REVIEW pubs.acs.org/crt
Improving Drug Candidates by Design: A Focus on Physicochemical Properties As a Means of Improving Compound Disposition and Safety Nicholas A. Meanwell* Department of Medicinal Chemistry, Bristol Myers Squibb Research and Development, 5 Research Parkway, Wallingford, Connecticut 06492, United States ABSTRACT: The development of small molecule drug candidates from the discovery phase to a marketed product continues to be a challenging enterprise with very low success rates that have fostered the perception of poor productivity by the pharmaceutical industry. Although there have been significant advances in preclinical profiling that have improved compound triaging and altered the underlying reasons for compound attrition, the failure rates have not appreciably changed. As part of an effort to more deeply understand the reasons for candidate failure, there has been considerable interest in analyzing the physicochemical properties of marketed drugs for the purpose of comparing with drugs in discovery and development as a means capturing recent trends in drug design. The scenario that has emerged is one in which contemporary drug discovery is thought to be focused too heavily on advancing candidates with profiles that are most easily satisfied by molecules with increased molecular weight and higher overall lipophilicity. The preponderance of molecules expressing these properties is frequently a function of increased aromatic ring count when compared with that of the drugs launched in the latter half of the 20th century and may reflect a preoccupation with maximizing target affinity rather than taking a more holistic approach to drug design. These attributes not only present challenges for formulation and absorption but also may influence the manifestation of toxicity during development. By providing some definition around the optimal physicochemical properties associated with marketed drugs, guidelines for drug design have been developed that are based largely on calculated parameters and which may readily be applied by medicinal chemists as an aid to understanding candidate quality. The physicochemical properties of a molecule that are consistent with the potential for good oral absorption were initially defined by Lipinski, with additional insights allowing further refinement, while deeper analyses have explored the correlation with metabolic stability and toxicity. These insights have been augmented by careful analyses of physicochemical aspects of drugtarget interactions, with thermodynamic profiling indicating that the signature of best-in-class drugs is a dependence on enthalpy to drive binding energetics rather than entropy, which is dependent on lipophilicity. Optimization of the entropic contribution to the binding energy of a ligand to its target is generally much easier than refining the enthalpic element. Consequently, in the absence of a fundamental understanding of the thermodynamic complexion of an interaction, the design of molecules with increased lipophilicity becomes almost inevitable. The application of ligand efficiency, a measure of affinity per heavy atom, group efficiency, which assesses affinity in the context of structural changes, and lipophilic ligand efficiency, which relates potency to lipophilicity, offer less sophisticated but practically useful analytical algorithms to assess the quality of drugtarget interactions. These parameters are readily calculated and can be applied to lead optimization programs in a fashion that helps to maximize potency while minimizing the kind of lipophilic burden that has been dubbed “molecular obesity”. Several recently described lead optimization campaigns provide illustrative, informative, and productive examples of the effect of paying close attention to carefully controlling physicochemical properties by monitoring ligand efficiency and lipophilic ligand efficiency. However, to be successful during the lead optimization phase, drug candidate identification programs will need to adopt a holistic approach that integrates multiple parameters, many of which will have unique dependencies on both the drug target and the specific chemotype under prosecution. Nevertheless, there are many important drug targets that necessitate working in space beyond that which has been defined by the retrospective analyses of marketed drugs and which will require adaptation of some of the guideposts that are useful in directing lead optimization.
