Top Banner
Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods) Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods) Prof. Michael Mascagni Department of Computer Science Department of Mathematics Department of Scientific Computing Graduate Program in Molecular Biophysics Florida State University, Tallahassee, FL 32306 USA E-mail: [email protected] or [email protected] URL: http://www.cs.fsu.edu/mascagni In collaboration with Drs. Marcia O. Fenley, Nikolai Simonov and Alexander Silalahi, and Messrs. Robert Harris, Travis Mackoy, and James McClain Research supported by ARO, DOE, NATO, and NSF
152

Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Mar 23, 2018

Download

Documents

trinhthuy
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Novel Stochastic Methods in BiochemicalElectrostatics: (Stochastic Methods for PDEs

Can Beat Deterministic Methods)

Prof. Michael Mascagni

Department of Computer ScienceDepartment of Mathematics

Department of Scientific ComputingGraduate Program in Molecular Biophysics

Florida State University, Tallahassee, FL 32306 USAE-mail: [email protected] or [email protected]

URL: http://www.cs.fsu.edu/∼mascagniIn collaboration with Drs. Marcia O. Fenley, Nikolai Simonov and Alexander Silalahi, and

Messrs. Robert Harris, Travis Mackoy, and James McClain

Research supported by ARO, DOE, NATO, and NSF

Page 2: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

IntroductionMotivation

Mathematical ModelElectrostatic Potential and EnergyThe Feynman-Kac Formula

Fast Exit Point Calculations‘Walk-on-Spheres’ AlgorithmWalk-in-SubdomainsMonte Carlo Treatment of Boundary Conditions

Monte Carlo EstimatesMonte Carlo EstimatesComputational GeometryCorrelated and Uncorrelated Sampling

Computational Results

Conclusions and Future Work

Page 3: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

IntroductionMotivation

Mathematical ModelElectrostatic Potential and EnergyThe Feynman-Kac Formula

Fast Exit Point Calculations‘Walk-on-Spheres’ AlgorithmWalk-in-SubdomainsMonte Carlo Treatment of Boundary Conditions

Monte Carlo EstimatesMonte Carlo EstimatesComputational GeometryCorrelated and Uncorrelated Sampling

Computational Results

Conclusions and Future Work

Page 4: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

IntroductionMotivation

Mathematical ModelElectrostatic Potential and EnergyThe Feynman-Kac Formula

Fast Exit Point Calculations‘Walk-on-Spheres’ AlgorithmWalk-in-SubdomainsMonte Carlo Treatment of Boundary Conditions

Monte Carlo EstimatesMonte Carlo EstimatesComputational GeometryCorrelated and Uncorrelated Sampling

Computational Results

Conclusions and Future Work

Page 5: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

IntroductionMotivation

Mathematical ModelElectrostatic Potential and EnergyThe Feynman-Kac Formula

Fast Exit Point Calculations‘Walk-on-Spheres’ AlgorithmWalk-in-SubdomainsMonte Carlo Treatment of Boundary Conditions

Monte Carlo EstimatesMonte Carlo EstimatesComputational GeometryCorrelated and Uncorrelated Sampling

Computational Results

Conclusions and Future Work

Page 6: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

IntroductionMotivation

Mathematical ModelElectrostatic Potential and EnergyThe Feynman-Kac Formula

Fast Exit Point Calculations‘Walk-on-Spheres’ AlgorithmWalk-in-SubdomainsMonte Carlo Treatment of Boundary Conditions

Monte Carlo EstimatesMonte Carlo EstimatesComputational GeometryCorrelated and Uncorrelated Sampling

Computational Results

Conclusions and Future Work

Page 7: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

IntroductionMotivation

Mathematical ModelElectrostatic Potential and EnergyThe Feynman-Kac Formula

Fast Exit Point Calculations‘Walk-on-Spheres’ AlgorithmWalk-in-SubdomainsMonte Carlo Treatment of Boundary Conditions

Monte Carlo EstimatesMonte Carlo EstimatesComputational GeometryCorrelated and Uncorrelated Sampling

Computational Results

Conclusions and Future Work

Page 8: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 9: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 10: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 11: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 12: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 13: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 14: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 15: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

IntroductionMotivation

I Experimental Data: Folding, stability & binding behavior ofbiomolecules can be modulated by changes in salt concentration

I Physical Model: Implicit solvent-based Poisson-Boltzmann modelcan provide accurate predictions of salt dependent behavior ofbiomolecules

I Mathematical Model: Elliptic boundary-value problemsSpecific Problems

I Electrostatic free energy for linear case: only finite number ofelectrostatic potential point values

I Dependence of energy on geometry: needs accurate treatmentI Singularities in solution: have to be taken into account

analyticallyI Behavior at infinity: must be exactly enforcedI Functional dependence on salt concentration: needs accurate

estimate

Page 16: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 17: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 18: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 19: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 20: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 21: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 22: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 23: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Introduction

Monte Carlo Methods: PropertiesI Monte Carlo method basics: I =

∫ 10 f (x)dx

1. Random/stochastic process: xi ∼ U[0, 1)2. Random variable: f (xi ) where E [f (xi )] = I and E [f 2(xi )] <∞

I Monte Carlo methods for solving Poisson and linearizedPoisson-Boltzmann equations (PBEs)

1. Analytical treatment of geometry, singularities, behavior at infinity2. Capability to compute point values of solution (energies) and its

spatial derivatives (forces)3. New methods for the flux boundary conditions (exact integral

formulation)4. Simultaneous correlated computation of values at different salt

concentrations

Page 24: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Motivation

Mathematical Model: Molecular Geometry

Figure: Biomolecule with dielectric εi and region region Gi is in solution withdielectric εe and region Ge. On the boundary of the biomolecule, electrostaticpotential and normal component of dielectric displacement continue

Page 25: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Partial Differential EquationsI Poisson equation for the electrostatic potential, Φi , and point

charges, Qm, inside a molecule (in CGS units):

εi ∆Φi (x) + 4πM∑

m=1

Qmδ(x − x (m)) = 0 , x ∈ Gi

I For 1-1 salt (such as NaCl) Poisson-Boltzmann equation (PBE):

∆Φe(x)− κ2 sinh(Φe(x)) = 0 , x ∈ Ge ,

but we only consider the linearized PBE:

∆Φe(x)− κ2Φe(x) = 0 , x ∈ Ge

I For one-surface model: continuity condition on the dielectricboundary

Φi = Φe , εi∂Φi

∂n(y)= εe

∂Φe

∂n(y), y ∈ Γ

Page 26: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Partial Differential EquationsI Poisson equation for the electrostatic potential, Φi , and point

charges, Qm, inside a molecule (in CGS units):

εi ∆Φi (x) + 4πM∑

m=1

Qmδ(x − x (m)) = 0 , x ∈ Gi

I For 1-1 salt (such as NaCl) Poisson-Boltzmann equation (PBE):

∆Φe(x)− κ2 sinh(Φe(x)) = 0 , x ∈ Ge ,

but we only consider the linearized PBE:

