Prompts for BioGPT embeddings

[miquelduranfrigola]

We need to define prompts for BioGPT. Embedding extraction is already in place.

We can enumerate prompts and produce embeddings correspondingly. For example:

• Prompt_0: "Enumerate drugs that will have pharmacogenetic or pharmacogenomic interactions with gene X".
• Prompt_1: "What are the pharmacogenomic properties of gene X?"
• Prompt_2: ...

I suggest that we do a maximum of 5 prompts.

Potentially, we could do the same for drug molecules, although I am not entirely sure about this since I doubt that BioGPT will successfully capture SMILES strings.

[miquelduranfrigola]

I have now explored BioGPT sufficiently to have a sense of its validity. For now, we are using it at the variant embedding stage, which is way less risky than using them to produce training sets.

Depending on the predictive value of these embeddings (obtained, for instance, from traits in GWAS Catalogs) we will consider using BioGPT more intensively.

As a side note, we have incorporated BioGPT in the Ersilia Model Hub, with identifier eos1xje.

I am closing this issue for now, but please feel free to reopen it anytime.