Advanced Drug Delivery Reviews 46 (2001) 3–26
L
www.elsevier.com / locate / drugdeliv
Experimental and computational approaches to estimate
solubility and permeability in drug discovery and development
settings q
Christopher A. Lipinski*, Franco Lombardo, Beryl W. Dominy, Paul J. Feeney
Central Research Division, Pfizer Inc., Groton, CT 06340, USA
Received 9 August 1996; accepted 14 August 1996
Abstract
Experimental and computational approaches to estimate solubility and permeability in discovery and development settings
are described. In the discovery setting ‘the rule of 5’ 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 methodology 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, solubility calculations focus on exact value prediction and are difficult because of
polymorphism. Recent work on linear free energy relationships and Log P approaches are critically reviewed. Useful
predictions are possible in closely related analog series when coupled with experimental thermodynamic solubility
measurements. 2001 Elsevier Science B.V. All rights reserved.
Keywords: Rule of 5; Computational alert; Poor absorption or permeation; MWT; MLogP; H-Bond donors and acceptors; Turbidimetric
solubility; Thermodynamic solubility; Solubility calculation
Contents
1. Introduction ............................................................................................................................................................................
2. The drug discovery setting .......................................................................................................................................................
2.1. Changes in drug leads and physico-chemical properties ......................................................................................................
2.2. Factors affecting physico-chemical lead profiles .................................................................................................................
2.3. Identifying a library with favorable physico-chemical properties..........................................................................................
2.4. The target audience — medicinal chemists.........................................................................................................................
2.5. Calculated properties of the ‘USAN’ library.......................................................................................................................
2.6. The ‘rule of 5’ and its implementation ...............................................................................................................................
2.7. Orally active drugs outside the ‘rule of 5’ mnemonic and biologic transporters .....................................................................
q
4
4
4
5
6
7
7
9
9
PII 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.: 11-860-4413561.
E-mail address: LIPINSKI@PFIZER.COM (C.A. Lipinski).
0169-409X / 01 / $ – see front matter 2001 Elsevier Science B.V. All rights reserved
PII: S0169-409X( 00 )00129-0
4
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
2.8. High MWT USANs and the trend in MLogP......................................................................................................................
2.9. New chemical entities, calculations ...................................................................................................................................
2.10. Drugs in absorption and permeability studies, calculations ................................................................................................
2.11. Validating the computational alert....................................................................................................................................
2.12. Changes in calculated physical property profiles at Pfizer ..................................................................................................
2.13. The rationale for measuring drug solubility in a discovery setting ......................................................................................
2.14. Drugs have high turbidimetric solubility ..........................................................................................................................
2.15. High throughput screening hits, calculations and solubility measurements ..........................................................................
2.16. The triad of potency, solubility and permeability ..............................................................................................................
2.17. Protocols for measuring drug solubility in a discovery setting............................................................................................
2.18. Technical considerations and signal processing.................................................................................................................
3. Calculation of absorption parameters ........................................................................................................................................
3.1. Overall approach..............................................................................................................................................................
3.2. MLogP. Log P by the method of Moriguchi .......................................................................................................................
3.3. MLogP calculations .........................................................................................................................................................
4. The development setting: prediction of aqueous thermodynamic solubility ..................................................................................
4.1. General considerations .....................................................................................................................................................
4.2. LSERs and TLSER methods .............................................................................................................................................
4.3. LogP and AQUAFAC methods .........................................................................................................................................
4.4. Other calculation methods ................................................................................................................................................
5. Conclusion .............................................................................................................................................................................
References ..................................................................................................................................................................................
1. Introduction
2. The drug discovery setting
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 absorption problems may
be significantly different at other drug discovery
locations, especially if chemistry focuses on peptidiclike 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 bioavailability 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.
2.1. Changes in drug leads and physico-chemical
properties
10
10
10
11
11
13
14
15
15
15
16
17
17
17
18
18
18
20
21
22
23
24
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 screening programs became less and less important in the drug
industry as the knowledge base grew for rational
drug design [1]. Leads in this era were discovered
using both in vitro and primary in vivo screening
assays and came from sources other than massive
primary 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 scientific meetings; 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.
