Introduction to networks and
pathways
Introduction to multi-omics data integration and
visualisation 2024
Satyadhyan Chickerur, PhD
KLE Technological University
Ref: Marton Olbei & Dezso Modos & Tamas Korcsmaros
A brief
introduction
to networks
Figure: @randomgraphs
Behind every complex system (comprised of many interconnected parts, with a complicated
arrangement) lies a network:
Networks encoding the interactions between genes, proteins, etc. the cellular network
The wiring diagram capturing the interactions between neurons, the neural network
The interactions of family, friends and colleagues, a social network
Communication networks, describing how communication devices interact with other
Power grid, trade networks and many - many more.
Figure: @randomgraphs
This is a node
In network biology nodes can
represent various components:
Genes
Proteins
RNAs
and other biological entities
This is an edge
node 1 node 2
Edges represent the interactions occurring between our
nodes
They can be: Directed or undirected
Weighted or unweighted
Stimulatory or inhibitory
Networks are collections of nodes and edges
In network biology we use them to:
Visualize our biological entities
See overrepresented functions and their interactions
Make predictions about potential interactions
Characteristics of a
network
Figure: @randomgraphs
Degree
a b c
The degree of a node
represents the number of
1 2 1
links it has to other nodes,
e.g. number of citations on a
paper in a research graph,
how many friends you have.
Ka = 1 Kb = 2 Kc = 1
As networks can be
directed or undirected for a b c
the former, in addition we
differentiate between in-
degree and out-degree 1 2 1
Kaout = 1 Kbout = Kcout = 1
Kain = 0 0
K in = 2 Kcin = 0
b
Degree
The degree of a node
represents the number of
1 2 1
links it has to other nodes,
e.g. number of citations on a
paper in a research graph,
how many friends you have. a
1
Hubs
Hubs are the largest
degree nodes in a network.
d e b
1
They are often very 4 1
important biologically and
their deletion can be lethal.
Can anyone think of an c
example? 1
Paths
d
A path is a sequence of
nodes in which each node a b c
is adjacent to the next one
a→b→c→d
a→b→c→e
e
Paths
d
A path is a sequence of
nodes in which each node
is adjacent to the next one
a b c
The distance between two
nodes is defined as the
number of edges along the a→b→c→d
shortest path connecting
them
a→b→c→e
e
Paths
a
A path is a sequence of c
nodes in which each node
is adjacent to the next one
The distance between two
nodes is defined as the b
number of edges along the d
Diameter:
shortest path connecting longest of all the
them. calculated a to e
shortest paths in 3 steps
a network
The shortest path is 4 steps
the path with the
shortest length Characteristic
(distance) between path length:
average length of f e
two nodes. shortest paths in
a network
Figure: network science book
Local network structures
Graphlet: A local unique (non-isomorphic) structure of a
network
Network motif: An overrepresented local structure of a
network (a common graphlet)
Example network motifs
Feed forward Feedback loop Three chain
a a a
b b
b
c c
c
Milo, et al (2002). Science
Connectedness
A graph is connected if
any two nodes can be
joined by a path.
We refer to the largest
connected component as
the giant component and
the rest as isolates.
Bridges connect the
various components - if
you erase them the graph
becomes disconnected
Figure: network science book
Clustering
i i
Clustering coefficient:
what fraction of your
neighbors are connected?
Watts & Strogatz, Nature
1998.
10 edges total 7 edges total
6 between 3 between
neighbours neighbours
Ci = 1 Ci = 0.5
Figure: network science book
Clustering Calculating clustering
coefficient
i ei = edges between the
neighbours of node i
ki = degree of node i
Clustering coefficient:
what fraction of your
neighbors are connected
(value between 0 and 1)?
7 edges total
Watts & Strogatz, Nature 3 between 2*3
1998. neighbours
Ci = 0.5 4 * (4-1)
6
= 0.5
12
Figure: network science book
Betweenness centrality
How much influence does
my node have over the
information flow in the
network?
