Structures, Molecular Descriptors, Model Development and Biopharmaceutical Property Estimation

Size: px

Start display at page:

Download "Structures, Molecular Descriptors, Model Development and Biopharmaceutical Property Estimation"

Irene Collins
6 years ago
Views:

1 Structures, Molecular Descriptors, Model Development and Biopharmaceutical Property Estimation Michael B. Bolger, Ph.D. Chief Scientist Simulations Plus, Inc. utline Biopharmaceutical Properties (What s important?) Human Jejunal Effective Permeability (Peff), Cell Culture Permeability (Papp) pka, log P, ative Solubility, Salt Solubility BBB penetration, Plasma Protein Binding, Volume of Distribution Estrogen receptor binding, carcinogenicity, mutagenicity, herg IC 50 Structures and Molecular Descriptor Generation Selected Biopharmaceutical Property Models logp Permeability Solubility (and pka and ph dependence) 1

2 Steps in Model Building utline Descriptor Pruning Selection of Training, Verification, and Test datasets. Model Architecture Descriptor Sensitivity Model Training Applicability Domain / ptimum Prediction Space Computational Alerts Rule of 5 ADMET Risk Quantitative Structure-Property Relationships and GI Simulation Models QSPR => Peff & Papp Solubility Diffusivity Frac. Unbound GI Simulation Reproduced by permission of Gordon Amidon & Sally Choe 2

GAL: Virtual ADME Screening IPUT Cl H 3 C 2D or 3D STRUCTURE Predictor: molecular descriptors 1D (molecular formula is sufficient) 16 (MWt, _Atoms, _Halogen, M_,.

3 GAL: Virtual ADME Screening IPUT Cl H 3 C 2D or 3D STRUCTURE Predictor: molecular descriptors 1D (molecular formula is sufficient) 16 (MWt, _Atoms, _Halogen, M_,...) 2D (connectivity table required) 206 (_FrRotB, PriAmine, HBD, Χ0...) 3D (Cartesian atom coordinates required) 24 (RgGrav, PolASA, SolvE, Dipole,...) Predictive Modeling: A Continuous Cycle Compound Database Candidate Compounds Evaluation Experiments Descriptor Generation Training Data Build Descriptor Generation Predict Model f(x) Single f n (x) f(x) f(x) f 1 (x) or Ensemble utput Candidate Predictions Good compound Poor compound Good compound Averaged utput 3

Key Biopharmaceutical Properties Partition Coefficient (oil / water) Solubility (how much can dissolve) Dissolution (how fast it dissolves) f(conc., Solub., Diffusion.

Fraction Absorbed to Portal Vein f (Dissolution, Permeability, Solubility, Formulation) Distribution vs.

4 Key Biopharmaceutical Properties Partition Coefficient (oil / water) Solubility (how much can dissolve) Dissolution (how fast it dissolves) f(conc., Solub., Diffusion., Density, Particle size) Permeability (how fast it gets absorbed) Human Peff (rate cm/s) Cell Culture - [Caco-2, MDCK] Papp (cm/s) Transcellular, Paracellular, Aqueous Boundary Fraction Absorbed to Portal Vein f (Dissolution, Permeability, Solubility, Formulation) Distribution vs. Partition Coefficient Hydrophobic phase ctanol ph = 2 log D ~ log P = 2 H Hydrophilic phase H 2 ph H 3 C ph = 8 Log D = -1.5 pka = 3.5 LogP LogD oct / water = log oct / water = [ HA] [ HA] C log C oct water aspirin oct aspirin water Log P is defined for the unionized species only H 3 C 4

5 Distribution Coefficient Hydrophobic & H-Bonding ctanol CycloHexane Hydrophobic Hydrophilic H 2 ph H 2 Hydrophilic Log D oct = log [ct]/[h 2 ] Log D CHex = log [CHex]/[H 2 ] Log P is ion corrected for the unionized species Permeability Efflux Transporter Influx Transporter Passive Diffusion Paracellular Transport Tight Junction Blood 5

log P on-polar solvent accessible surface area Inversely proportional to H-bonding Count

6 Passive Permeability Transcellular and Paracellular H H + H 3 f (Hydrophobicity, H-Bonding, Ionization) Passive Transcellular Permeability Proportional to lipophilicity log P on-polar solvent accessible surface area Inversely proportional to H-bonding Count of total number of and H-bond donors and acceptors Sum of partial charges on H-bonding groups. 6

7 Human Effective Permeability Cross-section Loc-I-Gut Perfusion utlet Bile Guide wire Balloon Inlet Balloon Bile Balloon Jejunal perfusion Lennernas, H, et al. Pharm. Res. 9: (1992) Epithelial Cell Permeability Assay Li, A.P., DDT, 6(7):

Paracellular Permeability f (Size, Shape, Ionization) H S H H + H 3 H H Pore size = 2.3-11.

