Content uploaded by Christopher Andrew Lipinski
Author content
All content in this area was uploaded by Christopher Andrew Lipinski on Jan 19, 2016
Content may be subject to copyright.
Experimental and computational approaches to estimate solubility and permeability
in drug discovery and development settingsi
Christopher A. Lipinski ⁎, Franco Lombardo, Beryl W. Dominy, Paul J. Feeney
Central Research Division, Pfizer Inc., Groton, CT 06340, USA
abstractarticle info
Article history:
Received 9 August 1996
Accepted 14 August 1996
Available online 13 September 2012
Keywords:
Rule of 5
Computational alert
Poor absorption or permeation
MWT
MLogP
H-Bond donors and acceptors
Turbidimetric solubility
Thermodynamic solubility
Solubility calculation
Experimental and computational approaches to estimate solubility and permeability in discovery and develop-
ment settings are described. In the discovery setting ‘theruleof5’predicts that poor absorption or permeation
is more likely when there are more than 5 H-bond donors, 10 H-bond acceptors, the molecular weight (MWT)
is greater than 500 and the calculated Log P (CLogP) is greater than 5 (or MlogP> 4.15). Computational method-
ology for the rule-based Moriguchi Log P (MLogP) calculation is described. Turbidimetric solubility measurement
is described and applied to known drugs. High throughput screening (HTS) leads tend to have higher MWT and
Log P and lower turbidimetric solubility than leads in the pre-HTS era. In the development setting, solubili ty calcu-
lations focus on exact value prediction and are difficult because of polymorphism. Recent work on linear free
energyrelationships and Log P approaches are critically reviewed. Useful predictions are possible in closely related
analog series when coupled with experimental thermodynamic solubility measurements.
© 2012 Published by Elsevier B.V.
Contents
1. Introduction ............................................................... 5
2. The drug discovery setting ........................................................ 5
2.1. Changes in drug leads and physico-chemical properties........................................ 5
2.2. Factors affecting physico-chemical lead profiles........................................... 5
2.3. Identifying a library with favorable physico-chemical properties ................................... 6
2.4. The target audience —medicinal chemists .............................................. 6
2.5. Calculated properties of the ‘USAN’library .............................................. 7
2.6. The ‘rule of 5’and its implementation ................................................ 7
2.7. Orally active drugs outside the ‘rule of 5’mnemonic and biologic transporters . ............................ 8
2.8. High MWT USANs and the trend in MLogP.............................................. 8
2.9. New chemical entities, calculations ................................................. 8
2.10. Drugs in absorption and permeability studies, calculations ...................................... 8
2.11. Validating the computational alert ................................................. 9
2.12. Changes in calculated physical property profiles at Pfizer....................................... 9
2.13. The rationale for measuring drug solubility in a discovery setting ................................... 10
2.14. Drugs have high turbidimetric solubility ...............................................10
2.15. High throughput screening hits, calcul ations and solubility me asurements . ............................11
2.16. The triad of potency, solubility and permeability ........................................... 11
2.17. Protocols for measuring drug solubility in a discovery setting ..................................... 11
2.18. Technical considerations and signal processing ............................................11
Advanced Drug Delivery Reviews 64 (2012) 4–17
iPII of original article: S0169-409X(96)00423-1. The article was originally published in Advanced Drug Delivery Reviews 23 (1997) 3–25.
⁎Corresponding author. Tel.: +1 860 4413561.
E-mail address: LIPINSKI@PFIZER.COM (C.A. Lipinski).
0169-409X/$ –see front matter © 2012 Published by Elsevier B.V.
http://dx.doi.org/10.1016/j.addr.2012.09.019
Contents lists available at SciVerse ScienceDirect
Advanced Drug Delivery Reviews
journal homepage: www.elsevier.com/locate/addr
3. Calculation of absorption parameters ................................................... 12
3.1. Overall approach ......................................................... 12
3.2. MLogP. Log P by the method of Moriguchi.............................................. 12
3.3. MLogP calculations ........................................................ 13
4. The development setting: prediction of aqueous thermodynamic solubility.................................. 13
4.1. General considerations....................................................... 13
4.2. LSERs and TLSER methods ..................................................... 14
4.3. LogP and AQUAFAC methods .................................................... 14
4.4. Other calculation methods ..................................................... 15
5. Conclusion................................................................ 15
References .................................................................. 16
1. Introduction
This review presents distinctly different but complementary
experimental and computational approaches to estimate solubility
and permeability in drug discovery and drug development settings.
In the discovery setting, we describe an experimental approach to
turbidimetric solubility measurement as well as computational
approaches to absorption and permeability. The absence of discovery
experimental approaches to permeation measurements reflects the
authors' experience at Pfizer Central Research. Accordingly, the
balance of poor solubility and poor permeation as a cause of absorp-
tion problems may be significantly different at other drug discovery
locations, especially if chemistry focuses on peptidic-like compounds.
This review deals only with solubility and permeability as barriers to
absorption. Intestinal wall active transporters and intestinal wall
metabolic events that influence the measurement of drug bioavail-
ability are beyond the scope of this review. We hope to spark lively
debate with our hypothesis that changes in recent years in medicinal
chemistry physical property profiles may be the result of leads
generated through high throughput screening. In the development
setting, computational approaches to estimate solubility are critically
reviewed based on current computational solubility research and
experimental solubility measurements.
2. The drug discovery setting
2.1. Changes in drug leads and physico-chemical properties
In recent years, the sources of drug leads in the pharmaceutical
industry have changed significantly. From about 1970 on, what
were considered at that time to be large empirically-based screen-
ing programs became less and less important in the drug industry
as the knowledge base grew for rational drug design [1].Leadsin
this era were discovered using both in vitro and primary in vivo
screening assays and came from sources other than massive prima-
ry in vitro screens. Lead sources were varied coming from natural
products; clinical observations of drug side effects [1];published
unexamined patents; presentations and posters at scientificmeet-
ings; published reports in scientific journals and collaborations
with academic investigators. Most of these lead sources had the
common theme that the ‘chemical lead’already had undergone
considerable scientific investigation prior to being identified as a
drug lead. From a physical property viewpoint, the most poorly
behaved compounds in an analogue series were eliminated and
most often the starting lead was in a range of physical properties
consistent with the previous historical record of discovering orally
active compounds.
This situation changed dramatically about 1989–1991. Prior to
1989, it was technically unfeasible to screen for in vitro activity
across hundreds of thousands of compounds, the volume of random
screening required to efficiently discover new leads. With the advent
of high throughput screening in the 1989–1991 time period, it
became technically feasible to screen hundreds of thousands of
compounds across in vitro assays [2–4]. Combinatorial chemistry
soon began
1
and allowed automated synthesis of massive numbers
of compounds for screening in the new HTS screens. The process
was accelerated by the rapid progress in molecular genetics which
made possible the expression of animal and human receptor
subtypes in cells lacking receptors that might interfere with an
assay and by the construction of receptor constructs to facilitate
signal detection. The screening of very large numbers of compounds
necessitated a radical departure from the traditional method of drug
solubilization. Compounds were no longer solubilized in aqueous
media under thermodynamic equilibrating conditions. Rather, com-
pounds were dissolved in dimethyl sulfoxide (DMSO) as stock solu-
tions, typically at about 20–30 mmol and then were serially diluted
into 96-well plates for assays (perhaps with some non ionic surfac-
tant to improve solubility). In this paradigm, even very insoluble
drugs could be tested because the kinetics of compound crystalliza-
tion determined the apparent ‘solubility’level. Moreover, com-
pounds could partition into assay components such as membrane
particulate material or cells or could bind to protein attached to
the walls of the wells in the assay plate. The net effect was a screen-
ing technology for compounds in the μM concentration range that
was largely divorced from the compounds true aqueous thermody-
namic solubility. The apparent ‘solubility’in the HTS screen is always
higher, sometimes dramatically so, than the true thermodynamic
solubility achieved by equilibration of a well characterized solid
with aqueous media. The in vitro HTS testing process is quite repro-
ducible and potential problems related to poor compound solubility
are often compensated for by the follow-up to the primary screen.
This is typically a more careful, more labor-intensive process of in
vitro retesting to determine IC50s from dose response curves with
more attention paid to solubilization. The net result of all these test-
ing changes is that in vitro activity is reliably detected in compounds
with very poor thermodynamic solubility properties. A corollary
result is that the measurement of the true thermodynamic aqueous
solubility is not very relevant to the screening manner in which
leads are detected.
2.2. Factors affecting physico-chemical lead profiles
The physico-chemical profile of current leads i.e. the ‘hits’in HTS
screens now no longer depends on compound solubility sufficient
for in vivo activity but depends on: (1) the medicinal chemistry prin-
ciples relating structure to in vitro activity; (2) the nature of the HTS
screen; (3) the physico-chemical profile of the compound set being
screened and (4) to human decision making, both overt and hidden
1
A search through SciSearch and Chemical Abstracts for references to combinatorial
chemistry in titles or descriptors using the truncated terms COMBIN? and CHEMISTR?
gave the following number of references respectively: 1990, 0 and 0; 1991, 2 and 1;
1993, 8 and 8; 1994, 12 and 11; 1995, 46 and 45.
5C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
as to the acceptability of compounds as starting points for medicinal
chemistry structure activity relationship (SAR) studies.
One of the most reliable methods in medicinal chemistry to
improve in vitro activity is to incorporate properly positioned
lipophilic groups. For example, addition of a single methyl group
that can occupy a receptor ‘pocket’improves binding by about
0.7 kcal/mol [6]. By way of contrast, it is generally difficult to improve
in vitro potency by manipulation of the polar groups that are involved
in ionic receptor interactions. The interaction of a polar group in a
drug with solvent versus interaction with the target receptor is a
‘wash’unless positioning of the polar group in the drug is precise.
The traditional lore is that the lead has the polar groups in the correct
(or almost correct) position and that in vitro potency is improved by
correctly positioned lipophilic groups that occupy receptor pockets.
Polar groups in the drug that are not required for binding can be
tolerated if they occupy solvent space but they do not add to receptor
binding. The net effect of these simple medicinal chemistry principles
is that, other factors being equal, compounds with correctly posi-
tioned polar functionality will be more readily detectable in HTS
screens if they are larger and more lipophilic.
The nature of the screen determines the physico-chemical profile
of the resultant ‘hits’. The larger the number of hits that are detected,
the more the physico-chemical profile of the ‘hits’resembles the
overall compound set being screened. Technical factors such as the
design of the screen and human cultural factors such as the stringen-
cy of the evaluation as to what is a suitable lead worth are major
determinants of the physico-chemical profiles of the eventual leads.
Screens designed with very high specificity, for example many recep-
tor based assays, generate small numbers of hits in the μM range. In
these types of screens the signal is easy to detect against background
noise, the hits are few or can be made few by altering potency criteria
and the physico-chemical profiles tend towards more lipophilic,
larger, less soluble compounds. Tight control of the criteria for
activity detection in the initial HTS screen minimizes labor-intensive
secondary evaluation and minimizes the effect of human biases. The
downside is that lower potency hits with more favorable physico-
chemical property profiles may be discarded.
