Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS...

49
Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110

Transcript of Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS...

Page 1: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Applications of Game Theory in the Computational Biology Domain

Richard Pelikan

April 16, 2008

CS 3110

Page 2: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Overview

• The evolution of populations• Understanding mechanisms for disease

and regulatory processes– Models of cancer development– Protein and drug interactions, resource

competition

• Many biological processes can be tied to game theory

Page 3: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Evolution

• Difficult process to describe

• Game theory seen as a way of formally modeling natural selection

Page 4: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Evolutionary Game Theory

• Evolution revolves around a fitness function– Fitness function is often unknown– Frequency based, success is measured

primitively by number present.– Strategies exist because of this function– Difficult to define the entire game with just the

strategy.

Page 5: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Prisoner’s Dilemma

• Earlier in the course, we knew just about everything about the game

• But we are so lucky to know this information!

Cooperate Defect

Cooperate 3/3 0/5Defect 5/0 1/1P

riso

ner

A

Prisoner B

Page 6: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Crocodile’s Dilemma

• V: The value of a resource• C: The cost to fight for a resource, C > V >0

• Negative payoff results in death– But who defines V and C? These variables are unclear for real-

life competitions.

Share Fight

Share / 0 / V

Fight V / 0 /Cro

cod

ile

A

Crocodile B

2

V

2

V

2

CV 2

CV

Page 7: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Population’s Dilemma

• Population members play against each other

• Natural selection favors the better strategists at the game

• Key: strategies are really genetically encoded and do not change

Page 8: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Evolutionary algorithm

• 1) Obtain strategy (at birth)• 2) Play strategy against environmental

opponents. • 3) Evaluate fitness based on value obtained

through strategy• 4) Convert fitness to replication, preserving the

phenotype

• The genetic code of a player can’t change, but their offspring can have mutated genes (and therefore a different strategy).

Page 9: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Population’s Dilemma

• Consider 2 scenarios from crocodile’s dilemma:– A population of purely aggressive crocodiles– A population of purely docile crocodiles

• In both scenarios, a mutation results in an “invasion” of better strategists.

Page 10: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Evolutionarily Stable Strategy (ESS)

• An ESS is a strategy used by a population of players

• Once established, it is not overtaken by rare (or “mutant”) strategies

• These are similar but not equivalent to Nash equilibria

Page 11: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Formal Definition of ESS

• Let S be an evolutionary strategy and T be any alternative strategy. S is an ESS if either of these conditions hold:

• Payoff(S,S) > Payoff(T,S) or• Payoff(S,S) = Payoff(T,S) and

Payoff(S,T) > Payoff(T,T)

• T is a neutral strategy against S, but S always maintains an advantage over T.

Page 12: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Difference between ESS and Nash

• In a Nash equilibrium, – Players know the structure of the game and

the potential strategies of opponents.

• In an ESS,– Strategies are not exhaustively defined– Payoffs are uncertain– Strategies can’t change– Everyone adopts the same strategy

Page 13: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Current applications of ESS to evolutionary theory

• Competition can, in general, be modeled as a search for an ESS

• ES strategies used to explain altruism, animal conflict, market competition, etc.

• Modeling evolution entirely through EES is hard.– On the smaller scale of cell populations, it’s easier to

see the practical applications.

Page 14: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Mechanisms of Disease

• In an organism, cells compete for various resources in their environment.

• Mutations occasionally occur in cell division due to various reasons

• Cancer is a disease where mutated (tumor) cells oust normal cells in a local population

Page 15: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Applied Game Theory for Cancer Therapeutics

• Paper:– Gatenby and Vincent, Application of quantitative models from

population biology and evolutionary game theory to tumor therapeutic strategies, Mol. Cancer Therapy, 2003; 2:919-927

• Claim: To effectively treat cancer, all system dynamics responsible for the tumor invasion must be controlled

• The problems:– Heterogeneity of cancer (i.e. different strategies)– Unfeasability of controlling all system dynamics

Page 16: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Modeling competition between tumor and normal cells

• Assume tumor and normal cells are players in a game

• Create equations which define a competition between normal and a certain type of tumor cells

• These equations incorporate system dynamics variables which can favor either normal or tumor cells

Page 17: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Lotka-Volterra Equations

• Used to model population competition

• Parameters: – x: number of prey (normal cells)– y: number of predators (tumor cells)– : parameters representing interaction btwn

species, open to design by user of model– Equations represent population growth rates over time

)( yaxdt

dx )( xydt

dy

,,,

Page 18: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Lotka-Volterra Equations

• Used to model population competition

• Basically,Rate of growth = # in population * (environmental help to population – rate of destruction by opponent)

• Parameters: – x: number of prey (normal cells)– y: number of predators (tumor cells)– : parameters representing interaction btwn species,

open to design by user of model– Equations represent population growth rates over time

,,,

Page 19: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

In the tumor vs. normal setting

• Lotka-Volterra equations formed as follows:

• If the populations play a pair of strategies, the possible outcomes at the stable state (where dx/dt = dy/dt = 0) are:

– x, y = 0• Trivial, non-relevant result

– x = kN, y = 0

• All normal cells, tumor completely recessed

– x = (kN - βkT)/(1 - βδ), y = (kT - δkN)/(1 - βδ)

• Normal and tumor cells living in equilibrium (benign tumor)

– x=0, y = kT

• All tumor cells, invasive cancer

Nk

yxx

dt

dx 1

Tk

xyy

dt

dy 1

Page 20: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Finding Equilibria

Recession Benign Invasive

Page 21: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Defining the multi-strategy case

• Until now, the tumor population had a constant strategy (mutation requires a different set of parameters)

• The new question is, where can the equilibria be when the strategy space is exhausted?