1 e m ) + (1 f κ κ κ e 1 Ref. A. Adson et al., J. Pharm. Sci., 83(11):1529 (1994) u ) P ± para P e =Effective Perm.

8 Paracellular Permeability f (Size, Shape, Ionization) H S H H + H 3 H H Pore size = Å (Colon < Caco-2 < MDCK < Jejunum) Billich, (1969) J Clin Invest 48(7): P P P P P e m Calculation of Paracellular Perm. = = o para + para para 1 P f abl u + ( P εdf r = δ εdf r = δ εdf r = δ 1 P o trans f + + P r R r R r R 1 P o para κ 1 e m ) + (1 f κ κ κ e 1 Ref. A. Adson et al., J. Pharm. Sci., 83(11):1529 (1994) u ) P ± para P e =Effective Perm. P abl =Aqueous Boundary P f = Filter P m = Membrane P trans = Transcellular P para = Paracellular ε/δ =Porosity / Pore Length = 1.22 D = Diffusion Coefficient f r (r/r) = Renkin equation κ = Electrochem. E. (0.6) Cation = 1.33 Anion =

9 % Paracellular vs. Total Permeability Paracellular Papp (% of Total) MDCK Permeability (cm/s x 10^4) Methods for Solubility Measurement ative Solubility (Best for modeling) Equilibrium Thermodynamic solubility in pure water. ph adjusts depending on saturated concentration and pka of Acids or Bases. Solubility in Buffers (Best for salts) Equilibrium Thermodynamic solubility at a known ph. Molecule could be ionized 0% - 100%. HTS-turbidity: (Rank order of solubility only) Precipitation of molecule from solution in organic solvent by adding ph 7 buffer. Hendriksen BA, Sanchez MF, and Bolger MB: (2003) The Composite Solubility Versus ph Profile and GPE its Kansas Role in 2006 Intestinal Absorption Prediction. AAPS Pharm. Sci. 5(1):Article 4,

Fallacy of trying to model solubility from data collected at fixed ph (eg. 2 or 7.4) Solubility (mg/ml) 1.0E+02 1.0E+01 1.0E+00 1.0E-01 1.0E-02 1.0E-03 1.0E-04 1.0E-05 1.0E-06 1.

10 Fallacy of trying to model solubility from data collected at fixed ph (eg. 2 or 7.4) Solubility (mg/ml) 1.0E E E E E E E E E E-07 Amiodarone and Promethazine Solubility Titration Solubility Factors = 7.56E+6 and 6.88E+2 Amiodarone Promethazine log P = 6.96 log P = ph H 3 C H 3 C I H 3 C CH 3 Data: Courtesy of Christel Bergstrom and Per Artursson I S Biopharmaceutical Classification (BCS) F a = f(permeability and Solubility) Hendriksen GPE Kansas BA, 2006Sanchez MF, and Bolger MB: (2003). AAPS Pharm. Sci. 5(1):Article 4,

11 Estimating Biopharmaceutical and Pharmacokinetic Properties Definition of molecular descriptor The molecular descriptor is the final result of a logic and mathematical procedure which transforms chemical information encoded within a symbolic representation of a molecule into a useful number or the result of some standardized experiment. Roberto Todeschini 11

