1 of 13

Normalization and Scaling

MONASH�GENOMICS and

BIOINFORMATICS

PLATFORM

2 of 13

Counts Matrix

2

https://www.monash.edu/researchinfrastructure/mgbp

3 of 13

Normalisation

Cells can have differences in sequencing depth

3

https://www.monash.edu/researchinfrastructure/mgbp

4 of 13

Normalisation

Normalization is required to make gene expression comparable across cells. We also log-transform.

​

LogNormalize:

Feature counts for each cell are divided by the total counts for that cell and multiplied by the scale.factor. This is then natural-log transformed using log1p.

​

Log( Counts of Gene_A / Library Size Per Cell * scale.factor + 1)

4

https://www.monash.edu/researchinfrastructure/mgbp

5 of 13

Getting ready for PCA

Now we have expression levels on a log scale that we can start to analyse.

​

Our first form of analysis will be PCA.

​

  • We need to select which genes to perform PCA on.

​

  • We need to center and (optionally) scale the data.

​

5

https://www.monash.edu/researchinfrastructure/mgbp

6 of 13

Feature Selection

Select genes that have high variability to focus on interesting biological signal for downstream steps that aggregate or cluster cells based on similarity

​

Not all genes necessarily contain useful information, some contain random noise

​

Seurat’s strategy is to pick the topmost variable 2000 genes �

6

https://www.monash.edu/researchinfrastructure/mgbp

7 of 13

Centering and scaling

  • PCA traditionally operates on centered and scaled data.��Centering: ensure each gene has mean=0��Scaling: ensure each gene has SD=0�
  • Centering is always necessary for PCA.
  • Scaling is optional. �Each gene is already on a comparable scale (log expression).
  • Seurat does scale, by default. �You can experiment with this!

7

https://www.monash.edu/researchinfrastructure/mgbp

8 of 13

Scaling

Expression of some selected genes

8

https://www.monash.edu/researchinfrastructure/mgbp

9 of 13

Extra slides

9

https://www.monash.edu/researchinfrastructure/mgbp

10 of 13

Scaling

To find markers in a cell population you need to compare genes between them

Comparing genes between them when they have different spectrums of expression is challenging

10

https://www.monash.edu/researchinfrastructure/mgbp

11 of 13

Normalisation

Normalizing single-cell RNA sequencing data: challenges and opportunities. Nat Methods 14, 565–571 (2017)

Differences in sequencing coverage between cells due to technical effects (e.g amount of mRNA captured), as well as the volumes of the cells as they grow and divide.

​

Library size normalisation: divide counts in a cell by the total counts for that cell

​

Typically performed by multiplying by 10000 and log transformed in Seurat

​

Seurat also has SCTransform - regularised negative binomial model. (but see this paper)

11

https://www.monash.edu/researchinfrastructure/mgbp

12 of 13

Normalisation

Normalisation corrects for differences in this direction

12

https://www.monash.edu/researchinfrastructure/mgbp

13 of 13

Scaling

Scaling equalizes the amount of variation in each gene

13

https://www.monash.edu/researchinfrastructure/mgbp