∆Φe(x)− κ2Φe(x) = 0 , x ∈ Ge

I For one-surface model: continuity condition on the dielectricboundary

Φi = Φe , εi∂Φi

∂n(y)= εe

∂Φe

∂n(y), y ∈ Γ

Page 27: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Partial Differential EquationsI Poisson equation for the electrostatic potential, Φi , and point

charges, Qm, inside a molecule (in CGS units):

εi ∆Φi (x) + 4πM∑

m=1

Qmδ(x − x (m)) = 0 , x ∈ Gi

I For 1-1 salt (such as NaCl) Poisson-Boltzmann equation (PBE):

∆Φe(x)− κ2 sinh(Φe(x)) = 0 , x ∈ Ge ,

but we only consider the linearized PBE:

∆Φe(x)− κ2Φe(x) = 0 , x ∈ Ge

I For one-surface model: continuity condition on the dielectricboundary

Φi = Φe , εi∂Φi

∂n(y)= εe

∂Φe

∂n(y), y ∈ Γ

Page 28: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Debye-Hückle Parameter

Dependence on salt in the Debye-Hückle parameter (units as perKirkwood):

κ2 =8πNAe2Cs

εe1000kBT, where

I Cs – concentration of ions (in moles)I NA – Avogadro’s numberI e – elementary protonic chargeI kB – Boltzmann’s constantI εe – dielectric permittivity outside the molecule

Page 29: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Debye-Hückle Parameter

Dependence on salt in the Debye-Hückle parameter (units as perKirkwood):

κ2 =8πNAe2Cs

εe1000kBT, where

I Cs – concentration of ions (in moles)I NA – Avogadro’s numberI e – elementary protonic chargeI kB – Boltzmann’s constantI εe – dielectric permittivity outside the molecule

Page 30: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Debye-Hückle Parameter

Dependence on salt in the Debye-Hückle parameter (units as perKirkwood):

κ2 =8πNAe2Cs

εe1000kBT, where

I Cs – concentration of ions (in moles)I NA – Avogadro’s numberI e – elementary protonic chargeI kB – Boltzmann’s constantI εe – dielectric permittivity outside the molecule

Page 31: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Debye-Hückle Parameter

Dependence on salt in the Debye-Hückle parameter (units as perKirkwood):

κ2 =8πNAe2Cs

εe1000kBT, where

I Cs – concentration of ions (in moles)I NA – Avogadro’s numberI e – elementary protonic chargeI kB – Boltzmann’s constantI εe – dielectric permittivity outside the molecule

Page 32: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Mathematical Model: Debye-Hückle Parameter

Dependence on salt in the Debye-Hückle parameter (units as perKirkwood):

κ2 =8πNAe2Cs

εe1000kBT, where

I Cs – concentration of ions (in moles)I NA – Avogadro’s numberI e – elementary protonic chargeI kB – Boltzmann’s constantI εe – dielectric permittivity outside the molecule

Page 33: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Electrostatic Potential and Energy

Electrostatic Potential and EnergyI Point values of the potential: Φ(x) = Φrf (x) + Φc(x)

Here, singular part of Φ:

Φc(x) =M∑

m=1

Qm

|x − x (m)|

I Reaction field electrostatic free energy of a molecule is linearcombination of point values of the regular part of the electrostaticpotential:

Wrf =12

M∑m=1

Φrf (x (m))Qm ,

I Electrostatic solvation free energy = difference between theenergy for a molecule in solvent with a given salt concentrationand the energy for the same molecule in vacuum:

∆Gelecsolv = Wrf (εi , εe, κ)−Wrf (εi ,1,0)

Page 34: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Electrostatic Potential and Energy

Electrostatic Potential and EnergyI Point values of the potential: Φ(x) = Φrf (x) + Φc(x)

Here, singular part of Φ:

Φc(x) =M∑

m=1

Qm

|x − x (m)|

I Reaction field electrostatic free energy of a molecule is linearcombination of point values of the regular part of the electrostaticpotential:

Wrf =12

M∑m=1

Φrf (x (m))Qm ,

I Electrostatic solvation free energy = difference between theenergy for a molecule in solvent with a given salt concentrationand the energy for the same molecule in vacuum:

∆Gelecsolv = Wrf (εi , εe, κ)−Wrf (εi ,1,0)

Page 35: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

Electrostatic Potential and Energy

Electrostatic Potential and EnergyI Point values of the potential: Φ(x) = Φrf (x) + Φc(x)

Here, singular part of Φ:

Φc(x) =M∑

m=1

Qm

|x − x (m)|

I Reaction field electrostatic free energy of a molecule is linearcombination of point values of the regular part of the electrostaticpotential:

Wrf =12

M∑m=1

Φrf (x (m))Qm ,

I Electrostatic solvation free energy = difference between theenergy for a molecule in solvent with a given salt concentrationand the energy for the same molecule in vacuum:

∆Gelecsolv = Wrf (εi , εe, κ)−Wrf (εi ,1,0)

Page 36: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

The Feynman-Kac Formula

The Feynman-Kac Formula

I Consider the Dirichlet problem for the Poisson equation in thedomain Ω ∈ Rd

−12

∆u(x) = g(x), x ∈ Ω, u(x) = f (x), x ∈ ∂Ω

I If we assume g(x) = 0, then we have the Laplace equation, andthe solution at the point y ∈ Ω is given as the following Brownianmotion expectation:

u(y) = E[f (βy (τ∂Ω))],

where βy (·) is Brownian motion starting at the point y , and τ∂Ω isthe first-passage time of this Brownian motion,i.e. τ∂Ω = inftβy (t) ∈ ∂Ω

Page 37: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

The Feynman-Kac Formula

The Feynman-Kac Formula

I Consider the Dirichlet problem for the Poisson equation in thedomain Ω ∈ Rd

−12

∆u(x) = g(x), x ∈ Ω, u(x) = f (x), x ∈ ∂Ω

I If we assume g(x) = 0, then we have the Laplace equation, andthe solution at the point y ∈ Ω is given as the following Brownianmotion expectation:

u(y) = E[f (βy (τ∂Ω))],

where βy (·) is Brownian motion starting at the point y , and τ∂Ω isthe first-passage time of this Brownian motion,i.e. τ∂Ω = inftβy (t) ∈ ∂Ω

Page 38: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

The Feynman-Kac Formula

The Feynman-Kac FormulaI If we set f (x) = 0 and have g(x) 6= 0, the solution is

u(y) = E[ ∫ τ∂Ω

0g(βy (s)) ds

]I By linear superposition, the solution to Poisson equation is given

probabilistically as

u(y) = E[ ∫ τ∂Ω

0g(βy (s)) ds + f (βy (τ∂Ω))

]I The linearized Poisson-Boltzmann equation is given by

∆u(x)−κ2u(x) = 0, x ∈ Ω, u(x) = f (x), x ∈ ∂Ω, u → 0 as |x | → ∞

and has Wiener integral representation:

u(y) = E[f (βy (τ∂Ω))e−

∫ τ∂Ω0 κ2 ds

]

