Prediction of Synthetic Genetic Interactions In Silico Using Protein Network Data

1
Prediction of Synthetic Genetic Interactions In
Silico Using Protein Network Data

• Sri R. Paladugu
• Keck Graduate Institute
• July 07, 2007

2
Outline

• What?
• Prediction of Synthetic sick or lethal (SSLs)
• Why?
• Goals of our study
• How?
• Computational Methods (SVMs)
• Results Conclusion

3
Protein Interaction Networks
Nodes represent Proteins Edges represent
Observed Interactions
4
Synthetic sick or lethal (SSL)

• Extension of single gene dispensability
  Dispensability of one gene versus, dispensability
  of a pair of genes.

Cells live(wild type)
Cells live
Cells live
Cells die or grow slowly
5
SSL Interaction Networks
Nodes represent Non-essential genes Edges
represent synthetic lethality or synthetic slow
growth
6
Scenarios giving rise to SSL interactions
Wong SL, Trends in Genetics. 21, 8 (2005)
7
SSL hubs might be good cancer targets
Cancer cells w/ random mutations
Normal cell
Alive
Dead
Dead
A. Kamb, J. Theor. Biol. 223, 205 (2003)
8
SSL Interactions ? Chemical Genetics

• SSL Interactions together with chemical genetics
  can help us predict the target pathways of magic
  bullets.

B. Stockwell, Nature Biotechnology. 22, 1 (2003)
9
Goals of our study

• Study the predictive power of protein interaction
  networks for synthetic genetic interactions
  (SGIs) in S. cerevisiae using machine learning
  methods
• Predict novel SGIs in yeast
• Validate the method and carry out novel
  predictions in other organisms

10
How?

• Experimental Methods
• Synthetic Genetic Array (SGA) analysis
• dSLAM
• Computational Methods
• Use machine learning analysis to predict
  synthetic lethal (SL) interactions
• Has been done previously by Wong et.al (2004),
  where they used lot of functional information
  (e.g. common upstream regulator, correlated mRNA
  expression, same MIPS function) to predict (SL)
  interactions in Saccharomyces
• Data for most of the predictor variables that
  were used in the above approach is not readily
  available for other species.

11
Can We Predict ?
Protein Interaction Network in S. Cerevesiae
Synthetic Lethal/Sick Interaction Network in S.
Cerevesiae
http//www.visualcomplexity.com/vc/project_details
.cfm?index19id184domainBiology
12
Network Data

• PIN data and the synthetic lethal/sick
  interactions for Saccharomyces were obtained from
  Reguly et.al (2006).
• Non-Synthetic Lethal/Non-Synthetic Sick
  interactions (Negatives) were constructed by
  generating all pairwise combinations of SGA baits
  (227) with all other genes in yeast and then
  subtracting the Lethal/Sick interactions
  confirmed by SGA and dSLAM methods. Pairs where
  either members are not part of PIN data are also
  removed.
• The genes that were used in SGA analysis was
  obtained from Tong et.al (2001).

Ref 1 Reguly et.al. Comprehensive curation and
analysis of global interaction networks in
Saccharomyces Cerevisiae. Journal of Biology 511
(2006). Ref 2 Tong et.al. Systemic Genetic
Analysis with Ordered Arrays of Yeast Deletion
Mutants. Science 294, 2364-68 (2001).
13
Input properties for SGI prediction

• For every pair of genes we have computed the
  following PIN properties, which were supplied as
  input to the SVM based classifier.
• Degree
• Clustering Co-efficient
• Betweenness Centrality
• Closeness Centrality
• Eigenvector Centrality
• Information Centrality
• Current-Flow Betweenness Centrality
• Bridge Centrality
• Stress Centrality
• Shortest Distance
• Mutual Neighborhood

14
Graph Properties
SSL Pairs
Non-SSL Pairs
15
Graph Properties
SSL Pairs
Non-SSL Pairs
16
Graph Properties
SSL Pairs
Non-SSL Pairs
17
Support Vector Machines
SVM Classifier
Graph properties of a of pair of proteins P1, P2
Optimal hyperplane
Average of Betweenness Centralities of pair P1-P2
Average of Closeness Centralities of pair P1-P2
Absolute Difference between Betweenness
Centralities of pair P1-P2
Output between 0 and 1 measures the propensity
for SL interaction
Absolute difference between Closeness
Centralities of pair of P1-P2
Shortest Distance P1 - P2
Mutual Neighborhood P1 - P2
support vectors
18
Results

• Receiver operating characteristic (ROC) curves
  for SVM classifier trained using various
  predictor variables.

• The accuracy of SVM Classification can be
  measured by the Area Under the ROC Curve (AUC).
• AUC for Classifier trained using All Features
  0.7960

19
Synthetic lethal network has many ?s
20
2Hop Characteristics

• 2Hop S-S is 1 if there is SL interaction between
  the pair (Gene A, Gene C) and also between the
  pair (Gene B, Gene C)
• 2Hop S-P is 1 if there is SL interaction between
  the Pair (Gene A, Gene C) and Physical
  interaction between the pair (Gene B, Gene C) or
  vice versa.

