Dimensionality Reduction

37. Dimensionality Reduction#

37.1. Motivation#

Feature matrices of surface protein markers are hard to grasp for humans as raw tables. Therefore, we resort to low dimensional embeddings that allow us to visualize the ADTs in commonly two dimensions. The approaches that we use and recommend for ADT data do not differ from the ones for transcriptomics data. All aforementioned limitations of visualizations obtained through methods like t-SNE and UMAP also apply to ADT data.

ADT data generally does not require any sophisticated feature selection, because features have already been selected a priori during experimental design. All selected ADTs should correspond to biologically relevant features. Nevertheless, large datasets may benefit from PCA to reduce the dataset from several hundred features to a few principal components. This is especially advisable if computational resources are limited.

In this and the following two chapters, we decided to focus on the ADT data and do not use the RNA data of the study. In the Paired integration chapter, we will explore how we can make use of both modalities jointly, which allows for a more detailed cell type annotation.

37.2. Environment setup#

import warnings

import muon as mu
import scanpy as sc

warnings.filterwarnings("ignore")
mu.set_options(pull_on_update=False)
sc.settings.verbosity = 0
sc.set_figure_params(
    dpi=80,
    facecolor="white",
    frameon=False,
)

import lamindb as ln

ln.track()

→ found notebook dimensionality_reduction.ipynb, making new version
→ created Transform('liGMVGre4G5H0004', key='dimensionality_reduction.ipynb'), started new Run('2jUW2z0ijgULitLv') at 2026-04-10 17:28:52 UTC
→ notebook imports: lamindb-core==2.3.1 muon==0.1.7 scanpy==1.12
• recommendation: to identify the notebook across renames, pass the uid: ln.track("liGMVGre4G5H")

37.3. Loading the data#

We load the MuData object we saved at the end of the previous chapter Doublet detection:

af = ln.Artifact.connect("theislab/sc-best-practices").get(
    key="surface-protein/cite_doublet_detection.h5mu", is_latest=True
)
mdata = af.load()
mdata

MuData object with n_obs × n_vars = 117951 × 36741
  var:	'gene_ids', 'feature_types'
  2 modalities
    rna:	117951 x 36601
      obs:	'donor', 'batch'
      var:	'gene_ids', 'feature_types'
    prot:	117951 x 140
      obs:	'donor', 'batch', 'n_genes_by_counts', 'log1p_n_genes_by_counts', 'total_counts', 'log1p_total_counts', 'n_counts', 'outliers', 'doublets_markers'
      var:	'gene_ids', 'feature_types', 'n_cells_by_counts', 'mean_counts', 'log1p_mean_counts', 'pct_dropout_by_counts', 'total_counts', 'log1p_total_counts'
      uns:	'doublets_markers_colors'
      layers:	'counts'

We remove the counts layer containing the raw data since we do not need it anymore.

del mdata["prot"].layers["counts"]

Isotype controls do not contain any biological information since their only purpose is to use them for dsb normalization, see the Normalization section. We can therefore remove them from our data.

mdata["prot"].var.index[:50]

Index(['CD86-1', 'CD274-1', 'CD270', 'CD155', 'CD112', 'CD47-1', 'CD48-1',
       'CD40-1', 'CD154', 'CD52-1', 'CD3', 'CD8', 'CD56', 'CD19-1', 'CD33-1',
       'CD11c', 'HLA-A-B-C', 'CD45RA', 'CD123', 'CD7-1', 'CD105', 'CD49f',
       'CD194', 'CD4-1', 'CD44-1', 'CD14-1', 'CD16', 'CD25', 'CD45RO', 'CD279',
       'TIGIT-1', 'Mouse-IgG1', 'Mouse-IgG2a', 'Mouse-IgG2b', 'Rat-IgG2b',
       'CD20', 'CD335', 'CD31', 'Podoplanin', 'CD146', 'IgM', 'CD5-1', 'CD195',
       'CD32', 'CD196', 'CD185', 'CD103', 'CD69-1', 'CD62L', 'CD161'],
      dtype='object')

isotype_controls = ["Mouse-IgG1", "Mouse-IgG2a", "Mouse-IgG2b", "Rat-IgG2b"]
temp = (
    mdata["prot"].var.loc[~mdata["prot"].var.index.isin(isotype_controls), :].index
)  # Select all proteins except isotype controls.

