JavaScript isn't enabled in your browser, so this file can't be opened. Enable and reload.

1 of 34

Inferring regulatory networks from scRNA-seq datasets

Oct 5^th 2021

BMI 826-23 Computational Network Biology�Fall 2021

Sushmita Roy

https://compnetbiocourse.discovery.wisc.edu

2 of 34

Plan for this section

Overview of network inference (Sep 21^st)
Bayesian networks (Sep 21^st,Sep 23rd)
Dependency networks (Sep 23^rd)
Integrative inference of gene regulatory networks (Sep 28^th,Sep 30^th)
Inference of regulatory networks from single cell datasets (Oct 5^th)

3 of 34

Plan for today

Brief background into scRNA-seq
Algorithms for scRNA-seq

PIDC
SCENIC

Benchmarking scRNA-seq network inference algorithms

4 of 34

scRNA-seq: a powerful technology to understand heterogeneous cell populations

Bulk

Single-cell

Figures: 10X Genomics; Shalek AK et al., Nature 2014

To start, as opposed to bulk RNA sequencing, scRNA-seq permits transcriptomic comparison of individual cells by first isolating individual cells using one of a variety of techniques [[[ POINT ]]]
Which allows you to assess heterogeneity within a sample of cells…
For example, this could be examining heterogenous cells in a tumor, or for cellular reprogramming.
A major use of scRNA-seq data has been to assess transcriptional similarities and differences within a population of cells.

Discovery

To do this, clustering methods can be applied to group cells according to their expression profile
Here, this figure is showing that further analysis of the transcriptional profiles shows that there is heterogeneity within this orange population of cells.
This orange cluster can be considered a subpopulation. It has a different transcriptional profile than the green, blue, or purple clusters.

Cell subpopulations can also be characterized by different temporal or spatial gene expression analyses.
For example, once you define cell subpopulations, you can do downstream analyses such as gene network inference to further understand the transcriptional state of a cell type and transcriptional differences between subpopulations of cells.
Therefore, elucidating cellular heterogeneity remains a core reason for embarking on scRNA-seq studies.
The results of new analysis discoveries can then be followed up by further laboratory validation and experimentation.

^{____________________________________}

https://genomemedicine.biomedcentral.com/articles/10.1186/s13073-017-0467-4

5 of 34

Computational problems with scRNA-seq data

Pre-processing and normalization
Visualization
Cell type identification
Trajectory inference:

Single cell ordering: pseudo time
Cell population structure relationships

Network inference

6 of 34

Inferring networks from scRNA-seq data

Lots of measurements from a single sample
Natural cell-to-cell variation should be valuable for network inference

7 of 34

Pseudotime

Bergen et al, 2021

If we can think of each cell to capture a snapshot of a dynamic process, can we order cells based on their transcriptional signatures?

8 of 34

Classes of network inference algorithms

Information theoretic methods

PIDC
knnDREMI
SCRIBE

ODE based methods

SCODE
Ocone et al

Boolean models

Graphical models & Dependency networks

SCENIC
SCHiRM
HurdleNormal
SILGGM

Correlation-based methods

LEAP

Adapted from Babtie et al., Current Opinions in Systems Biology 2017, Stone et al unpublished

Incorporate pseudotime

LEAP
SCRIBE
SCODE
Ocone et al
SINGE

Don’t use pseudotime

knnDREMI
SCENIC
SCHiRM
HurdleNormal
SILGGM
PIDC

How to model an edge?

Pseudo time?

9 of 34

Plan for today

Brief background into scRNA-seq
Overview of algorithms for scRNA-seq

PIDC
SCENIC

Benchmarking scRNA-seq network inference algorithms

10 of 34

PIDC

T. E. Chan, M. P. H. Stumpf, and A. C. Babtie, “Gene Regulatory Network Inference from Single-Cell Data Using Multivariate Information Measures.,” Cell systems, vol. 5, no. 3, Sep. 2017, [Online]. Available: http://view.ncbi.nlm.nih.gov/pubmed/28957658.
Based on information theory

Multivariate information (MVI)
Partial Information decomposition (PID)

Learns an undirected network by using PID to assess the information for three variables

This is better than mutual information (MI) based methods

11 of 34

A few information theoretic concepts

Mutual information

Conditional mutual information

Multivariate information

A way to measure information between more than two variables
PID and Interaction information are two ways to assess this

PID: Partial information decomposition is computed using

Redundant information
Synergistic information
Unique information

12 of 34

RECAP: Mutual Information

I(X;Y) is the mutual information between two variables

How much information do we have for one variable about the other?