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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, compounds
were dissolved in dimethyl sulfoxide (DMSO) as
stock solutions, typically at about 20–30 mmol and
then were serially diluted into 96-well plates for
assays (perhaps with some non ionic surfactant to
improve solubility). In this paradigm, even very
insoluble drugs could be tested because the kinetics
of compound crystallization determined the apparent
‘solubility’ level. Moreover, compounds 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 screening technology for
compounds in the mM concentration range that was
largely divorced from the compounds true aqueous
thermodynamic 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 reproducible and potential
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.
5
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 testing 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 principles 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 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
6
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
positioned polar functionality will be more readily
detectable in HTS screens if they are larger and more
lipophilic.
The nature of the screen determines the physicochemical 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 stringency 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 receptor based assays,
generate small numbers of hits in the mM 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 quantify 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 physicochemical 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 process 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
compound 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 followup is decreased but less potent, perhaps physicochemically more attractive 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 compounds, 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 physicochemical 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
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
and the most poorly behaved compounds are weeded
out early. Our hypothesis was that poorer physicochemical properties would predominate in the many
compounds that enter into and fail to survive preclinical 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 database of about 50 000 drugs from
the Derwent Co. The process used to select a subset
of 2245 compounds from this database that are likely
to have superior physico-chemical properties is as
follows: From the 50 427 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
compounds 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 5 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
regression coefficients. Experience had taught us that
7
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
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 50 427 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
explainable 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
8
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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 (x axis) versus MLogP ( y
axis) 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 hydrophilic 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 bilayer [14,15]. Hydrogen donor ability can be measured indirectly by the partition coefficient between
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 a parameter of a donor group
with perhaps a steric modifier to allow for the
interactions between donor and acceptor moieties.
Experimental a values for hydrogen bond donors and
b values 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 2 4.0 for a very strong
donor to 2 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 literature links with better permeability.
Too many hydrogen bond acceptor groups also
hinder permeability 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 b
parameter which measures acceptor ability varies far
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
more on a per nitrogen or oxygen atom basis than the
corresponding a parameter. 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 molecular 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:
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 parameters in the USAN data set, we found
that combinations of two parameters outside the
desirable range did not exceed 10%. The exact
values from the USAN set are: sum of N and
O 1 sum of NH and OH — 10%; sum of N and
O 1 MWT — 7%; sum of NH and OH 1 MWT —
4% and sum of MWT 1 Log P — 1%. The rarity
9
(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 exceptions can 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 antibiotics. 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
10
