1 of 23

Canadian Bioinformatics Workshops

The Brain Single Cell Initiative

Panoramics - A Vision

www.bioinformatics.ca

bioinformaticsdotca.github.io

1

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

2 of 23

2

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

3 of 23

Module #:2

Normalization and transforming your data

Introductory Spatial ‘Omics Analysis

July 9-10, 2024

Introductory Spatial ‘Omics Analysis

February 20-21, 2025

Savannah Kilpatrick, PhD

Scientist III, Cell Therapy

Genentech

3

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

4 of 23

Learning Objectives

By the end of this lecture, you will:

Become familiar with the Visium HD data output files.

Understand which files are most relevant for your analysis.

Employ QC filters and cut-offs to better interpret your data.
Understand various normalization approaches
(Lecture only) Data transformations and integration options for spatial ‘omics.

4

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

5 of 23

Data walk-through: Overview of data outputs, web summary file interpretation, etc.

5

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

6 of 23

Visium data output overview

Details on file formats

6

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

7 of 23

Visium data output: web summary files

7

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

8 of 23

Bins and filtering options

Bin size considerations:

Granularity vs sensitivity
Subcellular localization
Signal loss vs enrichment

8

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

9 of 23

Bins and filtering options

9

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

10 of 23

Data overview: image files

10

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

11 of 23

QC assessment: counts

10x Genomic’s Data repository

11

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

12 of 23

Data filtering: spatially-derived artifacts

Toddy, Hicks, and Guo 2024, BioRxIV Preprint

•Generic pipelines apply scRNA Seq QC measures, but often fail to appreciate spatially-derived artifacts.

•Artifacts can be introduced through user error.

12

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

13 of 23

Data filtering: spatially-derived artifacts

•Generic pipelines apply scRNA Seq QC measures, but often fail to appreciate spatially-derived artifacts.

•Artifacts can be introduced through user error, or have biologically-relevant origins (tissue permeability, cell density, size, and content).

13

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

14 of 23

Importance of choosing the right filters

Schuster et al., 2024 Nat Comm

14

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

15 of 23

Filtering considerations: what’s best for you

Maciejewski and Czerwinska, 2024 Cancers

15

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

16 of 23

Correcting for technical variation: count normalization

Normalization is essential for the removal of technical artifacts and for comparisons between populations.

16

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

Normalization is meant to remove the effects of technical variation brough on from the processing and sequencing of your samples. This includes sequencing depth variations that can be introduced when each tissue area is not exposed to the same amount of library prep reagents, or sequencing artifacts that limit the detection of particular transcripts.

In RNA Seq practices, Normalization was introduced with the idea that ALL variations in depth were a result of technical error, and as a result simply divides the sample counts by the library depth, and applies a scalar factor to a fixed total such as counts per million. This is described as library size normalization, and is the simplest strategy for performing scaling normalization. We define the library size as the total sum of counts across all genes for each cell, the expected value of which is assumed to scale with any cell-specific biases. The “library size factor” for each cell is then directly proportional to its library size where the proportionality constant is defined such that the mean size factor across all cells is equal to 1. This definition ensures that the normalized expression values are on the same scale as the original counts, which is useful for interpretation - especially when dealing with transformed data (see Section 2.5).

May introduce bias towards highly expressed transcripts while masking lower abundant transcripts. This bias is exacerbated by the prevalence of excessive zeros in single-cell/spatial data, compromising the accuracy of scaling factor calculations

17 of 23

Generic normalization strategies

Library size normalization

Scales raw counts by the total number of counts per spot (sequencing depth) and multiplies by a constant scalar factor:

•May introduce bias towards highly expressed transcripts

•Exacerbated by sparse/dispersed matrices as with ST

17

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

AKA Size factor normalization

Apply by a constant scaling factor like a million to convert our observed gene expression counts to a normalized gene expression magnitudes, assuming all cells are sequenced.

If you’re proceeding with log normalization, a log transformation is introduced to reduced the skewness of the data.