Cell-based assays, by their very nature tend to produce more ‘hits’
than receptor-based screens. These types of assays monitor a
functional event, for example a change in the level of a signaling
intermediate or the expression level of M-RNA or protein. Multiple
mechanisms may lead to the measured end point and only a few of
these mechanisms may be desirable. This leads to a larger number
of hits and therefore their physico-chemical profile will more closely
resemble that of the compound set being screened. Perhaps, equally
importantly, a larger volume of secondary evaluation allows for a
greater expression of human bias. Bias is especially difficult to quan-
tify in the chemists perception of a desirable lead structure.
The physico-chemical profile of the compound set being screened
is the first filter in the physico-chemical profile of an HTS ‘hit’.
Obviously high molecular weight, high lipophilicity compounds will
not be detected by a screen if they are not present in the library. In
the real world, trade-offs occur in the choice of profiles for compound
sets. An exclusively low molecular weight, low lipophilicity library
likely increases the difficulty of detecting ‘hits’but simplifies the pro-
cess of discovering an orally active drug once the lead is identified.
The converse is true of a high molecular weight high lipophilicity
library. In our experience, commercially available (non combinatorial)
compounds like those available from chemical supply houses tend
towards lower molecular weights and lipophilicities.
Human decision making, both overt and hidden can play a large
part in the profile of HTS ‘hits’. For example, a requirement that
‘hits’possess an acceptable range of measured or calculated
physico-chemical properties will obviously affect the starting com-
pound profiles for medicinal chemistry SAR. Less obvious are hidden
biases. Are the criteria for a ‘hit’changing to higher potency (lower
IC50) as the HTS screen runs? Labor-intensive secondary follow-up
is decreased but less potent, perhaps physico-chemically more attrac-
tive leads, may be eliminated. How do chemists react to potential lead
structures? In an interesting experiment, we presented a panel of our
most experienced medicinal chemists with a group of theoretical lead
structures —all containing literature ‘toxic’moieties. Our chemists
split into two very divergent groups; those who saw the toxic
moieties as a bar to lead pursuit and those who recognized the toxic
moiety but thought they might be able to replace the offending
moiety. An easy way to illustrate the complexity of the chemists
perception of lead attractiveness is to examine the remarkably
diverse structures of the new chemical entities (NCEs) introduced to
market that appear at the back of recent volumes of Annual Reports
in Medicinal Chemistry. No single pharmaceutical company can
conduct research in all therapeutic areas and so some of these com-
pounds, which are all marketed drugs, will inevitably be less familiar
and potentially less desirable to the medicinal chemist at one
research location, but may be familiar and desirable to a chemist at
another research site.
2.3. Identifying a library with favorable physico-chemical properties
The idea in selecting a library with good absorption properties is to
use the clinical Phase II selection process as a filter. Drug development
is expensive and the most poorly behaved compounds are weeded
out early. Our hypothesis was that poorer physico-chemical properties
would predominate in the many compounds that enter into and fail to
survive pre-clinical stages and Phase I safety evaluation. We expected
that the most insoluble and poorly permeable compounds would have
been eliminated in those compounds that survived to enter Phase II
efficacy studies. We could use the presence of United States Adopted
Name (USAN) or International Non-proprietary Name (INN) names to
identify compounds entering Phase II since most drug companies
(including Pfizer) apply for these names at entry to Phase II.
The (WDI) World Drug Index is a very large computerized data-
base of about 50 000 drugs from the Derwent Co. The process used
to select a subset of 2245 compounds from this database that are like-
ly to have superior physico-chemical properties is as follows: From
the 50427 compounds in the WDI File, 7894 with a data field for a
USAN name were selected as were 6320 with a data field for an
INN. From the two lists, 8548 compounds had one or both USAN or
INN names. These were searched for a data field ‘indications and
usage’suggesting clinical exposure, resulting in 3704 entries. From
the 3704 using a substructure data field we eliminated 1176 com-
pounds with the text string ‘POLY’, 87 with the text string ‘PEPTIDE’
and 101 with the text string ‘QUAT’. Also eliminated were 53
compounds containing the fragment O = P-O. We coined the term
‘USAN’library for this collection of drugs.
2.4. The target audience —medicinal chemists
Having identified a library of drugs selected by the economics of
entry to the Phase II process we sought to identify calculable
parameters for that library that were likely related to absorption or
permeability. Our approach and choice of parameters was dictated
by very pragmatic considerations. We wanted to set up an absorption–
permeability alert procedure to guide our medicinal chemists. Keeping
in mind our target audience of organic chemists we wanted to focus on
the chemists very strong pattern recognition and chemical structure
recognition skills. If our target audience had been pharmaceutical
scientists we would not have deliberately excluded equations or regres-
sion coefficients. Experience had taught us that a focus on the chemists
very strong skills in pattern recognition and their outstanding chemistry
structural recognition skills was likely to enhance information transfer.
In effect, we deliberately emphasized enhanced educational effectiveness
towards a well defined target audience at the expense of a loss of
6C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
detail. Tailoring the message to the audience is a basic communications
principle. One has only to look at the popular chemistry abstracting
booklets with their page after page of chemistry structures and minimal
text to appreciate the chemists structural recognition skills. We believe
that our chemists have accepted our calculations at least in part because
the calculated parameters are very readily visualized structurally and
are presented in a pattern recognition format.
2.5. Calculated properties of the ‘USAN’library
Molecular weight (formula weight in the case of a salt) is an
obvious choice because of the literature relating poorer intestinal
and blood brain barrier permeability to increasing molecular weight
[7,8] and the more rapid decline in permeation time as a function of
molecular weight in lipid bi-layers as opposed to aqueous media [9].
The molecular weights of compounds in the 2245 USANs were
lower than those in the whole 50427 WDI data set. In the USAN set
11% had MWTs> 500 compared to 22% in the entire data set.
Compounds with MWT>600 were present at 8% in the USAN set
compared to 14% in the entire data set. This difference is not explain-
able by the elimination of the very high MWTs in the USAN selection
process. Rather it reflects the fact that higher MWT compounds are in
general less likely to be orally active than lower MWTs.
Lipophilicity expressed as a ratio of octanol solubility to aqueous
solubility appears in some form in almost every analysis of physico-
chemical properties related to absorption [10]. The computational
problem is that an operationally useful computational alert to
possible absorption–permeability problems must have a no fail log P
calculation. In our experience, the widely used and accurate Pomona
College Medicinal Chemistry program applied to our compound file
failed to provide a calculated log P (CLogP) value because of missing
fragments for at least 25% of compounds. The problem is not an
inordinate number of ‘strange fragments’in our chemistry libraries
but rather lies in the direction of the trade off between accuracy
and ability to calculate all compounds adopted by the Pomona
College team. The CLogP calculation emphasizes high accuracy over
breadth of calculation coverage. The fragmental CLogP value is
defined with reference to five types of intervening isolating carbons
between the polar fragments. As common a polar fragment as a
sulfide (-S-) linkage generates missing fragments when flanked by
rare combinations of the isolating carbon types. Polar fragments as
defined by the CLogP calculation can be very large and are not
calculated as the sum of smaller, more common, polar fragments.
This approach enhances accuracy but increases the number of
missing fragments.
We implemented the log P calculation (MLogP) as described by
Moriguchi et al. [11] within the Molecular Design Limited MACCS
and ISIS base programs to avoid the missing fragment problem. As a
rule-based system, the Moriguchi calculation always gives an answer.
The pros and cons of the Moriguchi algorithm have been debated in
the literature [12,13]. We recommend that, within analog series, our
medicinal chemists use the more accurate Pomona CLogP calculation
if possible. For calculation or tracking of library properties the less
accurate MLogP program is used.
Only about 10% of USAN compounds have a CLogP over 5. The
CLogP value of 5 calculated on the USAN data set corresponds to an
MLogP of 4.15. The slope of CLogP (xaxis) versus MLogP (yaxis) is
less than unity. At the high log P end, the Moriguchi MLogP is
somewhat lower than the MedChem CLogP. In the middle log P
range at about 2, the two scales are similar. Experimentally there is
almost certainly a lower (hydrophilic) log P limit to absorption and
permeation. Operationally, we have ignored a lower limit because of
the errors in the MLogP calculation and because excessively hydro-
philic compounds are not a problem in compounds originating in
our medicinal chemistry laboratories.
An excessive number of hydrogen bond donor groups impairs
permeability across a membrane bi-layer [14,15]. Hydrogen donor
ability can be measured indirectly by the partition coefficient be-
tween strongly hydrogen bonding solvents like water or ethylene
glycol and a non hydrogen bond accepting solvent like a hydrocarbon
[15] or as the log of the ratio of octanol to hydrocarbon partitioning.
In vitro systems for studying intestinal drug absorption have been
recently reviewed [16]. Computationally, hydrogen donor ability
differences can be expressed by the solvatochromic αparameter of
a donor group with perhaps a steric modifier to allow for the interac-
tions between donor and acceptor moieties. Experimental αvalues
for hydrogen bond donors and βvalues for acceptor groups [17]
have been compiled by Professor Abraham in the UK and by the
Raevsky group in Russia [18,19]. Both research groups currently
express the hydrogen bond donor and acceptor properties of a moiety
on a thermodynamic free energy scale. In the Raevsky C scale, donors
range from about −4.0 for a very strong donor to −0.5 for a very
weak donor. Acceptors values in the Raevsky C scale are all positive
and range from about 4.0 for a strong acceptor to about 0.5 for a weak
acceptor. In the Abraham scale both donors and acceptors have positive
values that are about one-quarter of the absolute C values in the
Raevsky scale.
We found that simply adding the number of NH bonds and OH
bonds does remarkably well as an index of H bond donor character.
Importantly, this parameter has direct structural relevance to the
chemist. When one looks at the USAN library there is a sharp cutoff
in the number of compounds containing more than 5 OHs and NHs.
Only 8% have more than 5. So 92% of compounds have five or fewer
H bond donors and it is the smaller number of donors that the litera-
ture links with better permeability.
Too many hydrogen bond acceptor groups also hinder permeabil-
ity across a membrane bi-layer. The sum of Ns and Os is a rough
measure of H bond accepting ability. This very simple calculation is
not nearly as good as the OH and NH count (as a model for donor
ability) because there is far more variation in hydrogen bond acceptor
than donor ability across atom types. For example, a pyrrole and
pyridine nitrogen count equally as acceptors in the simple N O sum
calculation even though a pyridine nitrogen is a very good acceptor
(2.72 on the C scale) and the pyrrole nitrogen is an far poorer
acceptor (1.33 on the C scale). The more accurate solvatochromic β
parameter which measures acceptor ability varies far more on a per
nitrogen or oxygen atom basis than the corresponding αparameter.
When we examined the USAN library we found a fairly sharp cutoff
in profiles with only about 12% of compounds having more than 10
Ns and Os.