Page 39: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

The Feynman-Kac Formula

The Feynman-Kac FormulaI If we set f (x) = 0 and have g(x) 6= 0, the solution is

u(y) = E[ ∫ τ∂Ω

0g(βy (s)) ds

]I By linear superposition, the solution to Poisson equation is given

probabilistically as

u(y) = E[ ∫ τ∂Ω

0g(βy (s)) ds + f (βy (τ∂Ω))

]I The linearized Poisson-Boltzmann equation is given by

∆u(x)−κ2u(x) = 0, x ∈ Ω, u(x) = f (x), x ∈ ∂Ω, u → 0 as |x | → ∞

and has Wiener integral representation:

u(y) = E[f (βy (τ∂Ω))e−

∫ τ∂Ω0 κ2 ds

]

Page 40: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Mathematical Model

The Feynman-Kac Formula

The Feynman-Kac FormulaI If we set f (x) = 0 and have g(x) 6= 0, the solution is

u(y) = E[ ∫ τ∂Ω

0g(βy (s)) ds

]I By linear superposition, the solution to Poisson equation is given

probabilistically as

u(y) = E[ ∫ τ∂Ω

0g(βy (s)) ds + f (βy (τ∂Ω))

]I The linearized Poisson-Boltzmann equation is given by

∆u(x)−κ2u(x) = 0, x ∈ Ω, u(x) = f (x), x ∈ ∂Ω, u → 0 as |x | → ∞

and has Wiener integral representation:

u(y) = E[f (βy (τ∂Ω))e−

∫ τ∂Ω0 κ2 ds

]

Page 41: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

‘Walk-on-Spheres’ Algorithm

‘Walk-on-Spheres’ Algorithm

I Walk-on-spheres (WOS) algorithm for general domains with aregular boundary

I Define a Markov chain xi , i = 1,2, . . .I Set x0 = x (m) for some m, xi = xi−1 + diωi , i = 1,2, . . ., where

1. di = d(xi−1) is distance from xi−1 to Γ2. ωi is sequence of independent unit isotropic vectors3. xi is the exit point from the ball, B(xi−1, d(xi−1)), for a Brownian

motion starting at xi−1

I Outside the molecule, on every step, walk-on-spheres terminates

with probability 1− q(κ,di ), where q(κ,di ) =κdi

sinh(κdi )to deal

with LPBE

Page 42: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

‘Walk-on-Spheres’ Algorithm

‘Walk-on-Spheres’ Algorithm

I Walk-on-spheres (WOS) algorithm for general domains with aregular boundary

I Define a Markov chain xi , i = 1,2, . . .I Set x0 = x (m) for some m, xi = xi−1 + diωi , i = 1,2, . . ., where

1. di = d(xi−1) is distance from xi−1 to Γ2. ωi is sequence of independent unit isotropic vectors3. xi is the exit point from the ball, B(xi−1, d(xi−1)), for a Brownian

motion starting at xi−1

I Outside the molecule, on every step, walk-on-spheres terminates

with probability 1− q(κ,di ), where q(κ,di ) =κdi

sinh(κdi )to deal

with LPBE

Page 43: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

‘Walk-on-Spheres’ Algorithm

‘Walk-on-Spheres’ Algorithm

I Walk-on-spheres (WOS) algorithm for general domains with aregular boundary

I Define a Markov chain xi , i = 1,2, . . .I Set x0 = x (m) for some m, xi = xi−1 + diωi , i = 1,2, . . ., where

1. di = d(xi−1) is distance from xi−1 to Γ2. ωi is sequence of independent unit isotropic vectors3. xi is the exit point from the ball, B(xi−1, d(xi−1)), for a Brownian

motion starting at xi−1

I Outside the molecule, on every step, walk-on-spheres terminates

with probability 1− q(κ,di ), where q(κ,di ) =κdi

sinh(κdi )to deal

with LPBE

Page 44: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

‘Walk-on-Spheres’ Algorithm

‘Walk-on-Spheres’ Algorithm

I Walk-on-spheres (WOS) algorithm for general domains with aregular boundary

I Define a Markov chain xi , i = 1,2, . . .I Set x0 = x (m) for some m, xi = xi−1 + diωi , i = 1,2, . . ., where

1. di = d(xi−1) is distance from xi−1 to Γ2. ωi is sequence of independent unit isotropic vectors3. xi is the exit point from the ball, B(xi−1, d(xi−1)), for a Brownian

motion starting at xi−1

I Outside the molecule, on every step, walk-on-spheres terminates

with probability 1− q(κ,di ), where q(κ,di ) =κdi

sinh(κdi )to deal

with LPBE

Page 45: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

‘Walk-on-Spheres’ Algorithm

‘Walk-on-Spheres’ Algorithm

I Walk-on-spheres (WOS) algorithm for general domains with aregular boundary

I Define a Markov chain xi , i = 1,2, . . .I Set x0 = x (m) for some m, xi = xi−1 + diωi , i = 1,2, . . ., where

1. di = d(xi−1) is distance from xi−1 to Γ2. ωi is sequence of independent unit isotropic vectors3. xi is the exit point from the ball, B(xi−1, d(xi−1)), for a Brownian

motion starting at xi−1

I Outside the molecule, on every step, walk-on-spheres terminates

with probability 1− q(κ,di ), where q(κ,di ) =κdi

sinh(κdi )to deal

with LPBE

Page 46: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

‘Walk-on-Spheres’ Algorithm

‘Walk-on-Spheres’ Algorithm

I Walk-on-spheres (WOS) algorithm for general domains with aregular boundary

I Define a Markov chain xi , i = 1,2, . . .I Set x0 = x (m) for some m, xi = xi−1 + diωi , i = 1,2, . . ., where

1. di = d(xi−1) is distance from xi−1 to Γ2. ωi is sequence of independent unit isotropic vectors3. xi is the exit point from the ball, B(xi−1, d(xi−1)), for a Brownian

motion starting at xi−1

I Outside the molecule, on every step, walk-on-spheres terminates

with probability 1− q(κ,di ), where q(κ,di ) =κdi

sinh(κdi )to deal

with LPBE

Page 47: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

‘Walk-on-Spheres’ Algorithm

‘Walk-on-Spheres’ Algorithm

I Walk-on-spheres (WOS) algorithm for general domains with aregular boundary

I Define a Markov chain xi , i = 1,2, . . .I Set x0 = x (m) for some m, xi = xi−1 + diωi , i = 1,2, . . ., where

1. di = d(xi−1) is distance from xi−1 to Γ2. ωi is sequence of independent unit isotropic vectors3. xi is the exit point from the ball, B(xi−1, d(xi−1)), for a Brownian

motion starting at xi−1

I Outside the molecule, on every step, walk-on-spheres terminates

with probability 1− q(κ,di ), where q(κ,di ) =κdi

sinh(κdi )to deal

with LPBE

Page 48: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’I For general domains, an efficient way to simulate exit points is a

combination of1. Inside the molecule: ‘walk-in-subdomains’2. Outside the molecule ‘walk-on-spheres’

