Feature cv training

[HolEv]

What

• Implement cross-validation like training pipelines/cv_training

  • train the model on n-1 folds of the samples
  • compute the burdens on the held-out fold using this models
  • repeat this for all n-folds
  • for this, allow a sample file to be passed to dense_gt.py

• Allow different orders of samples in phenotype_df and genotypes.h5 in dense_gt.py

  • so far, samples in phenotype_df and genotypes.h5 had to be in the same order.
  • this is changed now by introducing an additional index map for the genotypes.h5, which retrieves samples in the order of self.samples

• Average burdens from multiple repeats and run association testing afterwards (deeprvat/associate.py) (as opposed to running the association testing on each cv individually)

• Restructure the snakefiles as it had been already started in the main branch

  • require baseline result only for training phenotypes

• Update evaluate.py

  • no repeats required any more
  • make bonferroni correction the default multiple testing correction
  • don't combine baseline discoveries with DeepRVAT discoveries
  • Re-test also the 'seed genes' since we don't evaluate on the same sample-gene combinations any more as we trained on (thanks to the cv-based training procedure)

• use additional covariates age2 and age*sex and correct for statin usage

  • updated the example data to have these fields

• update example data to have bit sample ids in genotypes.h5 and string sample ids in phenotype_df

• implement conditional analysis for common variants pipelines/association_testing_control_for_common_variants.snakefile

Testing

• quite extensively tested (many reasonable experiments done) but still need to check why github tests fail so far