• In practice, tumor cells from many different populations are already present; can the progress be reversed?

Page 22: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Heterogeneity of Cancer

• Parameter changes can affect the equilibria reached. This suggests an easy cure for cancer, just by changing parameters.

• In reality, the tumor population mutates quickly and changes strategy, making it independent from the previous system of equations

Page 23: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Heterogeneity of Cancer

• Basic idea: Assume n different populations of tumor cells can arise– Each population gets its own fitness function (i.e. own set of

Lotka-Volterra functions)

• Parameters:– αi: maximum rate of proliferation for ith population– ui: strategy of ith population– β(ui,uj): competitive effect of ui versus uj

– k(ui): maximum size of ith population

)( Nu,iii HNN

n

j jjii

iii Nuu

ukH

1),(

)()( Nu,

Page 24: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Tumor Evolution

• A strategy evolves according to:

• σi= chance for mutation in ith population

• v = auxillary variable over strategy space

• The strategy for normal cells has σi= 0

iuvii v

NuHu

|

),(

Page 25: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Tumor Evolution vs. Normal

• Normal cells don’t evolve (bottom) and continue to die, being pressured by tumor cells (top)

• The tumor cells appear to reach a steady state. Can they be treated at this point with a cell-specific drug?

Page 26: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Augmenting system with specific drug targets

• Extend fitness functions with a Gaussian, drug-specific term

• Parameters:– dh: dosage of drug h– σh: variance in effectiveness of drug h– : strategy (cell type) weakest against drug h

2

1 2exp),(

)()(

hh

n

j jjii

iii

uvdNuu

ukH

Nu,

u

Page 27: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

A Bleak Outcome

• Cell-specific treatment is effective at first, but evolving cells become resistant and invade

Page 28: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

In Summary

• Population fitness functions can be designed using the Lotka-Volterra functions

• Paper claims:– Targeted drug therapies alone won’t work– Trajectories of tumor evolution need to be

changed by systemic, outside factors– Angiogenesis inhibitors, TNF, etc.

Page 29: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Following this,

• Lots of interest regarding drug interactions and how they affect cells

• Usually dependent on how much of, or for how long, a drug molecule is in contact (binds) with a cell structure

• Computational approaches can be used to conduct drug simulations in silico

– Paper: Perez-Breva et. al, Game theoretic algorithms for protein-DNA binding, NIPS 2006

Page 30: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Game Theory in Molecular Biology

• Binding game– Inputs:

• Protein classes (players)• Sites (other set of players) which compete and coordinate for

proteins

– Players decide how much protein is allocated to each site, based on:

• How occupied sites are• Availability of proteins• Chemical equilibrium (sites have affinities for particular

proteins up to a certain constant)

– Output: allocation of proteins to sites

Page 31: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Formal definition of binding game

• fj = concentration of protein i

• pij= amount of protein i allocated to site j

• sij = amount of protein I bound to site j

• Eij = affinity of protein i to site j

• Utility of protein assignment is defined as:

)()1()('

, ij i

ijijijii pHsEpspu

Page 32: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Formal definition of binding game

• fj = concentration of protein i

• pij= amount of protein i allocated to site j

• sij = amount protein i bound to site j

• Eij = affinity of protein i to site j

• Utility of protein assignment to set of sites s:

)()1()('

, ij i

ijijijii pHsEpspu

Amount of time that site j is available for

protein i

Controls the mixing proportions of bound proteins

Page 33: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Formal definition of binding game

• fj = concentration of protein i

• pij= amount of protein i allocated to site j

• sij = amount of protein i bound to site j

• Eij = affinity of protein i to site j

• Kij = chemical equilibrium constant between protein i and site j

• Utility of site player j binding to a set of proteins p

i iijijiijijijj

sj ssfpKspsu )1)((),(

'

Amount of protein i bound to site j

Proportion of protein i that’s just

floating around

Page 34: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Finding the equilibrium

• It turns out, finding the equilibrium between protein and site player’s utilities reduces to finding site occupancies αj

• The equilibrium condition is expressed in terms of just αj, so that overall occupancy is determined by which proteins are currently bound elsewhere

i

ijj as )(

Page 35: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Algorithm

• Start with all sites empty (αj =0; j = 1…n)

• Repeat until convergence:– pick one site – maximize its occupancy time in the context of

available proteins and sites

• algorithm is monotone and guaranteed to find equilibrium

Page 36: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2 Switch Sites

Virus genes are embedded in a cell’s DNA

Page 37: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

RNA

Switch Sites

During normal function, cell requires RNA to transcribe genes to proteins

Page 38: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

RNA

Switch Sites

RNA unknowingly transcribes viral genes, producing virus proteins

Page 39: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

RNA

Virus proteins are produced by first gene

Page 40: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

RNA

Virus proteins bind to available sites

Page 41: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

Virus proteins prevent transcription of later genes, keeping virus dormant.

Page 42: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

Virus proteins bind and block transcription

Page 43: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

Stress changes the affinities of binding sites

Page 44: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

RNA is free to bind to later genes

Page 45: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

“clearing” virus proteins are produced

Page 46: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

Clearing proteins release viral proteins from the switches

Page 47: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Simulation model for λ-phage virus

gene CRogene CI2

Replicated virus

At this stage, cell breaks open and releases the replicated virus

Page 48: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Validation of simulated model

• Increasing concentration at different receptors leads to different equilibrium

• validated using studied concentrations in literature (shaded region)

Page 49: Applications of Game Theory in the Computational Biology Domain Richard Pelikan April 16, 2008 CS 3110.

Summary

• Many potential applications of game theory to biological domain

• Most methods include intuitive and simplistic reasoning about how biological entities compete

• Despite simplicity, the models often explain initial beliefs about behavior