I The whole domain, Gi , is represented as a union of intersectingsubdomains:

Gi =M⋃

m=1

Gm

I ‘Walk-in-Subdomains’: Simulate exit point separately in every Gm

1. x0 = x , x1, . . . , xN – Markov chain, every xi+1 is an exit point fromthe corresponding subdomain for Brownian motion starting at xi

2. For spherical subdomains, B(xmi ,R

mi ), exit points are distributed in

accordance with the Poisson kernel:

14πRm

i

|xi − xmi |2 − (Rm

i )2

|xi − xi+1|3

Page 49: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’I For general domains, an efficient way to simulate exit points is a

combination of1. Inside the molecule: ‘walk-in-subdomains’2. Outside the molecule ‘walk-on-spheres’

I The whole domain, Gi , is represented as a union of intersectingsubdomains:

Gi =M⋃

m=1

Gm

I ‘Walk-in-Subdomains’: Simulate exit point separately in every Gm

1. x0 = x , x1, . . . , xN – Markov chain, every xi+1 is an exit point fromthe corresponding subdomain for Brownian motion starting at xi

2. For spherical subdomains, B(xmi ,R

mi ), exit points are distributed in

accordance with the Poisson kernel:

14πRm

i

|xi − xmi |2 − (Rm

i )2

|xi − xi+1|3

Page 50: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’I For general domains, an efficient way to simulate exit points is a

combination of1. Inside the molecule: ‘walk-in-subdomains’2. Outside the molecule ‘walk-on-spheres’

I The whole domain, Gi , is represented as a union of intersectingsubdomains:

Gi =M⋃

m=1

Gm

I ‘Walk-in-Subdomains’: Simulate exit point separately in every Gm

1. x0 = x , x1, . . . , xN – Markov chain, every xi+1 is an exit point fromthe corresponding subdomain for Brownian motion starting at xi

2. For spherical subdomains, B(xmi ,R

mi ), exit points are distributed in

accordance with the Poisson kernel:

14πRm

i

|xi − xmi |2 − (Rm

i )2

|xi − xi+1|3

Page 51: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’I For general domains, an efficient way to simulate exit points is a

combination of1. Inside the molecule: ‘walk-in-subdomains’2. Outside the molecule ‘walk-on-spheres’

I The whole domain, Gi , is represented as a union of intersectingsubdomains:

Gi =M⋃

m=1

Gm

I ‘Walk-in-Subdomains’: Simulate exit point separately in every Gm

1. x0 = x , x1, . . . , xN – Markov chain, every xi+1 is an exit point fromthe corresponding subdomain for Brownian motion starting at xi

2. For spherical subdomains, B(xmi ,R

mi ), exit points are distributed in

accordance with the Poisson kernel:

14πRm

i

|xi − xmi |2 − (Rm

i )2

|xi − xi+1|3

Page 52: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’I For general domains, an efficient way to simulate exit points is a

combination of1. Inside the molecule: ‘walk-in-subdomains’2. Outside the molecule ‘walk-on-spheres’

I The whole domain, Gi , is represented as a union of intersectingsubdomains:

Gi =M⋃

m=1

Gm

I ‘Walk-in-Subdomains’: Simulate exit point separately in every Gm

1. x0 = x , x1, . . . , xN – Markov chain, every xi+1 is an exit point fromthe corresponding subdomain for Brownian motion starting at xi

2. For spherical subdomains, B(xmi ,R

mi ), exit points are distributed in

accordance with the Poisson kernel:

14πRm

i

|xi − xmi |2 − (Rm

i )2

|xi − xi+1|3

Page 53: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’I For general domains, an efficient way to simulate exit points is a

combination of1. Inside the molecule: ‘walk-in-subdomains’2. Outside the molecule ‘walk-on-spheres’

I The whole domain, Gi , is represented as a union of intersectingsubdomains:

Gi =M⋃

m=1

Gm

I ‘Walk-in-Subdomains’: Simulate exit point separately in every Gm

1. x0 = x , x1, . . . , xN – Markov chain, every xi+1 is an exit point fromthe corresponding subdomain for Brownian motion starting at xi

2. For spherical subdomains, B(xmi ,R

mi ), exit points are distributed in

accordance with the Poisson kernel:

14πRm

i

|xi − xmi |2 − (Rm

i )2

|xi − xi+1|3

Page 54: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’I For general domains, an efficient way to simulate exit points is a

combination of1. Inside the molecule: ‘walk-in-subdomains’2. Outside the molecule ‘walk-on-spheres’

I The whole domain, Gi , is represented as a union of intersectingsubdomains:

Gi =M⋃

m=1

Gm

I ‘Walk-in-Subdomains’: Simulate exit point separately in every Gm

1. x0 = x , x1, . . . , xN – Markov chain, every xi+1 is an exit point fromthe corresponding subdomain for Brownian motion starting at xi

2. For spherical subdomains, B(xmi ,R

mi ), exit points are distributed in

accordance with the Poisson kernel:

14πRm

i

|xi − xmi |2 − (Rm

i )2

|xi − xi+1|3

Page 55: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Walk-in-Subdomains

‘Walk-on-Spheres’ and ‘Walk-in-Subdomains’

Figure: Walk in subdomains example

Page 56: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

I Randomization of finite-difference approximation with step, h.u(y) = Eu(x) + O(h2)

I Exact treatment of boundary conditions (mean-value theorem)for boundary point, y , in the ball B(y ,a) with surface S(y ,a):

u(y) =εe

εe + εi

∫Se(y,a)

12πa2

κasinh(κa)

ue

+εi

εe + εi

∫Si (y,a)

12πa2

κasinh(κa)

ui (1)

− εe − εiεe + εi

∫Γ

⋂B(y,a)\y

cosϕyx

2π|y − x |2Qκ,au

+εi

εe + εi

∫Bi (y,a)

[−2κ2Φκ]ui

Page 57: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

I Randomization of finite-difference approximation with step, h.u(y) = Eu(x) + O(h2)

I Exact treatment of boundary conditions (mean-value theorem)for boundary point, y , in the ball B(y ,a) with surface S(y ,a):

u(y) =εe

εe + εi

∫Se(y,a)

12πa2

κasinh(κa)

ue

+εi

εe + εi

∫Si (y,a)

12πa2

κasinh(κa)

ui (1)

− εe − εiεe + εi

∫Γ

⋂B(y,a)\y

cosϕyx

2π|y − x |2Qκ,au

+εi

εe + εi

∫Bi (y,a)

[−2κ2Φκ]ui

Page 58: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 59: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 60: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 61: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 62: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 63: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 64: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 65: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

Randomized approximation to (1): u(y) = Eu(x) + O((a/2R)3):I With probability pe exit to solvent:

1. x is chosen isotropically on the surface of auxiliary sphere,S+(y , a), that lies above tangent plane

2. Walker survives with probabilityκa

sinh(κa)