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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 2 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 2 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 physicochemical 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
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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).
2.11. Validating the computational alert
Validating a computational alert for poor absorption or permeation in a discovery setting is quite
different than validating a quantitative 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 laboratories, 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
11
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 detection. 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 timedependent change in physical properties and we can
compare the profiles of internally synthesized compounds with the profiles of compounds purchased
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 percentage
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
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
12
Table 1
Partial list of drugs in absorption and permeability studies
Drug name
a,b
MLogP
OH1NH c
MWT
N1O d
Aciclovir a
20.09
4
225.21
8
Alprazolam
4.74
0
308.77
4
b
Aspirin a,b
1.70
1
180.16
4
Atenolol
0.92
4
266.34
5
b
Azithromycin
0.14
5
749.00
14
a
AZT
24.38
2
267.25
9
b
Benzyl-penicillin
1.82
2
334.40
6
b
Caffeine b
0.20
0
194.19
6
Candoxatril
3.03
2
515.65
8
Captopril a
0.64
1
217.29
4
a
Carbamazepine b
3.53
2
236.28
3
Chloramphenicol
1.23
3
323.14
7
a,b
Cimetidineb
0.82
3
252.34
6
Clonidine a
3.47
2
230.10
3
Cyclosporinea,b
20.32
5
1202.64
23
Desipramine b
3.64
1
266.39
2
Dexamethasone
1.85
3
392.47
5
b
Diazepam a
3.36
0
284.75
3
Diclofenac
3.99
2
296.15
3
a
Diltiazem-HCl
2.67
0
414.53
6
b
Doxorubicin
21.33
7
543.53
12
a
Enalapril-maleate
1.64
2
376.46
7
b
Erythromycin
20.14
5
733.95
14
a
Famotidinea,b
20.18
8
337.45
9
Felodipine b
3.22
1
384.26
5
Fluorouracil a
20.63
2
130.08
4
Flurbiprofena
3.90
1
244.27
2
Furosemide
0.95
4
330.75
7
b
Glycine
23.44
3
75.07
3
a
Hydrochlorthiazide
21.08
4
297.74
7
Ibuprofen b b
3.23
1
206.29
2
Imipramine a
3.88
0
280.42
2
Itraconazole a
5.53
0
705.65
12
Ketaconazole
4.45
0
380.92
1
Ketoprofen a a
3.37
1
254.29
3
Labetalol-HCl
2.67
5
328.42
5
Lisinoprilba
1.11
5
405.50
8
Mannitol
22.50
6
182.18
6
Methotrexate b
1.60
7
454.45
13
a,b
Metoprolol-tartrate
1.65
2
267.37
4
Nadolol a b
0.97
4
309.41
5
Naloxone
1.53
2
327.38
5
Naproxen-sodium a,b
2.76
1
230.27
3
a
Nortriptylene-HCl
4.14
1
263.39
1
a
Omeprazole
24.38
2
267.25
9
a
Phenytoin a
2.20
2
451.49
10
Piroxicam
0.00
2
331.35
7
b
Prazosin
2.05
2
383.41
9
a,b
Propranolol-HCl
2.53
2
259.35
3
Quinidine b
2.19
1
324.43
4
a
Ranitidine-HCl
0.66
2
314.41
7
b
Scopolamine
1.42
1
303.36
5
b
Tenidap
1.95
2
320.76
5
a
Terfenadine b
4.94
2
471.69
3
Testosterone b
3.70
1
288.43
2
Trovafloxacinb
2.81
3
416.36
7
Valproic-acid
2.06
1
144.22
2
b
Vinblastine b
2.96
3
811.00
13
Ziprasidone
3.71
1
412.95
5
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.
Alert e
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
0
1
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
Table 2
Percent of compounds with MWT (including salt) above 500
Year registered
Synthetic compounds
Commercial compounds
Pre-1984
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
16.0
18.9
12.1
12.6
13.4
14.6
23.4
21.1
25.4
34.2
33.2
32.7
5.4
14.7
15.5
5.5
5.8
8.2
4.1
3.3
1.8
6.8
8.4
7.9
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 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
Table 3
Synthetic compound properties in 1986 (pre-HTS) and 1994
(post-HTS)
Percentile
90th
75th
50th
MLogP
MWT
1986
1994
1986
1994
4.30
3.48
2.60
4.76
3.90
2.86
514
415
352
726
535
412
13
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, compounds 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 dissolution
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 thermodynamically supersaturated solution is kinetically determined
and to our knowledge is not predictable by computational methods. Solubility 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 struc-
14
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
tural 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 melting
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 pessimistic 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 crystalline 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
measurement at a later stage in the discovery process.
2.14. Drugs have high turbidimetric solubility
Measuring solubility by turbidimetry violates almost every precept taught in the pharmaceutical
sciences about ‘proper’ thermodynamic solubility
measurement. Accordingly, we have been profiling
known marketed drugs since our initial presentation
on turbidimetric solubility measurement [23] and
have measured turbidimetric solubilities 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
turbidimetric solubility of greater than 65 mg / ml.
