Skip to content
Toyofumi Fujiwara edited this page Sep 12, 2019 · 28 revisions

Phenotype / disease (nanbyo) / pedigree

Writethon

Members

  • lead: Toyofumi Fujiwara
  • lead: Orion Buske
  • Mariel Marshall
  • Ryota Yamada
  • Kota Ninomiya
  • Jae-moon Shin
  • Yui Asano
  • Yasunori Yamamoto
  • Atsuko yamaguchi
  • Soichi Ogishima

Projects

  1. Rare Disease Doctors
  2. Mapping Nanbyo diseases to MONDO diseases
  3. Building PubCaseFinder RDF
  4. HPO Japanese update

1. Rare Disease Doctors

State the problem you worked on

When conditions of patients are rare, it can be challenging to find experts with experience or knowledge of their particular condition.

Give the state-of-the art/plan

We can search for experts for rare diseases in Orphanet which provides a searchable directory of rare disease experts. Althogh the directory of Orphanet have been manually built by curators, it inherently suffer from limited coverage. In this hackathon, we automatically identified rare disease experts based on two million published case reports in PubMed and assigned scores to the experts using machine learning methods for the prioritization.

Describe what you have done/results starting with The working group created...

  • Made a dataset of rare disease experts / institutions: Diseases: 7,759 / Authors: 736,415

    • Made expert-disease pairs form two million published case reports
      • Extracted authors and their afficiations
      • Annotated those abstracts with OMIM / Orphanet diseases
    • Made a list of 148,940 Japanese doctors who are officially allowed to diagnose Nanbyo patients
    • Made a list of 37,819 Japanese pediatricians who are officially allowd to diagnose Nanbyo children
  • Made four datesets for evaluation of machine liearning methods

    • Scraped 2013 expert-disease pairs from GeneReviews
    • Scraped 2876 expert-disease pairs from all the Orphanet disease articles
    • Scraped 192 expert-disease pairs from MHLW Nanbyo research group data (including only Japanese)
    • Scraped 270 expert-disease pairs from AMED Nanbyo research group data (including only Japanese)
  • Assigned scores to doctors using machine learning

    • Annotated with 7 features, based on http://proceedings.mlr.press/v68/wang17a.html
    • For validation, created 5 separate datasets, each with all positive examples from GeneReviews (N=428) and 10 times as many randomly-selected unclassified examples to serve as negative examples (N=4280)
    • Logistic regression was applied to each validation set using 5-fold cross validation, and the results plotted in Figure
    • A model was then trained on the full dataset of positive (N=428) and unclassified examples serving as negative examples (N=2.7M), and the posterior probabilities were used as the final specialist scores
  • Implemented a new web service

Write a conclusion

We made a dataset of 736,415 authors - 7,759 diseases relations and four datasets for the evaluation of machine learning methods. We released Rare Disease Doctors (https://rarediseasedoctors.dbcls.jp/), a new searchable directory of rare disease experts in a freely available web application.

Write up any future work

  • add up/down-vote feature
  • evaluate w/ Japanese expert datasets
  • add country filter, nanbyo

2. Mapping Nanbyo diseases to MONDO diseases

State the problem you worked on

Japan has been dealing with the problems associated with rare and intractable diseases which are called "Nanbyo" in Japanese by providing the subsidy of medical expense to the patient.Every year, the number of the designated diseases whose patients can get the subsidy has been increasing and it amounts up to 333. However, the classification of nanbyo is not unified. For example, mithochondrial disease is designated as a disease, but includes 83 diseases such as Niemann-pick disease type C. On the other hand, systemic lupus erythematosus includes nothing else. Therefore, we need to look into the classification in detail, traslate the names of Nanbyo, and map them to MONDO or something like it in order to link the information of Nanbyo to the world.

Give the state-of-the art/plan

Previously we tried exact match to find the counterparts of the Nanbyo diseases. It found only 32.28% of the counterparts because the same disease are written different names in Nanbyo list and MONDO list. First, we tried to reduce the variants by using classical natural language processing techniques. We applied the following preprocessings before finding mathces.

  • lowercase letters
  • remove not word characters
  • remove "'s" of possessive case
  • replace greek letters to alphabets (ex. "α" to "alpha")
  • replace roman numerals to arabic numerals (ex. "I" to "1")
  • remove prepositions
  • remove stop words (ex. syndrome, disease, type)

Second, we also challenged to find the candidate MONDO IDs for the Nanbyo disease which could not be matched the first approach. At this step we used the relationship of MONDO (diseases in MONDO has parents and children respectively).

Describe what you have done/results starting with The working group created...

In oreder to get the hierarchical structure of Nanbyo,we looked into the medical certificates of Nanbyo formatted by the government. We got lower-layer diseases based on whether the entry to write about subtypes exists.Then, I found 949 different diseases in 331 diseases. 949

By applying those preprocessing, we found the conterparts for 79.43% (753 diseases) of the Nanbyo diseases (Figure 1). Figure 1

And found the candidates for the 12.24% (116 diseases) of remaining 20.57%. In figure 2 we show the candidates for Nanbyo:64-2 which the system suggested. And MONDO:0019740 is exactly the counterpart of the Nanbyo. Figure 2

Write a conclusion

We made a dataset of 949 Nanbyo diseases which mostly mapped to MONDO. This will en

Write up any future work

The experts will check the results and create a completely map for Nanbyo diseases and MONDO diseases. And we plan to provide the results to MONDO data.

3. Building PubCaseFinder RDF

  • Jae-moon Shin
  • Atsuko yamaguchi
  • Toyofumi Fujiwara

State the problem you worked on

There are many Phenotype-Disease relationships and Phenotype-Gene relationships, but there are not many Phenotype-Gene relationships, only manual information.

Give the state-of-the art/plan

The goal was to create a SPARQL endpoint using information from OMIM, Orphanet, HPO, and PubCaseFinder. In addition, the goal was to utilize the results of automatic extraction as well as manual by utilizing the information of DisGeNET and PubCaseFinder.

Describe what you have done/results starting with The working group created...

  1. Created RDF Schema(Figure 1).

  1. Converted relations OMIM-HPO, OMIM-Gene, Orpha-HPO, Orpha-Gene stored inn RDB for PubCaseFinder to RDF.

  2. Created a sparql endpoint of PubCaseFinder RDF(Installed Virtuoso and uploaded the RDF datasets above into it). SPARQL endpoint : https://pubcasefinder-rdf.dbcls.jp/sparql

  3. Compared Phenotype-Gene relations derived from our RDF data with Phenotype-Gene relations provided at the HPO website(Figure 2).

Write a conclusion

The Phenotype-Disease-Gene schema was created and the SPARQL endpoint was created.

Write up any future work

We're going to continue analysis of Phenotype-Gene relations using PubCaseFinder RDF.

Clone this wiki locally