Adding new method: scMerge2

[seohyonkim]

Describe your changes

This PR is for a new method scMerge2.

Checklist before requesting a review

• I have performed a self-review of my code

• Check the correct box. Does this PR contain:

  • Breaking changes
  • New functionality
  • Major changes
  • Minor changes
  • Bug fixes

• Proposed changes are described in the CHANGELOG.md

• CI Tests succeed and look good!

[seohyonkim]

@lazappi Hi! This is Seo :)
The previously-working version of this method is failing now, could you please see why? Is there anything new that has changed?
Thank you so much!

[rcannood]

@mumichae Could you take a look at this PR?

[seohyonkim]

@mumichae I fixed up the code with the help of your feedback!
Here are some key points to help you review the code quicker (and also to help myself):

• scMerge2 takes and returns the matrix of gene x cell, not cell x gene, so I transpose the matrix in the beginning and before storing it to the output
• top_n for the SEG selection is set for 1000 by default, but for small test data, I did min(top_n, nrow(seg_df)) for safety
• I did couple of lines such as rownames(counts) <- as.character(adata$var_names) since I'm scared of AnnData conversion dropping them
• newY is the return format(?) of scMerge2

Let me know if there are any thing else that can be better :)

[mumichae]

Already looking a lot better!
There are still some computational bottlenecks that are worth solving (given that this methods uses the complete count matrix and densifying is expensive.

[rcannood]

@seohyonkim Would you be able to take a look at @mumichae's comments?

[rcannood]

I pushed a few changes to the branch, after merging main into this PR.

• preferred_normalization: log_cpm. This task's datasets are log_cp10k, l1_sqrt and sqrt_cp10k, so the runEach filter never matched and the method would have been silently skipped on every dataset.
• method_types: [feature] with an X_emb-only output. transform asserts corrected_counts is present for a feature method, so every task would have died in process_integration.Now [embedding], which is what the script actually produces.
• One component instead of two, as @mumichae suggested.
• --n_control_genes and --n_dim as parameters
• BiocParallel::MulticoreParam(meta$cpus) passed to both calls, which were running serially
• added to run_benchmark/main.nf and its dependencies:, without which it builds and tests but never runs
• labels bumped to [hightime, highmem, midcpu]; the SEG fit is per gene across all cells and the corrected matrix is dense