2.6. The ‘rule of 5’and its implementation
At this point we had four parameters that we thought should be
globally associated with solubility and permeability; namely molecu-
lar weight; Log P; the number of H-bond donors and the number of
H-bond acceptors. In a manner similar to setting the confidence
level of an assay at 90 or 95% we asked how these four parameters
needed to be set so that about 90% of the USAN compounds had
parameters in a calculated range associated with better solubility or
permeability. This analysis led to a simple mnemonic which we called
the ‘rule of 5’[20] because the cutoffs for each of the four parameters
were all close to 5 or a multiple of 5. In the USAN set we found that
the sum of Ns and Os in the molecular formula was greater than 10
in 12% of the compounds. Eleven percent of compounds had a MWT
of over 500. Ten percent of compounds had a CLogP larger than 5
(or an MLogP larger than 4.15) and in 8% of compounds the sum of
OHs and NHs in the chemical structure was larger than 5. The
‘rule of 5’states that: poor absorption or permeation are more
likely when:
7C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
There are more than 5 H-bond donors (expressed as the sum of
OHs and NHs);
The MWT is over 500;
The Log P is over 5 (or MLogP is over 4.15);
There are more than 10 H-bond acceptors (expressed as the sum
of Ns and Os);
Compound classes that are substrates for biological transporters
are exceptions to the rule.
When we examined combinations of any two of the four parame-
ters in the USAN data set, we found that combinations of two param-
eters outside the desirable range did not exceed 10%. The exact values
from the USAN set are: sum of N and O + sum of NH and OH —10%;
sum of N and O + MWT —7%; sum of NH and OH +MWT —4% and
sum of MWT+ Log P —1%. The rarity (1%) among USAN drugs of
the combination of high MWT and high log P was striking because
this particular combination of physico-chemical properties in the
USAN list is enhanced in the leads resulting from high throughput
screening.
The rule of 5 is now implemented in our registration system for
new compounds synthesized in our medicinal chemistry laboratories
and the calculation program runs automatically as the chemist
registers a new compound. If two parameters are out of range, a
‘poor absorption or permeability is possible’alert appears on the
registration screen. All new compounds are registered and so the
alert is a very visible educational tool for the chemist and serves as
a tracking tool for the research organization. No chemist is prevented
from registering a compound because of the alert calculation.
2.7. Orally active drugs outside the ‘rule of 5’mnemonic and
biologic transporters
The ‘rule of 5’is based on a distribution of calculated properties
among several thousand drugs. Therefore by definition, some drugs
will lie outside the parameter cutoffs in the rule. Interestingly, only
a small number of therapeutic categories account for most of the
USAN drugs with properties falling outside our parameter cutoffs.
These orally active therapeutic classes outside the ‘rule of 5’are:
antibiotics, antifungals, vitamins and cardiac glycosides. We suggest
that these few therapeutic classes contain orally active drugs that
violate the ‘rule of 5’because members of these classes have structural
features that allow the drugs to act as substrates for naturally occurring
transporters. When the ‘rule of 5’is modified to exclude these few drug
categories only a very few exceptionscan be found. For example, among
the NCEs between 1990 and 1993 falling outside the double cutoffs in
‘the rule of 5’, there were nine non-orally active drugs and the only
orally active compounds outside the double cutoffs were seven antibi-
otics. Fungicides–protoazocides–antiseptics also fall outside the rule.
For example, among the 41 USAN drugs with MWT> 500 and
MLogP> 4.15 there were nine drugs in this class. Vitamins are another
orally active class drug with parameter values outside the double
cutoffs. Close to 100 vitamins fell into this category. Cardiac glycosides,
an orally active drug class also fall outside the parameter limits of the
rule of 5. For example among 90 USANs with high MWT and low
MLogP there were two cardiac glycosides.
2.8. High MWT USANs and the trend in MLogP
In our USAN data set we plotted MLogP against MWT and
examined the compound distributions as defined by the 50 and 90%
probability ellipses. A large number of USAN compounds had MLogP
more negative than −0.5. Among the USAN compounds there was a
trend for higher MWT to correlate with lower MLogP. This type of
trend is distinctly different from the positive correlation between
MLogP and MWT found in most SAR data sets. Usually as MWT
increases, compound lipophilicity increases and MLogP becomes
larger (more positive). From among the 2641 USANs, we selected
the 405 with MLogP more negative than −0.5 and from among
these selected those with MWT in excess of 500 and mapped the
resulting 90 against therapeutic activity fields in the MACCS WDI
database. About one half (44 of 90) of these high MWT, low MLogP
USANs were orally inactive consisting of 26 peptide agonists or
antagonists, 11 quaternary ammonium salts and seven miscellaneous
non-orally active agents.
Among the USAN compounds in our list fewer than 10% of
compounds had either high MLogP or high MWT. The combination
of both these properties in the same compound was even rarer.
Among 2641 USANs there were only 41 drugs with MWT > 500 and
MLogP> 4.15, about one-half (21) were orally inactive. Among the
remainder there were only six orally active compounds not in the
fungicide and vitamin classes.
2.9. New chemical entities, calculations
New chemical entities introduced between 1990 and 1993
were identified from a summary listing in vol. 29 of Annual Reports
in Medicinal Chemistry. All our computer programs for calculating
physico-chemical properties require that the compound be described
in computer-readable format. We mapped compound names and
used structural searches to identify 133 of the NCEs in the Derwent
World Drug to give us the computer-readable formats to calculate
the rule of 5. The means of calculated properties were well within
the acceptable range. The average Moriguchi log P was 1.80, the
sum of H-bond donors was 2.53, the molecular weight was 408 and
the sum of Ns and Os was 6.95. The incidence of alerts for possible
poor absorption or permeation was 12%.
2.10. Drugs in absorption and permeability studies, calculations
Very biased data sets are encountered in the types of drugs that
are reported in the absorption or permeability literature. Calculated
properties are quite favorable when compared to the profiles of
compounds detected by high throughput screening. Compounds
that are studied are usually orally active marketed drugs and
therefore by definition have properties within the acceptable range.
What is generally not appreciated is that absorption and permeability
are mostly reported for the older drugs. For example, our list of
compounds with published literature on absorption or permeability,
studied internally for validation purposes, is highly biased against
NCEs. Only one drug in our list of 73 was introduced in the period
1990 to date. In part this reflects drug availability, since drugs under
patent are not sold by third parties. Drugs studied in absorption or
permeability models tend to be those with value for assay validation
purposes, i.e. those with considerable pre-existing literature. In
addition, some of the newer studies are driven by a regulatory agency
interest in the permeability properties of generic drugs. In our listing
of 73 drugs in absorption or permeability studies there are 33 generic
drugs whose properties the FDA is currently profiling. Our list
includes an additional 23 drugs with CACO-2 cell permeation data.
Most of these are from the speakers' handouts at a recent meeting
on permeation prediction [21]; a few are from internal Pfizer
CACO-2 studies. A final 12 drugs are those with zwitterionic or very
hydrophilic properties for which there are either literature citations
or internal Pfizer data. The means of calculated properties for
compounds in this list are well within the acceptable range. The
average Moriguchi log P was 1.60, the sum of H-bond donors
was 2.49, the molecular weight was 361 and the sum of Ns and Os
was 6.27. The incidence of alerts for possible poor absorption or
permeation was 12% (Table 1).
8C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
2.11. Validating the computational alert
Validating a computational alert for poor absorption or perme-
ation in a discovery setting is quite different than validating a quanti-
tative prediction calculation in a developmental setting. In effect, a
discovery alert is a very coarse filter that identifies compounds lying
in a region of property space where the probability of useful oral
activity is very low. The goal is to move chemistry SAR towards the
region of property space where oral activity is reasonably possible
(but not assured) and where the more labor-intensive techniques of
drug metabolism and the pharmaceutical sciences can be more
efficiently employed. A compound that fails the computational alert
will likely be poorly bio-available because of poor absorption or
permeation and lies within that region of property space where
good absorption or solubility is unlikely. We believe the alert has its
primary value in identifying problem compounds. In our experience,
most compounds failing the alert also will prove troublesome if
they progress far enough to be studied experimentally. However,
the converse is not true. Compounds passing the alert still can prove
troublesome in experimental studies.
In this perspective, a useful computational alert correctly
identifies drug projects with known absorption problems. Drugs in
human therapy, whether poorly or well absorbed from the viewpoint
of the pharmaceutical scientist, should profile as ‘drugs’, i.e. as having
reasonable prospects for oral activity. The larger the computational
and experimental difference between drugs in human therapy and
those which are currently being made in medicinal chemistry labora-
tories, the greater the confidence that the differences are meaningful.
We assert that absorption problems have recently become worse in
the pharmaceutical industry as attested to by recent meetings and
symposia on this subject [22] and by the informal but industry-wide
concern of pharmaceutical scientists about drug candidates with
less than optimal physical properties. If we are correct, within any
drug organization, one should be able to quantify by calculation
whether time-dependent changes that might impair absorption
have occurred in medicinal chemistry. If these changes have occurred
one can try to correlate these with changes in screening strategy.
2.12. Changes in calculated physical property profiles at Pfizer
How relevant is our experience at the Pfizer Central Research
laboratories in Groton to what may be expected to be observed in
other drug discovery organizations? The physical property profiles
of drug leads discovered through HTS will be similar industry-wide
to the extent that testing methodology, selection criteria and the
compounds being screened are similar. Changes in physical property
profiles of synthetic compounds, made in follow-up of HTS leads by
medicinal laboratories, depend on the timing of a major change
towards HTS screening. The Pfizer laboratories in Groton were one
of the first to realize and implement the benefits of HTS in lead detec-
tion. As a consequence, we also have been one of the first to deal with
the effects of this change in screening strategy on physico-chemical
properties. In Groton, 1989 marked the beginning of a significant
change towards HTS screening. This process was largely completed
by 1992 and currently HTS is now the major, rich source of drug
discovery leads and has largely supplanted the pre-1989 pattern of
lead generation.
At the Pfizer Groton site, we have retrospectively examined the
MWT distributions of compounds made in the pre-1989 era and
since 1989. Since our registration systems unambiguously identify
the source of each compound, we can identify any time-dependent
change in physical properties and we can compare the profiles of in-
ternally synthesized compounds with the profiles of compounds pur-
chased from external commercial sources.
Before 1989, the percentage of internally synthesized high MWT
compounds oscillated in a range very similar to the USAN library
(Table 2). Starting in 1989, there was an upward jump in the percent-
age of high MWT compounds and a further jump in 1992 to a new
stable MWT plateau that is higher than in the USAN library and higher
than any yearly oscillation in the pre-1989 era. By contrast, there
was no change in the MWT profiles of commercially purchased
compounds over the same time period. A comparison of the MWT
and MLogP percentiles of synthetic compounds for a year before the
Table 1
Partial list of drugs in absorption and permeability studies.
