Add support for calculating percentile ranks on a consistent set of reference peptides

[malachig]

Currently some tools provide a percentile rank and other do not. The percentile ranks provided by each tool may be determined based on different (undocumented) reference peptide lists.

Create reference results for all the algorithms supported by pVAC so that we can calculate consistent percentile rank scores for any test peptide.

[susannasiebert]

The goal is to calculate rank scores for all prediction algorithms supported by pVACtools and a prioritized set of alleles and lengths. It has been discussed to use the hdf5 file format to store the results. The current though is to have one file per algorithm to store all results. The files would need to contain the following data: peptide, allele, peptide length, score. The data would need to be sorted on the score for efficient read-in of the results.

For using these scores, a new option should be added. Something like --use-normalized-percentiles or --use-pvactools-percentiles (parameter name tbd). When this option is set, percentiles will be calculated for algorithms that don't provide them natively and the percentiles for algorithms that provide them natively will be overwritten. This is done was follows:

• The pipeline will have been run with a specific set of prediction algorithms and alleles. For each neoantigen predicted by the pipeline, each algorithm the pipeline was run with and each hla allele a prediction score would've been calculated.
• To calculate the normalized percentile, look up the sorted list of reference scores for the same prediction algorithm, allele, and peptide length and calculate in which percentile the calculate score falls within the list of reference scores.
• Tbd: decide how to handle ties, e.g. when there are multiple reference peptides in this list with the same score or when the calculated score is identical to one of the reference ones.

Most of this work will need to happen in the output_parser.py module. For prediction algorithms that don't support percentiles, we don't currently add a column for writing the per-algorithm percentile score. I think it will be easiest to write a percentile score column for all algorithms the pipeline was run with. If the new parameter is not set, these columns would simply be filled in with NA. This will require updates to the following methods:

• output_headers: add percentile headers for all used algorithms, not just the ones that natively support them.
• get_scores: This method is used to parse the results from each prediction algorithm into a unified data structure. Here, we want to make sure that percentile entries are added for all algorithms. Set these to NA when the new parameter is not set. If it is set, calculate the normalized percentile and use that value.
• add_prediction_scores: This method adds per-algorithm scores to each output entry. Update this to write percentile results for all used algorithms, not just the ones that natively support them.

With these updates, calculating the median/best summary metrics should work without any additional changes.