SAMPL6pKa Challenge

MehtapIsikChoderaLab,MSKCC

D3R2018WorkshopFebruary22nd,2018

0

Why did we decide to organize a blind pKa prediction challenge?

SAMPL5 logD challenge indicated theimpact of prediction of the ionization statedistribution on the accuracy of logDpredictions.

Pickard,F.C.etal.BlindpredictionofdistributionintheSAMPL5challengewithQMbasedprotomer andpKa corrections.JournalofComputer-AidedMolecularDesign30,1087–1100(2016).

cyclohexane

1

pKa predictions contribute to the errors in binding free energy predictions.

Case1:FailingtocorrectbindingfreeenergywithpKa penalty

P:LH+ LH+

L

ΔGbind,LH+

ΔGprotΔGbind =?

dominantaq.microstate

ΔGbind =ΔGbind,LH++ΔGprot

1.38kcal/mol errortofreeenergyper1unitoferrorinpKa

Case2:Failingtopredictligandprotonationstateinthecomplex

P:LH+ LH+

L

ΔGbind,LH+

ΔGprot

dominantaq.microstateP:L

Modelingthewrongionizationstateincomplexcancausesevereerrors.

ΔGbind,L

2

Accurate prediction of pKas is a useful tool for computer aided drug design and lead optimization.

DuringleadoptimizationpKa valuesguide§ Improvingtargetpotency§ Reducingpotencyagainstundesiredtarget§ Modulatingsolubilityandlipophilicity§ ImprovingADMEproperties

PredictingpKas ofofdrug-likecompoundsarechallengingdueto

§ multipleprotonationsites

§ conjugatedsystemsandheterocycles

2nmicrostates

3

We selected fragment-like molecules with heterocyclescommon in kinase inhibitors for SAMPL6.

Startingpoint:ZINC15kinasesubsetandanodyne compounds

Frequentheterocycles foundinFDA-approvedkinaseinhibitors

Murcko ringfragmentation N

pyridine

N

N

quinazoline

N

quinoline

N

N

pyrimidine

NHN

indazoleN

HN

imidazole

NHN

pyrazole

Otherheterocycles intheSAMPL6set

S

N

NH2

aminothiazole

S

thiophene

S

NN

1,3,4-thiadiazoleS

HN

OO

5-methylenethiazolidine-2,4-dione 4

24 compounds are present in SAMPL6 pKa challenge

5

In multiprotic compounds it is important to differentiate between macroscopic and microscopic pKas.

Fraczkiewicz R.(2013)InSilico PredictionofIonization.In:Reedijk,J.(Ed.)ElsevierReferenceModuleinChemistry,MolecularSciencesandChemicalEngineering.Waltham,MA:Elsevier.RuppM.etal.PredictingthepKa ofSmallMoleculesCombinatorial Chemistry&HighThroughputScreening,2011,14,307-327.

MacroscopicpKas

UV-metricpKasmayfailtocaptureallmacroscopicpKas.

MicroscopicpKas

+2+10-1

6

Experimental pKa values for SAMPL6 were measured with Sirius T3.

§ Method:UV-absorbancespectrabasedpKa measurement§ Measurementrange:2-12§ 24smallkinaseinhibitorfragment-likemolecules§ Temperature:25°C§ Ionicstrength:150mM KCl solution§ 3independentreplicates(fromthesameDMSOstock)

DorothyLevorseTimothyRhodes

7

PyridoxineHCl

UV-metric pKa measurements of multiprotic compounds lead to determination of macroscopic pKa values.

8

pKas of water insoluble compounds were determined by extrapolation from multiple cosolvent experiments.

pKa isdeterminedbyYasuda-Shedlovsky extrapolation to0%cosolvent.

Acid/baseassignmentbasedonpKa shiftwithcosolvent doesnotprovidereliableevidenceforassigningpKa valuestoionizable groups,especiallyinmultiproticcompounds.

ApparentpKa ismeasuredat3cosolvent concentrations:30%,40%,and50%MeOH

Avdeef,A.etal.PH-metriclogP 11.pK adeterminationofwater-insolubledrugsinorganicsolvent–watermixtures.Journalofpharmaceuticalandbiomedicalanalysis20,631–641(1999). 9

Suggestions for future pKa experimental data collection

§ ForfuturepKa challengeswithmultiprotic compounds,itisidealtouseexperimentalmethodsthatcanmeasuremicroscopicpKas,suchasNMR.

§ UV-metricpKa measurementswithSiriusT3donotprovideanystructuralinformationaboutmicrostates.