I With probability pi = 1− pe:1. x is chosen isotropically in the solid angle below tangent plane;

with probability −2κ2Φκ & sampled in Bi (y , a) (reenter molecule)2. With the complementary probability x is sampled on the surface of

auxiliary sphere, S−(y , a), that lies below tangent plane3. x reenters molecule with conditional probability 1− a/2R and4. x exits to solvent with conditional probability a/2R

Page 66: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

In the exterior, probability of terminating Markov chain dependslinearly on the initial distance to the boundary, d0. Therefore,⇒Mean number of returns to the boundary is O(d0)−1

I Finite-difference approximation of boundary conditions, ε = h2

Mean number of steps in the algorithm is O(h−1 log(h) f (κ)), f isa decreasing function (f (κ) = O(log(κ)) for small κ). Estimatesfor point values of the potential and free energy are O(h)-biased

I New treatment of boundary conditions provides O(a)2-biasedand more efficient Monte Carlo algorithm. Mean number ofsteps is O((a)−1 log(a) f (κ)), a = a/2R.

I More subtle approximation to (1) will provide even more efficientMonte Carlo estimates

Page 67: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

In the exterior, probability of terminating Markov chain dependslinearly on the initial distance to the boundary, d0. Therefore,⇒Mean number of returns to the boundary is O(d0)−1

I Finite-difference approximation of boundary conditions, ε = h2

Mean number of steps in the algorithm is O(h−1 log(h) f (κ)), f isa decreasing function (f (κ) = O(log(κ)) for small κ). Estimatesfor point values of the potential and free energy are O(h)-biased

I New treatment of boundary conditions provides O(a)2-biasedand more efficient Monte Carlo algorithm. Mean number ofsteps is O((a)−1 log(a) f (κ)), a = a/2R.

I More subtle approximation to (1) will provide even more efficientMonte Carlo estimates

Page 68: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Fast Exit Point Calculations

Monte Carlo Treatment of Boundary Conditions

Monte Carlo Treatment of Boundary Conditions

In the exterior, probability of terminating Markov chain dependslinearly on the initial distance to the boundary, d0. Therefore,⇒Mean number of returns to the boundary is O(d0)−1

I Finite-difference approximation of boundary conditions, ε = h2

Mean number of steps in the algorithm is O(h−1 log(h) f (κ)), f isa decreasing function (f (κ) = O(log(κ)) for small κ). Estimatesfor point values of the potential and free energy are O(h)-biased

I New treatment of boundary conditions provides O(a)2-biasedand more efficient Monte Carlo algorithm. Mean number ofsteps is O((a)−1 log(a) f (κ)), a = a/2R.

I More subtle approximation to (1) will provide even more efficientMonte Carlo estimates

Page 69: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

Monte Carlo Estimates

I The estimate for the reaction-field potential point value:ξ[Φrf ](x (m)) = −Φc(x∗1 )

+

Nins∑j=2

Fj (κ) (Φc(x insj )− Φc(x∗j,ins)) (2)

I Here x∗j,ins is a sequence of boundary points, after which therandom walker moves inside the domain, Gi , to x ins

jI The estimate for the reaction-field energy:

ξ[Wrf ] =12

M∑m=1

Qm ξ[Φrf ](x (m)) (3)

Page 70: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

Monte Carlo Estimates

I The estimate for the reaction-field potential point value:ξ[Φrf ](x (m)) = −Φc(x∗1 )

+

Nins∑j=2

Fj (κ) (Φc(x insj )− Φc(x∗j,ins)) (2)

I Here x∗j,ins is a sequence of boundary points, after which therandom walker moves inside the domain, Gi , to x ins

jI The estimate for the reaction-field energy:

ξ[Wrf ] =12

M∑m=1

Qm ξ[Φrf ](x (m)) (3)

Page 71: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

Monte Carlo Estimates

I The estimate for the reaction-field potential point value:ξ[Φrf ](x (m)) = −Φc(x∗1 )

+

Nins∑j=2

Fj (κ) (Φc(x insj )− Φc(x∗j,ins)) (2)

I Here x∗j,ins is a sequence of boundary points, after which therandom walker moves inside the domain, Gi , to x ins

jI The estimate for the reaction-field energy:

ξ[Wrf ] =12

M∑m=1

Qm ξ[Φrf ](x (m)) (3)

Page 72: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

A Picture: The Algorithm for a Single Spherical Atom

Page 73: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

The Algorithm in Pictures: Walk Inside

Page 74: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

The Algorithm in Pictures: Walk Inside

Page 75: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

The Algorithm in Pictures: Walk Outside

Page 76: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

The Algorithm in Pictures: Walk Outside

Page 77: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

The Algorithm in Pictures: Walk to∞ in One Step

Figure: κ = 0, p∞ = 1− REnclosed/dist

Page 78: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

Monte Carlo Algorithm’s Computational ComplexityCost of a single trajectory

I Number of steps is random walk is not dependent on M, thenumber of atoms

I The cost of finding the nearest sphere is M log2(M) due tooptimizations

0

0.5

1

1.5

2

0 1000 2000 3000 4000 5000

CP

UTi

me(

sec)

Number of Atoms

"Simulation""Theory-prediction"

Figure: The CPU time per atom per trajectory is plotted as function of number of atoms. For smallnumber of atoms the CPU time scales linearly and for large number of atoms it asymptoticallyscales logarithmically

Page 79: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Monte Carlo Estimates

Monte Carlo Algorithm’s Computational ComplexityCost of a single trajectory

I Number of steps is random walk is not dependent on M, thenumber of atoms

I The cost of finding the nearest sphere is M log2(M) due tooptimizations

0

0.5

1

1.5

2

0 1000 2000 3000 4000 5000

CP

UTi

me(

sec)

Number of Atoms

"Simulation""Theory-prediction"

Figure: The CPU time per atom per trajectory is plotted as function of number of atoms. For smallnumber of atoms the CPU time scales linearly and for large number of atoms it asymptoticallyscales logarithmically

Page 80: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Computational Geometry

Geometry: Problem Descriptions

There are many geometric problems that arise in this algorithm:

I Efficiently determining if a point is on the surface of the moleculeor inside of it (for interior walks)

I Efficiently determining the closest sphere to a given exterior point(for walks outside molecule)

I Efficiently determining if a query point is inside of the convex hullof the molecule

I Efficiently finding the largest possible sphere enclosing a querypoint for external walks

Page 81: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Computational Geometry

Geometry: Problem Descriptions

There are many geometric problems that arise in this algorithm:

I Efficiently determining if a point is on the surface of the moleculeor inside of it (for interior walks)

I Efficiently determining the closest sphere to a given exterior point(for walks outside molecule)

I Efficiently determining if a query point is inside of the convex hullof the molecule

I Efficiently finding the largest possible sphere enclosing a querypoint for external walks

Page 82: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Computational Geometry

Geometry: Problem Descriptions

There are many geometric problems that arise in this algorithm:

I Efficiently determining if a point is on the surface of the moleculeor inside of it (for interior walks)

I Efficiently determining the closest sphere to a given exterior point(for walks outside molecule)

I Efficiently determining if a query point is inside of the convex hullof the molecule

I Efficiently finding the largest possible sphere enclosing a querypoint for external walks

Page 83: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Computational Geometry

Geometry: Problem Descriptions

There are many geometric problems that arise in this algorithm:

I Efficiently determining if a point is on the surface of the moleculeor inside of it (for interior walks)

I Efficiently determining the closest sphere to a given exterior point(for walks outside molecule)

I Efficiently determining if a query point is inside of the convex hullof the molecule

I Efficiently finding the largest possible sphere enclosing a querypoint for external walks

Page 84: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated and Uncorrelated Sampling

I Correlated sampling in Monte Carlo is essential for two importantreasons

1. To obtain smooth curves with a minimum of sampling(function-wise vs. point-wise sampling)

2. To obtain accurate results from quantities defined as thedifferences of Monte Carlo estimates

I With this correlated sampling sampling you can get a “smoothcurve" with three orders of magnitude less sampling, note: youstill have O(N−1/2) errors, just in “curve space," not point by point

Page 85: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated and Uncorrelated Sampling

I Correlated sampling in Monte Carlo is essential for two importantreasons

1. To obtain smooth curves with a minimum of sampling(function-wise vs. point-wise sampling)

2. To obtain accurate results from quantities defined as thedifferences of Monte Carlo estimates

I With this correlated sampling sampling you can get a “smoothcurve" with three orders of magnitude less sampling, note: youstill have O(N−1/2) errors, just in “curve space," not point by point

Page 86: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated and Uncorrelated Sampling

I Correlated sampling in Monte Carlo is essential for two importantreasons

1. To obtain smooth curves with a minimum of sampling(function-wise vs. point-wise sampling)

2. To obtain accurate results from quantities defined as thedifferences of Monte Carlo estimates

I With this correlated sampling sampling you can get a “smoothcurve" with three orders of magnitude less sampling, note: youstill have O(N−1/2) errors, just in “curve space," not point by point

Page 87: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated and Uncorrelated Sampling

I Correlated sampling in Monte Carlo is essential for two importantreasons

1. To obtain smooth curves with a minimum of sampling(function-wise vs. point-wise sampling)

2. To obtain accurate results from quantities defined as thedifferences of Monte Carlo estimates

I With this correlated sampling sampling you can get a “smoothcurve" with three orders of magnitude less sampling, note: youstill have O(N−1/2) errors, just in “curve space," not point by point

Page 88: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Salt Concentration

-2970

-2965

-2960

-2955

-2950

-2945

-2940

-2935

-2930

-2925

-4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0

∆G

elec

solv

(kca

l/mol

e)

log(Salt Concentration(M))

uncorr 500uncorr 1500

corr 500uncorr 4500

Figure: Electrostatic Solvation free Energy of 3icb calculated with three four conditions:uncorrelated sampling with 500 number of trajectories per concentration, uncorrelated samplingwith 1500 number of trajectories per concentration, uncorrelated sampling with 4500 number ofiterations, and correlated sampling with 500 number of trajectories

Page 89: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Dependence on Salt Concentration

I Values of scalar energies as a function of external saltconcentration are important

1. Smooth curves of internal energy vs. salt concentration (see above)2. Numerical estimate of the derivative as salt concentration vanishes

I For κ used in simulations, Fj (κ) = 1I For an arbitrary κ′ > κ:

Fj (κ′) is multiplied by the ratio

q(κ′,d)

q(κ,d)on every step of the WOS

in the exteriorI The results obtained with the estimates (2) and (3) for different

values of κ are highly correlated

Page 90: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Dependence on Salt Concentration

I Values of scalar energies as a function of external saltconcentration are important

1. Smooth curves of internal energy vs. salt concentration (see above)2. Numerical estimate of the derivative as salt concentration vanishes

I For κ used in simulations, Fj (κ) = 1I For an arbitrary κ′ > κ:

Fj (κ′) is multiplied by the ratio

q(κ′,d)

q(κ,d)on every step of the WOS

in the exteriorI The results obtained with the estimates (2) and (3) for different

values of κ are highly correlated

Page 91: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Dependence on Salt Concentration

I Values of scalar energies as a function of external saltconcentration are important

1. Smooth curves of internal energy vs. salt concentration (see above)2. Numerical estimate of the derivative as salt concentration vanishes

I For κ used in simulations, Fj (κ) = 1I For an arbitrary κ′ > κ:

Fj (κ′) is multiplied by the ratio

q(κ′,d)

q(κ,d)on every step of the WOS

in the exteriorI The results obtained with the estimates (2) and (3) for different

values of κ are highly correlated

Page 92: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Dependence on Salt Concentration

I Values of scalar energies as a function of external saltconcentration are important

1. Smooth curves of internal energy vs. salt concentration (see above)2. Numerical estimate of the derivative as salt concentration vanishes

I For κ used in simulations, Fj (κ) = 1I For an arbitrary κ′ > κ:

Fj (κ′) is multiplied by the ratio

q(κ′,d)

q(κ,d)on every step of the WOS

in the exteriorI The results obtained with the estimates (2) and (3) for different

values of κ are highly correlated

Page 93: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Dependence on Salt Concentration

I Values of scalar energies as a function of external saltconcentration are important

1. Smooth curves of internal energy vs. salt concentration (see above)2. Numerical estimate of the derivative as salt concentration vanishes

I For κ used in simulations, Fj (κ) = 1I For an arbitrary κ′ > κ:

Fj (κ′) is multiplied by the ratio

q(κ′,d)

q(κ,d)on every step of the WOS

in the exteriorI The results obtained with the estimates (2) and (3) for different

values of κ are highly correlated

Page 94: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Dependence on Salt Concentration

I Values of scalar energies as a function of external saltconcentration are important

1. Smooth curves of internal energy vs. salt concentration (see above)2. Numerical estimate of the derivative as salt concentration vanishes

I For κ used in simulations, Fj (κ) = 1I For an arbitrary κ′ > κ:

Fj (κ′) is multiplied by the ratio

q(κ′,d)

q(κ,d)on every step of the WOS

in the exteriorI The results obtained with the estimates (2) and (3) for different

values of κ are highly correlated

Page 95: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Binding Calculations

I Binding computation requires three energy computationsE(A + B)− E(A)− E(B)

I Monte Carlo requires “help" when differencingI We use the reproducibility in SPRNG to do this effectively

1. Unbound: when exiting the molecule the seed is stored usingSPRNG tools

2. Bound: walks resume at the exit points with the same randomnumber streams and reusing

3. At this exit point, only the exit point information is requiredI The leads to correlation between unbound and bound energy

computations that decreases as the walk length increases (κ2

decreases)

Page 96: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Binding Calculations

I Binding computation requires three energy computationsE(A + B)− E(A)− E(B)

