Fix: categorical association dataframe output

[aron0093]

1. The results dataframe should contain "program_name" column. Currently it is the index which is inconsistent from the output from other evaluations.
2. The p-values are automatically rounded off. Explicitly setting dtype should fix this.

[adamklie]

After discussion, we are thinking of moving to a different strategy for the categorical association evaluation than is currently implemented.

Evaluation side

By default, we will treat each level of the category separately and run the following procedure:

1. Binarize the category across cells (1 for category==level, 0 if category!=level) --> can probably use Narges's function here
2. For each program, calculate a Pearson correlation between this binary vector and the program scores. Adjust the pvals using https://www.statsmodels.org/dev/generated/statsmodels.stats.multitest.fdrcorrection.html
3. For each program, calculate a log2FC of (mean program scores of category==level)/(mean program scores of category!=level)

We will then combine these results across levels into a single wide dataframe:

program_name	batch_level_1_pearsonr_stat	batch_level_1_pearsonr_pval	batch_level_1_pearsonr_adj_pval	batch_level_1_log2FC	batch_level_2_pearsonr_stat	batch_level_2_pearsonr_pval	batch_level_2_pearsonr_adj_pval	batch_level_2_log2FC
0	0.82	0.004	0.012	1.4	0.75	0.008	0.02	1.3
1	0.65	0.03	0.09	0.7	0.58	0.035	0.12	0.8
2	0.90	0.002	0.006	1.9	0.81	0.005	0.01	1.7

We will then save this as a txt file for each categorical_key in the config with the following file name: {prog_key}_{categorical_key}_association_results.txt

• prog_key comes from the config
• categorical_key comes from the config
• must end in association_results.txt suffix
  Example: cNMF_batch_association_results.txt

Dashapp side

1. The dashapp will parse the output directory for all files with this suffix and will parse the appropriate categorical_key (which it can match to the eval config it is also passed in)
2. It will then plot a level x program by heatmap of the correlations (update to plotly)
3. It will also compute an aggregate statistic across levels (e.g. max) and generate a plot that shows the distribution of the statistic (without pval for now*)

• This is due to the inflation of pvals seen when running a test on this statistic. Also gaven't fully decided on whether to rank-order this or not.

[adamklie]

I came up with a really hacky solution to this. I really couldn't figure out an easy way to implement this as an option with the current codebase. Basically, it looks for the combination of using test="pearsonr" and mode=one_vs_all and runs a separate set of code for that one case.

Obviously this won't do for the future, but it was the only way I could manage to get it to work while still preserving most of the old functionality.

TL;DR hopefully this works, but we will likely need to revisit

[adamklie]

• Hacked the implementation
• Update the eval pipeline
• Reconfigure the dashapp