Drug name MLogP OH+ NH
c
MWT N+ O
d
Alert
e
Aciclovir
a,b
−0.09 4 225.21 8 0
Alprazolam
a
4.74 0 308.77 4 0
Aspirin
b
1.70 1 180.16 4 0
Atenolol
a,b
0.92 4 266.34 5 0
Azithromycin
b
0.14 5 749.00 14 1
AZT
a
−4.38 2 267.25 9 0
Benzyl-penicillin
b
1.82 2 334.40 6 0
Caffeine
b
0.20 0 194.19 6 0
Candoxatril
b
3.03 2 515.65 8 0
Captopril
a
0.64 1 217.29 4 0
Carbamazepine
a
3.53 2 236.28 3 0
Chloramphenicol
b
1.23 3 323.14 7 0
Cimetidine
a,b
0.82 3 252.34 6 0
Clonidine
b
3.47 2 230.10 3 0
Cyclosporine
a
−0.32 5 1202.64 23 1
Desipramine
a,b
3.64 1 266.39 2 0
Dexamethasone
b
1.85 3 392.47 5 0
Diazepam
b
3.36 0 284.75 3 0
Diclofenac
a
3.99 2 296.15 3 0
Diltiazem-HCl
a
2.67 0 414.53 6 0
Doxorubicin
b
−1.33 7 543.53 12 1
Enalapril-maleate
a
1.64 2 376.46 7 0
Erythromycin
b
−0.14 5 733.95 14 1
Famotidine
a
−0.18 8 337.45 9 0
Felodipine
a,b
3.22 1 384.26 5 0
Fluorouracil
b
−0.63 2 130.08 4 0
Flurbiprofen
a
3.90 1 244.27 2 0
Furosemide
a
0.95 4 330.75 7 0
Glycine
b
−3.44 3 75.07 3 0
Hydrochlorthiazide
a
−1.08 4 297.74 7 0
Ibuprofen
b
3.23 1 206.29 2 0
Imipramine
b
3.88 0 280.42 2 0
Itraconazole
a
5.53 0 705.65 12 1
Ketaconazole
a
4.45 0 380.92 1 0
Ketoprofen
a
3.37 1 254.29 3 0
Labetalol-HCl
a
2.67 5 328.42 5 0
Lisinopril
a
1.11 5 405.50 8 0
Mannitol
b
−2.50 6 182.18 6 0
Methotrexate
b
1.60 7 454.45 13 1
Metoprolol-tartrate
a,b
1.65 2 267.37 4 0
Nadolol
a
0.97 4 309.41 5 0
Naloxone
b
1.53 2 327.38 5 0
Naproxen-sodium
a,b
2.76 1 230.27 3 0
Nortriptylene-HCl
a
4.14 1 263.39 1 0
Omeprazole
a
−4.38 2 267.25 9 0
Phenytoin
a
2.20 2 451.49 10 0
Piroxicam
a
0.00 2 331.35 7 0
Prazosin
b
2.05 2 383.41 9 0
Propranolol-HCl
a,b
2.53 2 259.35 3 0
Quinidine
b
2.19 1 324.43 4 0
Ranitidine-HCl
a
0.66 2 314.41 7 0
Scopolamine
b
1.42 1 303.36 5 0
Tenidap
b
1.95 2 320.76 5 0
Terfenadine
a
4.94 2 471.69 3 0
Testosterone
b
3.70 1 288.43 2 0
Trovafloxacin
b
2.81 3 416.36 7 0
Valproic-acid
b
2.06 1 144.22 2 0
Vinblastine
b
2.96 3 811.00 13 1
Ziprasidone
b
3.71 1 412.95 5 0
a
Standard or drug in FDA bioequivalence study.
b
Studied in CACO-2 permeation.
c
Sum of OH and NH H-bond donors.
d
Sum of N and O H-bond acceptors.
e
Computational alert according to the rule of 5; 0, no problem detected; 1, poor
absorption or permeation are more likely.
9C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
advent of HTS and for 1994 in the post-HTS era shows a similar
pattern (Table 3). The upper range percentiles for MWT and MLogP
properties are skewed towards physical properties less favorable for
oral absorption in the more recent time period.
The trend towards higher MWT and LogP is in the direction of the
property mix that is least populated in the USAN library. There was no
change over time in the population of compounds with high numbers
of H-bond donors or acceptors.
2.13. The rationale for measuring drug solubility in a discovery setting
In recent years, we have been exploring experimental protocols in a
discovery setting that measure drug solubility in a manner as close
as possible to the actual solubilization process used in our biological
laboratories. The rationale is that the physical forms of the compounds
solubilized and the methods used to solubilize compounds in discovery
are very different from those used by our pharmaceutical scientists and
that mimicking the discovery process will lead to the best prediction of
in vivo SAR.
In discovery, the focus is on keeping a drug solubilized for an assay
rather than on determining the solubility limit. Moreover, there is
no known automated methodology that can efficiently solubilize
hundreds of thousands of sometimes very poorly soluble compounds
under thermodynamic conditions. In our biological laboratories, com-
pounds that are not obviously soluble in water or by pH adjustment
are pre-dissolved in a water miscible solvent (most often DMSO)
and then added to a well stirred aqueous medium. The equivalent
of a thermodynamic solubilization, i.e. equilibrating a solid compound
for 24–48 h, separating the phases, measuring the soluble aqueous
concentration and then using the aqueous in an assay, is not done.
When compounds are diluted into aqueous media from a DMSO
stock solution, the apparent solubility is largely kinetically driven.
The influence of crystal lattice energy and the effect of polymorphic
forms on solubility is, of course, completely lost in the DMSO dissolu-
tion process. Drug added in DMSO solution to an aqueous medium is
delivered in a very high energy state which enhances the apparent
solubility. The appearance of precipitate (if any) from a thermody-
namically supersaturated solution is kinetically determined and to
our knowledge is not predictable by computational methods. Solubil-
ity may also be perturbed from the true thermodynamic value in
purely aqueous media by the presence of a low level of residual
DMSO.
The physical form of the first experimental lot of a compound
made in a medicinal chemistry lab can be very different from that
seen by the pharmaceutical scientist at a later stage of development.
Solution spectra, HPLC purity criteria and mass spectral analysis are
quite adequate to support a structural assignment when the chemist's
priority is on efficiently making as many well selected compounds as
possible in sufficient quantity for in vitro and in vivo screening. All the
measurements that support structural assignment are unaffected by
the energy state (polymorphic form) of the solid. Indeed, depending
on the therapeutic area, samples may not be crystalline and most
compounds synthesized for the first time are unlikely to be in lower
energy crystalline forms. Attempts to compute solubility using melt-
ing point information are not useful if samples do not have well
defined melting points. Well characterized, low energy physical
form (from a pharmaceutics viewpoint) reduces aqueous solubility
and may actually be counter productive to the discovery chemists
priority of detecting in vivo SAR.
In this setting, thermodynamic solubility data can be overly pessi-
mistic and may mislead the chemist who is trying to relate chemical
structural changes to absorption and oral activity in the primary in
vivo assay. Our goal is to provide a relevant experimental solubility
measurement so that chemistry can move from the pool of poorly
soluble, orally inactive compounds towards those with some degree
of oral activity. For maximum relevance to the in vivo biological
assay our solubility measurement protocol is as close as possible to
the biological assay ‘solubilization’. In this paradigm, any problems
that might be related to the poor absorption of a low energy crystal-
line solid under thermodynamic conditions are postponed and not
solved. The efficiency gain in an early discovery stage solubility
assay lies in the SAR direction provided to chemistry and in the
more efficient application of drug metabolism and pharmaceutical
sciences resources once oral activity is detected. The value of this
type of assay is very stage-dependent and the discovery type of
assay is not a replacement for a thermodynamic solubility measure-
ment at a later stage in the discovery process.
2.14. Drugs have high turbidimetric solubility
Measuring solubility by turbidimetry violates almost every pre-
cept taught in the pharmaceutical sciences about ‘proper’thermody-
namic solubility measurement. Accordingly, we have been profiling
known marketed drugs since our initial presentation on turbidimetric
solubility measurement [23] and have measured turbidimetric solu-
bilities on over 350 drugs from among those listed in the Derwent
World Drug Index. The calculated properties of these drugs are well
within the favorable range for oral absorption. The average of the
calculated properties are: MLogP, 1.79; the sum of OH and NH, 2.01;
MWT, 295.4; the sum of N and O, 4.69. Without regard to the
therapeutic class, only 4% of these drugs would have been flagged as
having an increased probability of poor absorption or permeability
in our computational alert. Of the 353 drugs, 305 (87%) had a turbidi-
metric solubility of greater than 65 μg/ml. There were only 20 drugs
(7%) with a turbidimetric solubility of 20 μg/ml or less. If turbidimet-
ric solubility values lie in this low range, we suggest to our chemists
that the probability of useful oral activity is very low unless the com-
pound is unusually potent (e.g. projected clinical dose of 0.1 mg/kg)
or unusually permeable (top tenth percentile in absorption rate
constant) or unless the compound is a member of a drug class that
is a substrate for a biological transporter.
Our drug list was compiled without regard to literature thermody-
namic solubilities but does contain many of the types of compounds
studied in the absorption literature. Of the 353 drugs studied in the
discovery solubility assay, 171 are drugs from four sources. There
are 77 drugs from the compilation of 200 drugs by Andrews et al.
[6]. This compilation is biased towards drugs with reliable measured
in vitro receptor affinity and with interesting functionality and not
Table 3
Synthetic compound properties in 1986 (pre-HTS) and 1994 (post-HTS).
Percentile MLogP MWT
1986 1994 1986 1994
90th 4.30 4.76 514 726
75th 3.48 3.90 415 535
50th 2.60 2.86 352 412
Table 2
Percent of compounds with MWT (including salt) above 500.
Year registered Synthetic compounds Commercial compounds
Pre-1984 16.0 5.4
1984 18.9 14.7
1985 12.1 15.5
1986 12.6 5.5
1987 13.4 5.8
1988 14.6 8.2
1989 23.4 4.1
1990 21.1 3.3
1991 25.4 1.8
1992 34.2 6.8
1993 33.2 8.4
1994 32.7 7.9
10 C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
necessarily towards drugs with good absorption or permeation char-
acteristics. There are 23 drugs from a list of generics whose properties
FDA is currently profiling for bio-equivalency standards. In addition,
there are 42 NCEs introduced between 1983 and 1993 and 37 entries
are for drugs with CACO-2 cell permeation data.
The profile of drug turbidimetric solubilities serves as a useful
benchmark. Compounds that are drugs have a very low computation-
al alert rate for absorption or permeability problems and a low
measured incidence of poor turbidimetric solubility of about 10%.
The calculated profiles and alert rates of compounds made in
medicinal chemistry laboratories can be compared to those of drugs
and the profiles can be compared on a project by project basis.
Within the physical property manifold of ‘marketed drugs’we
would expect a poor correlation of our turbidimetric solubility data
with literature thermodynamic solubility data since the properties
of ‘drugs’occupy only a small region of property space relative to
what is possible in synthetic compounds and HTS ‘hits’. Our turbidi-
metric solubilities for drugs are almost entirely at the top end of a
relatively narrow solubility range, whereas from a thermodynamic
viewpoint the drugs in our list cover a wide spectrum of solubility.
We caution that turbidimetric solubility measurements are most
definitely not a substitute for careful thermodynamic solubility
measurements on well characterized crystalline drugs and should
not be used for decision making in a development setting.
2.15. High throughput screening hits, calculations and solubility
measurements
Calculated properties and measured turbidimetric solubilities for
the best compounds identified as ‘hits’in our HTS screens are in accord
with the hypothesis that the physico-chemical profiles of leads have
changes from those in the pre-1989 time period. Nearly 100 of the
most potent ‘hits’from our high throughput screens were examined
computationally and their turbidimetric solubilities were measured.
