[integration/peerlab] feature: adding a reference gene correlation

[ananya-nandula]

Revised the gene correlation reference feature so that we check to see if it contains raw counts, and then normalize counts the same way segger does. Also, added an assertion to make sure that all genes in the reference anndata pass the genes min counts threshold set by segger. Example use case: if you have multiple samples you want to segment but want to use the same gene-gene correlation prior, pool the desired samples together into one anndata, ensuring that the anndata contains raw counts, all genes pass the min counts threshold, and that the anndata contains all genes present in the samples you want to segment.

New parameters:

• gene_corr_reference_path (--gene-corr-reference-path): Path to an .h5ad reference dataset. Used to compute gene-gene PCA embeddings. .X must be raw counts.
• gene_missing_strategy (--gene-missing-strategy): What to do when genes in the data are missing in the reference. Defaults to "error". Alternative use "remove" to remove these from the data. Work in progress is "fill" which will use correlations from the data for missing genes only.

[Tobiaspk]

Good PR.

Here's what should be changed before we can merge:

• Don't test if norm is a layer. Instead, ensure that inputs are raw counts and normalise the same way as anndata
• Add a parameter fill_missing_gene_correlations (or similar) to ignore errors when genes are missing in the reference
• Remove "# ADDED" and "# NEWLY ADDED" comments. Instead, add those in the PR description.

Creating some tests now to double check these changes

[Tobiaspk]

Thanks @ananya-nandula and @nkalfus for your contributions. A custom gene correlation can now be provided. Missing feature for next PR: filling genes if they are missing from the reference.