Putting genetic interactions in
context through a global
modular decomposition
Jamal
Motivation
• Genetic interaction provide powerful
perspective how gene functions
• specific mechanisms that give rise to these
interactions not well understood
• Requires a thorough study of genetic
interaction networks  understand the
structure of the network.
This study
• This study uses a datamining approach to
explore all block structure with in this
network.
Characteristics
• Genetic interaction:
“Multiple genetic perturbations whose combination result in a
phenotype that is unexpected given the phenotypes of the
individual perturbations”
The redundancies and dependencies within genetic network can
provide powerful means for functional characterization.
• Unlike the PPI network, there is no obvious
functional interpretation of a single genetic
interaction, either negative or positive.
• The genetic interaction of two genes does not
imply that they interact physically, it simply
suggest that they share some kind of
functional interaction.
Modular hypothesis
• Gene membership falls into different type of
functional modules
• For example:
• Protein complexes, pathways, etc.
Negative between pathway Model
• Defines Negative interactions: which are
thought to arise between functionally
redundant pathways such that deleting any
pair of genes spanning across the pathways
results in a significant reduction of fitness
Positive within pathway Model
• defines Positive interactions: If the second
deletion in that same compromised pathway
does not result in any additional fitness
defect.
Bi-Clusters as block pattern in network
• Can be over-lapping or disjoint sets of genes
• Every gene in one set is connected to every
other gene in other set.
• Pu et al.(2008) specifically designed an
algorithm that randomly start with an initial
bi-cluster and then rediscover the prominent
bi-cluster many times.
• In this study authors employed an approach
based on an algorithm from field association
rule mining to find all biclusters of sufficient
size.
Approach Summary--bi-cluster
Discovery
• Recent data from Costanzo et al. (2010) was
used in this study and the developed
approach utilizes the apriori algorithm from
the field of association rule mining to discover
all biclusters.
• and the biclusters that can be expressed by
degree distribution alone were filter out using
non-parametric statistical assessment.
XMOD
• This approach XMOD (eXhaustive Modular
Discovery) guaranteed to find all bi-partite
graphs : Where 1 part of bi-partite acts as a
functional unit
Presence of degree distribution based
Bi-clusters
• Edges were randomized and still bi-partite graphs
were obtained suggesting that biologically
meaningless bipartite graphs can exist.
• score for each bi-cluster lower for biologically
meaningful
• Score: “ the product of probabilities of each edge
occurring independently conditioned on the
degree of two interacting genes”
• Filtered Biclusters: using the independence
score a cutt off is applied to separate the ones
with less independence score
• Condensed Biclusters: after removing
Biclusters with >40% overlap
Comparison with other techniques
Dataset
• The dataset in Costanzoo et al. in 2010 was
used.
• 85,714 negative interactions and 35,858
interactions were used.
Association rule Mining
• Apriori Algorithm in Agrawal (1993) was used.
• Its standard available implementation from a
website was used.
• Apriori was run on a binary set of positive
interactions and also on a set of negative
interactions
Randomizing the Genetic Interaction
network
• The number of edges for each gene was
preserved but the targets were randomized.
• A gene cannot have an edge with itself
Filtering Random bi-clusters
• We found that 50% of the real negative
biclusters and 6% of real positive biclusters
have scores below the 0.01 percentile of
biclusters of the same size from the random
networks. This resulted in 256,502 negative
biclusters and 2194 positive biclusters.
Removing overlap from Biclusters
• we first arranged the biclusters in descending
order by area.
• Then, beginning with the first bicluster A, we
removed all biclusters whose area overlap
with A was greater than 0.4, where overlap
between biclusters A and B was calculated
using the following formula:
Evaluation of Functional Coherence
MEFIT network is based on coexpression
data and does not use genetic interaction
datasets
• Improvements?