Betweenness: fraction of
shortest paths which
goes through a given
node
Figure: network science book
Networks vs Pathways
Bipartite networks
Bipartite networks are
networks whose nodes can
be divided into two distinct
sets U and V in such a way
that every link connects a
node in U to one in V
https://www.nature.com/articles/srep00196
Multilayered networks
Multilayered networks
contain multiple types
of edges, changing the
context of the
interaction.
Representing networks
One of the most commonly used
representation due to its ease of
use, and the ability to use it
directly for measurements is the
adjacency matrix.
For an undirected network, if the
position A12 equals to one, that
means there exists an interaction
between nodes 1 and 2.
For directed networks the
direction of the interaction
carries information on its own: 1
→ 3 is not the same as 1 ← 3,
and their adjacency matrices
behave accordingly.
Figure: network science book
Representing networks
a
network.csv
Another very commonly used
format is the edge list.
a,e,
b,e,
The edge list is more human c,e, d e b
readable, and is easy to write up d,e
on the fly.
The example uses a .csv format,
but different software of course
can accept different delimiters.
c
Representing networks
Standards are very important to
every field, and network biology
is no different.
One attempt at this, to
standardize the efforts of the
molecular interaction field is the
PSI-MI TAB format, which uses a
controlled vocabulary to
describe interactions.
https://academic.oup.com/bioinformatics/article/35/19/3779/5355053
Where do networks come
from?
Figure: @randomgraphs
Network databases
Correlation/similarity networks:
You can generate from any omics
experiment (transcriptomics, proteomics
etc.) and build them through various
metrics e.g. correlation.
Pros:
● Condition specific network
Cons:
● Undirected
Tools: WGCNA
Physical interactions
● “Strict” definition: captures the physical
interactions of proteins
● “Loose” definition: captures the functionally
interacting proteins (e.g. includes members of a
complex)
● Main experimental methods: Y2H, AP-MS,
fragment complementation, BioID
Large scale protein interaction networks -
Y2H
Pros:
● High throughput
● Well established technique
Cons:
● Cannot detect genes which not go
into cell nucleus
● Bait-prey interactions have to work
● Context dependent
Example databases: HuRI
Common methods for creating large scale
protein interaction networks - manual
curation Pros:
● Usually more accurate
Save to
database
Cons:
● Labour intensive
● Publication/curator bias
Example datasets:
From: https://phdcomics.com/
SignaLink, Reactome,
WikiPathways
Common methods for creating large scale
protein interaction networks - text mining
Pros:
Searches the whole
literature
Cons:
Potential bias from the
algorithm
Examples: STRING
Common methods for creating large scale
protein interaction networks - secondary
databases Pros:
● Database of databases
● Many interactions
Cons:
● Bias of all the various datasets
● Need to know what is the aim
of your research
E.g. OmniPath, STRING ,
Türei et al 2021
ConsensusPathDB
Available tools
Figure: @randomgraphs
Tools
Easy to use, user friendly
Visualization, analysis and data sharing (NDEx)
Has its own ecosystem of apps, greatly extending its usability
https://apps.cytoscape.org/
Automation: RCy3, py4cytoscape
R / Python (and other languages) library for network
manipulation and visualization
Powerful tool, intermediate difficulty
Great introduction: https://kateto.net/network-
visualization
ggraph / tidygraph
Two R libraries made to help users manipulate
and visualize network data using the tidyverse
conventions (i.e. ggplot, dplyr)
Tidygraph provides manipulations and
analysis grammar for network data, while
ggraph offers visualization grammar for
network data
A bit easier to use than igraph, lacks some
functions, but has other important ones, e.g.
easy faceting based on attributes
Network analysis and discovery
Figure: @randomgraphs
● Let’s say you constructed, queried, curated the network you are
interested in
● Let’s say it is the differentially expressed proteins in a disease
using proteomics and their interaction partners in the SignaLink 3
network resource
● What now?
