Running MA-FOCUS with customised weights always results in NA values for pop2.

[rodrigoduarte88]

Hey guys,

Thanks for developing FOCUS! I am having issues running multi-ancestry FOCUS using custom weights imported from FUSION. I can run MA-focus with my sumstats, own gencode file, and ref data, but with the toy db files provided by the authors (genoa files). The plots and focus.tsv table are generated as expected. However, when I run using my own weights, the run never works for pop2 (independent of the order of the files I set up in the command, i.e. pop1 = EUR, pop2 = AFR, or pop1 = AFR, pop2 = EUR): the plot for pop2 never contains dots, and values like pip_pop2 and twas_z_pop2 are assigned to NA to most non-NULL models. Only one or two genes for each run get non-NA values for those columns.

Here's a summary of what I did and my log file. Please note that I edited locally my focus files to try to debug this, so some extra warnings are provided:

1. The weights for ma-FOCUS need to be created using an extra DIR column in the pos file when importing. See here: Error: Import weights from predixcan and fusion #20 (comment)

2. The weights for a multi-ancestry run need to be created using the same schema, which means that the way they are read by focus finemap requires them to be created using the exact same flags by focus import:

focus import weights_wrapped_2a.pos fusion --tissue X --name Y --assay Z --output weights_EUR

focus import weights_wrapped_2b.pos fusion --tissue X --name Y --assay Z --output weights_AFR

3. The logs and example output for these runs are here:
   https://www.dropbox.com/scl/fo/2ihlfp1zgqiqb056p46fx/AKHuNIfwSdrbCex1O1UwCE8?rlkey=p3rnivqh8p2b7zfjx5v7rvwwe&dl=0

I was wondering if the developers could share the code used to create the genoa weights in the wikipedia page, so we can try to reproduce using our own weights? Or if you have any clues what part of the code could be giving this error, it would really help!

[rodrigoduarte88]

I think I managed to make it work, but I need to triple check I didn't do anything wrong.

When joining pop1 statistics with pop2, there was a part of the code (in create_output) that was joining the dataframes by model_id, which could change in two different populations (e.g., blup may be better to predict expression in one pop, but enet in another). Somehow, after removing this "join by model_id", things are working now (although I still cannot explain why the genoa weights were working... )

I'll work more on this this week and post a more definitive solution.

Thanks!