Identifying Candidate Pathways to Explain Phenotypes in

Genome-Wide Mutant Screens

Mark CravenDepartment of Biostatistics & Medical Informatics

University of Wisconsin-MadisonU.S.A.

joint work with: Deborah Chasman, Paul Ahlquist, Brandi Gancarz, Linhui Hao

Viruses take advantage of host cell genes

Figure from: C. E. Samuel, Journal of Biological Chemistry 285, 2010.

Genome-wide mutant screens

HOS1 SPE3

MED1 YPR071W NOT5

LTP1

HOS1 SPE3

MED1 YPR071W NOT5

LTP1

HOS1 SPE3

MED1 YPR071W NOT5

LTP1

HOS1 SPE3

MED1 YPR071W NOT5

LTP1

mutant phenotype

Example: determining which host genes affect viral replication [Kushner et al., PNAS 2003; Gancarz et al., PLoS One 2011]

Genome-wide mutant screensThe output of such screens are sets of genes that either inhibit or stimulate viral processes

Characterizing virus-host interactionsgiven such interaction data, we want to

• identify pathways that provide consistent explanations for the genome-wide measurements

• predict the interfaces to the virus

before inference after inference

Some interactions are deemed not consistent with the measurements

Directions and signs of interactions are specified

Interfaces to the virus are hypothesized

Integer programming approach

1. Collect candidate pathways for each “hit”

2. Use IP to identify a globally consistent subnetwork

Collecting candidate pathways for a hit

generate candidate pathways, up to a specified length, that link a hit to the virus

Variables in integer programming approach

€

σ p

€

x2, s2, d2€

x1

€

x3, s3

Variable Descriptionxe is edge e active?

se sign of edge e (up- or down-regulating)

de direction of edge e

tg phenotype (effect) of knocking out gene g

σp is pathway p active?€

tg

Constraints in integer programming approach

€

∀n ∈ hits σ pp:n∈ nodes ( p )

∑ ⎛

⎝ ⎜ ⎜

⎞

⎠ ⎟ ⎟> 0

€

σ p

€

x2, s2, d2€

x1

€

x3, s3

€

tg

all significant measurements (hits) are explained by at least one pathway

Constraints in integer programming approach

€

σ p

€

x2, s2, d2€

x1

€

x3, s3

€

∀p σ p = 0 ∨ consistent (p)( )

€

conσiσtent(p) =

∧e∈ edges ( p ){ i, j}= nodes (e )

xe =1 ∧ de = dir(p,e) ∧ se = tit j( )

€

tg

all active pathways are consistent, with edges directed toward the interface

Current objective function ininteger programming approach

minimize the number of interfaces

€

min I σ p > 0p:n∈ nodes ( p )

∑ ⎛

⎝ ⎜ ⎜

⎞

⎠ ⎟ ⎟

n∈ interfaces∑

€

σ p

€

x2, s2, d2€

x1

€

x3, s3

€

tg

before inference

after inference

How to evaluate the IP approach?

hold a measurement asidesee if we can correctly predict it using inferred networks

?

Baseline predictors

?

neighbor voting further neighbor voting

?

predict

1 neighbor votes

2 neighbors vote

predict

3 neighbors vote

2 neighbors vote

Baseline predictors

consistency neighbor voting

predict

1 neighbor votes

2 neighbors vote

?

This gene votes because it has a repressing interaction with query gene

Markov network approach

• variables are the same as in the IP

• potential functions represent• the constraints• uncertainty associated with specific interactions• the preference for a small number of interfaces

• inference done using Gibbs sampling

Predictive Accuracy (BMV)

Predictive Accuracy (FHV)

Future work

• taking into account additional sources of information• quantitative values from assays• genetic interactions• interactions automatically extracted from the scientific literature

• adapting approach to RNAi screens in mammalian cells• more genes• lower density of known interactions• more uncertainty in measurements

• devising methods that use these models to determine which follow-up experiments would be most informative