The idea is to compare expression of the target in the relevant tissue with expression across usual suspects for side effects (blood, liver, heart) or surrounding organs.

The problem is defining what a surrounding organ is.

• - cellontology_utils - functions to handle cell ontologies
• - preprocessing_utils - things like anndata2pseudobulk, adding similarity btw cell types to the pseudobulk objects, functions to save and read pseudobulk objects?
• - cxg_utils - functions to download data and metadata using cxg census
• - plotting_utils - functions to make diagnostic plots
• - de_utils - functions for DE analysis (cleaning data, running DE analysis with different regimes, saving outputs)
• - sc_evidence - transform DE analysis results to evidence for drug targets
• - opentargets_utils - functions to clean OT datasets

• Identify terms to blacklist (e.g. native cell)
• The depth_level should be proportional to the total number of ancestors

• Volcano plots
• ECDF of significant cell types x lfcThreshold
• Expression of genes that are significant in a large number of cell types for the cell type specificity evidence
• Number of significant cell types

• Compare 2 diseases with and without genetic evidence (e.g. COPD and pulmonary fibrosis) - which successful targets are prioritized by single-cell evidence?
• Why is there no genetic evidence for some of these diseases? Is it lack of GWAS studies or no significant hits?
• Are targets with sc evidence associated with GWAS variants for lung function?

Compare targets with single-cell evidence for coarse vs fine uniformed cell annotations using lung samples/diseases, comparing ontology based coarse annotations with annotations from Extended Lung cell atlas.

• Are the same genes identified as cell type markers and disease cell type markers?
• Are more successful targets identified when testing on fine grained annotation?

Brain files fail because of insufficient RAM even when requesting 700GB. Download of stomach files from cxg hangs forever, looks like something is wrong in the census side.

Try downloading datasets directly from cellxgene website or sfaira, then filtering and pseudobulking.

• Download and pseudobulk
• Run DE analysis
• Compute OR

A bunch of diseases gets excluded in DE in disease analysis because no disease pseudobulk is left after filtering for low quality annotation, although in prev version of the pipeline there were cell types to do comparison:

To check

MONDO_0005575
MONDO_0006156
MONDO_0006249
MONDO_0024660
MONDO_0024661
MONDO_0024885
MONDO_0001056 # gastric cancer

Check output plots for each disease in /nfs/team205/ed6/bin/sc_target_evidence/data/plots/ (or here). What each figure shows is briefly explained in DATA_INFO.

• Check whether we need to blacklist more terms - e.g. lymphocyte
• Check whether we need to exclude certain diseases
• Check expression of correct marker genes - the full list of cell ontology terms used across diseases is here. See DATA_INFO for location of .h5ad files.
• Flagging general odd things

• Fix install with rpy2 and friends
• Celltype marker test
• Disease marker test

Thanks for this amazing work. I was wondering how can I get the original cxg dataset. I try to use the "sc_target_evidence/src/process_sc_data.py", but it seems "TargetDiseasePairs_OpenTargets_cellXgeneID_12072023.csv" is missing.

Using total expression levels, run dendrogram with scanpy. Ideally make a separate function for this, to use in diagnostic plots.