The Ping-Pong Algorithm

Kutalik, Beckmann, Bergmann NBT 08

Some slides from Sven Bergmann http://serverdgm.unil.ch/bergmann

ABDBM © Ron Shamir

The NCI-60 drug response and GE db Scherf et al. Nat Gen 00

ABDBM © Ron Shamir

R E Rij:

Response of cell line j to drug i

Eij: Exp of gene i in

untreated cell line j

Clustering the cell lines

ABDBM © Ron Shamir

Based on E

Based on R

Very different pictures. Can we reconciliate?

Approaches • ISA(E) : get transcription modules. Ignores R. • ISA (E·RT) : compute drug-gene correlation across

the cell lines, generate drug-gene modules by compiling info across all cell lines. Misses signals on few cell lines

• ISA(E)&ISA(T): Suppose modules are represented as row/col weights. Get modules {(di,ci)} in D and {(gi,c*i)}. Select highly correlated module pairs: corr(ci,c*j)>α “late integration”

• Cluster (E·RT) : Clusters the genes using HC, Set cutoff to get desired no. of clusters. Cluster the drugs too. Match gene and drug clusters by mean bic signal. “late integration”

• … ABDBM © Ron Shamir

Generalizing ISA

Co-modules Gene-modules Drug-modules

C3

F4

C4

F3

G3

G4

[AGF]

[AGF]

[BFC]

[BFC]

C5 D3

F6

C6 D4

F5

[BFC]

[BFC]

[CDF]

[CDF]

C1 D1

F2

C2 D2

F1

G1

G2

[AGF] [CDF] [BFC]

[AGF] [CDF] [BFC]

ABDBM © Ron Shamir

ABDBM © Ron Shamir

G

D

C

G4

D4

C3 C4

Drug-modules

Gene-modules

C5 C6

Modules and Co-modules

D3

G3

Co-modules

G2

G1

C1 C2

D1

D2

ABDBM © Ron Shamir

E C G R D MED

Iteratively refine genes, cell-lines and drugs to get co-modules

The Ping-Pong algorithm!

1

2

3

4

ABDBM © Ron Shamir

Results on NCI-60 (859 m

odules) ABDBM © Ron Shamir

Co-modules have predictive power for drug-gene associations

ABDBM © Ron Shamir

Co-modules analysis provides biological focus through data integration

ABDBM © Ron Shamir

Dots: GO BP categories. X: ISA module

score. Y: PPA score

• Analysis of large-scale expression data bears great potential to understand global transcription programs and their evolution

• Innovative analysis tools needed to extract information from such data

• ISA & Ping-Pong Algorithms: – decomposes data into “transcription

modules” – integrates external information – allows for interspecies comparative analysis

Sven’s Take-home Messages

ABDBM © Ron Shamir