There were only 20 drugs (7%) with a turbidimetric
solubility of 20 mg / ml or less. If turbidimetric
solubility values lie in this low range, we suggest to
our chemists that the probability of useful oral
activity is very low unless the compound 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 thermodynamic 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 necessarily towards drugs with good absorption or permeation characteristics. 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 computational 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
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
‘drugs’ occupy only a small region of property space
relative to what is possible in synthetic compounds
and HTS ‘hits’. Our turbidimetric 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 product 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 1 mM 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.
15
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 computational 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 21 . 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 mg / 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 aqueous media are key elements in our
discovery solubility protocol. Drug is dissolved in
DMSO at a concentration of 10 mg / ml of DMSO
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 temperature. 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 physiological
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
16
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
groups in replacing 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
experiment. The additions of DMSO are spaced a
minute apart. A total of 14 additions are made. These
correspond to solubility increments of , 5 mg / ml to
a top value of . 65 mg / 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
consecutive 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 instruments.
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 fit to the
Absorbance ( y axis) vs. ml of DMSO (x axis) plot.
The coordinates of the intersect point of the two line
segments are termed X crit and Y crit. X crit is the
microlitres of DMSO added when precipitation
occurs and Y crit is the UV Absorbance at the
precipitation point. The concentration of drug in
DMSO (10 mg / ml) is known. The volume of
aqueous buffer (typically 2.5 ml in a cuvette) is
known so the drug concentration expressed as mg of
drug per ml 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 mg / 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 mg / 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 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 mg / 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 scattering 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 5 k 3 nm 2n , where k
is a constant, nm 5 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 n ranged 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 absorbance 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
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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 curvature
can be used to differentiate color from light scattering. The even wavelength spacing in our diode array
UV means that the absorbance 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 experience 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 disadvantage 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 3 12 mm. The
use of even smaller tubes and the resultant advantage
of reduced volume is precluded by light scattering
from the more sharply curved surface of a smaller
diameter tube.
Using nephelometric turbidity unit (NTU) standards, the threshold 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 ana-
17
lytical 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 programming 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 discussion. 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
18
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
experimental log Ps. When a ‘saturation correction’
is applied by raising the lipophilic parameter value to
the 0.6 power and the hydrophilic 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
experimental 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;
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
3.
4.
5.
6.
7.
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;
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;
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;
NO2, the number of nitro groups in the molecule;
NCS, a variable that adds 1.0 for each isothiocyanate group and 0.5 for each thiocyanate group;
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
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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 solubility 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 summarized by considering the three basic
quantities governing the solubility (S) of a given
solid solute:
19
S 5 f(Crystal Packing Energy 1 CavitationEnergy
1 Solvation Energy)
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 , a first 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
structures 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
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 solubility of crystalline solids.
3
The program PROMET is available from Professor Gavezzotti,
University of Milan, Italy.
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
20
prediction of solubility, if the target of the prediction
is the solubility of complex pharmaceutical drug
candidates. Although the judicious application of
some these approaches might be useful for ‘rankordering’ 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.
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 properties,
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 5 c 1 rR 2 1 aSa 2H 1 bSb 2H 1 sp H2
1 nVx
where c is a constant, R 2 is an excess molar
refractivity, Sa 2H and Sb 2H are the (summation or
‘effective’) solute hydrogen-bond acidity and basicity, respectively, p H2 is the solute dipolarity-polarizability and Vx is McGowan’s characteristic volume
[29]. The main problem encountered when using
parameterized equations is that such quantities (parameters or descriptors) cannot easily be estimated,
from structures 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
successfully applied to the correlation between the
solvatochromic parameters described above and the
aqueous solubility of relatively simple organic nonelectrolytes [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:
0.24 2 5.28VI
log Sw (aromatics) 5 ]]]] 1 4.03bm
100
1 1.53am 2 0.0099(m.p.
2 25)
where VI is the intrinsic (van der Waals) molar
volume of the solute, the other parameters are
defined as above and the subscript m indicates a non
self-associating solute monomer. It is interesting to
note that the term 0.0099(m.p.225) is used, in the
words of the author, ‘to account for the process of
conversion of the solid solute to super-cooled liquid
at 258C.’ 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 1008C.
