Training datasets for ML/AI - publication centric

[ValWood]

Create a "publication centric" file containing all entities / annotations (all datatypes) for each publication.

Json?

[kimrutherford]

JSON makes sense. How urgent is this?

[ValWood]

It would be good to have it in a few of weeks I think to keep the ball rolling. I'm meeting the ePMC ML person on Monday if you want to join (forwarded the invite)
v

[kimrutherford]

I'll start this on Monday. It might take a couple of days because the existing code needs improving first. A lot was written in a hurry for PomBase v2. Now I've had time (7 years?) to think about it, there are better ways to do things.

Proposed JSON structure (work in progress):

PMID:

• MF annotations
  • gene
  • term
  • evidence
  • extension (as JSON)
  • comment
• BP annotations
  • (same details)
• CC annotations
  • (same details)
• FYPO annotations
  • genotype
    • loci (one or more)
      • alleles (one or more):
        • type
        • name
        • description
  • term
  • evidence
  • conditions (as JSON)
  • extension (as JSON)
  • comment
• Modification annotations
  • ...

[kimrutherford]

From Zoom: make sure to include annotation comments in the output.

[ValWood]

related: #1185 we'll discuss this on the next call....

[kimrutherford]

After the chat with ePMC a while ago, I'm wondering if it's useful to create a file like this in advance. It sounded like there are particular file formats that each group uses. So perhaps we should create files when asked? Unless it's a especially wacky format I think I could create files on request with a 24 turn-around.

[ValWood]

OK keep this on the back burner.