Set up an automated ML task to predict drug -> drug target edges

[justaddcoffee]

Describe the desired behavior

It would be useful for Drug Central to have inferred drug -> drug target edges, which we could formulate as a graph ML link prediction task.

We maybe could/should do this in an automated way, on each build of KG-IDG.

A possible roadmap:

• Write Jupyter notebook for link prediction task
• Write some tooling to run this code from Jenkins
• Write some code to push the results to KGHub

Additional context

Per convo with Tudor et al on IDG just now

[justaddcoffee]

Discussing this more with Harry - some thoughts.

We can use NEAT for this (i.e. spin up a Google Cloud instance from Jenkins, pass off a YAML file to the instance describing the graph ML task, wait for it to complete and upload artifacts to S3).

This strategy has the advantage of:

• helping us get NEAT updated
• helping us flesh out the general use case of getting automatic embedding and graph ML working with KGHub and other projects (Monarch, eco-kg, others)

[caufieldjh]

As per IDG meeting on Jan 20, following drug categories would be good for evaluation:

• Those targeting GPCRs
• antihistaminics (H1 blockers)
• antipsychotics (multi-target)
• antivirals (no human targets)
• antiestrogens (mostly ERalpha)

[caufieldjh]

I don't think we include any viral proteins right now, so we'd expect the antivirals to have few/no human targets, especially because we aren't doing any inference based off sequence or structure.