logd and pka enhancements: applications in adme profiling€¦ · · 2011-02-09logd and pka...
TRANSCRIPT
LogD and pKa enhancements: Applications in ADME profiling
ACD UK users meeting, Windsor, 2007
Jordi Munoz-Muriedas, GSK, Stevenage
logD and pKa in ADME
Related with almost every ADME process
Increases absorption, volume of distribution, plasma protein binding, Pgpefflux, CNS penetration Cytochrome inhibition, intrinsic clearanceDecreases solubility
Hydrophobicity generally:
Acidity generally:
Increases plasma protein binding, solubilityDecreases in vivo clearance, CNS penetration, PGP efflux, volume of distribution
Basicity generally:
Increases in vivo clearance, solubilityDecreases plasma protein binding
pKa predictors performance
ACD pKA vs. other software (n=1000)
0
20
40
60
80
100
+/-0.25 +/-0.5 +/-1 +/-2 >+/-2 not_calcMeasured - Calculated pKa within specified range
Cum
ulat
ive
%
external 1 ACD_v4.1 external 2 ACD_v7 external 3
logD predictors performance
ACD logD vs. other software (n=4693)
0
10
20
30
40
50
60
70
80
90
100
+/-0.5 +/-1 +/-2 >+/-2 not_calcMeasured - Calculated logD within specified range
Cum
ulat
ive
%
in-house method ACD_v4.1_logd external 1 ACD_v7_logd external 2
logD predictors performance
0
20
40
60
80
100
"+/- 0.5" "+/- 1.0" "+/- 2.0" "+/- >2.0" "not_calc"
ACD v8 ACD v9
ACD logD prediction: logD range [-1.5,3.0] n=17006
logD predictors performance
ACD logD prediction n=22903
Quick implementation in ADMET models
Absorption predicted by logD/CMR model
-12
-8
-4
0
4
8
0 5 10 15 20 25cmr (Daylight)
AC
D_L
ogD
pH
7.4
OK Poor
Quick Quick implementation in ADMET modelsin ADMET models
Necessity of good logD/pKa predictions
Expt Absorb: High
Pred. Absorb: Fail (high confidence)
Expt. logD= 1.89
ACD logD= -1.5
Expt Absorb: High
Pred. Absorb: Fail (Borderline)
Expt. logD= 0.75
ACD logD= -3.8
5 10 15 20
-2
-1
0
1
2
3
4
CMR
Exo.
logD
Colours based on predicted Absorption:
Green: Good
Yellow: Good (borderline)
Orange: Poor borderline
Red: Poor
Training Training pKapKa, , logPlogP and and logDlogD
ACD physchem and pKa accuracy extenders
pKalogPSolubility
pKA and logP trained values can be used together to increase accuracy in logD predcition
pKa training
pKa training
ApKa1 Prediction
62
84
93100
40
62
89
100
0
10
20
30
40
50
60
70
80
90
100
<0.5 <1 <2 >=2
Error (log units)
Cum
ulat
ive
%
V9_UserV9
Global dataset training
BpKa1 Prediction
55
71
90
100
49
67
90
100
0
10
20
30
40
50
60
70
80
90
100
<0.5 <1 <2 >=2
Error (log units)
Cum
ulat
ive
%
V9_UserV9
N=369 N=684
pKa training: Basic compounds
0 2 4 6 8 10 12 14
2
4
6
8
10
12
Global dataset training: Basic pKas
K b10 2 4 6 8 10 12 14
2
4
6
8
10
12
Before Training After training
pKa training
Most acidic pKa: Measured vs. ACD v9
0 2 4 6 8 10 12 14
2
4
6
8
10
12
Most acidic pKa: Measured vs. ACD v9_user
K 10 2 4 6 8 10 12 14
2
4
6
8
10
12
Acidic compounds: New reaction center
R2R1N
OOH
R3
Before Training
After Training
pKa training
Expt_LogD vs. CMR
5 10 15 20
-2
-1
0
1
2
3
4
R2R1N
OOH
R3
Expt Absorb: High
Pred. Absorb: Fail (Borderline)
Expt. logD= 2.25
ACD logD= 1.20
ACD logD (pKa trained) = 2.84
New absorption prediction: High
Acidic compounds: Absortion predictionColours based on predicted Absorption:
Green: Good
Yellow: Good (borderline)
Orange: Poor borderline
Red: Poor
cmr
Exp_
logD
pKa training
logD correction using trained pKas (n=17006)
LogD prediction
37
64
90
100
36
63
90
100
0
10
20
30
40
50
60
70
80
90
100
<0.5 <1 <2 >=2
Error (log units)
Cum
ulat
ive
%
V9_UserV9
pKa training
Solubility correction using trained pKas (n=909)
40.0%
50.0%
60.0%
70.0%
80.0%
90.0%
100.0%
Error<0
.5Erro
r<1.0
Error<1
.5Erro
r<2.0
Error<3
.0Erro
r>3.0
v9v9_pKa_user
100.0%96.4%87.6%78.2%63.9%42.5%
909876796711581386v9_pKa_user
100.0%95.4%86.2%76.7%63.3%42.4%
909867784697575385v9
Error>3.0
Error<3.0
Error<2.0
Error<1.5
Error<1.0
Error<0.5
100.0%96.4%87.6%78.2%63.9%42.5%
909876796711581386v9_pKa_user
100.0%95.4%86.2%76.7%63.3%42.4%
909867784697575385v9
Error>3.0
Error<3.0
Error<2.0
Error<1.5
Error<1.0
Error<0.5
Error Sol(pH7)-Error Sol(pH1.2) / Error Sol(pH7)*- Error Sol(pH1.2)*
pKa training
Application to projects
ACD K i6.5 7 7.5 8 8.5 9 9.5 10 1...