I Monte Carlo requires “help" when differencingI We use the reproducibility in SPRNG to do this effectively

1. Unbound: when exiting the molecule the seed is stored usingSPRNG tools

2. Bound: walks resume at the exit points with the same randomnumber streams and reusing

3. At this exit point, only the exit point information is requiredI The leads to correlation between unbound and bound energy

computations that decreases as the walk length increases (κ2

decreases)

Page 97: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Binding Calculations

I Binding computation requires three energy computationsE(A + B)− E(A)− E(B)

I Monte Carlo requires “help" when differencingI We use the reproducibility in SPRNG to do this effectively

1. Unbound: when exiting the molecule the seed is stored usingSPRNG tools

2. Bound: walks resume at the exit points with the same randomnumber streams and reusing

3. At this exit point, only the exit point information is requiredI The leads to correlation between unbound and bound energy

computations that decreases as the walk length increases (κ2

decreases)

Page 98: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Binding Calculations

I Binding computation requires three energy computationsE(A + B)− E(A)− E(B)

I Monte Carlo requires “help" when differencingI We use the reproducibility in SPRNG to do this effectively

1. Unbound: when exiting the molecule the seed is stored usingSPRNG tools

2. Bound: walks resume at the exit points with the same randomnumber streams and reusing

3. At this exit point, only the exit point information is requiredI The leads to correlation between unbound and bound energy

computations that decreases as the walk length increases (κ2

decreases)

Page 99: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Binding Calculations

I Binding computation requires three energy computationsE(A + B)− E(A)− E(B)

I Monte Carlo requires “help" when differencingI We use the reproducibility in SPRNG to do this effectively

1. Unbound: when exiting the molecule the seed is stored usingSPRNG tools

2. Bound: walks resume at the exit points with the same randomnumber streams and reusing

3. At this exit point, only the exit point information is requiredI The leads to correlation between unbound and bound energy

computations that decreases as the walk length increases (κ2

decreases)

Page 100: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Binding Calculations

I Binding computation requires three energy computationsE(A + B)− E(A)− E(B)

I Monte Carlo requires “help" when differencingI We use the reproducibility in SPRNG to do this effectively

1. Unbound: when exiting the molecule the seed is stored usingSPRNG tools

2. Bound: walks resume at the exit points with the same randomnumber streams and reusing

3. At this exit point, only the exit point information is requiredI The leads to correlation between unbound and bound energy

computations that decreases as the walk length increases (κ2

decreases)

Page 101: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Monte Carlo Estimates

Correlated and Uncorrelated Sampling

Correlated Sampling: Binding Calculations

I Binding computation requires three energy computationsE(A + B)− E(A)− E(B)

I Monte Carlo requires “help" when differencingI We use the reproducibility in SPRNG to do this effectively

1. Unbound: when exiting the molecule the seed is stored usingSPRNG tools

2. Bound: walks resume at the exit points with the same randomnumber streams and reusing

3. At this exit point, only the exit point information is requiredI The leads to correlation between unbound and bound energy

computations that decreases as the walk length increases (κ2

decreases)

Page 102: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Accuracy: Monte Carlo vs. Deterministic

Page 103: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 104: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 105: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 106: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 107: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 108: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 109: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 110: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 111: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 112: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 113: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 114: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Sampling Error and Bias

I In Monte Carlo there are biases (errors) and sampling error1. Sampling error is based on standard error O(N−1/2)2. Difference between expected value and PDE solution is bias

I Capture thickness (ε): bias is O(ε)I Auxiliary sphere radius (a): bias is O(a3)I Effective Van der Waals sphere radius, RI Overall bias:

( a2R

)3+

2R

)3. Var [

∑i qi Φ(xi )] =

∑i q2

i Var [Φ(xi )]4. Given a desired variance, divide it evenly over this sum5. Running time ∝ | ln(ε)|

a6. Can reduce running time by 2 orders of magnitude by bias/variance

balancing and using larger ε, a and ANN7. Large ANN means errors in drawing the largest sphere outside the

molecule for WOS

Page 115: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Computational Results

Timing: Better Than Expected

Figure: O(M log M)?

Page 116: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Conclusions

I We have developed a novel stochastic linear PBE solver that canprovide highly accurate salt-dependent electrostatic properties ofbiomolecules in a single PBE calculation

I Advantages of the stochastic linear PBE solver over the moremature deterministic methods include: the subtle geometricfeatures of the biomolecule can be treated with higher precision,the continuity and outer boundary conditions are accounted forexactly, a singularity free scheme is employed andstraightforward implementation on parallel computer platform ispossible

I Codes provide higher accuracy (on demand) and do not sufferlosses in accuracy near the boundary

I Only way to handle large (M >> 10000) molecules

Page 117: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Conclusions

I We have developed a novel stochastic linear PBE solver that canprovide highly accurate salt-dependent electrostatic properties ofbiomolecules in a single PBE calculation

I Advantages of the stochastic linear PBE solver over the moremature deterministic methods include: the subtle geometricfeatures of the biomolecule can be treated with higher precision,the continuity and outer boundary conditions are accounted forexactly, a singularity free scheme is employed andstraightforward implementation on parallel computer platform ispossible

I Codes provide higher accuracy (on demand) and do not sufferlosses in accuracy near the boundary

I Only way to handle large (M >> 10000) molecules

Page 118: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Conclusions

I We have developed a novel stochastic linear PBE solver that canprovide highly accurate salt-dependent electrostatic properties ofbiomolecules in a single PBE calculation

I Advantages of the stochastic linear PBE solver over the moremature deterministic methods include: the subtle geometricfeatures of the biomolecule can be treated with higher precision,the continuity and outer boundary conditions are accounted forexactly, a singularity free scheme is employed andstraightforward implementation on parallel computer platform ispossible

I Codes provide higher accuracy (on demand) and do not sufferlosses in accuracy near the boundary

I Only way to handle large (M >> 10000) molecules

Page 119: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Conclusions

I We have developed a novel stochastic linear PBE solver that canprovide highly accurate salt-dependent electrostatic properties ofbiomolecules in a single PBE calculation

I Advantages of the stochastic linear PBE solver over the moremature deterministic methods include: the subtle geometricfeatures of the biomolecule can be treated with higher precision,the continuity and outer boundary conditions are accounted forexactly, a singularity free scheme is employed andstraightforward implementation on parallel computer platform ispossible

I Codes provide higher accuracy (on demand) and do not sufferlosses in accuracy near the boundary

I Only way to handle large (M >> 10000) molecules

Page 120: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 121: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 122: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 123: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 124: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 125: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 126: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 127: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 128: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 129: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 130: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future WorkI Binding computations: using correlated sampling by directly

reprocessing walksI Simple code interface for distribution with

1. Desired accuracy as input that allows a precalculation of thenumber of needed trajectories

2. Importance sampling for optimal estimation of scalar energy values3. Built-in CONDOR support for distribution of concurrent tasks4. Multicore distributed computing support for the code:

OpenMP/OpenMPI5. Precompiled code module distribution to protect IP6. Webpage to describe the method and the mathematical

background and applicationI Exploit the implicit inverse computation this methods provides

1. Can do computation without knowing charges until the end (aninverse)

2. Simple to examine many charge distributions in a perfectlycorrelated setting

Page 131: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 132: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 133: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 134: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 135: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 136: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 137: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 138: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 139: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 140: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 141: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 142: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 143: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 144: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Future Work

I Further algorithmic development1. Computation of gradients using existing Markov chains2. Global computation of field variables and their visualization3. Nonlinear BVPs perhaps via branching processes4. Using “Walk-on-the-Boundary" (WOB) techniques

I Geometric Issues1. Computation of the three region model problem2. More complicated surfaces (solvent-excluded and ion-excluded)3. Accuracy issues related to the Van der Waals surface

I Optimize the performance1. Error/bias/variance balancing2. Importance sampling and the outer walks3. WOB to eliminate walks outside4. QMC methods

Page 145: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Bibliography

[M. Fenley, M. Mascagni, J. McClain, A. Silalahi and N. Simonov(2010)] Using Correlated Monte Carlo Sampling for EfficientlySolving the Linearized Poisson-Boltzmann Equation Over aBroad Range of Salt Concentrations Journal of Chemical Theoryand Computation, 6(1): 300–314.

[N. Simonov and M. Mascagni and M. O. Fenley (2007)] MonteCarlo Based Linear Poisson-Boltzmann Approach MakesAccurate Salt-Dependent Solvation Energy Predictions PossibleJournal of Chemical Physics, 187(18), article #185105, 6 pages.

[M. Mascagni and N. A. Simonov (2004)] Monte Carlo Methodsfor Calculating Some Physical Properties of Large MoleculesSIAM Journal on Scientific Computing, 26(1): 339–357.

Page 146: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Bibliography

[M. Fenley, M. Mascagni, J. McClain, A. Silalahi and N. Simonov(2010)] Using Correlated Monte Carlo Sampling for EfficientlySolving the Linearized Poisson-Boltzmann Equation Over aBroad Range of Salt Concentrations Journal of Chemical Theoryand Computation, 6(1): 300–314.

[N. Simonov and M. Mascagni and M. O. Fenley (2007)] MonteCarlo Based Linear Poisson-Boltzmann Approach MakesAccurate Salt-Dependent Solvation Energy Predictions PossibleJournal of Chemical Physics, 187(18), article #185105, 6 pages.

[M. Mascagni and N. A. Simonov (2004)] Monte Carlo Methodsfor Calculating Some Physical Properties of Large MoleculesSIAM Journal on Scientific Computing, 26(1): 339–357.

Page 147: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Bibliography

[M. Fenley, M. Mascagni, J. McClain, A. Silalahi and N. Simonov(2010)] Using Correlated Monte Carlo Sampling for EfficientlySolving the Linearized Poisson-Boltzmann Equation Over aBroad Range of Salt Concentrations Journal of Chemical Theoryand Computation, 6(1): 300–314.

[N. Simonov and M. Mascagni and M. O. Fenley (2007)] MonteCarlo Based Linear Poisson-Boltzmann Approach MakesAccurate Salt-Dependent Solvation Energy Predictions PossibleJournal of Chemical Physics, 187(18), article #185105, 6 pages.

[M. Mascagni and N. A. Simonov (2004)] Monte Carlo Methodsfor Calculating Some Physical Properties of Large MoleculesSIAM Journal on Scientific Computing, 26(1): 339–357.

Page 148: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Bibliography

[N. A. Simonov and M. Mascagni (2004)] Random WalkAlgorithms for Estimating Effective Properties of Digitized PorousMedia Monte Carlo Methods and Applications, 10: 599–608.

[M. Mascagni and N. A. Simonov (2004)] The Random Walk onthe Boundary Method for Calculating Capacitance Journal ofComputational Physics, 195: 465–473.

[A. Karaivanova, N. A. Simonov and M. Mascagni(2004)] ParallelQuasirandom Walks on the Boundary Monte Carlo Methods andApplications, 11: 311–320.

[C.-O. Hwang and M. Mascagni (2001)] Efficient modified “walkon spheres" algortihm for the linearized Poisson-Boltzmannequation Applied Physics Letters, 78: 787–789.

Page 149: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Bibliography

[N. A. Simonov and M. Mascagni (2004)] Random WalkAlgorithms for Estimating Effective Properties of Digitized PorousMedia Monte Carlo Methods and Applications, 10: 599–608.

[M. Mascagni and N. A. Simonov (2004)] The Random Walk onthe Boundary Method for Calculating Capacitance Journal ofComputational Physics, 195: 465–473.

[A. Karaivanova, N. A. Simonov and M. Mascagni(2004)] ParallelQuasirandom Walks on the Boundary Monte Carlo Methods andApplications, 11: 311–320.

[C.-O. Hwang and M. Mascagni (2001)] Efficient modified “walkon spheres" algortihm for the linearized Poisson-Boltzmannequation Applied Physics Letters, 78: 787–789.

Page 150: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Bibliography

[N. A. Simonov and M. Mascagni (2004)] Random WalkAlgorithms for Estimating Effective Properties of Digitized PorousMedia Monte Carlo Methods and Applications, 10: 599–608.

[M. Mascagni and N. A. Simonov (2004)] The Random Walk onthe Boundary Method for Calculating Capacitance Journal ofComputational Physics, 195: 465–473.

[A. Karaivanova, N. A. Simonov and M. Mascagni(2004)] ParallelQuasirandom Walks on the Boundary Monte Carlo Methods andApplications, 11: 311–320.

[C.-O. Hwang and M. Mascagni (2001)] Efficient modified “walkon spheres" algortihm for the linearized Poisson-Boltzmannequation Applied Physics Letters, 78: 787–789.

Page 151: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

Bibliography

[N. A. Simonov and M. Mascagni (2004)] Random WalkAlgorithms for Estimating Effective Properties of Digitized PorousMedia Monte Carlo Methods and Applications, 10: 599–608.

[M. Mascagni and N. A. Simonov (2004)] The Random Walk onthe Boundary Method for Calculating Capacitance Journal ofComputational Physics, 195: 465–473.

[A. Karaivanova, N. A. Simonov and M. Mascagni(2004)] ParallelQuasirandom Walks on the Boundary Monte Carlo Methods andApplications, 11: 311–320.

[C.-O. Hwang and M. Mascagni (2001)] Efficient modified “walkon spheres" algortihm for the linearized Poisson-Boltzmannequation Applied Physics Letters, 78: 787–789.

Page 152: Novel Stochastic Methods in Biochemical Electrostatics ...people.sc.fsu.edu/~jburkardt/presentations/sem_2012_mascagni2.pdf · Novel Stochastic Methods in Biochemical Electrostatics:

Novel Stochastic Methods in Biochemical Electrostatics: (Stochastic Methods for PDEs Can Beat Deterministic Methods)

Conclusions and Future Work

c© Michael Mascagni, 2012