12 ADMET Predictor Descriptors Classified by Chemistry Simple Constitutional Descriptors 27 (MWt, _Atoms, _Bonds, _Rings,...) Heteroatomic Functional Groups 79 (AlHdrxl, Sulfonmd, ArCbxyl, Barbitur,...) Topological Indices 3 (X0, X1, Wiener) Atom-type Electrotopological State Indices 1 68 (SsCH3, Sds, SsH, SddssS, SHsH2,...) Ionization in Water 10 (_IoAcAt, FAnion, FCation, FZwitter,...) Molecular Pattern Flags 5 (AlphaAA, AlphaAE, Steroid,...) Electronic Properties 8 (ABSQ, MaxQ, Dipole, PolarizG,...) 1. Kier, L.B.; Hall, L.H.; Molecular Structure Description ; Academic Press; ADMET Predictor Descriptors Classified by Chemistry Hydrogen Bonding 14 (HBA, HBD, HBAch, IHB,...) Molecular Size and Shape (3D) 14 (RgGrav, DStokes, TotASA, MIRxx,...) Solvation Effects (3D) 8 (PolASA, SolvE, SolvEMt, HBAwsa,...) Protein Recognition 1 5 (PEoED, PEoEDIa, PEoEDIb,...) Moriguchi Descriptors for MlogP 2 13 (M_CX, M_, M_PRX, M_UB,...) Meylan Flags for MH Sw Models 3 15 (H_AlAlco, H_AlPyri, H_Falkan,...) 1. Seelig, A., R. Gottschlich, and R.M. Devant; Proc. atl. Acad. Sci. USA, : p Moriguchi, I., et al., Chem. Pharm. Bull., : p Meylan, W.M., P.H. Howard, and R.S. Boethling; Env. Tox. Chem., : p

13 Descriptor Generation & Academic Prices Dragon ( ~$450 MolConn-Z ( ~$550 ME (CCG) ( ADMET Predictor / Modeler ( $3500 PreADMET( Free, one molecule at a time. Virtual Computational Chemistry Laboratory ( Free, plus it includes web-based model building software. Methods for Modeling Multiple Linear Regression (MLR - Classical QSAR) Simple, fast calculation Results have quantitative meaning to chemists Lack of interactions between independent variables Genetic Algorithm Partial Least Squares (GA-PLS) Good for small datasets when many descriptors are needed ot as robust as AE or SVM Artificial eural etwork Ensembles (AE) Highly accurate within the chemical space of training set Support Vector Machines (SVM) Excellent for classification problems Sometimes outperform AEs 13

14 Artificial eural etworks (A) MW DIFFUSIV MLALVL UMCX UM PRX UMUB ITRAHB PLARM AMPHTER ALKAEES UMRG UM2 UM1 UM2 ALIPHH ALIPHACD ALIPHAM PHEL AZ ITRILE HC ITR PARMHC MULTI AA SPHS How to avoid over-training. Training Set Verification Set Early Stopping 14

15 Partition Coefficient Estimation 3D S+log P Model (=9658) 11 9 Train/Verify Test Linear (Test) 7 bserved logp Train/Verify Set Test Set = 8362 = 1296 RMSE = RMSE = MAE = MAE = y = 1.013x y = 0.978x R^2 = R^2 = Predicted logp 15

16 Independent Comparison of log P models Permeability Estimation 16

17 Example of Simple MLR model for Human Peff log Simple Peff = 0.25 * MlogP * IHB * HBA 1.0 y = 1.00x R 2 = Log bserved Peff Log Predicted Peff Pause to build Peff Model using Excel 17

18 2D Model of Human Jejunal Permeability Train/Verify Test Linear (Train/Verify) bserved log(peff) Train/Verify Set = 35 RMSE = MAE = y = 1.058x R^2 = Test Set = 13-1 RMSE = MAE = y = 0.958x R^2 = Predicted log(peff) How well do S+Peff in silico results compare with in vitro measurements of permeability? MDCK P app Human P eff S+P eff Methyldopa Hydrochlorothiazide Hydrochlorothiazide Amoxicillin Furosemide Furosemide Terbutaline Methyldopa Amoxicillin Atenolol Amoxicillin Terbutaline Furosemide Terbutaline Atenolol Hydrochlorothiazide Atenolol Methyldopa Fluvastatin Metoprolol Metoprolol Ketoprofen Fluvastatin Piroxicam aproxen Propranolol Propranolol Verapamil Carbamazepine Fluvastatin Metoprolol Desipramine Verapamil Piroxicam Antipyrine Desipramine Carbamazepine Verapamil Carbamazepine Desipramine Piroxicam Antipyrine Propranolol aproxen Ketoprofen Antipyrine Ketoprofen aproxen 18