The profiles are strikingly different from those of the 353 drugs we
studied. The HTS hits are on average more lipophilic and less soluble
than the drugs. The 96 compounds we measured were the end prod-
uct of detection in HTS screens and secondary in vitro evaluation.
These were the compounds highlighted in summaries and which
captured the chemist's interest with many IC50s clustered in the
1μM range. As such, they are the product of a biological testing
process and a chemistry evaluation as to interesting subject matter.
Average MLogP for the HTS hits was a full log unit higher than for
the drugs and the average MWT was nearly 50 Da higher. By contrast,
there was little difference in the number of hydrogen bond donors and
acceptors. The distribution curves for MLogP and MWT are roughly the
same shape for the HTS hits and drugs but the means are shifted
upwards in the HTS hits with a higher distribution of compounds
towards the unfavorable range of physico-chemical properties. The
actual averages, HTS vs. Drug are: MLogP, 2.81 vs. 1.79; MWT, 366
vs. 295; sum of OH NH, 1.80 vs. 2.01; sum of N and O, 5.4 vs. 4.69.
2.16. The triad of potency, solubility and permeability
Acceptable drug absorption depends on the triad of dose, solubility
and permeability. Our computational alert does not factor in dose,
i.e. drug potency. It only addresses properties that are related to
potential solubility and permeation problems and it does not allow
for a very favorable value of one parameter to compensate for a less
favorable value of another parameter. In a successful marketed drug,
one parameter can compensate for another. For example, a computa-
tional alert is calculated for azithromycin, a successful marketed
antibiotic. In azithromycin, which has excellent oral activity, a very
high aqueous solubility of 50 mg/ml more than counterbalances a
very low absorption rate in the rat intestinal loop of 0.001 min
−1
.
Poorer permeability in orally active peptidic-like drugs is usually
compensated by very high solubility. Our solubility guidelines to our
chemists suggest a minimum thermodynamic solubility of 50 μg/ml
for a compound that has a mid-range permeability and an average
potency of 1.0 mg/kg. These solubility guidelines would be markedly
higher if the average compound had low permeability.
2.17. Protocols for measuring drug solubility in a discovery setting
The method and timing of introduction of the drug into the aque-
ous media are key elements in our discovery solubility protocol.
Drug is dissolved in DMSO at a concentration of 10 μg/μlofDMSO
which is close to the 30 mM DMSO stock concentration used in
our own biology laboratories. This is added a microlitre at a time
to a non-chloride containing pH 7 phosphate buffer at room temper-
ature. The decision to avoid the presence of chloride was a tradeoff
between two opposing considerations. Biology laboratories with
requirements for iso-osmotic media use vehicles containing physio-
logical levels of saline (e.g. Dulbecco's phosphate buffered saline)
with the indirect result that the solubility of HCl salts (by far the
most frequent amine salt from our chemistry laboratories) can be
depressed by the common ion effect. Counter to this consideration,
is the near 100% success rate of our pharmaceutical groups in replac-
ing problematical HCl salts with other salts not subject to a chloride
common ion effect. We chose the non-chloride containing medium
to avoid pessimistic solubility values resulting from a historically
very solvable problem.
The appearance of precipitate is kinetically driven and so we avoid a
short time course experiment where we might miss precipitation that
occurs on the type of time scale that would affect a biological experi-
ment. The additions of DMSO are spaced a minute apart. A total of 14
additions are made. These correspond to solubility increments of
b5μg/ml to a top value of > 65 μg/ml if the buffer volume is 2.5 ml
(as in a UV cuvette). If it is clear that precipitation is occurring early in
the addition sequence, we stop the addition so that we have two con-
secutive readings after the precipitate is first detected. Precipitation
can be quantified by an absorbance increase due to light scattering by
precipitated particulate material in a dedicated diode array UV machine.
The sensitivity to light scattering is a function of the placement of the
diode array detector relative to the cuvette and differs among instru-
ments. We found that the array placement in a Hewlett Packard
HP8452A diode array gives high sensitivity to light scattering. Increased
UV absorbance from light scattering is measured in the 600–820 nm
range because most drugs have UV absorbance well below this range.
In its simplest implementation, the precipitation point is calculated
from a bilinear curve fittotheAbsorbance(yaxis) vs. μlofDMSO
(xaxis) plot. The coordinates of the intersect point of the two line
segments are termed Xcrit and Ycrit. Xcrit is the microlitres of DMSO
added when precipitation occurs and Ycrit is the UV Absorbance at
the precipitation point. The concentration of drug in DMSO (10 μg/ml)
is known. The volume of aqueous buffer (typically 2.5 ml in a cuvette)
is known so the drug concentration expressed as μgofdrugperml
buffer at the precipitation point is readily calculated. The volume
percent aqueous DMSO at the precipitation point is also reported.
Under our assay conditions it does not exceed 0.67% for a turbidimetric
solubility of >65 μg/ml. The upper solubility limit is based on the
premise that for most projects permeability is not a major problem
and that solubility assays will most often be requested for poorly
soluble compounds. In the absence of poor permeability, solubilities
above 65 μg/ml suggest that if bio-availability is poor, solubility is not
the problem.
2.18. Technical considerations and signal processing
In our experience, most UV active compounds made in our Medicinal
Chemistry labs have UV peak maxima below 400 nm. Approximation to
a Gaussian form for absorbance peaks allows an estimate for the UV
11C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
absorbance at long wavelength from the peak maximum and peak width
at half height. A soluble compound with maximum absorbance at 400
nm and extinction coefficient of 10 000 and peak width at half height
of 100 nm at a concentration of 400 μg/ml (well above the maximum
for our assay) has calculated absorbance of 0.000151 at 600 nm.
The sensitivity of UV absorbance measurements to light scatter-
ing is largely a function of how closely the diode array is positioned
to the UV cuvette and varies among manufacturers. The HP89532
DOS software detects a curve due to light scattering by fitting the
absorbance over a wavelength range to a power curve of the form.
Abs=k × nm
−n
, where k is a constant, nm = wavelength.
Values for ‘n’were examined in a total of 45 solubility experiments.
The last scan in each solubility series was examined since precipitation
is most likely at the highest drug concentration. In this 45 assay series
precipitation was not observed in 10 assays (as assessed by values of
n>0). Positive values of nranged as high as 5.054 in the 35 assays in
which precipitation occurred. Once precipitation occurred, all scans
in an assay sequence could be fit with a power curve. The overall
absorbance increase due to light scattering can be quite low. In most
of the 45 assays, the total absorbance increase at 690 nm (due to
precipitate formation) was in the OD range 0–0.01. Half the absor-
bance increases were in the range 0–0.001. Measurements within
these very small ranges quantitate the precipitation point.
Problems in determining the precipitation point occur when a
compound is intensely colored since colored compounds may be
miscalled as insoluble. In collaboration with Professor Chris Brown
at the University of Rhode Island, we implemented a fast fourier
transform (FFT) signal processing procedure to enhance assay
sensitivity and to avoid false positive solubility values due to colored
compounds [20]. The absorbance curve due to light scattering has an
apparent peak width at half height which is much wider than the
apparent peak width at half height for a typical UV absorption
curve. An analysis procedure that is sensitive to the degree of curva-
ture can be used to differentiate color from light scattering. The
even wavelength spacing in our diode array UV means that the absor-
bance vs. wavelength matrix in each scan can be treated as if it were a
time series (which it really is not). In a time series, the early terms in
an FFT describe components of low curvature (low frequency). An FFT
over a 256 nm range (566–820 nm) generates 128 absorbance values
which in turn generates 128 FFT terms. FFT term 1 describes the
baseline shift. By plotting the real component of FFT term 1 or term
2 vs. DMSO addition, the false positive rate from color is much
reduced and we detect the onset of precipitation as if we were
plotting absorbance at a single wavelength vs. absorbance.
An alternative to the use of a dedicated diode array UV is to use
one of a number of relatively inexpensive commercially available
nephelometers. The solubility protocol using a nephelometer as the
signal detector is identical to that using a UV machine. We have expe-
rience using a HACH AN2100 as a turbidity detector. A nephelometer
has the advantage that colored impurities do not cause a false positive
precipitation signal and so signal processing is avoided. The disadvan-
tage is the larger volume requirement relative to a UV cuvette. The
HACH unit uses inexpensive disposable glass test tubes that can be
as small as 100 mm × 12 mm. The use of even smaller tubes and the
resultant advantage of reduced volume is precluded by light scatter-
ing from the more sharply curved surface of a smaller diameter tube.
Using nephelometric turbidity unit (NTU) standards, the thresh-
old for detection using a UV detector-based assay is 0.2 NTUs and a
0.4 NTU standard can be reliably detected vs. a water blank. Turbidity
standards in the range 0.2–2 NTU units suffice to cover the scattering
range likely to be detected in a solubility assay. Some type of signal
detector is necessary if light scattering is the analytical signal used
to detect precipitation. For example, a 1.0 NTU standard was our
lower visual detection limit using a fiber optic illuminator to visualize
Tyndall light scattering. The European Pharmacopoeia defines the
lowest category of turbidity —‘slight opalescence’on the basis of
measured optical density changes in the range 0.0005–0.0156 at
340–360 nm. These optical density readings correspond to NTU
standards well below 1.0 (in the 0.2–0.4 range) in our equipment.
3. Calculation of absorption parameters
3.1. Overall approach
The four parameters used for the prediction of potential absorption
problems can be easily calculated with any computer and a program-
ming language that supports or facilitates the analysis of molecular
topology. At Pfizer, we began our programming efforts using MDL's
sequence and MEDIT languages for MACCS and have since successfully
ported the algorithms to Tripos' SPL and MDL's ISIS PL languages
without difficulty.
The parameters of molecular weight and sum of nitrogen and
oxygen atoms are very simple to calculate and require no further dis-
cussion. Likewise, the calculation of the number of hydrogen-bond
acceptors is simply the number of nitrogen and oxygen atoms
attached to at least one hydrogen atom in their neutral state.
3.2. MLogP. Log P by the method of Moriguchi
The calculation of log P via the method of Moriguchi et al. [11]
required us to make some assumptions that were not clear from the
rules and examples in the two papers describing the method
[11,12]. Therefore, more detailed discussion on how we implemented
this method is necessary.
The method begins with a straightforward counting of lipophilic
atoms (all carbons and halogens with a multiplier rule for normalizing
their contributions) and hydrophilic atoms (all nitrogen and oxygen
atoms). Using a collection of 1230 compounds, Moriguchi et al. found
that these two parameters alone account for 73% of the variance in
the experimental log Ps. When a ‘saturation correction’is applied by
raising the lipophilic parameter value to the 0.6 power and the hydro-
philic parameter to the 0.9 power, the regression model accounted for
75% of the variance.
