NetBioSIG2014 at ISMB in Boston, MA, USA on July 11, 2014

1. TimeXNet: Identifying active
gene sub-networks using time-
course gene expression profiles
Ashwini Patil
Institute of Medical Science
University of Tokyo
NetBio SIG, ISMB 2014

2. Goal
• Comprehensive computational analysis of the innate
immune response
Mouse Interaction network
103218 protein-protein, protein-DNA,
post-translational modifications
Time-course gene expression
RNA-seq expression levels in dendritic
cells on LPS stimulus at 8 time points

3. Innate immune system
Kawai & Akira, Nat. Immunology, 2010

4. Method - TimeXNet
Partition differentially expressed
genes into 3 time-based groups
Identify most probable paths in the
network connecting the three groups
Patil et al., PLOS Comp. Biol., 2013

5. Minimum cost flow optimization
• ResponseNet
• Identifies paths between two groups of genes (genetic hits and differentially
expressed genes in yeast)
- Yeger-Lotem et l., Nat. Genetics, 2009

6. TimeXNet methodology
• Edge cost: inversely proportional to edge reliability
• Edge capacity: directly proportional to
• Fold change in expression of adjacent gene(s)
• Absolute tag counts of adjacent gene(s)
• Objective function
Minimize cost of flow through the network from T1 to
T3 genes
• Constraint
Flow must pass through intermediate nodes (T2 genes)
Most probable paths connecting T1->T2->T3 genes
2681 scored interactions among 1225 proteins

7. Candidate genes
Early genes
(0.5-1 hour)
Intermediate genes
(2-4 hours)
Late genes
(6-8 hours)
Genes with no change
in expression
Gene Flow Gene Flow Gene Flow Gene Flow
Jun 13.68 Socs3 85.85 Cxcl10 10.91 Stat1 8.74
Fos 10.34 Nfκb1 76.87 Ddx58 9.33 Mapk8 8.72
Il1b 9.86 Jak2 54.44 Stat2 8.65 Irf5 7.60
Tnf 9.36 Src 38.30 Atf3 8.29 Adcy5 7.43
Cxcl2 7.59 Pik3r5 27.86 Isg15 8.15 Mapk1 7.40
Il1a 7.40 Rela 23.35 Irf7 7.30 Sp1 7.37
Akt1 6.43 Stat5a 20.40 Nos2 6.91 Stat6 7.17
Atf4 5.49 Met 18.94 Ifnar2 5.20 Sp3 7.13

8. Candidate networks
Gm13305
Ifnar1
Il12rb1
Il13ra2
Ifngr2
Gm2002
Il13ra1
Il11ra1
Stat5b
Stat4
Irs3
Irs4
Lifr
Jak2
Cxcr4
Stat6
Il9r
Nck1
Il20ra
Il22ra11
Il22ra2
Il7r
Il2rg
Il4ra2
Il28raa Il2ra
Il6ra
Ifnar2
Il21r
Stat2
Il3ra
Crlf2
Ifngr1
Il15ra
Ddx58
Fos
Rela
Nfkb1
Stat5a
Bcl10
Il10rb
JunStat1
Sp3
CR974586.2
Socs3
Foxo3
CT868723.4 Csf2rb
Gfi1b
CT868723.4CT868723.4CT868723.4
Csf2rb2
Cntfr
bb Creb1
• Socs3
• Suppressor of cytokine signaling 3
• Induced by Nfkb and inhibits a large number of proteins, specifically the
interleukin receptors

9. Candidate networks

10. Method evaluation
• Comparison with experimentally identified regulators
• Amit et al., Science 2009: 49.6% previously unknown genes identified
• Chevrier et al., Cell 2011: 69.8% regulators (novel and known) and 54.9% TLR target
genes identified
• Overlap with KEGG pathways
• Directed paths of 3 to 7 edges identified in 13 KEGG pathways
• Jak-STAT signaling pathway, Chemokine signaling pathway, Toll-like receptor pathway,
MAPK signaling pathway

11. Noise in the interaction network

12. Comparison with other methods
Method
Experimentally confirmed
regulators (3 datasets)
KEGG Pathways
with predicted
paths (max length)
Execution
time (4 CPUs,
2.4Ghz, 12Gb
RAM)
Prior knowledge
required
Time-
course
data
TimeXNet 49.6%1 69.8%2 54.9%3 13 (7 edges) 3 min None Yes
ResponseNet* 39.2%1 53.5%2 39.2%3 0 (3 edges) 1 min None No
SDREM 12.0%1 32.6%2 11.8%3 2 (4 edges) ~10 days Initial genes Yes
1 Regulatory genes from Amit et al., Science, 2009
2 Regulatory genes from Chevrier et al., Cell, 2011
3 Target genes from Chevrier et al., Cell, 2011
*Local implementation using GLPK

13. Yeast osmotic stress response
• Time-course gene expression (min) in yeast on hyperosmotic stress
- Romero-Santacreu et al., RNA 2009
• Previously used to evaluate SDREM and ResponseNet
- Gitter et al., Genome Research 2013
• Genes with 1.5 fold change in expression
• Initial response genes: 2-4 min
• Intermediate regulators: 6-8 min
• Final effectors: 10-15 min

14. Predicted osmotic stress response network
• 2-4 min
• 6-8 min
• 10-15 min
• Predicted
Method
Gold
Standard* TFs* Hog1 Runtime
TimeXNet 19 5 Yes 5 sec
SDREM* 10 4 Yes -
ResponseNet* 3 2 No -*Taken from Gitter et al., Genome Research 2013