19 3D ative Solubility Model without Melting Point bserved log(solubility) [mg/ml] Train/Verify Test Linear (Test) Train/Verify Set = 1131 RMSE = MAE = y = 0.998x R^2 = Test Set = 206 RMSE = MAE = y = 0.974x R^2 = Predicted log(solubility) [mg/ml] Deardon J., Expert pinion Drug Discov. -1(1):31 (2006) 19

20 pka and ph Dependence of Solubility Predicting Salt Solubility Composite Solubility vs. ph Profile Base Solubility S WB = S B 10 pka ph ( 1+ ) ph of intersection between ionized and unionized when titrating from low to high ph ph = S pka log ConjA S B Hendriksen BA, Sanchez MF, and Bolger MB: (2003). AAPS Pharm. Sci. 5(1):Article 4,

21 Atomic Descriptors H 54 Special descriptors localized on a specified atom Partial charge, E-state, Access,... Excellent for predicting localized properties (pk a, specific binding, etc.) Prediction of multiprotic pk a with artificial neural networks H Cl 21

22 QMPRPlus A Models of Single pk a bserved pka Train/Verify Test Linear (Test) Test Set 9 = 288 RMSE = MAE = y = 1.015x R^2 = Train/Verify Set 3 = 1202 RMSE = MAE = y = 1.006x R^2 = Predicted pka bserved pka Train/Verify Test Linear (Test) Test Set = 127 RMSE = MAE = y = 0.996x R^2 = Train/Verify Set 4 = 162 RMSE = MAE = y = 1.028x R^2 = Predicted pka bserved pka 14 Train/Verify Test 12 Linear (Test) Test Set Train/Verify Set -4 = 524 = 2120 RMSE = RMSE = MAE = MAE = y = 1.002x y = 1.006x R^2 = R^2 = Predicted pka bserved pka Train/Verify Test Linear (Test) Test Set = 298 RMSE = MAE = y = 0.987x y = 1.019x R^2 = R^2 = Predicted pka Train/Verify Set = 1047 RMSE = MAE = D Model of pka (=10674 multiprotic pka) bserved pka Train/Verify Test Linear (Test) Train/Verify Set = 8531 RMSE = MAE = y = 1.016x R^2 = Test Set = 2143 RMSE = MAE = y = 0.998x R^2 = Predicted pka 22

23 Solubility Factor Intrinsic Solubility and Solubility Factor 1 μg/ml 1 ng/ml log Sol. Factor S+ Internal Data Bergstrom & Artursson 1 mg/ml log Intrinstic Sol (mg/ml) Data: Courtesy of Christel Bergstrom and Per Artursson SF ~ 20/Intrinsic Solubility in mg/ml Data: Courtesy of Christel Bergstrom and Per Artursson Propafenone (S+pKa = 9.3) Solubility Titration S+Solubility Factor = Solubility (mg/ml) BaseSw SaltSol Composite Propafenone QMPR Est. QMPRProfile H H ph 23

Steps in Model Building Analyze data set Identical or Low Variance Descriptors Under Representation High Correlation verlap Cluster data in descriptor space by Kohonen self-organizing map Divide into

Partition of training pool into different training and verification sets for every network in an ensemble Efficient partitioning algorithm (Tetko & Villa) Train matrix of artificial neural network

24 Steps in Model Building Analyze data set Identical or Low Variance Descriptors Under Representation High Correlation verlap Cluster data in descriptor space by Kohonen self-organizing map Divide into training pool and test set Sensitivity analysis to identify most influential descriptors for a particular network architecture Changing number of nodes will change most sensitive descriptors Partition of training pool into different training and verification sets for every network in an ensemble Efficient partitioning algorithm (Tetko & Villa) Train matrix of artificial neural network ensembles Select best architecture as final model Kohonen Self-rganizing Feature Map For Selection of Training, Verification, and Test The Kohonen self-organizing feature map (SFM) maps multidimensional data onto a 2-dimensional plane From Kohonen, T. (1984). Self-organization and associative memory. Springer Verlag. 24

