Generation of pseudobulk in Diff. exp. chapter

[Claptar]

Hello! Your book is really a tresure for a novice bioinformatician like myself.

I have a question about diff. exp. chapter. When you describe the proccess of pseudobulk aggregation you write

Since edgeR was introduced as a method for DE analysis for bulk data, we first need to create pseudobulk samples from our single-cell dataset. For each patient we create 1 pseudobulk sample per cell type by aggregating the cell from each subpopulation and taking the mean gene expression within that subpopulation.

I am confused about "taking the mean" part. As I understand you aggregate (sum?) all cells with the same cell type withing for each sample. And then ... ?

I tried to find this "taking the mean" part in code but I only see the part where you sum in aggregate_and_filter() function.

def aggregate_and_filter(
    adata,
    cell_identity,
    donor_key="sample",
    condition_key="label",
    cell_identity_key="cell_type",
    obs_to_keep=[],  # which additional metadata to keep, e.g. gender, age, etc.
    replicates_per_patient=1,
):
    # subset adata to the given cell identity
    adata_cell_pop = adata[adata.obs[cell_identity_key] == cell_identity].copy()
    # check which donors to keep according to the number of cells specified with NUM_OF_CELL_PER_DONOR
    size_by_donor = adata_cell_pop.obs.groupby([donor_key]).size()
    donors_to_drop = [
        donor
        for donor in size_by_donor.index
        if size_by_donor[donor] <= NUM_OF_CELL_PER_DONOR
    ]
    if len(donors_to_drop) > 0:
        print("Dropping the following samples:")
        print(donors_to_drop)
    df = pd.DataFrame(columns=[*adata_cell_pop.var_names, *obs_to_keep])

    adata_cell_pop.obs[donor_key] = adata_cell_pop.obs[donor_key].astype("category")
    for i, donor in enumerate(donors := adata_cell_pop.obs[donor_key].cat.categories):
        print(f"\tProcessing donor {i+1} out of {len(donors)}...", end="\r")
        if donor not in donors_to_drop:
            adata_donor = adata_cell_pop[adata_cell_pop.obs[donor_key] == donor]
            # create replicates for each donor
            indices = list(adata_donor.obs_names)
            random.shuffle(indices)
            indices = np.array_split(np.array(indices), replicates_per_patient)
            for i, rep_idx in enumerate(indices):
                adata_replicate = adata_donor[rep_idx]
                # specify how to aggregate: sum gene expression for each gene for each donor and also keep the condition information
                agg_dict = {gene: "sum" for gene in adata_replicate.var_names}
                for obs in obs_to_keep:
                    agg_dict[obs] = "first"
                # create a df with all genes, donor and condition info
                df_donor = pd.DataFrame(adata_replicate.X.A)
                df_donor.index = adata_replicate.obs_names
                df_donor.columns = adata_replicate.var_names
                df_donor = df_donor.join(adata_replicate.obs[obs_to_keep])
                # aggregate
                df_donor = df_donor.groupby(donor_key).agg(agg_dict)
                df_donor[donor_key] = donor
                df.loc[f"donor_{donor}_{i}"] = df_donor.loc[donor]
    print("\n")
    # create AnnData object from the df
    adata_cell_pop = sc.AnnData(
        df[adata_cell_pop.var_names], obs=df.drop(columns=adata_cell_pop.var_names)
    )
    return adata_cell_pop

[Zethson]

Ahh, this is a bit imprecise. There's two ways of creating pseudobulks. One is via mean aggregation and the other is via sum aggregation. The code tells the story of what is actually being used here as you found it. We'll revise the text a bit to be more precise.

Thank you!

[Claptar]

Oh, great, thank you!