This finding might be exploited in a series of close
structural analogs where a large variation in melting
points (.1008C) 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.
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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
very judicious choice of the ‘disconnection 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 protonacceptor solutes and proton-donor solvents, and the
H-bond formation between amphiphilic 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
21
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 cancellation of some term, can be made but, if an attempt to
predict the solubility 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 published 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 crystal lattice
disruption. However Yalkowski’s work is largely
based on the prediction or estimation of the solubility
of halogenated aromatic and polycyclic halogenated
aromatic hydrocarbons [42], due to their great environmental importance. The general solubility equation, for organic non-electrolytes is reported below.
DSm (m.p. 2 25)
log Spred 5 2 ]]]]] 2 log P 1 0.80
1364
In this equation, DSm is the entropy of melting and
m.p. is the melting point in 8C. The signs of 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 Sm . 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 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
22
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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
constants 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 molecules. The melting point prediction is also discussed
and a computational 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 biphenyls, but would not easily be
applicable for complex drug candidates.
A similar approach to solubility predictions using
a group-contribution method has been implemented
in the CHEMICALC-2 program [47], which calculates LogP and log 1 /S where S is the molar aqueous
solubility. 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 . If T m is not
available, the program calculates the solubility of the
super-cooled liquid at 258C. 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 DSm
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 partition
coefficients are rather insensitive to the hydrogenbond 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.
DSm (m.p. 2 25)
log Spred 5 2 ]]]]] 2 S n i qi
1364
where qi 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. OH5 21.81) 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 functional 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 predic-
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
tions 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 compounds, 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 258C. 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 coefficients, 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 techniques 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
23
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 compounds with in vitro potency
suitable to a potential drug candidate. This stage of
the discovery process is highly predictable. However,
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 compound 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 quantities and in non-traditional formats. Early
discovery methods deal with probabilities and not
exact value predictions. They enhance productivity
24
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
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 diversity so as to maximize coverage of
three-dimensional receptor space. Assuming this goal
is not compromised by physico-chemical calculations, 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
probabilities are important, and because the prediction of crystal packing energies is at present extremely difficult. The problem of polymorphism, common
in pharmaceutical research, which may have been
deferred in the discovery setting has to be addressed
in the development 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 possible studies in the drug developmental setting
means that even small changes towards poorer, but
still acceptable, physico-chemical properties in compounds approaching candidacy can translate to higher developmental 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 chemical
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 difference between failure
and the development of a successful oral drug
product.
References
[1] Kubinyi, H. (1995) Strategies and recent technologies in
drug discovery. Pharmazie 50, 647–662.
[2] Moos, W.H. and Green, G.D. (1993) Chapter 33. Recent
Advances in the Generation of Molecular Diversity. Annu.
Rep. Med. Chem. 28, 315–324.
[3] Patel, D.V. and Gordon, E.M. (1996) Applications of smallmolecule combinatorial chemistry to drug discovery. Drug
Discov. Today 1, 134–144.
[4] Baum, R.M. (1994) Combinatorial approaches provide fresh
leads for medicinal chemistry. Chem. Eng. News February 7,
20–26.
[5] Chow, H., Chen, H., Ng, T., Nyrdal, P. and Yalkowski, S.H.
(1995) Using backpropagation networks for the estimation of
aqueous activity coefficients of aromatic compounds. J.
Chem. Inf. Comput. Sci., 35, 723–728.
[6] Andrews, P.R., Craik, D.J. and Martin, J.L. (1984) Functional group contributions to drug-receptor interactions. J.
Med. Chem. 27, 1648–1657.
[7] Navia, M.A. and Chaturvedi, P.R. (1996) Design principles
for orally bioavailable drugs. Drug. Dev. Today 1, 179–189.
[8] Pardridge, W.M. (1995) Transport of small molecules
through the blood-brain barrier: biology and methodology.
Adv. Drug Deliv. Rev. 15, 5–36.