25 H H 3 C H H H 3 C H H 3 C H H 3 C H H 3 C H 3 C H H H H H H HH H H H H H H H H H S H S S H H S H H H H H 3 C H C H Kohonen SM clusters like compounds H H H H H CH H 3 H H H CH H 3 H CH H H H H 3 H H H H CH H H 3 H H H H H H H H H H H H H H H CH 3 CH 3 H H H H H H SM of 351 compounds in MDCK P app data set CH H 3 CH H H CH CH 3 3 CH H CH H CH 3 3 CH H 3 H CH H H 3 H H H H H F H H F F CH H 3 C CH H 3 C CH H 3 C H H H H H H H H H H H H H H H H 3 C H 3 C S S Steroid cluster MDCK P app Low High Kohonen SM clusters like compounds H 2 H 2 H H H H S H H S H 2 S H 2 H 2 H H H SM of 351 compounds in MDCK P app data set H H H HS H HS CH 3 CH S 3 CH 3 H H H H 3 Lactam cluster MDCK P app Low High 25

26 Kohonen SM clusters like compounds x19 Kohonen map for log MDCK Papp data selects 74 compounds for test set, 187 for training pool CPU Time = 3 minutes umber of neurons containing compounds = 185 umber of training cases for genetic algorithm step = 185 umber of verify cases for genetic algorithm step = 102 umber of test cases sequestered = 74 Selecting Descriptors: Sensitivity Analysis Some descriptors have more influence on the predicted property than others. Train the model with a given architecture A: Sequentially remove each descriptor and calculate the increase in error for predictions. Rank descriptors in order of sensitivity 26

27 MDCK P app Sensitivity Analysis Training an Artificial eural etwork Descriptor 1 Descriptor 2 Descriptor 3 Descriptor 4 Descriptor 5 C1*D1 C13 (bias) C2*D2 C3*D3 C4*D4 C5*D5 C6*D6 C8*D1 C9*D2 C10*D3 C11*D4 C12*D5 1 = 1 / (1 + e Σ(C1*D1.. C6*D6)+C13 ) ode 1 ode 2 C15*1 C16*2 2 = 1 / (1 + e Σ[C8*D1.. C13*D6]+C14 ) (Predicted Property) utput = C15*1 + C16*2 Descriptor 6 C13*D6 C14 (bias) Training = Adjust C1-C16 until Σ(Predicted-bserved) 2 is minimized 27

28 How to avoid over-training. Training Set Verification Set Early Stopping Applicability Domain / ptimum Prediction Space Shen M., et al., J. Med. Chem. 45:2811 (2002) Gombar V.K., et al., J. Chem. Inf. Comput. Sci. 36:1127 (1996) 28

29 Computational Alerts Lipinski s Rule of Five (R5) Poor absorption is more likely when: Hb: HBD > 5 Mw: MWt > 500 LP: MlogP > 4.15 : M_ > 10 where: HBD = number of hydrogen bond donors MWt = molecular weight (in Daltons) MlogP = log P calculated by Moriguchi s method (Moriguchi; 1992) M_ = sum of nitrogen and oxygen atoms 29

$Rule of Five correlates poorly (at R 2 =0.171) with experimentally determined fraction absorbed. Data obtained from Zhao, Y. H.; Le, J.; et al.; J. Pharm.$ 406 PC(+2): MolQ > 0 where: MlogP = log P calculated by Moriguchi s method (Moriguchi; 1992) S+Peff = human jejunal permeability (Simulations Plus model;

406 PC(+2): MolQ > 0 where: MlogP = log P calculated by Moriguchi s method (Moriguchi; 1992) S+Peff = human jejunal permeability (Simulations Plus model;

30 Rule of Five correlates poorly (at R 2 =0.171) with experimentally determined fraction absorbed. Data obtained from Zhao, Y. H.; Le, J.; et al.; J. Pharm. Sci. 2001; 90(6): Default ADMET Risk Poor absorption is more likely when: LP: MlogP < Pr: S+Peff < Ha: HBAoch < PC(+2): MolQ > 0 where: MlogP = log P calculated by Moriguchi s method (Moriguchi; 1992) S+Peff = human jejunal permeability (Simulations Plus model; μm/s) HBAoch = partial atomic charge on H-bond accepting oxygens MolQ = number of functional groups bearing permanent charge, e.g., quaternary amine, sulfonium, diazo, etc. 30