Assumption: while there are differences in gene expression between cells for a handful of genes (these being your biologically relevant genes), it assumes that all cells have similar TOTAL gene expression. But the caveat is that if a gene is super highly expressed, it might incorrectly inflate the total counts and as a result dilute out the lowly expressed genes. In this way, while the highly expressed genes retain their variance post-normalization, small differences between low-expressed genes become less distinguishable. Downstream effects to this bias could include false positives amongst the more highly expressed genes, whereas rare or lowly expressed genes are underrepresented.

Due to sequencing depth limitations at the single cell level, datasets with many dropout effects (that is, when the counts become zero) become exacerbated for the bias towards the genes that ARE highly expressed.

Despite these caveats, in our tutorial later we will be be applying library normalization in part for its ease and scalability for data-intensive datasets such as spatial transcriptomics.

18 of 23

Generic normalization strategies

Binomial regression normalization (Seurat SCTransform):

⇒ Accounts for the relationship between sequencing depth and gene expression while accounting for technical noise.

⇒ Does not assume each spot (“cell”) should have the same count�

Hafemeister and Satija, 2019 Genome Biol

18

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

The SCtransform method models the cell or UMI counts using a regularized negative binomial model to remove the variation due to sequencing depth (total number of UMIs per cell), while adjusting the variance based on pooling information across genes with similar abundances (similar to some bulk RNA-seq methods).

We utilize the Pearson residuals (response residuals divided by the expected standard deviation), effectively representing a variance-stabilizing transformation (VST), where both lowly and highly expressed genes are transformed onto a common scale.

They propose that the Pearson residuals from “regularized negative binomial regression,” where cellular sequencing depth is utilized as a covariate in a generalized linear model, successfully remove the influence of technical characteristics from downstream analyses while preserving biological heterogeneity.

A major caveat to this application is how computationally intensive it is, requiring more ram than we could currently provide you all with today. In addition, SCT Transform doesn't have the luxury of using a training dataset and projecting onto your dataset. I encourage you to check out this application when you do have the chance, however.

Some additional normalization strategies include Scale factor normalization (DESeq2, scran):

Adjusts for sequencing depth by computing size factors that normalize counts without assuming equal distributions of gene expression.

19 of 23

Spatially-informed normalization approaches

Spatial smoothing: applies a smoothing kernel based on neighboring expression to “smooth out” gene expression spatially

•Gaussian Kernel smoothing: weighted average with higher influence of closer spots

•K-Nearest Neighbors (K-NN): averages the values (“K”) of nearest neighbors (more uniform)

•Disc smoothing (SME): Determines neighborhoods based on cell morphology and proximity.

19

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

Gaussian Kernel Smoothing: Averages gene expression over neighboring spots, weighted by distance.Laplacian Smoothing: Uses a graph-based method where spots are nodes, and edges represent spatial proximity.

K-NN may not work well with neural data where there ARE distinct patterns where cell density varies (such as upper vs lower layers in the cortex)

Examples include:

stLearn-SME: Neighborhood information and morphological distance incorporated

Disc smoothing: assumes that cells sharing morphological similarities also have more similar transcriptional profiles.

Image: 10x Xenium datasets features the mouse olfactory bulb. Left is a traditional log normalization approach, the middle is the same but with reconstruction of the missing values in a sparse dataset, and the last is a spatially deconstructed approach wherein spatial smoothing (Gaussian kernel) is applied.

20 of 23

Transforming and integrating datasets

Data integration can mitigate unwanted covariates like batch, sex, or family cohort effects

Corrective/integrating programs vary but include: �ComBat-Seq, Harmony, Liger, Seurat

20

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

21 of 23

Thank you!

Contact info

Savannah Kilpatrick

e-mail: sav.kilpatrick@gmail.com

LinkedIn: linkedin.com/in/savannah-kilpatrick-phd-432737296

21

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

22 of 23

Lab Practice

Using a single dataset, perform the following:

Data filtering and QC
Normalization
Gene expression visualization

22

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision

23 of 23

We are on a Coffee Break & Networking Session

Workshop Sponsors:

23

bioinformatics.ca | The Brain Single Cell Initiative | Panoramics - A Vision