Develop Drug-Variant-Gene triplets pipeline

[miquelduranfrigola]

A drug-variant-gene triplet predictor needs to be developed. Unfortunately, the gene-drug pair pipeline can only be used partially, since in this case there are a few key differences, including:

• Out-of-sample African variants: The majority of variants will not be in PharmGKB. Therefore, variant embeddings need to work at prediction time.
• Train-test splits: since variants depend on genes, splits and sampling of negatives cannot be done randomly. Rather, this needs to be done at the gene level.
• Embeddings for variants are less clear and well-defined, especially for intron variants which are not identifiable at the sequence level. As it turns out, intron variants are abundant and PharmGKB (training set) and, therefore, we need to include them as part of our encoding.

Therefore, the tasks we need to address are the following:

• [ ] Refactor the pair prediction pipeline to work on triplets
• [ ] Triplet sampler.
• [ ] Triplet splitter.
• [ ] Fit-Transform capabilities for the variant embeddings, since they are out-of-sample.

[miquelduranfrigola]

We have found many intron variants in PharmGKB and we consider that variant embeddings would not be very informative. Therefore, we decided to aim at predicting drug-gene pairs for now. I am closing this issue. We may want to reopen it in the future.