Thermodynamic Integration with EnhancedSampling (TIES)

A. P. Bhati, S. Wan, D. W. Wright and P. V. Coveneyagastya.bhati.14@ucl.ac.uk

Centre for Computational ScienceDepartment of ChemistryUniversity College London

16th May 2017

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 1 21

Credits due

Peter V. Coveney

Shunzhou Wan

David W. Wright

UCL Overseas ResearchScholarship

Inlaks Shivdasani Foundation

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 2 21

Overview

Brief introduction to the binding affinity and methods to calculate it

Importance of being able to predict binding affinities reliably

Issues with the traditional in silico approaches

Ensemble simulation based approach: TIES

Success story of TIES: Case studies

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 3 21

Overview

Brief introduction to the binding affinity and methods to calculate it

Importance of being able to predict binding affinities reliably

Issues with the traditional in silico approaches

Ensemble simulation based approach: TIES

Success story of TIES: Case studies

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 3 21

Overview

Brief introduction to the binding affinity and methods to calculate it

Importance of being able to predict binding affinities reliably

Issues with the traditional in silico approaches

Ensemble simulation based approach: TIES

Success story of TIES: Case studies

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 3 21

Overview

Brief introduction to the binding affinity and methods to calculate it

Importance of being able to predict binding affinities reliably

Issues with the traditional in silico approaches

Ensemble simulation based approach: TIES

Success story of TIES: Case studies

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 3 21

Overview

Brief introduction to the binding affinity and methods to calculate it

Importance of being able to predict binding affinities reliably

Issues with the traditional in silico approaches

Ensemble simulation based approach: TIES

Success story of TIES: Case studies

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 3 21

Molecular Dynamics (MD) Simulations

An experiment is usually made on a macroscopic sample that containsan extremely large number of atoms or molecules sampling anenormous number of conformations.

In MD simulation, macroscopic properties corresponding to experimentalobservables are defined in terms of ensemble averages.1

Free energy is such a measurement

1

Coveney & Wan, PCCP, 2016, 18, 30236-30240, DOI: 10.1039/C6CP02349E

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 4 21

Molecular Dynamics (MD) Simulations

An experiment is usually made on a macroscopic sample that containsan extremely large number of atoms or molecules sampling anenormous number of conformations.In MD simulation, macroscopic properties corresponding to experimentalobservables are defined in terms of ensemble averages.1

Free energy is such a measurement

1Coveney & Wan, PCCP, 2016, 18, 30236-30240, DOI: 10.1039/C6CP02349EAgastya P. Bhati (CCS, UCL) TIES 16th May 2017 4 21

Molecular Dynamics (MD) Simulations

An experiment is usually made on a macroscopic sample that containsan extremely large number of atoms or molecules sampling anenormous number of conformations.In MD simulation, macroscopic properties corresponding to experimentalobservables are defined in terms of ensemble averages.1

Free energy is such a measurement1Coveney & Wan, PCCP, 2016, 18, 30236-30240, DOI: 10.1039/C6CP02349E

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 4 21

Binding Affinity

How does it help us?

Ligand binding driven by changes in the Gibbs free energy

The more negative the ∆G the stronger the binding

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 5 21

Binding Affinity

How does it help us?

Ligand binding driven by changes in the Gibbs free energy

The more negative the ∆G the stronger the bindingAgastya P. Bhati (CCS, UCL) TIES 16th May 2017 5 21

In silico free energy calculation methods

Docking methods

Linear Interaction method

MMPBSA+NMODE

Thermodynamic integration

Free energy perturbation (EXP, BAR,MBAR)

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 6 21

Thermodynamic Integration

Hybrid potential function:

H(λ) = λHA + (1 − λ)HB

The coupling parameter, λ, defines theprogress of a system along the path, B toA, as λ is changed from 0 to 1.

∆G =

1∫0

⟨∂H(λ)

∂λ

⟩λ

dλ Alchemical mutation from B(left) to A (right)

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 7 21

Relative Binding Affinity

Relative binding affinity calculations with alchemical mutation: make useof thermodynamic cycle to calculate binding free energy differences

∆∆Gbinding = ∆Gbindingligand2 − ∆Gbinding

ligand1 = ∆Galchligand − ∆Galch

complex

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 8 21

Relative Binding Affinity

Relative binding affinity calculations with alchemical mutation: make useof thermodynamic cycle to calculate binding free energy differences

∆∆Gbinding = ∆Gbindingligand2 − ∆Gbinding

ligand1 = ∆Galchligand − ∆Galch

complex

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 8 21

Relative Binding Affinity

Relative binding affinity calculations with alchemical mutation: make useof thermodynamic cycle to calculate binding free energy differences

∆∆Gbinding = ∆Gbindingligand2 − ∆Gbinding

ligand1 = ∆Galchligand − ∆Galch

complex

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 8 21

Application of binding affinity prediction

Drug designing: Lead optimisation

Searching for a needle in a haystack

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 9 21

Application of binding affinity prediction

Drug designing: Lead optimisationSearching for a needle in a haystack

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 9 21

Application of binding affinity prediction

Drug designing: Lead optimisationSearching for a needle in a haystack

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 9 21

Application of binding affinity prediction

Drug designing: Lead optimisationSearching for a needle in a haystack

High-ThroughputScreening (HTS)

HTS can test thousandsof compounds per day

Cost of HTS issubstantial: $1-10 percompound

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 9 21

Application of binding affinity prediction

Drug designing: Lead optimisationSearching for a needle in a haystack

Virtual Screening:Systematic computer-basedprediction of binding affinityof compounds to proteins

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 9 21

Issues with the available in silico methods

Theories exist - Then predictions are possible, and in principle, weshould be able to apply existing methods in drug screening domains

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 10 21

Issues with the available in silico methods

Theories exist - Then predictions are possible, and in principle, weshould be able to apply existing methods in drug screening domains

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 10 21

Single vs Ensemble MD simulations

The binding free energy and potential energy derivative can vary widely(up to 12 kcal/mol) between two single simulations.