’ CONTENTS 1. Introduction 2. Characterizing the Perceived Problems in Contemporary Drug Design 2.1. Analyses of the Structural Elements Associated with Oral Bioavailability 2.2. Time-Related Differences in the Physical Properties of Oral Drugs r 2011 American Chemical Society
2.3. Correlates between Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) Profiles and Calculated Physical Properties 2.4. Physical Properties and Toxicity
1421 1422
1426 1426
1422 1423
Received: May 18, 2011 Published: July 26, 2011 1420
dx.doi.org/10.1021/tx200211v | Chem. Res. Toxicol. 2011, 24, 1420–1456
Chemical Research in Toxicology 2.5. Aromatic Ring Count and the Probability of Successful Drug Development 2.6. Physicochemical Properties and Drug Success: The Effect of sp2 Atom Count 2.7. Physical Properties and Solubility 3. Incorporating Physicochemical Properties into Drug Design 3.1. Molecular Properties Influencing Oral Absorption: Rotatable Bonds 3.2. Physicochemical Properties: Lipophilicity and Oral Absorption 3.3. Golden Triangle Observations: Optimizing Oral Absorption and Clearance 3.4. Physicochemical Properties of CNS Drugs: A Special Case 3.5. Further Defining CNS Drug Space and Candidate Success: A CNS Multi-Parameter Optimization Tool 3.6. Optimizing DrugTarget Interactions: Thermodynamics of LigandProtein Binding 3.7. Thermodynamic Signatures and EnthalpyOptimized Drug Candidates 3.8. H-Bonding in DrugProtein Interactions 3.9. Ligand Efficiency, Binding Efficiency Index, and Surface Efficiency Index 3.10. Lipophilic Ligand Efficiency: LLE or LipE 3.11. Analysis of the BEI and SEI for 92 Marketed Oral Drugs 3.12. Ligand Efficiency and Molecular Size 3.13. Group Efficiency: An Assessment of the LE of Drug Fragments Used in Lead Optimization 3.14. Analysis of the Ligand Efficiency of Leads and the Resultant Drugs 4. Some Recent Examples of the Application of LE and LLE in Drug Optimization 4.1. Cyclin-Dependent Kinase-2 Inhibitors (Astex) 4.2. Protein Kinase B Inhibitors (Astex) 4.3. Soluble Epoxide Hydrolase Inhibitors (Sumitomo) 4.4. CB2 Agonists (Pfizer) 4.5. CB2 Agonists/CB1 Inverse Agonists (Solvay Pharmaceuticals) 4.6. ATP-Competitive Akt Inhibitors (Pfizer) 4.7. Dual PI3K/mTOR Inhibitors (Pfizer) 4.8. HIV Non-Nucleoside Reverse Transcriptase Inhibitors: The Discovery of Lersivirine (Pfizer) 5. Epilogue Author Information Acknowledgment Abbreviations References
REVIEW
1429 1430 1432 1433 1434 1434 1434 1436
1436 1438 1438 1440 1440 1441 1442 1442
1443 1444 1444 1444 1447 1447 1448 1449 1449 1450
1450 1451 1453 1453 1453 1453
1. INTRODUCTION Attrition rates for small molecule drug candidates at various stages of development remain stubbornly high despite several scientific advances that have improved compound profiling and candidate triaging at the preclinical stage.1 Although the pattern of underlying causes of candidate attrition has changed, the failure rate has not, leading to the current perception of low productivity by the pharmaceutical industry. This phenomenon is set against the backdrop of a rapidly rising investment in research and development activities that has not been translated into a commensurate increase in drug approvals, a paradigm that has driven the continuous increase in the estimated costs of developing a drug to launch (data summarized in Table 1). These statistics have raised questions and considerable concern about the sustainability of such an enterprise.2,3 Poor pharmacokinetic (PK) properties were estimated to be an important contributor to clinical failure in 1991 (data summarized in Table 2).1 However, within 10 years, poor human PKs had declined significantly as a source of candidate failure, contributing to the demise of just 8% of candidates in 2000 compared to 39% in 1991 (Table 2), a statistic that reflects advances in preclinical profiling and improved methods of extrapolating in vitro and in vivo PK data to humans. Interestingly, commercial reasons and cost of goods have emerged as more significant sources of drug failure, presumably a consequence of a more competitive business environment. Formulation issues and toxicity have also increased as sources of attrition, the former reflecting the challenging physicochemical properties of contemporary drug candidates, while the latter may be due to an increase in the number of compounds advancing beyond phase 1 PK studies. As part of an initiative by the medicinal chemistry community to more deeply understand the reasons behind drug failure, there has been considerable interest over the past decade in trying to define the physicochemical properties of drug candidates that predict long-term viability. These studies have been conducted with a view to incorporating the insights prospectively into contemporary drug design programs as a means of improving candidate quality. Several analyses of marketed drugs have attempted to equate physicochemical properties with success, by necessity a retrospective analysis of drugs approved in the latter half of the 20th century.410 What has transpired is a scenario in which contemporary drug design appears to be focused too heavily on addressing ever more challenging targets by relying upon structural fragments that display a high dependence on sp2-based ring systems and which are all too easily assembled into molecules with increased molecular weight (MW) and high overall lipophilicity.412 This movement, which has been characterized as driving potency by relying on molecular obesity,11 appears to be a function of identifying drug candidates that are more heavily dependent on entropic rather than enthalpic contributions to the thermodynamics of drugtarget interactions.1113 Although there is a heightened awareness of the importance of the physicochemical properties of a drug candidate to its long-term success, the integration of multiple parameters into drug design will be essential if improved molecules with greater potential to succeed are to be identified. However, it remains to be seen just how successful this enterprise will be to pharmaceutical industry productivity, and given the length of time it takes to develop a drug, it may be a decade or more before the impact of any changes in practices in drug design may be felt. 1421
dx.doi.org/10.1021/tx200211v |Chem. Res. Toxicol. 2011, 24, 1420–1456
Chemical Research in Toxicology
REVIEW
Table 1. Research and Development Spending, Drug Approvals, and Estimated Costs of Developing a Drug research and development spending (millions of $)
drug approvals estimated cost of
nonyear
PhRMA
total
NMEsa
BLAsb
(millions)
1979
$100
1991
$300
1995
15.2
1996
16.5
53
3
1997
18.8
39
6
1998
21.0
30
7
1999
22.3
35
3
2000 2001
26.0 29.8
27 24
2 5
2002
31.0
17
7
2003
34.5
21
6
2004
37.0
10.8
47.8
31
5
2005
39.9
11.9
51.8
18
2
2006
42.4
15.7
58.1
18
4
2007
47.9
13.3
61.2
16
2
2008 2009
47.4 46.8
14.3 19.5
61.7 66.3
21 19
2 6
15
6
2010 a
PhRMA
developing a drug
Figure 1. Lipinski’s “rule of 5” for predicting drug permeability.
