PyProphet 2

[grosenberger]

This PR contains the following changes:

General:

• CLI implemented using Python Click
• Support for Python 3
• Integration with Travis-CI

Data formats:

• Support for OSW SQLite files
• Legacy support for OpenSWATH TSV files: Only MS2-level scoring available
• Export of legacy TSV files
• Export of matrix files
• Export of score plots

Semi-supervised learning:

• Learning and scoring on MS2-, MS1- and transition-levels supported.

Statistics:

• Estimation of pi0 using Storey's approach by sampling a specified lambda range (validated against Bioconductor/qvalue)
• Estimation of q-value using Storey's approach (numerically slightly different to mProphet; validated against Bioconductor/qvalue)

IPF:

• IPF scoring
• Native computation of posterior error probabilities (QVALITY dependency removed)

Jumbo-PyProphet:

• Merging and/or subsampling of multiple OSW files
• Assessement of peptide and protein-level error rates in run-specific, experiment-wide and global contexts

Other:

• PyProphet scoring can be switched against Percolator using PercolatorAdapter from OpenMS/develop.

Limitations and future improvements:

• Merging can be slow and needs to be improved
• Improve robustness for exotic combinations
• Some components (e.g. sqMass filtering) need more testing and validation

[hroest]

============== 7 failed, 35 passed, 1 warnings in 878.86 seconds ===============

[grosenberger]

The regression tests need some further fixes...

[hroest]

After subsampled learning, the main limitations remains the application of the scores. Currently, the full table, which can contain hundreds of runs and millions of features, is loaded into memory and scored. This requires lots of memory and would be the first thing to optimize, e.g. by iterative scoring of the runs.

I agree, once we have the weights why not load chunks of data (and only the scores, not the whole meta data) into memory?

[hroest]

This is a really huge PR, basically it changes everything. Maybe instead we can try this: Can you comment on what does not change, e.g. are simple legacy workflows still supported and give the same results (e.g. running pyprophet on a single file) ?

[grosenberger]

I agree, once we have the weights why not load chunks of data (and only the scores, not the whole meta data) into memory?

At least the skipping of metadata optimization is already implemented. However, the tables are still too big, which is why further optimization is required.

This is a really huge PR, basically it changes everything. Maybe instead we can try this: Can you comment on what does not change, e.g. are simple legacy workflows still supported and give the same results (e.g. running pyprophet on a single file) ?

Legacy TSV-based workflows are still supported. OpenSWATH TSV input will result in very similar results as before. However, there are some slight changes, due to the new functions, including lambda and PEP estimation. Further, I changed the implementation of the statistic functions, resulting in numerical differences. I validated the new functions against Bioconductor/qvalue, but I think I should point out the differences to the old implementation better.

[hroest]

I am basically in favour, however I did not do a in-depth code review of all the changes (there are too many). However this looks like a great improvement and from my side we can go ahead and merge

[grosenberger]

Great thanks. If there are no further comments, I will merge tonight.