[9] Cohen, B.E. and Bangham, A.D. (1972) Diffusion of small
non-electrolytes across liposome membranes. Nature 236,
173–174.
[10] Testa, B., Carrupt, P-A., Gaillard, P., Billois, F. and Weber,
P. (1996) Lipophilicity in molecular modelling. Pharm. Res.
13, 335–343.
[11] Moriguchi, I., Hirono, S., Liu, Q., Nakagome, Y. and
Matsushita, Y. (1992) Simple method of calculating octanol /
water partition coefficient. Chem. Pharm. Bull. 40, 127–130.
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
[12] Moriguchi, I., Hirono, S., Nakagome, I. and Hirano, H.
(1994) Comparison of log P values for drugs calculated by
several methods. Chem. Pharm. Bull. 42, 976–978.
[13] Leo, A.J. (1995) Critique of recent comparison of log P
calculation methods. Chem. Pharm. Bull. 43, 512–513.
[14] Abraham, M.H., Chadha, S.H., Whiting, G.S. and Mitchell,
R.C. (1994) Hydrogen bonding. 32. An analysis of wateroctanol and water-alkane partitioning and the delta log P
parameter of Seiler. J. Pharm. Sci. 83, 1085–1100.
[15] Paterson, D.A., Conradi, R.A., Hilgers, A.R., Vidmar, T.J.
and Burton, P.S. (1994) A non-aqueous partitioning system
for predicting the oral absorption potential of peptides.
Quant. Struct.-Act. Relatsh. 13, 4–10.
[16] Hillgren, K.M., Kato, A. and Borchardt, R.T. (1995) In vitro
systems for studying intestinal drug absorption. Med. Res.
Rev. 15, 83–109.
[17] Abraham, M.H. (1993) Hydrogen bonding. 31. Construction
of a scale of solute effective or summation hydrogen-bond
basicity. J. Phys. Org. Chem. 6, 660–684.
[18] Raevsky, O.A., Grigor’ev, V.Y., Kireev, D.B. and Zefirov,
N.S. (1992) Complete thermodynamic description of Hbonding in the framework of multiplicative approach. Quant.
Struct.-Act. Relatsh. 11, 49–63.
[19] Raevsky, O.A., Schaper, K-J. and Seydel, J.K. (1995) HBond contribution to octanol-water partition coefficients of
polar compounds. Quant. Struct.-Act. Relatsh. 14, 433–436.
[20] Lipinski, C.A. (1995) 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.
[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] Lipinski, C.A., Davis, A.L., Aldridge, P.K. and Brown, C.W.
(1994) Biologically relevant UV-based solubility assays.
Pharmaceutical Analysis Session, 45th Pittsburgh Conf.,
Chicago, Oral Presentation 959.
[24] Reichardt, C. (1988) Solvents and Solvent Effects in Organic
Chemistry. 2nd edn. VCH, Weinheim, pp. 27–35.
[25] Bodor, N., Harget, A. and Huang, N.-J. (1991) Neural
network studies. 1. Estimation of the aqueous solubility of
organic compounds. J. Am. Chem. Soc. 113, 9480–9483.
[26] Gavezzotti, A. (1994) Are crystal structures predictable?
Acc. Chem. Res. 27, 309–314.
[27] Polymorph Predictor, Molecular Simulations Inc., Burlington, MA.
[28] Abraham, M.H. (1993) Scales of solute hydrogen-bonding:
their construction and application to physicochemical and
biochemical processes. Chem. Soc. Rev. 22, 73–83.
[29] Abraham, M.H. and McGowan, J.C. (1987) The use of
characteristic volumes to measure cavity terms in reversed
phase liquid chromatography. Chromatographia 23, 243–
246.
[30] Taft, R.W., Abraham, N.H., Doherty, R.M. and Kamlet, M.J.
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
25
(1985) The molecular properties governing solubilities of
organic nonelectrolytes in water. Nature 315, 384–386.
