L2G feature matrix revision

[xyg123]

As a developer I want to revise and refactor some of the logic in the L2G feature matrix generation step

Background

tbd.

Tasks

• [ ] Refactor logic for assigning potential target genes to studyLocusIds, to include genes in proximity to all variants in the studyLocus, instead of just the top hit.
• [ ] Implement logic to assign potential target genes which have a colocalisation result with a studyLocus, but is out of range from the studyLocus.
• [ ] Include additional distance to footprint feature in the L2G model, using the latest VEP variant index.
• [ ] Implement logic to save the feature matrix when it is generated.
• [ ] Implement logic to merge features across studies which have the same EFO, to address the missingness rate of features in L2G.
• [ ] Refactor logic for gold standard Negative selection process, to prioritise keeping GSNs with non-distance features, followed by prioritising the closest genes.
• [ ] Implement logic to remove studyLocus where all potential target genes have only distance-based features from the training set.

[addramir]

As a general plan looks good to me. @ireneisdoomed and @DSuveges please have a look to, but we need more granular plan on this. In general, the top priority for us is to reach the level of clean credible set level feature matrix as a first step implementing suggestions above.

[project-defiant]

@ireneisdoomed which tasks from the list above are going to be implemented in the l2g fm refactoring?

[ireneisdoomed]

None. The refactoring I am doing is to solve #3252

[addramir]

 @xyg123

Refactor logic for assigning potential target genes to studyLocusIds, to include genes in proximity to all variants in the studyLocus, instead of just the top hit.

• I guess it should be covered by issue adding tss and footprint based features? https://github.com/opentargets/issues/issues/3258

Implement logic to assign potential target genes which have a colocalisation result with a studyLocus, but is out of range from the studyLocus.

• I guess it should be covered by issue adding coloc based features? https://github.com/opentargets/issues/issues/3252

Include additional distance to footprint feature in the L2G model, using the latest VEP variant index. https://github.com/opentargets/issues/issues/3258

The rest is ok.

[addramir]

We need to re-write this issue next week.

[addramir]

The issue was spllited on several issues.