§ Monoprotic compoundsshouldbeprefered ifUV-metricorpotentiometricmethodsforpKa measurementswillbeused.

§ Acid/baseassignmentbasedonpKa shiftwithcosolvent isnotreliableinmultiprotic compounds.

§ Compoundpurityiscriticalforaccuracy.

§ CompoundsolubilityisthelimitingfactorforpKa measurementswithSiriusT3.

10

TypeI- microscopicpKas andmicrostatesPredictingmicroscopicpKa valuesandrelatedmicrostatestructures.

32submissions

TypeII- microstatepopulationsasafunctionofpHPredictingfractionalmicrostatepopulationsbetweenpH2to12in0.1pHincrements.

27submissions

TypeIII- macroscopicpKasPredictingthevalueofmacroscopicpKas between2and12.

34submissions

Submission types and participation to pKa prediction challenge

11

Analysis of macroscopic pKa predictions requires mapping of experimental pKas to predicted pKas.

ClosestMethod§ EachpredictedpKa ismatchedtoclosestexperimentalpKa value(minabsoluteerror).§ WhenmorethanonepredictedpKa matchtothesameexperimentalpKa,onlythe

predictedpka thathasthelowestabsoluteerroriskept.§ ExtrapredictedorexperimentalpKas areignored.

12

HungarianMethod§ ExperimentalpKas andpredictedpKas arematchedfollowingHungarianalgorithm.§ Optimumglobalassignmentthatminimizeslinearsumofsquarederrorsofallpairwise

matches.

Kiril Lanevskij

ExperimentalpKas2.159.5811.02

PredictedpKas0.51.84

ExperimentalpKas2.159.5811.02

PredictedpKas0.51.84

Overall performance of macroscopic pKa predictions

SubmissionID

RMSEvaluesspantherangeof0.7-5pKa units.

13

Analysis of pKa predictions and future directions

MacroscopicpKa analysisresultscanbefoundinSAMPL6GitHub repository:

https://github.com/MobleyLab/SAMPL6

– Experimentalvs predictedpKa valuecorrelationplots– Errordistributionplotsforeachmolecule– Performancestatistics(RMSE,MAE,ME,R2,slope)

WewillkeepupdatingtheSAMPL6repositorywith:• AdditionalperformancecriteriaforpKa predictions• AnalysisofmicroscopicpKa valuesandmicrostatepopulations

14

4 participants of the pKa challenge will present us their perspectives.

SubmissionID15

Qiao Zeng

NEXTSamarjeet Prasad(TypeI)

MarvinWaldman

Bogdan IorgaTOMORROW

SAMPL6organizersandadvisors

DavidMobleyJohnChoderaAndreaRizziCaitlinBannanBasRustenburgMichaelChiuMichaelGilsonMichaelShirtsPaulCzodrowski

MerckPreformulation Department

TimothyRhodesDorothyLevorseBradSherborneHeatherWang

Acknowledgments

ParticipantsofpKa challengeCaitlinBannanRobertFraczkiewiczBogdan IorgaStefanKastKiril LanevskijChrisLoschenPhilippPrachtSamarjeet PrasadGeoffSkillmanRainerWilckenQiao Zeng

Tri-InstitutionalPhDPrograminChemicalBiologyDorisJ.HutchisonFellowship

16

Chemical and Availability Criteria

Fragment-like Drug-like

• Tier1 compound• Availability of least 100 mg• Cheaper compounds in logP bins are prioritized.• Non-hazardous.• Anodyne (PAINs and reactive groups removed)

• At least 1 pKa in the interval 3 ≤ pKa ≤ 11• Multiple pKa’s at least 1 log unit part in selected pKa interval.• Minimum number of UV-chromophore unit: 8 • -1 < XlogP < 6

Number of rotatable bonds ≤ 3• 150 ≤ mw < 350

• Number of rotatable bonds ≤ 8• 350 ≤ mw ≤ 500

SMARTS: [n,o,c][c,n,o]cc

Selecting kinase inhibitor-like compounds

17

pKas of water insoluble compounds were determined by extrapolation from multiple cosolvent experiments.

50%MeOH

40%MeOH

30%MeOH

pKa isdeterminedbyYasuda-Shedlovskyextrapolation to0%cosolvent.

Acid/baseassignmentbasedonpKa shiftwithcosolvent doesnotprovidereliableevidenceforassigningpKa valuestoionizable groups,especiallyinmultiprotic compounds.

18

Analysis of macroscopic pKa predictionsOverall performance of macroscopic pKa predictions

19