Single simulation: not reproducible, unscientific!

L1Q-LI9 ligand transformation bound to CDK22 Drug-HIV1 protease3

The energy/energy derivatives from ensemble simulations follow welldefined Gaussian distributions.

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b009792Wright, Hall, Kenway, Jha & Coveney, JCTC, (2014), DOI: 10.1021/ct4007037

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 11 21

Single vs Ensemble MD simulations

The binding free energy and potential energy derivative can vary widely(up to 12 kcal/mol) between two single simulations.

Single simulation: not reproducible, unscientific!

L1Q-LI9 ligand transformation bound to CDK22 Drug-HIV1 protease3

The energy/energy derivatives from ensemble simulations follow welldefined Gaussian distributions.

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b009792Wright, Hall, Kenway, Jha & Coveney, JCTC, (2014), DOI: 10.1021/ct4007037

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 11 21

Thermodynamic Integration with EnhancedSampling (TIES)

Binding Affinity Calculator (BAC) is a software toolkit which automatesthe implementation of TIES (and ESMACS) methods for binding affinitycalculations

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b00979Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 12 21

Thermodynamic Integration with EnhancedSampling (TIES)

Binding Affinity Calculator (BAC) is a software toolkit which automatesthe implementation of TIES (and ESMACS) methods for binding affinitycalculations

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b00979Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 12 21

TIES: Convergence of errors

Variation of error with different parameters

L1Q-LI9 ligand transformation bound to CDK2

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b00979Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 13 21

TIES: Biomolecular systems studiedCrystal structures of the protein from the Protein Data Bank

CDK2 MCL1PTP1B

Thrombin TYK2Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 14 21

TIES predictions

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b00979Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 15 21

TIES predictions

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b00979Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 15 21

TIES reproducibility

Reproducibility of TIES predictions for transformation between ligandsbound to CDK2

1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b00979Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 16 21

Thrombin S1 pocket: water inflow

Captured successfully by TIES1Bhati, Wan, Wright & Coveney, JCTC, (2017), DOI: 10.1021/acs.jctc.6b00979

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 17 21

Blind study with Pfizer: Tropomysin receptorkinase A

The TIES study gives the same ranking of thebinding free energies as the experimental data forthe compounds studied

Pearson correlation of 0.88

10 out of the 14 pairsdirectional agreements

A better agreement might beachieved when the error barsof the experimental ∆∆G arealso taken into account

1Wan, Bhati, Skerratt, Omoto, Shanmugasundaram, Bagal, Coveney, JCIM (2017),DOI: 10.1021/acs.jcim.6b00780

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 18 21

Blind study with Pfizer: Tropomysin receptorkinase A

The TIES study gives the same ranking of thebinding free energies as the experimental data forthe compounds studied

Pearson correlation of 0.88

10 out of the 14 pairsdirectional agreements

A better agreement might beachieved when the error barsof the experimental ∆∆G arealso taken into account

1Wan, Bhati, Skerratt, Omoto, Shanmugasundaram, Bagal, Coveney, JCIM (2017),DOI: 10.1021/acs.jcim.6b00780

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 18 21

Blind study with Pfizer: Tropomysin receptorkinase A

The TIES study gives the same ranking of thebinding free energies as the experimental data forthe compounds studied

Pearson correlation of 0.88

10 out of the 14 pairsdirectional agreements

A better agreement might beachieved when the error barsof the experimental ∆∆G arealso taken into account

1Wan, Bhati, Skerratt, Omoto, Shanmugasundaram, Bagal, Coveney, JCIM (2017),DOI: 10.1021/acs.jcim.6b00780

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 18 21

Blind study with GSK: Bromodomain 4

1Wan, Bhati, Zasada, Wall, Green, Bamborough, Coveney, JCTC (2017), DOI:10.1021/acs.jctc.6b00794

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 19 21

Excellent scalability

TIES workflow is perfectlyscalable

Giant run on PRACE’s Tier0SuperMUC

LRZ press releaseLondon Science Museum blog post

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 20 21

Conclusions

The traditional in silico methods to calculate binding affinity are notreliable, and hence, not widely applicable in pharmaceutical domain

Ensemble simulation based method is the way out

TIES substantially improves the accuracy, precision and reliability ofcalculated relative binding affinities

Ligand-protein binding affinity predictions made with TIES arereproducible

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 21 21

Conclusions

The traditional in silico methods to calculate binding affinity are notreliable, and hence, not widely applicable in pharmaceutical domain

Ensemble simulation based method is the way out

TIES substantially improves the accuracy, precision and reliability ofcalculated relative binding affinities

Ligand-protein binding affinity predictions made with TIES arereproducible

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 21 21

Conclusions

The traditional in silico methods to calculate binding affinity are notreliable, and hence, not widely applicable in pharmaceutical domain

Ensemble simulation based method is the way out

TIES substantially improves the accuracy, precision and reliability ofcalculated relative binding affinities

Ligand-protein binding affinity predictions made with TIES arereproducible

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 21 21

Conclusions

The traditional in silico methods to calculate binding affinity are notreliable, and hence, not widely applicable in pharmaceutical domain

Ensemble simulation based method is the way out

TIES substantially improves the accuracy, precision and reliability ofcalculated relative binding affinities

Ligand-protein binding affinity predictions made with TIES arereproducible

Agastya P. Bhati (CCS, UCL) TIES 16th May 2017 21 21