Network clustering
● One common way of adding
value to your network is to
cluster modularise partition
them
● Module: A part of the network
where the nodes are more
connected with the nodes within
the module then outside of the
module
● There are many approaches to
do so, including in Cytoscape
(MCODE, clusterMaker) https://mr.schochastics.net/
How to measure modularity?
Module: A part of the network where the nodes are more connected with the nodes within
the module then outside of the module
Weighted degree of node i and j
Kronecker delta which is 1 if ci = cj and 0 otherwise
Community of node i and j
Sum of all edge weights Edge weight between
nodes i and j
-½ - not modular network
0 - random network
1 - completely modular network
How to find communities, modules?
Girvan-Newman clustering (top-down):
1. Calculate the betweenness centrality of each edge.
2. Remove the highest betweenness centrality edge(s)
3. Recalculate the betweenness.
4. Repeat till no edges remain
Best partition where the modularity is the highest
Louvain clustering (bottom up):
1. Each nodes assigned to it’s own community
2. Adding a node to its neighbours community than calculate the modularity change for each node
3. Assign each node to that neighbour where the modularity change is the largest.
4. Create communities and do from step one again where the nodes are the communities
The modularity plateaus - multiple equally good
partition
Gene ontology enrichment
● H0: In our network module the genes are as common as in the
background
● H1: The genes in the network module are more common than in
the background
● What should be the background?
● How to know what is more common?
What should be the background?
● Whole human genome? Signalling Process, Intracellular
○ It is not in the network! communication, kinase,
○ We do not measure it in Enzyme
proteomics
● The whole network?
○ Not all IDs are mapped to
● The intersection of the whole
network and the omics data
○ We do not find anything
How to know what is more common?
● Statistical test Signalling Process, Intracellular
○ Fisher exact test communication, kinase,
○ Hypergeometric test Enzyme
○ Chi square test
● You do for each Gene
Ontology function a
statistical test! More GO
function then gene!
● Use false discovery rate
correction!
Network clustering and enrichment Cell cycle
● As we’ve discussed
before, closely connected
nodes often work together
in shared tasks
● To find out what these are
in an unsupervised way we
can run enrichment
analyses on them (BinGO)
https://mr.schochastics.net/
Visualising enrichment results
Dotplot Treemap plot
Top n enriched biological functions in All significant biological process mapped by the GO
annotation database tree.
Network propagation
● Network propagation
methods can highlight
nodes of interest in
addition to already known
ones
● Example algorithms:
PageRank, HotNet2
Hristov et al 2020
Footprint based methods
● Footprint methods can
estimate activity of kinases
or transcription factors
using prior knowledge, eg.
highlighting the most
active TFs in a given RNA-
Seq readout
● Examples: VIPER,
decoupleR Vandereyken et al 2018
Dangers and pitfalls
Figure: @randomgraphs
Study bias
• Biological networks almost always have to deal with a huge
study bias
• some proteins / interactions will be more studied, as they are
more relevant for health for example
• as such they will show up more frequently in our network
resources
Trivial predictions
• Trivial predictions can dominate your results, especially if
you are familiar with the biological domain (this is also a
result of study bias)
• As such, you will probably make a long list of predictions
(that will be correct) that you have to go through to find novel
information
Path fragility
• Biological networks can have simultaneously many false
negative / false positive hits
• Depending on your experimental method, studied organism
etc. the coverage of your interaction networks can vary quite
a lot
• Therefore using path based metrics such as shortest paths
or betweenness centrality can be misleading, as even the
addition / loss of one edge can shift these results a lot
ID mapping in Cytoscape