6.5
7
7.5
8
8.5
9
9.5NR1
R2
6.5 7 7.5 8 8.5 9 9.5 10 1...
6.5
7
7.5
8
8.5
9
9.5Before training After training
pKa= 1.62pKa_acd -7.00
R2=0.72pKa= 1.07pKa_acd – 0.57
R2=0.87
N=16
pKa training
XX
X
XX
NR2
R1
5.8 6 6.2 6.4 6.6 6.8 7 7.2
6
6.5
7
7.5
Application to ADME
Bioavailability and CYP inhibition issues related with the pKa of the imidazopyridine groupPoor pKa prediction by ACDResults improved significantly after pKa AE training
pKa= 0.09pKa_acd + 6.03
R2=0.02
5 6 7 8 9 10
6
6.5
7
7.5
pKa= 1.14pKa_acd – 0.93
R2=0.63
Validation set
N=28
Before training After training
pKa training
2C19 pIC50 vs. Basic pKa (n=77)
4.5 5 5.5 6 6.5 7 7.5
4
4.5
5
5.5
6
6.5
7
4 5 6 7 8 9 10
4
4.5
5
5.5
6
6.5
7
5 5 5 6 6 5 7
4
4.5
5
5.5
6
6.5
7
R2=0.40 R2=0.09 R2=0.33
Experimental pKa Before Training After Training
Expt pKa ACD pKa ACD trained pKa
pIC
50
“logD training”
logD training
-1 0 1 2 3 4
-1
0
1
2
3
-2 -1 0 1 2 3
-1
0
1
2
3
Using logP training to improve logD prediction
N
NHR1
R2
Expt LogD = 0.68 ACD logD + 0.23R2 = 0.39Average error 0.17Overprediction
Expt LogD = 0.69 ACD logD + 0.57R2 = 0.64Average error -0.35Underprediction
N=445
Training process leads to a better correlation but to a systematic underestimation of logD
logD training
Exp LogD vs T rained ACD logD
-2 -1 0 1 2 3
-1
0
1
2
3
Relating error with specific chemical groups
N
R2O
NHR1
N=26
Red: Amides
logD training
Relating error with specific chemical groups
-1 0 1 2 3 4
-1
0
1
2
3
-2 -1 0 1 2 3
-1
0
1
2
3
-1 0 1 2 3 4
-1
0
1
2
3
-2 -1 0 1 2 3 4
-1
0
1
2
3
N
R2O
NHR1
logP underpredicted for series with amidesTraining process corrected this partially
ACD_logD ACD_logD
ACD_logP ACD_logP
Exp_
logD
Exp_
logD
Exp_
logD
Exp_
logD
Before Training After Training
logD training
N
R2O
NHR1
-2 -1 0 1 2 3
-1
0
1
2
3
-1 0 1 2 3 4
-1
0
1
2
3
Relating error with pKa prediction
The training process leads to a underprediction of the logP. Probably due to a underprediction of the basicityof the compounds
ACD_logD ACD_logD
ACD_logP ACD_logP
Exp_
logD
Exp_
logD
Exp_
logD
Exp_
logD
Before Training After Training
-1 0 1 2 3 4
-1
0
1
2
3
-2 -1 0 1 2 3 4
-1
0
1
2
3
N=419
logD training
pKa underprediction
7 7.5 8 8.5 9 9.5
7.6
7.8
8
8.2
8.4
8.6
8.8
9
9.2
N
NHR1
R2
Y=0.37x+5.46
R2=0.26
pKa influence in logP training
Experimental logD and predicted ACD pKa values are used to calculate estimation of “experimental” logPThese estimations are used to train logP fragmentsTrained logP can be used to calculate logDLogD predictions can be enhanced by using (indepently) trained pKasProblem: Errors in pKa prediction in the first step will affect logP training and logD prediction
It should be possible to introduce Experimental logD and experimental pKa (or at least trained pKas) to get a better prediction of logP in the first stepThis logP trained using pKa information and trained pKa should lead to a better logD prediction
∑∑=
−
Δ−
=−
−Δ
⎟⎟⎠
⎞⎜⎜⎝
⎛
++−⎟
⎟⎠
⎞⎜⎜⎝
⎛
++−=
Nb
ipHBpKa
BpKaNa
iApKapH
ApKa
i
Bii
i
iAi
PD11 101
101log101101logloglog
Current algorithm:
Ideal:
Conclusions
pKa accuracy extender is a very useful tool to enhance pKapredictions when applied to improve predictions for local sets.
pKa AE power would be increased and be of great help if it could beused simultaneously with the logP training algorithm.
logP training algorithm is very dependent of pKa prediction and should not be used with datasets where errors in pKa are expected unless experimental logP values could be provided.
Ackowledgements
ADMET In Silico Team, GSK, Stevenage:Anne Hersey, Paul Gleeson, Sandeep Modi
ACD Labs: Ed Kovolanov, Greg Pearl