The Moriguchi method then applies 11 correction factors, four that
increase the hydrophobicity and seven that increase the lipophilicity,
and the final equation accounts for 91% of the variance in the experi-
mental log Ps of the 1230 compounds. The correction factors that
increase hydrophobicity are:
1. UB, the number of unsaturated bonds except for those in nitro
groups. Aromatic compounds like benzene are analyzed as having
alternating single and double bonds so a benzene ring has 3 double
bonds for the UB correction factor, naphthalene has a value of 5;
2. AMP, the correction factor for amphoteric compounds where each
occurrence of an alpha amino acid structure adds 1.0 to the AMP
parameter, while each amino benzoic acid and each pyridine
carboxylic acid occurrence adds 0.5;
3. RNG, a dummy variable which has the value of 1.0 if the compound
has any rings other than benzene or benzene condensed with
other aromatic, hetero-aromatic, or hydrocarbon rings;
4. QN, the number of quaternary nitrogen atoms (if the nitrogen is
part of an N-oxide, only 0.5 is added).
The seven correction factors that increase lipophilicity are:
1. PRX, a proximity correction factor for nitrogen and oxygen atoms
that are close to one another topologically. For each two atoms
directly bonded to each other, add 2.0 and for each two atoms
connected via a carbon, sulfur, or phosphorus atom, add 1.0 unless
one of the two bonds connecting the two atoms is a double bond,
in which case, according to some examples in the papers, you
must add 2.0. In addition, for each carboxamide group, we add
an extra 1.0 and for each sulfonamide group, we add 2.0;
12 C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
2. HB, a dummy variable which is set to 1.0 if there are any structural
features that will create an internal hydrogen bond. We limited
our programs to search for just the examples given in the
Moriguchi paper [11] as it is hard to determine how strong a
hydrogen bond has to be to affect lipophilicity;
3. POL, the number of heteroatoms connected to an aromatic ring by
just one bond or the number of carbon atoms attached to two or
more heteroatoms which are also attached to an aromatic ring by
just one bond;
4. ALK, a dummy parameter that is set to 1.0 if the molecule contains
only carbon and hydrogen atoms and no more than one double
bond;
5. NO2, the number of nitro groups in the molecule;
6. NCS, a variable that adds 1.0 for each isothiocyanate group and 0.5
for each thiocyanate group;
7. BLM, a dummy parameter whose value is 1.0 if there is a beta
lactam ring in the molecule.
3.3. MLogP calculations
Log Ps, calculated by our Moriguchi-based computer program for a
set of 235 compounds were less accurate than the calculated log Ps
(CLogPs) from Hansch and Leo's Pomona College Medicinal Chemistry
Project MedChem software distributed by Biobyte. The set of 235 was
chosen so that the CLogP calculation would not fail because of missing
fragments. Our implementation of the Moriguchi method accounts
for 83% of the variance with a standard error of 0.6 whereas the
Hansch values account for 96% of the variance with a standard error
of 0.3. The advantages of the Moriguchi method are that it can be
easily programmed in any language so that it can be integrated with
other systems and it does not require a large database of parameter
values.
4. The development setting: prediction of aqueous
thermodynamic solubility
4.1. General considerations
The prediction of the aqueous solubility of drug candidates may
not be a primary concern in early screening stages, but the knowledge
of the thermodynamic solubility of drug candidates is of paramount
importance in assisting the discovery, as well as the development,
of new drug entities at later stages. A poor aqueous solubility is likely
to result in absorption problems, since the flux of drug across the
intestinal membrane is proportional to its concentration gradient
between the intestinal lumen and the blood. Therefore even in the
presence of a good permeation rate a low absorption is likely to be
the result. Conversely, a compound with high aqueous solubility
might be well absorbed, even if it possesses a moderate or low
permeation rate.
Formulation efforts can help in addressing these problems, but
there are severe limitations to the absorption enhancement that can
be realistically achieved. Stability and manufacturing problems also
have to be taken into account since it is likely that an insoluble drug
candidate may not be formulated as a conventional tablet or capsule,
and will require a less conventional approach such as, for example, a
soft gel capsule. Low solubility may have an even greater impact if an
i.v. dosage form is desired. Obviously, a method for predicting solubil-
ity of drug candidates at an early stage of discovery would have a
great impact on the overall discovery and development process.
Unfortunately the aqueous solubility of a given molecule is the
result of a complex interplay of several factors ranging from the
hydrogen-bond donor and acceptor properties of the molecule and
of water, to the energetic cost of disrupting the crystal lattice of the
solid in order to bring it into solution (‘fluidization’)[24].
In any given situation, not all the factors may play an important
role and it is difficult to predict the solubility of a complex drug
candidate, on the basis of the presence or absence of certain functional
groups. Conformational effects in solution may play a major role in the
outcome of the solubility and cannot be accounted for by a simple
summation of ‘contributing’groups.
Thus, any method which would aim at predicting the aqueous
solubility of a given molecule would have to take into account a
more comprehensive ‘description’of the molecule as the outcome of
the complex interplay of factors.
The brief discussion of the problem outlined above can be summa-
rized by considering the three basic quantities governing the solubility
(S) of a given solid solute:
S¼fðCrystal Packing Energy þCavitation Energy þSolvation EnergyÞ
In this equation, the crystal packing energy is a (endoergic) term
which accounts for energy necessary to disrupt the crystal packing
and to bring isolated molecules in gas phases, i.e. its enthalpy of
sublimation. The cavitation energy is a (endoergic) term which
accounts for the energy necessary to disrupt water (structured by
its hydrogen bonds) and to create a cavity into which to host the
solute molecule. Finally, the solvation energy might be defined as
the sum (exoergic term) of favorable interactions between the
solvent and the solute.
In dealing with the prediction of the solubility of crystalline
solids
2
,afirst major hurdle to overcome is the determination or
estimation of their melting point or, better, of their enthalpy of
sublimation. At present no accurate and efficient method is available
to predict these two quantities for the relatively complex molecules
which are encountered in the pharmaceutical research. Gavezzotti
3
[26] has discussed this point in a review article on the predictability
of crystal structures and he states that ‘...the melting point is one
of the most difficult crystal properties to predict.’This author has
pioneered the use of computational methods to predict crystal struc-
tures and polymorphs and, consequently, properties such as melting
point and enthalpy of sublimation. A commercially available program
has been recently developed [27] but the use of these approaches is
still far from being routine and from being useful in a screening
stage for a relatively large number of compounds, all of which possess
a relatively high conformational flexibility.
Thus, although there are several approaches to estimating and
predicting the solubility of organic compounds, the authors of this
article are of the opinion that none of the presently available methods
can truly be exploited for a relatively accurate prediction of solubility,
if the target of the prediction is the solubility of complex pharmaceu-
tical drug candidates. Although the judicious application of some
these approaches might be useful for ‘rank-ordering’of compounds
and prioritization of their synthesis, we are not aware of any such
systematic use of estimation methods.
The sections that follow will discuss available methods, taking into
account the second and third terms of the above relationship and the
feasibility of their assessment a priori, and they will be treated as one
term since the available methods consider the interactions in solution
as the (algebraic) sum of the two terms and their contributors. This
discussion is by no means exhaustive but it is rather intended as
an overview of the methods available as seen, in particular, from a
pharmaceutical perspective.
2
Since the vast majority of drug molecules and most substances of pharmaceutical
interest are crystalline solids, this discussion will focus on the prediction of the solubil-
ity of crystalline solids.
3
The program PROMET is available from Professor Gavezzotti, University of Milan,
Italy.
13C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
4.2. LSERs and TLSER methods
Linear Solvation Energy Relationships (LSERs), based upon
solvatochromic parameters, have the advantage of a good theoretical
background and offer a correlation between several molecular prop-
erties, and a solute property, SP. Several LSERs have been developed
over the past few years and they seem to work well for predicting a
generalized SP for a series of solutes in one or more (immiscible)
phases. Most notably, the work of Abraham et al. [28] has generated
an equation of the general type:
LogSP ¼cþrR2þa∑αH
2þb∑βH
2þsπH
2þnVx
where c is a constant, R
2
is an excess molar refractivity, Σα
2
H
and Σβ
2
H
are the (summation or ‘effective’) solute hydrogen-bond acidity and
basicity, respectively, π
2
H
is the solute dipolarity-polarizability and V
x
is McGowan's characteristic volume [29]. The main problem encoun-
tered when using parameterized equations is that such quantities
(parameters or descriptors) cannot easily be estimated, from struc-
tures only, for complex multi-functional molecules such as drug
candidates, especially if they are capable of intra-molecular hydrogen
bonding, as is often the case. Nevertheless, the method was success-
fully applied to the correlation between the solvatochromic parame-
ters described above and the aqueous solubility of relatively simple
organic non-electrolytes [30].
More recently, Kamlet [31] has published equations describing the
solubility of aromatic solutes including polycyclic and chlorinated
aromatic hydrocarbons. In these equations a term accounting for the
crystal packing energy was introduced, and the equation has the
general form:
log SwaromaticsðÞ¼
0:24−5:28VI
100 þ4:03βm
þ1:53αm−0:0099 m:p:−25
ðÞ
where V
I
is the intrinsic (van der Waals) molar volume of the solute,
the other parameters are defined as above and the subscript mindi-
cates a non self-associating solute monomer. It is interesting to note
that the term 0.0099(m.p.−25) is used, in the words of the author,
‘to account for the process of conversion of the solid solute to
super-cooled liquid at 25 °C.’This term is therefore related to the
crystal packing energy mentioned earlier, albeit representing the
conversion from a solid to a ‘super-cooled’liquid, not to isolated
molecules in gas phase. The author finds the above term ‘robust’in
its statistical significance and it should be noted that coefficient of
0.0099 implies that a variation of less than one order of magnitude
will be observed for variations in melting points of less than 100 °C.
This finding might be exploited in a series of close structural
analogs where a large variation in melting points (>100 °C) is not
expected (as might often be the case) and the ‘solution behavior’
could be estimated by solvatochromic parameters. Thus, with some
error, the prioritization of more soluble synthetic targets might be
achieved, since the relative (‘rank-order’) solubility of structurally
close analogs may be all that it is sought at an early stage. However
this prioritization would rely on the assumption that variations in
structural properties which bring about a (desired) lowering of the
crystal packing energy, would not significantly and adversely alter
the properties of a molecule with respect to its solvation in water. If
the lower crystal packing energy is the result, for example, of a
lower hydrogen-bond capability, a diminished solvation in water
may offset the lowering of the crystal packing energy.
Even with the assumption described above, the estimation of a
relatively good rank-ordering of aqueous solubilities would still
require the determination of solvatochromic parameters which is
generally achieved through the determination of several partition
coefficients. On the other hand, descriptor values for several fragments
(functional groups) are available and they may be used to calculate the
‘summation’parameters for the molecules of interest. This process is
not without caveats though, as a veryjudicious choice of the ‘disconnec-
tion pattern’must be made to obtain reliable results. In a recent paper
describing the partition of solutes across the blood-brain barrier,
Abraham et al. [32] reported the calculation and use of these descriptors
for compounds of pharmaceutical interest but he warned about the
possibility of inter-molecular hydrogen bonding, which may be a source
of error if not present in the ‘reference’compounds, and pointed out the
fact that these correlations are best used within the descriptors range
used to generate them.