Kamlet, M.J. (1993) 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, 295–317.
Abraham, M.H., Chadha, H.S. and Mitchell, R.C. (1994)
Hydrogen bonding. 33. Factors that influence the distribution
of solutes between blood and brain. J. Pharm. Sci. 83,
1257–1268.
Brinck, T., Murray, J.S. and Politzer, P. (1993) Octanol /
water partition coefficients expressed in terms of solute
molecular surface areas and electrostatic potentials. J. Org.
Chem. 58, 7070–7073.
Murray, J.S. and Politzer, P. (1991) Correlations between the
solvent hydrogen-bond-donating parameter a and the calculated molecular surface electrostatic potential. J. Org. Chem.
56, 6715–6717.
Murray, J.S., Ranganathan, S. and Politzer, P. (1991) Correlations between the solvent hydrogen bond acceptor parameter b and the calculated molecular electrostatic potential. J.
Org. Chem. 56, 3734–3737.
Famini, G.R., Penski, C.A. and Wilson, L.Y. (1992) Using
theoretical descriptors in quantitative structure activity relationships: some physicochemical properties. J. Phys. Org.
Chem. 5, 395–408.
Headley, A.D., Starnes, S.D., Wilson, L.Y. and Famini, G.R.
(1994) Analysis of solute / solvent interactions for the acidity
of acetic acids by theoretical descriptors. J. Org. Chem. 59,
8040–8046.
Ruelle, P., Sarraf, E. and Kesselring, U.W. (1994) Prediction
of carbazole solubility and its dependence upon the solvent
nature. Int. J. Pharm. 104, 125–133.
Ruelle, P., Rey-Mermet, C., Buchmann, M., Nam-Tran, H.,
Kesselring, U.W. and Huyskens, P.L. (1991) A new predictive equation for the solubility of drugs based on the
thermodynamics of mobile disorder. Pharm. Res. 8, 840–
850.
Ruelle, P. and Kesselring, U.W. (1994) Solubility predictions
for solid nitriles and tertiary amides based on the mobile
order theory. Pharm. Res. 11, 201–205.
Yalkowski, S.H. and Valvani, S.C. (1980) Solubility and
partitioning 1: solubility of nonelectrolytes in water. J.
Pharm. Sci. 69, 912–922.
Abramovitz, R. and Yalkowski, S.H. (1990) Estimation of
aqueous solubility and melting point of PCB congeners.
Chemosphere 21, 1221–1229.
Yalkowski, S.H. and Pinal, R. (1993) Estimation of the
solubility of complex organic compounds. Chemosphere 26,
1239–1261.
Leo, A. (1993) Calculating log Poct from structures. Chem.
Rev. 93, 1281–1306 (CLOGP version 3.5).
Abramovitz, R. and Yalkowski, S.H. (1990) Melting point,
boiling point, and symmetry. Pharm. Res. 7, 942–947.
Dannenfelser, R.M., Surendran, N. and Yalkowski, S.H.
(1993) Molecular symmetry and related properties. SAR &
QSAR Environ. Res. 1, 273–292.
26
C. A. Lipinski et al. / Advanced Drug Delivery Reviews 46 (2001) 3 – 26
[47] Suzuki, T. (1991) Development of an automatic estimation
system for both the partition coefficient and aqueous solubility. J. Comput. Aided Mol. Des. 5, 149–166
(CHEMICALC-2, v. 1.0).
[48] Myrdal, P.B., Manka, A.M. and Yalkowsky, S.H. (1995)
Aquafac 3: aqueous functional group activity coefficients;
application to the estimation of aqueous solubility. Chemosphere 30, 1619–1637.
[49] Bodor, N. and Huang, M.-J. (1992) A new method for the
estimation of the aqueous solubility of organic compounds. J.
Pharm. Sci. 81, 954–959.
[50] Nelson, T.M. and Jurs, P.C. (1994) Prediction of aqueous
solubility of organic compounds. J. Chem. Inf. Comput. Sci.
34, 601–609.