Now we actually remove isotype controls from the data.

mu.pp.filter_var(data=mdata["prot"], var=temp.tolist())

The data does not contain the isotype controls anymore.

mdata["prot"].var.index[:50]

Index(['CD86-1', 'CD274-1', 'CD270', 'CD155', 'CD112', 'CD47-1', 'CD48-1',
       'CD40-1', 'CD154', 'CD52-1', 'CD3', 'CD8', 'CD56', 'CD19-1', 'CD33-1',
       'CD11c', 'HLA-A-B-C', 'CD45RA', 'CD123', 'CD7-1', 'CD105', 'CD49f',
       'CD194', 'CD4-1', 'CD44-1', 'CD14-1', 'CD16', 'CD25', 'CD45RO', 'CD279',
       'TIGIT-1', 'CD20', 'CD335', 'CD31', 'Podoplanin', 'CD146', 'IgM',
       'CD5-1', 'CD195', 'CD32', 'CD196', 'CD185', 'CD103', 'CD69-1', 'CD62L',
       'CD161', 'CD152', 'CD223', 'KLRG1-1', 'CD27-1'],
      dtype='object')

mdata["prot"]

AnnData object with n_obs × n_vars = 117951 × 136
    obs: 'donor', 'batch', 'n_genes_by_counts', 'log1p_n_genes_by_counts', 'total_counts', 'log1p_total_counts', 'n_counts', 'outliers', 'doublets_markers'
    var: 'gene_ids', 'feature_types', 'n_cells_by_counts', 'mean_counts', 'log1p_mean_counts', 'pct_dropout_by_counts', 'total_counts', 'log1p_total_counts'
    uns: 'doublets_markers_colors'

37.4. PCA and UMAP#

We can now reduce the dimensionality of the data with PCA since our dataset is quite big (136 surface proteins).

sc.pp.pca(mdata["prot"], svd_solver="arpack", random_state=0)

We create an elbow plot in order to decide how many PCs we use:

sc.pl.pca_variance_ratio(mdata["prot"], n_pcs=50)

../_images/52f7ceefdd1cce64fd030ef52423409651fead4ad0930078665f5787ffe57854.png

We use 20 PCs because PCs 1-20 capture much of the variance in the data and PCs 20-50 capture little variance of the data and can thus be discarded. We now compute a neighborhood graph and a UMAP embedding to visualize the study’s variables.

sc.pp.neighbors(mdata["prot"], n_pcs=20, random_state=0)

sc.tl.umap(mdata["prot"], random_state=0)

Now we have our data compressed into 2 dimensions, which we can use to visualize the data. Let’s first visualize and evaluate if there are batch effects, that is, if different donors and different batches form separate clusters.

sc.pl.umap(mdata["prot"], color=["donor", "batch"])

../_images/91d552ebd8b02f5aa952200a4575311116aa8c901f2b1a8305c019882970f341.png

We indeed see that some donors form separate clusters. Also batches form separate clusters. Thus, it seems that batch correction is necessary. To confirm, we plot markers of CD4 and CD8 T cells:

sc.pl.umap(mdata["prot"], color=["CD4-1", "CD8", "CD3"])

../_images/773aa42a05f0d5b4ac92f0e4e85fd474ba76f60904e2b34dad6e5f85b1f95315.png

CD4 T cells fragment into donor-specific mini-clusters, meaning cells are grouping by donor identity rather than cell type. Ideally, CD4 T cells from all donors should cluster together regardless of their donor of origin. This donor-driven separation is a batch effect, and must be corrected before downstream analysis.

af_dimensionality_reduction = ln.Artifact.from_mudata(
    mdata,
    key="surface-protein/cite_dimensionality_reduction.h5mu",
    description="CITE-seq data after dimensionality reduction",
)
af_dimensionality_reduction.save()

ln.finish()

37.5. References#

37.6. Contributors#

We gratefully acknowledge the contributions of:

37.6.1. Authors#

Javier Marchena-Hurtado
Daniel Strobl
Ciro Ramírez-Suástegui

37.6.2. Reviewers#

Lukas Heumos
Anna Schaar