Title: Designing for solubility when lowering log P isnt enough
1Designing for solubility - when lowering log P
isnt enough
- A lot of modeling is how much crap you can
take.Lauren Hutton - Logic is useful for proof, but almost never for
making discoveries. - Vilfredo Pareto
- a model, like a good cartoon, captures
important features of the reality it is meant to
portray. - Norbet Muller
2Many facets of solubility
-Bergström, et al. Eur. J. Pharm. Sci. 22 (2004)
387-398.
3Solubility model development schema
Iterative approach brings significant risk of
overfitting.
4Intrinsic Solubility Model
- Data selection
- Non-ionizable compounds from PhysProp
- 250 MW lt 750
- Reasonable structures only selected by hand.
- 199 training cmpds, 20 test cmpds
- Structures
- OEChem Canoncial Smiles, Omega conformer,
minimized with OPLS2005/water using Bmin. - Descriptors
- clogP,Volsurf, H-bond acceptors donors
- Feature Selection
- Simulated annealing with MLR/PRESS
5Intrinsic Solubility Model(Solubility of neutral
form outside of crystal lattice)
- Training Set R20.88, RMSe0.61, N199
- Validation Set R20.93, RMSe0.50, N20
- - constant not used when Xtal packing effects
included
6Intrinsic Solubility Model(Solubility of neutral
form outside of crystal lattice)
Pharmaceutically relevant range
- Training Set R20.77, RMSe0.59
- Validation Set R20.87, RMSe0.50
- - constant not used when Xtal packing effects
included
7Crystal Lattice Stability
- Hypothesis The extent of disorder of the unit
cell in response to increasing energy is
indicative of the stability of a crystal lattice. - Ansatz Effect is non-linear, with functional
form
Griseofulvin
8Crystal Packing Melting Point
Mult. Crystal forms 10 crystal forms of same
compound Project 1 5 compounds from three
chemotypes Project 2 6 compounds from two
chemotypes
9Error model for pH-solubility curve
- Prediction errors for various terms in model
- Uses 90 confidence interval of intrinsic
solubility - 0.5 pKa units from ACD/pKa or Epik. (arbitrary)
- Std. dev. of 3 runs for crystal packing effect.
- 0.2 error for std. dev. of ionization slope.
(arbitrary) - Monte Carlo error simulation
- Each error term is considered independent
normally distributed with the error limits
(either 90 conf interval or std dev.) defined
above. - Predicted solubility generated 1000 times at each
pH, report median 90 interval.
10Intrinsic Solubility Predictions
RMSe0.86 Median err0.41
Intrinsic solubility data from Wassvik, et al
(Euro. J. Pharm. Sci., 29, 2006, 294305) (N25)
11Solubility Prediction Glyburide
- Crystal packing simulation based on CSD entry
DUNXAL - Predicted Xpack1.30.26, log So-5.90.35,
pKa4.70.5 - Measured log So-7.1, pKa5.4
12Project Example
- Early members of lead chemotype are potent and
selective antagonists - Characterized by poor solubility that is not
tightly coupled to high logP - In vivo efficacy established
- 313 compounds with measured solubility
- 30 cmpds sol. gt 1 ?g/ml
- 2 with Ki lt 50 nM
13Cmpd-1 Solubility Profile
- Model highlights logP size as major factors
- Other parameters have important impact as well
- Lack of significant HPLC LogP correlation hinted
that crystal packing may play significant role as
well.
14Cmpd-1 Solubility Profile
HPLClogP gt 7 Exp. Sol lt 0.1?g/mL Pred. Sol
0.0005 ?g/mL
- Small molecule X-ray shows tight packing into the
pocket formed by aromatic rings. - MD simulations show this crystal form to be
highly stable.