15. Circadian regulation of metabolism in mouse liver cells
- Unpublished
• Paths connecting genes showing rhythmic patterns of expression in 24 hours
• Network predicted by TimeXNet contains Sphk2, Pld1, Pld2, Glud1

16. TimeXNet Availability: http://timexnet.hgc.jp/

17. • Input
• 3 sets of genes with
scores
• Weighted interaction
network
• Parameters gamma1 and
2
• Location of glpsol
executable from the GLPK
• Directory where results
will be storedCytoscape
Running TimeXNet
• Standalone application
• Command line version
• Iterative command line version to
identify optimal parameters
Patil & Nakai, under review

18. Conclusion
• TimeXNet: A method to predict active gene sub-networks using time-
course gene expression profiles
• Advantages
• Accurate and fast
• Independent of biological system: Innate immune response, circadian regulation of
metabolism in mouse, yeast osmotic stress response
• Amenable to incorporation of other time-course data types: phosphorylation levels,
protein levels, epigenetic information
• Issues to be addressed
• Allowing path prediction between more than 3 groups of genes while maintaining
speed and accuracy
• Incorporating other forms of time-course information
• Enhancements: Automatic install of GLPK, allowing users to enter non-numeric gene
IDs
Patil et al., PLOS Comp. Biol., 2013

19. Acknowledgements
• Innate immune response
• Prof. Kenta Nakai - University of Tokyo
• Dr. Yutaro Kumagai – Osaka University
• Dr. Kuo-ching Liang – University of Tokyo
• Prof. Yutaka Suzuki – University of Tokyo
• Dr. Tomonao Inobe – Toyama University
• Yeast osmotic stress response
• Dr. Anthony Gitter – Microsoft Research
• Circadian regulation of metabolism
• Dr. Craig Jolley – RIKEN Center for
Developmental Biology, Kobe
• Funding
• Japan Society for the Promotion of
Science (JSPS) FIRST Program
• JSPS Grant-in-Aid for Young Scientists
• Takeda Science Foundation (with Dr.
Tomonao Inobe)
• Computational resources
• Supercomputer at the Human Genome
Center, Institute of Medical Science,
University of Tokyo

21. Edge Capacities
For edges between the auxiliary source, S, and the initial response genes GT1,
2 1log
/ /
imax i
Si T
imax ii i
fc e
C i G
fc N e N
   
 
(3)
For edges connected to the intermediate regulators GT2,
2 2 2log ,
/ /
imax i
ij T T
imax ii i
fc e
C i G j G
fc N e N
    
 
(4)
2 2
2
log log
/ // /
,
2
jmax jimax i
imax jmaxi ji ji j
ij T
fc efc e
fc N fc Ne N e N
C i j G
   
     
          
  
(5)
For edges between the late effectors, GT3, and the auxiliary sink T,
2 3log
/ /
imax i
iT T
imax ii i
fc e
C i G
fc N e N
   
 
(6)
2 2
2
log log
/ // /
,
2
jmax jimax i
imax jmaxi ji ji j
ij T
fc efc e
fc N fc Ne N e N
C i j G
   
     
          
  
(5
For edges between the late effectors, GT3, and the auxiliary sink T,
2 3log
/ /
imax i
iT T
imax ii i
fc e
C i G
fc N e N
   
 
(6
For edges between the auxiliary source, S, and the initial response genes GT1,
2 1log
/ /
imax i
Si T
imax ii i
fc e
C i G
fc N e N
   
 
(3)
For edges connected to the intermediate regulators GT2,
2 2 2log ,
/ /
imax i
ij T T
imax ii i
fc e
C i G j G
fc N e N
    
 
(4)
2 2
2
log log
/ // /
,
2
jmax jimax i
imax jmaxi ji ji j
ij T
fc efc e
fc N fc Ne N e N
C i j G
   
     
          
  
(5)
For edges between the late effectors, GT3, and the auxiliary sink T,
For edges connected to the intermediate regulators GT2,
• Graph G = (V, E) with E edges and V
nodes (containing S – auxiliary
source, T – auxiliary sink)
• fc = fold change
• 𝑒 = average expression level at all
time points
• N = number of genes with expression
values
• S = auxiliary source node
• T = auxiliary sink node
• GT1, GT2, GT3 = genes having
maximal fold change at times T1, T2
and T3
For all other edges, not connected to the intermediate regulators or the auxiliary source and s
21 ,ij TC i j S G T  

22. Edge costs
1Si Si Tw C i G   (8)
2ij ij Tw C i G   (9)
3iT iT Tw C i G   (10)
  2,ij ij Tw f s i j S G T   , as per equation (2)
The edge costs were calculated as:
Where ()f = scaling function
 likelihood ratio , HitPredictijs i j   ; 0.163 999ijs 
 999 , Innatedb, KEGGijs i j  
 , TRANSFACijs Transfacscore i j   ; 1 6ijs 
3iT iT Tw C i G  
  2,ij ij Tw f s i j S G T   , as per equation (2)
The edge costs were calculated as:
 10log ,ij ijA w i j E   
2ij ij Tw C i G  
3iT iT Tw C i G  
  2,ij ij Tw f s i j S G T   , as per equation (2)
The edge costs were calculated as:
 log ,A w i j E   
2ij ij Tw C i G  
3iT iT Tw C i G  
  2,ij ij Tw f s i j S G T   , as per equation (2)
The edge costs were calculated as:
 10log ,ij ijA w i j E   

23. Optimization problem