2Hop Y-Z
?
Gene A
Gene B
Y
Z
Gene C
21
2Hop Characteristics

• Addition of 2Hop input properties to the SVM
  Classifier improves the ROC curves significantly.

22
Positive Predictive Value (PPV)

Positive predictive value
Threshold (Cut-Off)
23
Robustness Analysis

• We added some random edges to protein network
  data and recomputed the input properties for each
  gene pair.
• The SVM classifier seems to be robust to bias in
  PIN data

24
Novel Predictions

• Trained the SVM classifier using all the known
  positives and an equal number of known negatives.
• Carried out predictions on the 2 million
  non-essential pairs that were never screened for
  synthetic lethality.
• Repeated the training and testing 5 times each
  time using a different training set.
• Based on a threshold level of 1, we selected the
  pairs that were predicted to be lethal in all 5
  Runs.

25
New Predictions
26
Graph Properties
True Positives
False Positives
27
Graph Properties
True Positives
False Positives
28
Graph Properties
True Positives
False Positives
29
Experimental validation of SSL pairs
30
New Predictions
Pairs not being validated as single gene knockout
strains are not available for one of the two
genes in the pair
Pairs which are being validated experimentally
31
rad1
dep1
end3
ras2
apn2
cti6
thi3
sap30
apn1
atx1
erf4/shr5
cis3
pir1
ccc2
erf2
thi2
spt4
hsp150
YPL183C
YKL050C
Known Interactions
Predicted Interactions
bug1
trm7
SS-SL interactions were color coded based on the
localization information.
32
trm82
pus7
dus3
ncl1
smm1/dus2
dus1
trm8
Known Interactions
Predicted Interactions
33
tsl1
rvs167
fps1
hsp82
tps1
rvs161
Nucleus
Cytoplasm
ER
Cell wall / Plasma Membrane
gpd1
tps3
Golgi Apparatus
Peroxisome
Vacuole
Other
npt1
rim15
Known Interactions
ypt6
ric1
Predicted Interactions
34
rad1
dep1
end3
ras2
apn1
apn2
cti6
thi3
sap30
atx1
erf4/shr5
cis3
ccc2
erf2
pir1
thi2
spt4
hsp150
YPL183C
YKL050C
Known Interactions
Predicted Interactions
bug1
trm7
SS-SL interactions were color coded based on the
biological process.
35
trm82
pus7
dus3
ncl1
smm1/dus2
dus1
trm8
Known Interactions
Predicted Interactions
36
tsl1
rvs167
fps1
hsp82
tps1
rvs161
gpd1
tps3
npt1
rim15
Known Interactions
ypt6
ric1
Predicted Interactions
37
Future Work

• Train the model on Saccharomyces Synthetic
  Genetic Interactions data and test on
  Caenorhabditis elegans data.
• Quantify the role of graph centrality properties
  in the prediction task.
• Identify misclassified gene pairs and understand
  where our methods go wrong.

38
Acknowledgements

• Shan Zhao (University of Rochester)
• Dr. Animesh Ray
• Dr. Alpan Raval
• NSF

39
References

• Ye P, Peyser BD, Pan X, Boeke JD, Spencer FA and
  Bader JS. Gene function prediction from congruent
  synthetic lethal interactions in yeast. Molecular
  Systems Biology 1(Nov 22), 2005 .
• Newman M. A measure of betweenness centrality
  based on random walks. Social Networks, 27
  (1)3954, 2005.
• Wong SL. et al. Combining biological networks to
  predict genetic interactions. Proc. Natl. Acad.
  Sci. USA, 101, 15682-15687, 2004.
• T. Reguly et al. Comprehensive curation and
  analysis of global interaction networks in
  Saccharomyces cerevisiae. Journal of Biology,
  511, 2006.

40
In Pursuit of Modularity

• Sri R Paladugu
• May 08, 2007

41
Modular structure
42
Where are the modules?
43
Algorithm

• 1. Choose a centrality measure (e.g., Edge
  Betweenness, Edge Information Centrality)
• 2. Calculate the centrality score for each of the
  edge
• 3. Remove the edge highest centrality measure
• 4. Perform an analysis of the components
  generated
• 5. Repeat steps 2- 4 until the desired number of
  components are generated

S. Fortunato, V. Latora, and M. Marchiori,
cond-mat/042522 (2004)
44
Modular structure
0.11
0.11
0.11
0.02
0.11
0.01
0.11
0.21
0.61
0.21
0.51
0.12
0.11
0.12
0.01
0.11
0.11
0.11
0.41
0.11
0.11
0.11
0.11
0.11
0.11
0.11
0.11
0.31
0.41
0.01
0.11
0.11
0.01
0.01
0.01
0.01