Xpack1.97
15Cmpd-2
EWG replaced by aliphatic group
HPLClogP3.8 Measured 4 ?g/mL Predicted 0.4
?g/mL
- Xpack simulation used Cmpd-1 x-ray structure as
a template. - Xpack 1.4
- Head group leaves aryl pocket in simulation at
higher T - Intrinsic solubility pred. 2.5 log units
improved over cmpd-1
16Cmpd-7
- Heterocycle replacement for phenyl leads to
unsymmetrical pocket - Simulation using modeled structure of Cmpd-7
based on Cmpd-1 shows that EWG not as stable in
the modified pocket
HPLCLogP5.2 Exp. Sol 0.2?g/mL Pred. Sol 0.1
?g/mL
17Cmpd-8
- Solubility _at_pH6.5
- 3.8 - 55 ?g/mL depending on salt/crystal form.
18Solubility Estimation
- General solubility prediction including
solvation, ionization, and local crystal packing
simulation. - Crystalline material only
- Most 1.0 ?g/mL 0.1 ?g/mL were recorded as soft
numbers (e.g. lt 1.0 ?g/mL)
19Trends in Solubility Potency
50-compound moving geometric mean
Cmpd-8 registered
Cmpd-7 X-ray supports strategy
Crystal packing strategy proposed
- Chemistry took knowledge from modeling and
derived structures to disrupt key crystal
interaction. - Improved solubility apparent, with maintained
potency.
20Primary Assumptions
- Crystal packing effect falls off with ionization
- Untrue possibly fall off as ionization state
differs from ionization state of crystal? At
what rate? - Crystal packing effect is independent of the free
energy of solvation. - Crystal packing can limit aqueous solubility by
up to 2 orders of magnitude. (arbitrary) - Ionization effect limit set to 4.25 log units for
single acid or base, 5.0 for multiple ionizable
centers. (incorrect) - Crystal form of a new analog is similar to that
of solved X-ray of a previous similar analog.
(very risky)
21Major sources of error
- Sensitive to quality logP estimate.
- Comparability of crystal packing effect
simulation is unclear - only relevant compared to
other compounds in a series? - No consideration of crystal rearrangement on
melting - Hemihydrate water leaves crystal (50 oc) prior
to melting at 174 oc. Simulation shows this
structure as unstable at low T. - Better for compounds with known x-ray unit cell.
- Most common source of error (by far!) are poor
pKa estimates - Jaguar calculated pKas can somewhat mitigate
this at significant added compute time. - 4-10 cpu hours typical vs. 1 min. for Epik or ACD
22Summary
- Reasonable performance over pH-solubility curve
when viewed in entirety, but - Point estimates of solubility at pH6.5 are often
dramatically off, usually due to errors in pKa
estimate - Individual components of the model can be
interpreted to provide predicted information on - Molecular features implicated in intrinsic
solubility - pKa
- Influence of crystal packing on limiting
solubility. - Complicated, but generally interpretable, model.
23Acknowledgements
- Collaborators
- Olafur Gudmundsson (Pharmaceutics/PCO)
- Xue-Qing Chen (Pharmaceutics/PCO)
- Denette Murphy (ARD)
- Chemistry
- Numerous colleagues
- Lexicon
- Terry Stouch
- Analytical RD
- Jack Gougoutas
- Mary Malley
- CADD
- Malcolm Davis
- Brian Claus
- Andy Pudzianowksi
- Arthur Doweyko
- Doree Sitkoff
- Stan Krystek
- Dan Cheney
24Solubility Prediction Algorithm
Effect of Ionization
Effect of Crystal Packing Forces
Solubility at a given pH
Aggregation Number
Intrinsic Solubility
Ionization effect Range
FI is the fraction ionized at pH
Where DT are the diffusion coefficients from an
escalating temperature molecular dynamics
simulation.
25Solubility Prediction Haloperidol
- HCl salt HALDOL01
- Mesylate modeled on YANMUW
26If we had a crystal structure
- we wouldnt need to predict solubility.
- Simulate crystal packing by overlaying molecule
of interest on known crystal structure of related
compound - Less reliable as compounds become less related.
- Only suggests if new compound can form the same
crystal form as the known crystal structure.