Some authors have reported the calculation of quantities related
to those descriptors, via ab initio [33–35] or semi-empirical methods
[36,37]. The equations stemming from computed values have been
termed TLSERs (Theoretical Linear Solvation Energy Relationships)
[36]. However, we are not aware of any application of this approach
to a series of complex multifunctional compounds, and these types
of correlations are likely to be difficult for these compounds, due to
the relatively high level of computation involved.
Ruelle and Kesselring and colleagues [38–40] reported a multi-
parameter equation, qualitatively similar to the LSERs described
above. This equation attempts to predict solubility by using terms
which account for the quantities that play a role in the process. It does
contain a solute ‘fluidization’term (endoergic cost of destroying the
crystal lattice of a solid) and other terms describing the hydrophobic
effect, hydrogen bond formation between proton-acceptor solutes and
proton-donor solvents, and the H-bond formation betweenamphiphilic
solutes and proton acceptor and/or proton-donor solvents as well as
the auto-association of the solute in solution.
Although this equation takes into account the free energy changes
involved in the dissolution process, in our opinion its complexity
prevents its use for multifunctional molecules. The examples reported
address simple hydrocarbons or mono-functional molecules and
much emphasis is placed on organic (associated and non-associated)
solvents. In many such cases, approximations leading to the cancella-
tion of some term, can be made but, if an attempt to predict the solu-
bility of complex drug candidates in water is made, all those terms
might be present at the same time and thus it would be very difficult
to treat solubility within the framework of this equation.
4.3. LogP and AQUAFAC methods
Prominent in this area is the work of Yalkowski [41] who has pub-
lished a series of papers describing the prediction of solubility using
LogP (the logarithm of the octanol/water partition coefficient) and a
term describing the energetic cost of the crystallattice disruption. How-
ever Yalkowski's work is largely based on the prediction or estimation
of the solubility of halogenated aromatic and polycyclichalogenated ar-
omatic hydrocarbons [42], due to their great environmental impor-
tance. The general solubility equation, for organic non-electrolytes is
reported below.
log Spred ¼−ΔSmm:p:−25ðÞ
1364 −log Pþ0:80
In this equation, ΔS
m
is the entropy of melting and m.p. is the
melting point in °C. The signsof the two terms considered are physically
reasonable, since an increase in either the first term (higher crystal
packing energy) or in LogP (more lipophilic compound), would cause
a decrease in the observed (molar) solubility S
m
. In a recent paper
[43], this author discusses the predictive use of the above equation
and, in particular, the prediction of activity coefficients. The latter is a
term which accounts for deviations from ideal solubility behavior due
to differences in size and shape, but also in hydrogen bonding ability,
between the solute and the solvent. The conclusion is that, among
methods based upon solvatochromic parameters, or simply based on
14 C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
molecular volume, molecular weight or regular solution theory, the
estimation of the activity coefficient is best achieved by using the LogP
method.
Many computational methods are indeed available to address the
prediction of LogP and the aqueous solubility of complex molecules.
A well known and widely used program to predict LogP values is
CLogP [44] which uses a group-contribution approach to yield a
LogP value. Another method, developed by Moriguchi et al. [11],
which uses atomic constants and correction factors to account for
different atom types is discussed in detail in Section 3.2. We have
observed that, in the daily practice of pharmaceutical sciences, both
methods have their ‘outliers’but methods based on fragmental con-
stants tend to fail, in the not infrequent instances where appropriate
constants are not available.
However, LogP prediction aside, the method reported by Yalkowski
was developed on a data set largely based upon rigid, polycyclic and
halogenated aromatic compounds and does not seem to easily yield
itself to the prediction of complex pharmaceutical compounds. The
basic difficulty is that while LogP could be estimated albeit with some
error by computational approaches, the melting point and entropy
of melting are still difficult to calculate or even simply to estimate.
Yalkowski discusses this point in several papers [42,45,46] and shows
the relationship between the entropy of fusion and the molecular
rotational and translational entropies. Some rules are offered for the
estimation of entropy, but the work is limited to relatively simple mol-
ecules. The melting point prediction is also discussed and a computa-
tional approach, based on molecular properties such as eccentricity
(the ratio between the maximum molecular length and the mean
molecular diameter) is proposed. However, the calculation of such
properties may be easy to perform on simple polychlorinated biphe-
nyls, but would not easily be applicable for complex drug candidates.
A similar approach to solubility predictions using a group-c ontribution
method has been implemented in the CHEMICALC-2 program [47],
which calculates LogP and log 1/Swhere Sis the molar aqueous solubil-
ity. This program uses several different algorithms to calculate log 1/S
depending on the complexity and nature of the molecule, and requires
knowledge of the melting point, T
m
.IfT
m
is not available, the program
calculates the solubility of the super-cooled liquid at 25 °C. In the case
of complex molecules, fragmental constants may be missing from its
database and poor results are obtained. We have used this program to
some extent and we are not encouraged by the correlation between
‘predicted’and experimental solubility.
Yalkowski and colleagues [48] have more recently discussed an
improvement of the AQUAFAC (AQUeous Functional group Activity
Coefficients) fragmental constant method. In this work, the authors
describe a correlation between the sum of fragmental constants of a
given molecule and the activity coefficient, defined as a measure of
the non-ideality of the solution. The knowledge or estimation of ΔS
m
and m.p. is necessary, but the method seems to be somewhat better
than the general solubility equation based on LogP values. Yalkowski
explains this by pointing out that these group contribution constants
were derived entirely from aqueous phase data and they should
perform better than octanol-water partition coefficients. We concur
with this explanation since it is known that the octanol-water parti-
tion coefficients are rather insensitive to the hydrogen-bond donor
capability of the solute. Furthermore, the authors point out the fact
that molecules like small carboxylic acids are likely to dimerize in
octanol, while in water they would not.
The solubility equation derived using the AQUAFAC coefficients is
reported below.
log Spred ¼−ΔSmm:p:−25ðÞ
1364 −∑niqi
where q
i
is the group contribution of the ith group and n
i
is the number of
times the ith group appears in the molecule. The negative sign of
the second term stems from the fact that the constant of polar groups
(e.g. OH= −1.81) has a negative sign and a net negative sign of the
summation of contributors would yield an overall positive contribution
to solubility. However, while this method might be of simple application,
its scope seems limited to molecule containing relatively simple func-
tional groups, and the objections to the use of group contribution
methods, which do not consider conformational effects, remain.
4.4. Other calculation methods
Bodor and Huang [49] and Nelson and Jurs [50] have reported
methods based entirely on calculated geometric, electronic and
topological descriptors, for a series of relatively simple liquid and
solid solutes.
We favor these methods as truly a priori predictions based on
molecular structures only, but some questions arise when the
compounds have conformational flexibility and multiple functional
groups, and some of the descriptors will depend upon the particular
conformation chosen. As it is generally true for many QSAR
approaches, there is uncertainty about the actual predictive value of a
test set which does not include a wide variety of compounds and, in
Bodor's training set of 331 compounds we fail to recognize with few
exceptions represented by rigid steroids, complex multifunctional
molecules. Furthermore a large number of the compounds used are
liquids or gases at ambient temperature.
Bodor's method involves the calculation of 18 descriptors, among
which are the ovality of the molecule, the calculated dipole moment,
and the square root of the sum of squared charges on oxygen atoms,
but it does yield a good correlation for the 331-compound set. The
predictive power of the model is illustrated by a table of 17 com-
pounds, but most of them are rigid aromatics, although a reasonably
good prediction is offered for dexamethasone. The latter however is
an epimer of betamethasone which is present in the training set,
and it is difficult to predict the robustness of the correlation with
regard to its application to a truly diverse set of molecules. Similar
considerations could be extended to the work by Nelson and Jurs,
which is also based on calculated descriptors and it does not seem
to involve any polyfunctional molecule or any solid compound at
25 °C. Overall the correlation is good but the caveats on its application
to drug-like compounds remain, as well as our objections on the ease
of calculation of the parameters for compounds of pharmaceutical
interest.
Finally, Bodor et al. [25] and Yalkowski and colleagues [5] have
reported the use of neural networks to develop correlations using
the calculated parameters discussed above or the AQUAFAC coeffi-
cients, respectively. While we have no direct experience with the
use of neural networks, we are of the opinion that it may not be a
trivial task to set up and ‘train’a neural network and the superiority
of this approach in comparison to ‘conventional’regression tech-
niques may be more apparent than real. Indeed Bodor reports a similar
standard deviation for the prediction using the neural network or
regression analysis [49] on the same data set, and the use of a neural
network does not appear to offer any advantage over the regression
analysis.
5. Conclusion
Combinatorial chemistry and high throughput screening (HTS)
techniques are used in drug research because they produce leads
with an efficiency that compares favorably with ‘rational’drug design
and, perhaps more importantly, because these techniques expand the
breadth of therapeutic opportunities and hence the leads for drug
discovery. Established methodology allows the medicinal chemist,
often in a relatively short time, to convert these novel leads to com-
pounds with in vitro potency suitable to a potential drug candidate.
This stage of the discovery process is highly predictable. However,
15C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
the majority of drugs are intended for oral therapy and introducing
oral activity is not predictable, is time and manning expensive and
can easily consume more resources than the optimization of in vitro
activity. The in vitro nature of HTS screening techniques on com-
pound sets with no bias towards properties favorable for oral activity
coupled with known medicinal chemistry principles tends to shift
HTS leads towards more lipophilic and therefore generally less
soluble profiles. This is the tradeoff in HTS screening. Efficiency of
lead generation is high, and therapeutic opportunities are much
expanded, but the physical profiles of the leads are worse and oral
activity is more difficult. Obtaining oral activity can easily become a
rate-limiting step and hence methods which allow physico-chemical
predictions from molecular structure are badly needed in both early
discovery and pharmaceutical development settings.
Computational methods in the early discovery setting need to deal
with large numbers of compounds and serve as filters which direct
chemistry SAR towards compounds with greater probability of oral
activity. These computational methods become particularly important
as experimental studies become more difficult because compounds
are available for physico-chemical screening in only very small quan-
tities and in non-traditional formats. Early discovery methods deal
with probabilities and not exact value predictions. They enhance
productivity by indicating which types of compounds are less likely
to be absorbed and which are more likely to require above average
manning expenditures to become orally active. Calculations, however
imprecise, are better than none when choices must be made in the
design or purchase of combinatorial libraries. Drug discovery requires
a starting point —a lead. Hence the current literature correctly focuses
on improving in vitro activity detection by optimizing chemical diver-
sity so as to maximize coverage of three-dimensional receptor space.
Assuming this goal is not compromised by physico-chemical calcula-
tions, we believe a competitive advantage accrues to the organization
that can identify compound sets likely to give leads more easily
converted to orally active drugs.
