2 of 14

scRNA-Seq pipeline overview

biological sample

sequencer output

unfiltered count matrix

filtered count matrix

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

Cells

HVG selection + scaling

Cells

Genes

Cells

We want a visual summary of thousands cells’ gene expression.

3 of 14

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

How ?

Why not ?

How ?

Why an intermediate step is done ?

We will summarize genes expression in few dimensions, before building the 2D projection.

http://cmdlinetips.com/wp-content/uploads/2018/03/Sparse_Matrix.png

scRNA-Seq data are sparse

> 70 % of the expression matrix is 0 : not very informative

Data are noisy

Some genes are more informative than some other.

There is biological / technical noise in gene expression.

Computational time

prop(expr_mat == 0)

4 of 14

Challenges

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

Cells

How to reduce the number of dimensions ?
How many dimensions ?

We want a visual summary of thousands cells’ gene expression.

How ?

How to identify cell populations ?

5 of 14

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

How ?

Why not ?

How ?

Dimensionality reduction

Overview

Commonly used dimensionality reduction methods

PCA Principal Component Analysis
BFA Binary Factor Analysis
ICA Independent Component Analysis
LSI Latent Semantic Indexing
LDA Linear Discriminant Analysis
…

Important parameters

information : number of variable genes (HVG)
number of dimensions to generate (signal / noise)
randomness : random seed
convergence criteria

HVG selection

Cells

Genes

(≈15,000)

HVG

“constant” genes

Cells

HVG

(≈3,000)

Cells

scaling

reduced space

Dimensions

(≈50)

Cells

HVG

(≈3,000)

6 of 14

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

How ?

Why not ?

How ?

Dimensionality reduction

Principal Component Analysis - principle

Input : X (≈ 2 000 - 5 000) HVG with scaled expression leveles
Goal : Group genes by dimensions when they have similar expression across cells

gene 1

gene 2

gene 3

gene 4

gene 5

gene 6

dim 1

dim 2

dim 3

Output : Z (≈ 50 - 100) dimensions “Principal Component”
Each PC summarizes a certain amount of the input data variability
First PC recapitulates the most part of information
Last PC can be considered as noise

7 of 14

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

How ?

Why not ?

How ?

Dimensionality reduction

Principal Component Analysis - visualization

Input : X most variable genes
Goal : Group genes by dimensions when they have similar expression across cells
Output : Z dimensions “Principal Component”
Each PC summarizes a certain amount of the input data variability

Now, we will use the reduced space to make a 2D representation.

8 of 14

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

How ?

Why not ?

How ?

2D space for cells visualization

Commonly used 2D space

UMAP
tSNE
Diffusion Map
…

The same cells can be represented using different 2D spaces.

Do not make to many interpretations from the 2D space, it is an over-simplified representation of cells.

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6417818/

Important parameters

input information : number of dimensions
cells neighborhood : number of neighbors, perplexity, distance method, …

9 of 14

normalized matrix

reduced space

cells visualization

Dimensions

Cells

Genes

Dim. 1

Dim. 2

How ?

Why not ?

How ?

Clustering

Clustering is made on expression matrix or reduced space, not on the 2D projection.

The 2D projection is not a clustering. A clustering is an annotation.

Commonly used methods

Louvain clustering
Leiden clustering
k-means
…

Important parameters

input information : number of dimensions
cells neighborhood parameters : number of neighbors, distance measurement method, resolution…

k-nearest neighbors (kNN)

k = 3

k = 6

shared nearest neighbors (SNN)

clustering

(from SNN graph)

10 of 14

Summary

HVG selection

normalized matrix

Cells

Axis 1

Axis 2

Cluster

Cells

Genes

(≈15,000)

HVG

Cells

HVG

(≈3,000)

Cells

scaled matrix

reduced space

Dimensions

(≈50)

Cells

HVG

(≈3,000)

Genes

(≈15,000)

UMAP

tSNE

others…

11 of 14

Take Home Messages

The number of variable genes impact the PCA, thus the 2D space. It depends on the expected number of cell populations in the dataset.
Number of dimensions = amount of information (not enough < - - > noisy data)
UMAP is suited to visualize several cell types and their global transcriptomic profile
tSNE is suited to visualize sub cell types and their local transcriptomic particularity
Diffusion Map is suited to visualize cell differentiation data
The resolution impacts the number of clusters : not enough clusters / not biologically interpretable clusters

Advice :

Make the analysis with all default settings :

2000 HVG
15 PC to generate a UMAP (or tSNE)
Resolution 1 for the clustering

Identify your cell populations
Change the settings to make the representation showing what you identified

The goal is to generate a quick representation for your cells. Run your favorite analyses and represent results on the representation. Do not make to many interpretations from the 2D representation itself.

12 of 14

Let’s go to practice

normalized matrix

reduced space

PCA

cells visualization

UMAP

Dimensions

Cells

Genes

Dim. 1

Dim. 2

Cells

Nb of variable features : 500, 2000, 5000
Nb of dimensions : 50

Nb of dimensions : 5, 15, 50, 100

Nb of dims : same as UMAP
Resolution : 0.1, 0.5, 1, 5

13 of 14

	500	2000	5000
5
15
50

Number of variable features

Number of PC (/50) to make the UMAP

14 of 14

0.1	0.5	1	5

Resolution