Fix: debug DeseqStats constrasts

[BorisMuzellec]

This PR aims to debug DeseqStats, which gave inconsistent results when changing the reference level using the constrast argument.

As an example, setting constrast = ["condition", "A", "B"] instead of constrast = ["condition", "B", "A"] led to different p-values, whereas same p-values but opposite lfcs and stats are expected.

More precisely, in this PR:

• A design_matrix attribute is added to the DeseqStats class,
• We make sure that the design_matrix and LFC fields of a DeseqStats object are consistent with its contrast attribute,
• We refactor the synthetic data that is used for pytests. The same data is now used for single and multi-factor tests, the only difference lying in the choice of design_factors,
• A test_contrast pytest was implemented to ensure that changing the reference level in the contrast yields expected results.

[maikia]

Thanks @BorisMuzellec nice PR. In general LGTM and ready to merge, just two minor points (not necessary for this PR, but good to keep in mind for the future):

1. As part of this PR there are a lot of non-relevant (to this PR) name changes, etc.. which makes more difficult PR checking itself. Next time it would be great if you could keep those to separate PR.
2. using your unit test I tried to make some tests on DeseqStats:

• when values: res_B_vs_A = DeseqStats(dds, contrast=["condition", "A", "A"]) are passed I get a message: *** KeyError: 'condition_A', perhaps more descriptive message could be given?
• res_B_vs_A = DeseqStats(dds, contrast=["group", "B", "A"]) message is: *** AssertionError: The contrast levels should correspond to design factors levels.
  Although this is correct a more descriptive message could be given (eg, design factor 'group' does not have 'A' level. The contrast levels should correspond to design factors levels.

[BorisMuzellec]

Thanks @maikia, I'll merge this PR then and open another one to improve error messages.

[hfl112]

I still got some error when using contrast:

stat_res = DeseqStats(dds, contrast=['condition', 'range_high', 'range_low'])

KeyError: "The tested level ('range_high') should correspond to one of the levels of 'condition'"

[BorisMuzellec]

Hi @hfl112, can you try replacing underscores with hyphens?

stat_res = DeseqStats(dds, contrast=['condition', 'range-high', 'range-low'])