Greg Carter
Galitski LabInstitute for Systems Biology (Seattle)
Maximal Extraction of Biological Information from
Genetic Interaction Data
Genetic Interaction
Pairwise perturbation
two genes combine to affect phenotype
Hereford & Hartwell 1974
Measure a phenotype for 4 strains:
1. Wild-type reference genotype
2. Perturbation of gene A
3. Perturbation of gene B
4. Double perturbation of A and B
• Loss-of-function, gain-of-function, dominant-negative, etc.
• Interaction depends on phenotype measured.
Example: flo11 and sfl1 for yeast invasion.
WT flo11 sfl1 flo11sfl1
pre-
was
hpo
st-w
ash
Invasion Assay
~2000 interactions measured
(Drees et al, 2005)
Genetic Interaction
45 possible phenotype inequalities
Classified into 9 rules (Drees, et al. 2005)
Classification of Interactions
WT=A=B=AB, WT=A<B=AB, A=B=WT<AB, A<B<WT=AB, AB<A<WT=B, WT=A=AB<B, WT=A=AB<B, A<B<WT<AB, etc…
Distribution of Rules
2000 interactions among 130 genes
Yeast Invasion Network
Extracting Biological Statements
Statistical associations of a gene interacting with a function
PhenotypeGenetics plug-in for Cytoscape
www.cytoscape.org
WT=A=B=AB, WT=A<B=AB, A=B<WT<AB, A<B<WT=AB, AB<A<WT=B, WT=A=AB<B, WT=A=AB<B, A<B<WT<AB, etc…
?
Can the 45 interactions be classified in a more informative way?
How many rules?
Distribution of interactions?
Classification Problem
Requirements for a complexity metric :
1. Adding a gene with random interactions adds no information
2. Duplicating a gene adds no information
3. Should depend on
(i) the information content of each gene’s interactions, and
(ii) the information content of all gene-gene relationships.
General requirements for biological information (see poster).
Context-dependent Complexity
= Ki mij (1 – mij )
Ki is the information of node i,
mij is the mutual information between i and j,
0 ≤ mij ≤ 1and
0 ≤ ≤ 1
Applied to (see poster):
• Sets of bit strings (sequences)• Network architecture• Dynamic Boolean networks• Genetic interaction networks…
pairs ij
Context-dependent Complexity
Genetic Interaction Networks
• Invasion network of Drees, et al. Genome Biology 2005
130 genes, 2000 interactions
• MMS fitness network of St Onge, et al. Nature Genetics 2007
26 genes, 325 interactions
Determined networks of maximum complexity .
Network Classification Scheme
Invasion Data MMS Fitness Data
biological
statements
biological statements
Drees, et al. 0.57 52 0.27 28Segré, et al. 0.52 47 0.32 19St Onge, et al. - - 0.16 10Maximum 0.79 72 0.62 32
Complexity and Biological Information
Number of biological statements is correlated with
115k possible MMS fitness networks, r = 0.80
Genetic Interaction Networks
Maximally complex MMS fitness network
Rule Frequency InequalitiesClassical Interpretation
(Drees et al. 2005)
1 120 PAB = PA < PB < PWTepistatic
2 55 PAB < PA = PB < PWTadditive
3 92 PAB < PA < PB < PWTadditive
4 30PAB = PA = PB < PWT
PAB = PA < PB = PWT
asynthetic
non-interactive
5 26
PAB < PA = PB = PWT
PA < PAB = PB < PWT
PAB = PA = PB = PWT
PAB < PA < PB = PWT
PA < PAB < PB < PWT
synthetic
epistatic
non-interactive
conditional
single-nonmonotonic
gene interacts via with genes P
SGS1 Rule 5 error-free DNA repair 0.00014
SWC5 Rule 2 error-free DNA repair 0.00056CSM2 Rule 4 error-free DNA repair 0.0026
SHU2 Rule 4 error-free DNA repair 0.0030SHU1 Rule 4 error-free DNA repair 0.0065
Genetic Interaction Networks
Biological statements from the
maximally complex MMS fitness network
gene interacts via with genes P
PSY3 Rule 1 meiotic recombination 0.0011
St Onge, et al. Figure 5d
Conclusion and Future Work
For a given data set, maximizing facilitates unsupervised, maximal information extraction by balancing over-generalized and over-specific classifications schemes.
Need network-based methods to interpret the maximally complex interaction rules. Interpretations will depend on the system, specific to phenotype measured and perturbations performed.
See poster for more details
Becky DreesAlex Rives
Marisa RaymondIliana Avila-Campillo
Paul ShannonJames TaylorSusanne Prinz
Vesteinn ThorssonTim Galitski
Matti NykterNathan Price
Ilya ShmulevichDavid Galas
Thanks to