Accelera'ng Drug Discovery with Free Energy Calcula'ons on GPUs - - PowerPoint PPT Presentation

Accelera'ng Drug Discovery with Free Energy Calcula'ons on GPUs

Robert Abel VP, Scien1fic Development

Drug Discovery

• “It’s damned tough to discover a drug.” –Eugene Cordes
• The mission of Schrödinger R&D is to make this much easier

Lead Op'miza'on is Profoundly Challenging

Phase of Drug Discovery Total Costs (USD millions) Target-to-hit $94 Hit-to-lead $166 Lead-op1miza1on $414 Preclinical $150 Phase I $273 Phase II $319 Phase III $314 Submission to launch $48 Total $1,778

Paul SM et al. Nat. Rev. Drug Disc. 9:203-214. 2010.

Free Energy Calcula'ons Should be able to Help

• Faster potency op1miza1on with

fewer synthesized compounds

• Be]er maintenance of potency

while tuning ADMET proper1es

• Account for other proper1es

relevant to lead op1miza1on

– Binding selec1vity – Muta1onal resistance – Solubility – Membrane permeability

Rela've Binding Free Energy Calcula'ons with FEP

• Compu1ng rela1ve free energies

has notable advantages

– Modeling of smaller perturba1ons should be more accurate – Rela1ve differences are o`en of greatest interest in lead op1miza1on

• Instead of modeling the full binding

process, we use FEP to compute

– the difference between ligand 1→2 in solu1on (A) – the difference between ligand 1→2 in the binding site (B)

ΔΔ ΔΔGbinding = ΔG1 – ΔG2 = ΔGA – ΔGB A B 1 2

Schrödinger’s Approach: FEP+

• Complete solu1on combining state-of-the-art MD engine,

sampling algorithm, force field, so`ware engineering, and GPU support for unparalleled accuracy, throughput, and ease of use in real-life prospec1ve drug discovery projects

– A rou1ne part of the porgolio of design tools used by internal Schrödinger drug discovery group – Significantly enriching 1ght binders in all prospec1ve studies – Con1nuously developed for accuracy and performance improvements – Validated across wide range of systems – Scoring each ligand requires roughly 1 GPU day of compute ?me

Desmond GPU

• GPU compu1ng provides a significant advantage

• Over 200 ligands scored w/ iden1cal protocol
• RMSE ≈ 1 kcal/mol, correla1ons appear predic1ve

Schrödinger FEP+ Retrospec've Accuracy

|ΔΔGFEP – ΔΔGExpt.| (kcal/mol) Percentage

46.2% 24.8% 15.4% 7.4% 6.2% 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% < 0.6 0.6-1.2 1.2-1.8 1.8-2.4 >2.4

• 15
• 14
• 13
• 12
• 11
• 10
• 9
• 8
• 7
• 6
• 5
• 4
• 15 -14 -13 -12 -11 -10 -9
• 8
• 7
• 6
• 5
• 4

BACE CDK2 JNK1 MCL1 P38 PTB1B THROM TYK2

ΔG FEP (kcal/mol) ΔG Expt. (kcal/mol)

FEP+ — L Wang, et al. JACS., 137:2695–2703. (2015) OPLS3 — E Harder, et al. JCTC., 12:281–296. (2016)

Referenced paper: L Wang, et al. JACS., 137:2695–2703. (2015)

• Our goal is not merely to

make accurate predic1ons

• Rather, our goal is to drive

discovery projects forward

• Can we do this?

Collabora'on A—Extremely rapid op'miza'on

• f a new lead compound

Rapid op'miza'on of a HTS hit with FEP+

• A high-throughput screen iden1fied a weak inhibitor (80 μM)
• f a high-value target
• Project team needed to decide if the inhibitor could be further
• p1mized to progress the project from lead iden1fica1on to

lead op1miza1on

• Collaborator had synthesized 73 molecules to improve the

potency of the HTS hit, but none were sa1sfactory

Rapid op'miza'on of a HTS hit with FEP+

• A crystal structure of the inhibitor showed a highly unusual

binding mode

• Collaborator suspected that unusual binding mode might give

the inhibitor high specificity, if potency could be op6mized

• Schrödinger was given 1 week to determine if there was any

viable route to improve the molecule

Rapid op'miza'on of a HTS hit with FEP+

• Using in-house and cloud compu1ng resource ~3500 synthe1cally

plausible deriva1ve molecules were scored

– ~100K GPU hours

• Only 23 of these molecules (0.6%) were predicted by FEP+ to boost

the binding potency of the HTS hit

• The collaborator chose to ini1ally only synthesize 3 of the

recommended 23 molecules

– Synthesis costs, if the molecules are challenging, can easily exceed $5,000 per molecule – Synthesizing all 3,500 would not be feasible for most discovery projects

• Expt. Ki (μM)

26* 3.8 2.2 80

Rapid op'miza'on of a HTS hit with FEP+

Molecule R-group FEP+ Ki (μM) FEP+ rank MMGBSA rank A 0.6 1 427 B 1.8 2 95 C 3.2 3 435 HTS hit