31 Performance of ADMET Risk in the training set = 115, R 2 = Performance of ADMET Risk in the external test set = 71, R 2 = Data obtained from Zhao, Y. H.; Le, J.; et al.; J. Pharm. Sci. 2001; 90(6): Fa = 1 e 1.47* CalcPeff 108 Comp.: Calc Peff Eqn. nly 85% within 25% of observed bserved Fa Sw > 8 ug/ml Sw < 8 ug/ml Predicted Fa% Fa data from: Zhao et al., J. Pharm. Sci. 90:(6), 749 (2001) 31

32 Fa = 1 e 1.47* CalcPeff 37 Comp.: Calc Peff Eqn. nly 51% within 25% of observed bserved Fa Influx Substrates Efflux Substrates Predicted Fa% Transported Compounds Effluxed Pgp amiloride -tr ef etoposide -tr ef nadolol -tr ef norfloxacin -tr -ef progesterone -tr -ef quinidine -tr -ef ranitidine -tr -ef trovofloxicin -tr -ef verapamil -tr ef lovastatin tr -ef Influxed PepT1 amoxicillin -tr -if ampicillin -tr -if benzylpenicillin -tr -if captopril -tr -if carfecillin -tr -if cefadroxil -tr -if cefatrizine -tr -if cephalexin -tr if enalapril -tr -if fosinopril -tr if lisinopril -tr if phenoxymethylpenicillin -tr -if Influxed Amino Acid α-difluoromethylornithine -tr if glycine -tr -if levodopa -tr -if cycloserine -tr if Influxed ther Transporters bumetanide -tr -if loracarbef -tr -if mercaptoethanesulfonic acid -tr -if methotrexate -tr -if nicotinic acid -tr -if nizatidine -tr -if pravastatin -tr -if sorivudine -tr -if theophylline -tr -if trimethoprim -tr -if zidovudine -tr -if 32

33 ther Pharmacokinetic and Toxicology Models Modeling Software for Early Discovery: Structure-Property Modeling Which compounds in a library are obviously high risk? Very low solubility Very low permeability Ionization effects Saturable absorption at higher doses in human and laboratory animals Very high volume of distribution Very high protein binding High/low blood-brain barrier penetration Various types of toxicity Combinations of the above 33

34 Which Ketoprofen Analogs Can Be Eliminated? H H 2 H SP0001 SP0005 SP0009 S H H H H 2 H H H C 3 + H 3 C H 3 C SP0002 SP0003 H Cl H H SP0006 H 2 S SP0007 H H 2 Br SP0010 SP0011 S H 2 H H H 2 H H 3 C SP0004 Cl H 2 SP0008 Br SP0012 Which nes Can Be Eliminated? 34

35 Which nes Can Be Eliminated? MRTD < 2, Tox_ER > 0.002, Tox_FHM < 5, Tox_BRM < 50 Customized ADMET Risk for Rank rdering LP MlogP < Pr S+Peff < Ha HBAoch < C1 QuaAmine_>[+]< Sw S+Sp-oMP < 8e-3 Bb S+BBB = High T1 TX_MRTD < 2 T2 TX_ER > T3 TX_FHM < 5 T4 TX_BRM_Rat < 50 T5 TX_BRM_Sal = Positive Pb S+PrUnbnd < 1 Vd S+Vd > 15 35

36 Acknowledgments Robert Fraczkiewicz, Ph.D. (ADMET Predictor) Dechuan Zhang, Ph.D. (ADMET Modeler) Boyd Steere, Ph.D. (ADMET Modeler) John Rose, Ph.D. (GastroPlus) Thome Gilman, Pharm.D. (GastroPlus) Balaji Agoram, Ph.D. (GastroPlus) Jason Chittenden, M.S. (GastroPlus) Viera Lukacova, Ph.D. (GastroPlus PBPK) John DiBella, M.S. (DDDPlus) Anand Prabhakaran, M.S. (DDDPlus) Grace Fraczkiewicz, Ph.D. candidate (Databases) Walter S. Woltosz, M.S. (CE) 36

GI Simulation Methods

Gastrointestinal Simulation Based on the Advanced Compartmental Absorption and Transit (ACAT) Michael B. Bolger, Ph.D. Founding Scientist Simulations Plus, Inc. GI Simulation Methods Dispersion Model Ho,