Methods in the pharmaceutical developmental setting deal with
much smaller numbers of compounds. Here, a more accurate prediction
is computationally complex because exact values rather than probabil-
ities are important, and because the prediction of crystal packing
energies is at present extremely difficult. The problem of polymor-
phism, common in pharmaceutical research, which may have been
deferred in the discovery setting has to be addressed in the develop-
ment setting. Currently, only approximate estimates of the solubility
of multifunctional and conformationally flexible drug candidates are
possible and these need to be supported by physical measurements
which provide experimental ‘feedback’on analogs in a particular class
of compounds. In our view, a priori solubility estimation methods like
Bodor's multi-parameter equation [49] are the current best choice,
but some of the required properties are not easily computed without
a preliminary optimization of preferred conformations and good
initial estimates. The accurate prediction of the solubility of complex
multifunctional compounds at the moment still remains an elusive
target. The requirements for high accuracy and the complexity of possi-
ble studies in the drug developmental setting means that even small
changes towards poorer, but still acceptable, physico-chemical proper-
ties in compounds approaching candidacy can translate to higher devel-
opmental time and manning requirements. Moreover, there has not
been the same level of efficiency improvement in many developmental
assays as there has been in discovery screening. For example, there
is not the same level of efficiency improvement in measuring
accurate equilibrium solubility as there has been in the efficiency of
detecting leads.
Medicinal chemists efficiently and predictably optimize in vitro
activity, especially when the lead has no key fragments missing.
This ability will likely be reinforced because the current focus on chem-
ical diversity should produce fewer leads with missing fragments.
Oral activity prospects are improved through increased potency, but
improvements in solubility or permeability can also achieve the same
goal. Despite increasingly sophisticated formulation approaches,
deficiencies in physico-chemical properties may represent the differ-
ence between failure and the development of a successful oral drug
product.
References
[1] H. Kubinyi, Strategies and recent technologies in drug discovery, Pharmazie 50
(1995) 647–662.
[2] W.H. Moos,G.D. Green, Chapter 33.Recent Advances in the Generationof Molecular
Diversity, Annu. Rep. Med. Chem. 28 (1993) 315–324.
[3] D.V. Patel, E.M. Gordon, Applications of small-molecule combinatorial chemistry
to drug discovery, Drug Discov. Today 1 (1996) 134–144.
[4] R.M. Baum, Combinatorial approaches provide fresh leads for medicinal chemistry,
Chem. Eng. News (February 7, 1994) 20–26.
[5] H. Chow, H. Chen, T. Ng, P. Nyrdal,S.H. Yalkowski, Using backpropagation networks
for the estimation of aqueous activity coefficients of aromatic compounds, J. Chem.
Inf. Comput. Sci. 35 (1995) 723–728.
[6] P.R. Andrews, D.J. Craik, J.L. Martin, Functional group contributions to
drug-receptor interactions, J. Med. Chem. 27 (1984) 1648–1657.
[7] M.A. Navia, P.R. Chaturvedi, Design principles for orally bioavailable drugs, Drug
Discov. Today 1 (1996) 179–189.
[8] W.M. Pardridge, Transport of small molecules through the blood-brain barrier:
biology and methodology, Adv. Drug Deliv. Rev. 15 (1995) 5–36.
[9] B.E. Cohen, A.D. Bangham, Diffusion of small non-electrolytes across liposome
membranes, Nature 236 (1972) 173–174.
[10] B. Testa, P.-A. Carrupt, P. Gaillard, F. Billois, P. Weber, Lipophilicity in molecular
modelling, Pharm. Res. 13 (1996) 335–343.
[11] I. Moriguchi, S. Hirono, Q. Liu, Y. Nakagome, Y. Matsushita, Simple method of
calculating octanol/water partition coefficient, Chem. Pharm. Bull. 40 (1992)
127–130.
[12] I. Moriguchi, S. Hirono, I. Nakagome, H. Hirano, Comparison of log P values for
drugs calculated by several methods, Chem. Pharm. Bull. 42 (1994) 976–978.
[13] A.J. Leo, Critique of recent comparison of log P calculation methods, Chem. Pharm.
Bull. 43 (1995) 512–513.
[14] M.H. Abraham, S.H. Chadha, G.S. Whiting, R.C. Mitchell, Hydrogen bonding. 32. An
analysis of water-octanol and water-alkane partitioning and the delta log P
parameter of Seiler, J. Pharm. Sci. 83 (1994) 1085–1100.
[15] D.A. Paterson, R.A. Conradi, A.R. Hilgers, T.J. Vidmar, P.S. Burton, A non-aqueous
partitioning system for predicting the oral absorption potential of peptides,
Quant. Struct.-Act. Relatsh. 13 (1994) 4–10.
[16] K.M. Hillgren, A. Kato, R.T. Borchardt, In vitro systems for studying intestinal drug
absorption, Med. Res. Rev. 15 (1995) 83–109.
[17] M.H. Abraham, Hydrogen bonding. 31. Construction of a scale of solute effective
or summation hydrogen-bond basicity, J. Phys. Org. Chem. 6 (1993) 660–684.
[18] O.A. Raevsky, V.Y. Grigor'ev, D.B. Kireev, N.S. Zefirov, Complete thermodynamic
description of H-bonding in the framework of multiplicative approach, Quant.
Struct.-Act. Relatsh. 11 (1992) 49–63.
[19] O.A. Raevsky, K.-J. Schaper, J.K. Seydel, H-Bond contribution to octanol-water
partition coefficients of polar compounds, Quant. Struct.-Act. Relatsh. 14 (1995)
433–436.
[20] C.A. Lipinski, Computational alerts for potential absorption problems: profiles of
clinically tested drugs, in: Tools for Oral Absorption. Part Two. Predicting
Human Absorption. BIOTEC, PDD symposium, AAPS, Miami, 1995.
[21] Prediction of Drug Transport Through Membranes. Chicago IAM Meeting, August
18, 1995.
[22] Workshop on Oral Drug Delivery: Interface Between Discovery and Development.
Report and Recommendations. National Institutes of Health, National Institute of
General Medical Sciences. Herndon, Virginia, December 5–7, 1993.
[23] C.A. Lipinski, A.L. Davis, P.K. Aldridge, C.W. Brown, Biologically relevant UV-based
solubility assays, in: Pharmaceutical Analysis Session, 45th Pittsburgh Conf.,
Chicago, Oral Presentation, 1994, p. 959.
[24] C. Reichardt, in: Solvents and Solvent Effects in Organic Chemistry, 2nd edn., VCH,
Weinheim, 1988, pp. 27–35.
[25] N. Bodor, A. Harget, N.-J. Huang, Neural network studies. 1. Estimation of the
aqueous solubility of organic compounds, J. Am. Chem. Soc. 113 (1991)
9480–9483.
[26] A. Gavezzotti, Are crystal structures predictable? Acc. Chem. Res. 27 (1994)
309–314.
[27] Polymorph Predictor, Molecular Simulations Inc., Burlington, MA.
[28] M.H. Abraham, Scales of solute hydrogen-bonding: their construction and
application to physicochemical and biochemical processes, Chem. Soc. Rev. 22
(1993) 73–83.
[29] M.H. Abraham, J.C. McGowan, The use of characteristic volumes to measure cavity
terms in reversed phase liquid chromatography, Chromatographia 23 (1987)
243–246.
[30] R.W. Taft, N.H. Abraham, R.M. Doherty, M.J. Kamlet, The molecular properties
governing solubilities of organic nonelectrolytes in water, Nature 315 (1985)
384–386.
[31] M.J. Kamlet, Linear solvation energy relationships: an improved equation for
correlation and prediction of aqueous solubilities of aromatic solutes including
polycyclic aromatic hydrocarbons and polychlorinated biphenyls, Prog. Phys.
Org. Chem. 19 (1993) 295–317.
16 C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17
[32] M.H. Abraham, H.S. Chadha, R.C. Mitchell, Hydrogen bonding. 33. Factors that
influence the distribution of solutes between blood and brain, J. Pharm. Sci. 83
(1994) 1257–1268.
[33] T. Brinck, J.S. Murray, P. Politzer, Octanol/water partition coefficients expressed in
terms of solute molecular surface areas and electrostatic potentials, J. Org. Chem.
58 (1993) 7070–7073.
[34] J.S. Murray, P. Politzer, Correlations between the solvent hydrogen-bond-donating
parameter αand the calculated molecular surface electrostatic potential, J. Org.
Chem. 56 (1991) 6715–6717.
[35] J.S. Murray, S. Ranganathan, P. Politzer, Correlations between the solvent hydrogen
bond acceptor parameter βand the calculated molecular electrostatic potential,
J. Org. Chem. 56 (1991) 3734–3737.
[36] G.R. Famini, C.A. Penski, L.Y. Wilson, Using theoretical descriptors in quantitative
structure activity relationships: some physicochemical properties, J. Phys. Org.
Chem. 5 (1992) 395–408.
[37] A.D. Headley, S.D. Starnes, L.Y. Wilson, G.R. Famini, Analysis of solute/solvent
interactions for the acidity of acetic acids by theoretical descriptors, J. Org.
Chem. 59 (1994) 8040–8046.
[38] P. Ruelle, E. Sarraf, U.W. Kesselring, Prediction of carbazole solubility and its
dependence upon the solvent nature, Int. J. Pharm. 104 (1994) 125–133.
[39] P. Ruelle, C. Rey-Mermet, M. Buchmann, H. Nam-Tran, U.W. Kesselring, P.L.
Huyskens, A new predictive equation for the solubility of drugs based on the
thermodynamics of mobile disorder, Pharm. Res. 8 (1991) 840–850.
[40] P. Ruelle, U.W. Kesselring, Solubility predictions for solid nitriles and tertiary
amides based on the mobile order theory, Pharm. Res. 11 (1994) 201–205.
[41] S.H. Yalkowski, S.C. Valvani, Solubility andpartitioning1: solubility of nonelectrolytes
in water, J. Pharm. Sci. 69 (1980) 912–922.
[42] R. Abramovitz, S.H. Yalkowski, Estimation of aqueous solubility and melting point
of PCB congeners, Chemosphere 21 (1990) 1221–1229.
[43] S.H. Yalkowski, R. Pinal,Estimation of the solubility of complex organic compounds,
Chemosphere 26 (1993) 1239–1261.
[44] A. Leo, Calculating log Poct from structures, Chem. Rev. 93 (1993) 1281–1306
(CLOGP version 3.5).
[45] R. Abramovitz, S.H. Yalkowski, Melting point, boiling point, and symmetry,
Pharm. Res. 7 (1990) 942–947.
[46] R.M. Dannenfelser, N. Surendran, S.H. Yalkowski, Molecular symmetry and related
properties, SAR & QSAR Environ. Res. 1 (1993) 273–292.
[47] T. Suzuki, Development of an automatic estimation system for both the partition
coefficient and aqueous solubility, J. Comput. Aided Mol. Des. 5 (1991) 149–166
(CHEMICALC-2, v. 1.0).
[48] P.B.Myrdal, A.M.Manka, S.H.Yalkowsky, Aquafac 3: aqueous functionalgroup activity
coefficients; application to the estimation of aqueous solubility, Chemosphere 30
(1995) 1619–1637.
[49] N. Bodor, M.-J. Huang, A new method for the estimation of the aqueous solubility
of organic compounds, J. Pharm. Sci. 81 (1992) 954–959.
[50] T.M. Nelson, P.C. Jurs, Prediction of aqueous solubility of organic compounds,
J. Chem. Inf. Comput. Sci. 34 (1994) 601–609.
17C.A. Lipinski et al. / Advanced Drug Delivery Reviews 64 (2012) 4–17