Mutual information is defined as

Measures the difference between the two distributions: joint and product of marginals

13 of 34

RECAP: Conditional mutual information

I(X;Y | Z) is defined as the mutual information between X and Y given a third variable Z

Conditional mutual information is defined as

14 of 34

Interaction information

Let X, Y, Z denote three random variables
Interaction Information (II) is defined as the additional information we get when we observe a third variable versus when we don’t
Defined as the difference between conditional mutual information and mutual information

Conditional mutual information

Mutual information

15 of 34

Partial Information Decomposition

A way to measure information between three variables, where two are “source” variables and the third is a target
Let S={X, Y} denote the source set and Z the target
PID is defined using the unique, synergistic and redundant information between Z and X,Y

16 of 34

Partial Information Decomposition

PID is defined as

Unique_Y(Z;X) is the unique information between source variable X and target variable Z when the other source variable is Y
For network inference, only the “Unique” and “Redundancy” information is important

17 of 34

Redundancy

Redundancy is defined based on the minimal information about Z’s state from the source variables.
First define for each source X

Redundancy is

18 of 34

Unique information

The mutual information between two variables can be decomposed into two parts

19 of 34

Synergistic information

Recall interaction information: II(X;Y;Z)

Synergistic information is defined as the information about target Z, that we have due to both X and Y

20 of 34

Unique information can discriminate between edges and non-edges

21 of 34

Using PID for network inference

22 of 34

PIDC on simulated data

23 of 34

Application of PIDC to single cell data

24 of 34

Plan for today

Brief background into scRNA-seq
Algorithms for scRNA-seq

PIDC
SCENIC

Benchmarking scRNA-seq network inference algorithms

25 of 34

SCENIC

S. Aibar et al., “SCENIC: single-cell regulatory network inference and clustering,” Nature Methods, vol. advance online publication, Oct. 2017, doi: 10.1038/nmeth.4463.
Based on three tools

GENIE3, RcisTarget, AUCcell

Incorporates TF motif binding information to infer physical interactions

26 of 34

SCENIC workflow

Use GENIE3 to predict target genes of each regulator.

These are likely co-expressed

Use RcisTarget to infer which targets have a binding motif support. Such filtered targets are called regulons

Use AUCcell to assess how active a regulon is in a cell.

27 of 34

Evaluation of SCENIC

Can SCENIC predict key regulators of different cell types
Does cell clustering based on SCENIC AUCcell score provide agree with known cell types?
Assess performance on a mouse brain scRNA-seq data with known cell types

28 of 34

Application of SCENIC on a dataset with known cell labels

SCENIC accurately predicts regulators for known cell types. Clustering of cell types is more accurate when using SCENIC regulons

29 of 34

Plan for today

Brief background into scRNA-seq
Algorithms for scRNA-seq

PIDC
SCENIC

Benchmarking scRNA-seq network inference algorithms

30 of 34

Benchmarking single-cell network inference algorithms

A. Pratapa, A. P. Jalihal, J. N. Law, A. Bharadwaj, and T. M. Murali, “Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data,” Nat Methods, vol. 17, no. 2, pp. 147–154, Feb. 2020, doi: 10.1038/s41592-019-0690-6.

31 of 34

Performance of algorithms on simulated data

Different network topologies

Number of genes: 7-18

32 of 34

Performance on real data

Each number is the Early Precision Ratio. Higher the better.

33 of 34

Overall findings from benchmarking

PIDC or GENIE3 might be good algorithms for scRNA-seq network inference

34 of 34

Take away points

Inference of regulatory networks from scRNA-seq data is in its early stages

Although a lot of algorithms exist, there is room for improvement

Algorithms differ based on

Pseudotime incorporation
Model of interaction

Information theoretic methods might be a promising direction to explore for scRNA-seq network inference