(and in general)
Dezso Modos
General Biological Genome Databases
Name of the Website Based on what? Species Example IDs
database
Uniprot www.uniprot.org/ Proteins You can Multiple (from P00533
find the protein bacteria till human) P00533-1
isoforms
Ensemble https://www.ensembl.org/ Genome base: Vertebrae ENSG00000146648
Genes/ transcript/ ENSP00000275493
proteins ENST00000455089.5
NCBI Gene https://www.ncbi.nlm.nih.gov/ Multiple genomes Multiple 1956
gene/
NCBI RefSeq www.ncbi.nlm.nih.gov/refseq/ Reference Multiple NM_001346897.2
sequences NP_001333826.1
Interaction/pathway databases
Database Website Content Species Example IDs
name
OmniPath omnipath.org A collection of multiple Human and P00533
protein-protein interaction mouse - mostly
databases but continuously
updated
STRING https://string-db.org/ Integrated protein-protein Many different 9606.ENSP00000275493
interaction database species mostly
based on
orthology
Reactome https://reactome.org/ Reaction based pathway Multiple species R-HSA-179837
database
IntAct https://www.ebi.ac.uk/intact/ Curated protein-protein and Multiple species P00533
other molecule interactions including viruses
SignaLink http://signalink.org/ Multilayer signalling database H. sapiens, D. P00533
containing proteins, miRNAs, rerio, C. elegans,
TF-TG interactions D. melanogaster
Many to many mapping - the biological rational
Interaction measurements
Gene Transcript 1 Protein 1 Yeast two hybrid
Transcript 2 Protein 2 Affinity purification
Gene
Probe from a microarrays Proteomic
RNA-seq read mapping measurements
Genomic analysis
(GWAS)
Many to many mapping - mapping table
Gene/ Value Gene/ Protein ID Protein 1 Protein 2
transcript Transcript ID
ID1 1 ID1 ID_A_1
ID_A_1 ID_A_2
ID2 1.5 ID2 ID_A_2
ID_A_2 ID_A_3
ID3 2.5 ID3 ID_A_3
ID_A_1 ID_A_4
ID4 1 ID4 ID_A_4
ID_A_3 ID_A_4
ID4 ID_A_5 ID_A_3 ID_A_5
Source of mapping tables
Ensembl BioMart - note you can access it programmatically through R -
https://www.ensembl.org/biomart/martview/
https://bioconductor.org/packages/release/bioc/html/biomaRt.html
Uniprot - https://www.uniprot.org/uploadlists/
ID mapping: what should I do during my project?
Know how do you map, document what have you done (e.g. How many IDs were
mapped to how many)
Try to map as many ID as possible - check the unmapped IDs
Use only the “sure” IDs - Uniprot Swiss Prot
Do not use gene names - the septins and androgen receptors will haunt you
How can we show this
in cytoscape?
Step 1 load in the network
Step 2 - the mapping table
Step 3 - the data
Real word example
You downloaded the interactions from Reactome through Cytoscape related to the
EGFR protein
You got from your bioinformatician the RNA-seq data using HGNC Gene names
Downloading data through PSICQUIC
Oh wait… we lost around 10% of our IDs!
Why do the identifiers do not mapped?
Common reasons:
● Different version between the databases
● Pseudogene and it is not in the
protein/transcript based database
● It is not in the database, typo in ID
● Wrong species the interaction is in rats
meanwhile we use humans
Let’s see a few examples: Example 1 wrong ID in the Reactome
database
One of the missing ones is KRAS - with a strange uniprot ID
If we search for that id we got back the normal uniprot ID of KRAS: P01116
In such case I suggest to make a separated column for our uniprot IDs maybe we will
need them for further work and adding the HGNC column the KRAS manually
Example 2: Not surely expressed protein TREMBL ID
This gene do not have a name in
Reactome - uniprot ID in
Human Readable Label column
Uniprot says it is Unreviewed protein
and only exist at transcript level
We can exclude from our analysis
Example 3 - wrong uniprot ID
You need to check the genes “manually” Basically building up your own mapping table
But this one do not have any uniprot entry regarding this ID
Search to the gene name and use uniprot mapping to map to HGNC gene name
Hands on the computer
1. Download the interactome of colorectal cancer associated proteins
(FAM123B ARID1A, ATM, CTNNB1, GNAS, PTEN, DNMT3A, NF1, MLL2,
PIK3R1)
2. Download a mapping table
3. Map transcriptomic data of p53 mutant and wild type adenocarcinoma
4. See the difference