• * Mixture with 3 stereocenters, frac1on of eutomer may be << 1/8

Rapid op'miza'on of a HTS hit with FEP+

• Using FEP+ to guide compound synthesis, the potency of the

HTS hit was improved 40x in a single round of chemistry

• Using simple MM-GB/SA scoring the top 435 compounds

would have needed to be synthesized to recover these top 3 compounds

• Further op1miza1on of molecules B and C is now also in

progress

Collabora'on B—simultaneous op'miza'on of potency, selec'vity, and solubility

FEP+ use case from discovery collabora'on B

• Prior to the ini1a1on of a large-scale FEP screening

campaign (June 2015), no molecules had been iden1fied which simultaneously achieved high potency, selec1vity, and solubility

– Many molecules achieved par1al success, but no molecules were sa1sfactory across all four criteria

• Star1ng in June 2015, an unprecedented FEP

scoring campaign was ini1ated to find sa1sfactory molecules

– ~4000 molecules scored by FEP to date (April 2016) – Equivalent to > 5 years of wet-lab experimental chemistry to test all scored idea molecules

• Goal was to marry expert molecular design with

predic1ve scoring to enable the undertaking of challenging synthe1c targets

Expt. End Point Mol. A Mol. B Mol. C Mol. D Potency (pKi > 9)

✔ ✔ ✔ ✔

Selec'vity 1 ( > 100x )

✔ ✔ ✗ ✗

Selec'vity 2 ( > 100x)

✔ ✗ ✔ ✔

Solubility ( > 20 uM)

✗ ✔ ✔ ✗

FEP+ use case from discovery collabora'on B

• The FEP scoring campaign has been hugely successful, with 10 molecules now simultaneously

mee1ng potency, selec1vity and solubility goals:

• 46 FEP-r

46 FEP-recommended compounds wer ecommended compounds were synthesized and 9 checked all four boxes e synthesized and 9 checked all four boxes

Molecule FEP Recommended Date Synthesized Potency (pKi > 9) Selectivity 1 ( > 100x ) Selectivity 2 ( > 100x) Solubility ( > 20 uM)

• Mol. E

Yes 11/4/15

✔ ✔ ✔ ✔

• Mol. F

Yes 12/10/15

✔ ✔ ✔ ✔

• Mol. G

Yes 12/10/15

✔ ✔ ✔ ✔

• Mol. H

Yes 12/17/15

✔ ✔ ✔ ✔

• Mol. I

No (Charged) 12/24/15

✔ ✔ ✔ ✔

• Mol. J

Yes 12/28/15

✔ ✔ ✔ ✔

• Mol. K

Yes 1/15/16

✔ ✔ ✔ ✔

• Mol. L

Yes 1/21/16

✔ ✔ ✔ ✔

• Mol. M

Yes 1/21/16

✔ ✔ ✔ ✔

• Mol. N

Yes 3/16/16

✔ ✔ ✔ ✔

FEP+ use case from discovery collabora'on B

• The parent compound (Mol. D) of molecules E, F, and G was neither highly

selec1ve nor soluble:

• Molecules A, B and C were from a different subseries, and could not be

related to molecules D, E, F or G through FEP+ perturba1ons

– Parallel large-scale op1miza1on of mul1ple subseries was essen1al to the iden1fica1on of molecules E, F, and G – These molecules were synthe1cally challenging due to the chemistry of the core, and were unlikely to have been synthesized without FEP+ scoring

Core Core Core

• Mol. D:

pKi > 9

• Selec. < 100x
• Solub. < 10 µM
• Mol. E, F and G:

pKi > 9

• Selec. > 100x
• Solub. > 20 µM

Core

FEP+ should dras'cally expand the chemical space explored by discovery project teams

• True impact of FEP+ will not be seen un1l investment in FEP is

comparable to other drug discovery project costs

– Ideally, many thousands of FEP calcula1ons per project – In effect, one is able to run a ~100,000 compound medicinal chemistry project in less than a year – 100x increase in throughput over a typical discovery project

• Such predic1ve scoring should enable much more rapid
• p1miza1on and balancing of ligand proper1es than would be
• therwise possible in lead op1miza1on
• Increasing the success of preclinical drug discovery may be the

most promising avenue to be]er meet the urgent need for more effec1ve drug therapies

Acknowledgements

Applica'ons Sciences Thijs Beuming Daniel Cappel Osamu Ichihara Roy Kimura Ana Negri Daniel Robinson Woody Sherman Thomas Steinbrecher Leadership Ramy Farid Drug Discovery Group Leah Frye Sarah Boyce Mark Brewer Sathesh Bhat Jonathan Gable Jeremy Greenwood Kyle Konze Shaughn Robinson Markus Dahlgren Fiona McRobb Scien'fic Development Wolfgang Damn Yuqing Deng Ed Harder Joe Kaus Byungchan Kim Jen Knight Goran Krilov David Lebard Dima Lupyan Sayan Mondal Lingle Wang Chuanjie Wu Yujie Wu Yutong Zhao Chongkai Zhu Scien'fic Advisors Bruce Berne John Chodera Rich Friesner Bill Jorgensen David Mobley Vijay Pande Mark Murcko

• D. E. Shaw Research

Michael Bergdorf Jus1n Gullingsrud Ross Lippert Charles Rendleman Danielle White Huafeng Xu

And enormous thanks to our collaborators!

Q / A