$ pip install -r requirements.txt
-
We released the PhenoTagger v1.2 with the pubmedbert model. you could download the released version on our GitHub.
-
If you would like to use other models. You need to create a model folder named "models" in the PhenoTagger folder, then download the model files ( four trained models for HPO concept recognition are released, i.e., CNN, Bioformer, BioBERT, PubMedBERT) into the model folder.
- First download original files of the pre-trained language models (PLMs): Bioformer, BioBERT, PubMedBERT
- Then download the fine-tuned model files for HPO in Here
-
The corpora used in the experiments are provided in /data/corpus.zip. Please unzip the file, if you need to use them.
You can use our trained PhenoTagger to identify the HPO concepts from biomedical texts by the PhenoTagger_tagging.py file.
The file requires 2 parameters:
- --input, -i, help="the folder with input files"
- --output, -o, help="output folder to save the tagged results"
The file format can be in BioC(xml) or PubTator(tab-delimited text file) (click here to see our format descriptions). There are some examples in the /example/ folder.
Example:
$ python PhenoTagger_tagging.py -i ../example/input/ -o ../example/output/
We also provide some optional parameters for the different requirements of users in the PhenoTagger_tagging.py file.
para_set={
'model_type':'bioformer', # three deep learning models are provided: cnn, bioformer, biobert or pubmedbert
'onlyLongest':False, # False: return overlapping concepts; True: only return the longgest concepts in the overlapping concepts
'abbrRecog':True, # False: don't identify abbreviation; True: identify abbreviations
'negation': True, # True: negation detection
'ML_Threshold':0.95, # the Threshold of deep learning model
}
When 'negation' is set to true, two output files (e.g., ex.neg2.PubTator (the output with negation results) and ex.PubTator (the output without negation results)) will be generated by the tool for each input file (ex.PubTator).
The file requires 3 parameters:
- --input, -i, help="input the ontology .obo file"
- --output, -o, help="the output folder of dictionary"
- --rootnode, -r, help="input the root node of the ontogyly"
Example:
$ python Build_dict.py -i ../ontology/hp.obo -o ../dict/ -r HP:0000118
After the program is finished, 6 files will be generated in the output folder.
- id_word_map.json
- lable.vocab
- noabb_lemma.dic
- obo.json
- word_id_map.json
- alt_hpoid.json
The file requires 4 parameters:
- --dict, -d, help="the input folder of the ontology dictionary"
- --fileneg, -f, help="the text file used to generate the negatives" (You can use our negative text "mutation_disease.txt" )
- --negnum, -n, help="the number of negatives, we suggest that the number is the same with the positives."
- --output, -o, help="the output folder of the distantly-supervised training dataset"
Example:
$ python Build_distant_corpus.py -d ../dict/ -f ../data/mutation_disease.txt -n 10000 -o ../data/distant_train_data/
After the program is finished, 3 files will be generated in the outpath:
- distant_train.conll (distantly-supervised training data)
- distant_train_pos.conll (distantly-supervised training positives)
- distant_train_neg.conll (distantly-supervised training negatives)
The file requires 4 parameters:
- --trainfile, -t, help="the training file"
- --devfile, -d, help="the development set file. If don't provide the dev file, the training will be stopped by the specified EPOCH"
- --modeltype, -m, help="the deep learning model type (cnn, biobert, pubmedbert or bioformer?)"
- --output, -o, help="the output folder of the model"
Example:
$ python PhenoTagger_training.py -t ../data/distant_train_data/distant_train.conll -d ../data/corpus/GSC/GSCplus_dev_gold.tsv -m bioformer -o ../models/
After the program is finished, 2 files will be generated in the output folder:
- cnn.h5/biobert.h5 (the trained model)
- cnn_dev_temp.tsv/biobert_dev_temp.tsv (the prediction results of the development set, if you input a development set file)
If you're using PhenoTagger, please cite:
- Ling Luo, Shankai Yan, Po-Ting Lai, Daniel Veltri, Andrew Oler, Sandhya Xirasagar, Rajarshi Ghosh, Morgan Similuk, Peter N Robinson, Zhiyong Lu. PhenoTagger: A Hybrid Method for Phenotype Concept Recognition using Human Phenotype Ontology. Bioinformatics, Volume 37, Issue 13, 1 July 2021, Pages 1884–1890.
This research is supported by the Intramural Research Programs of the National Institutes of Health, National Library of Medicine. Thanks to Dr. Chih-Hsuan Wei for his help with Web APIs.
This tool shows the results of research conducted in the Computational Biology Branch, NCBI. The information produced on this website is not intended for direct diagnostic use or medical decision-making without review and oversight by a clinical professional. Individuals should not change their health behavior solely on the basis of information produced on this website. NIH does not independently verify the validity or utility of the information produced by this tool. If you have questions about the information produced on this website, please see a health care professional. More information about NCBI's disclaimer policy is available.