Compound Developability; Dr. Stephen Johnson (Bristol-Myers Squibb), Molecular Matched Pairs Derived QSAR for the Optimization of ADMET Properties; and Dr. Travis T. Wager (Pfizer), Moving Beyond Rules: The Development of a Central Nervous System Multi-Parameter Optimization (CNS MPO) Approach to Enable Alignment of Drug-Like Properties.
$800
$1300
New molecular entity. b Biologic license application.
Table 2. Estimated Sources of Drug Candidate Failure in 1991 and 2000 1991
2000
lack of efficacy
30%
24%
PK/bioavailability clinical safety
39% 10%
8% 12%
toxicity
11%
19%
commercial
5%
19%
cost of goods
0%
8%
formulation
0%
4%
other/unknown
5%
6%
This review presents a synopsis of recent studies published in the medicinal chemistry literature and captures some of the more prominent physicochemical guideposts that have been developed as useful aids to decision making during lead optimization. The material summarized includes published elements of presentations made as part of a symposium entitled Improving Drug Candidates By Design: A Focus on Physical Properties to Improve Disposition and Safety convened by Nicholas A. Meanwell and F. Peter Guengerich under the joint sponsorship of the Divisions of Chemical Toxicology and Medicinal Chemistry at the 240th American Chemical Society National Meeting held in Boston, Massachusetts on Tuesday, August 24th, 2010.14 Speakers at the session, in order of appearance, were: Dr. James Empfield (Astra-Zeneca), Physicochemical and Pharmacological Properties as Predictors of Drug Safety and Success; Dr. Simon J. F. Macdonald (GlaxoSmithKline), Aromatic Ring Count and
2. CHARACTERIZING THE PERCEIVED PROBLEMS IN CONTEMPORARY DRUG DESIGN The properties of a molecule are inherent to its structure, and once synthesized, all further studies of a drug candidate during development are essentially focused on understanding its biological activities, metabolism and pharmacokinetics, toxicological profile, and pharmaceutical properties. The fundamental attributes of a molecule can only be managed as it progresses through the successive phases of drug development, and the factors involved in the long term success of a drug are still somewhat enigmatic. The rising awareness that decisions made during lead optimization are of critical importance to the ultimate success of a drug candidate has led to a developing belief that molecules can be designed more effectively if physicochemical principles are given due consideration and applied in a constructive fashion.9 This is dependent on embracing a deeper understanding of the physicochemical aspects of a molecule and its interaction with its biological receptor as well as the alternate proteins that give rise to off-target toxicities, metabolizing enzymes, and the biological membranes encountered in vivo that modulate drug delivery. At a fundamental level of drug design, this involves avoiding structural elements associated with poor outcomes (toxicophores) and maximizing drugtarget association in a fashion that reduces dependence on entropy (lipophilicity) by increasing enthalpybased interactions.12,13 However, the successful oral delivery of a drug candidate necessitates orchestrating a compromise between the properties that confer high potency, reasonable pharmaceutic properties, high membrane permeability, and acceptable metabolic stability.1522 Although there has been a significant and inevitable focus on understanding the latter elements, interest has begun to evolve toward understanding the role of physicochemical properties in predicting toxicological outcomes as means of enhancing overall candidate viability. 2.1. Analyses of the Structural Elements Associated with Oral Bioavailability. The landmark assessment of the physico-
chemical properties associated with the oral bioavailability of drugs and advanced candidates conducted by Christopher Lipinski and his colleagues at Pfizer that has been codified in the simple mnemonic known as the “rule of 5” stimulated considerable interest in providing a deeper and broader perspective on this algorithm.16 The “rule of 5” predicted the potential for a compound to exhibit good absorption and was based on an 1422
dx.doi.org/10.1021/tx200211v |Chem. Res. Toxicol. 2011, 24, 1420–1456
Chemical Research in Toxicology
REVIEW
Table 3. Time Related Changes in the Physicochemical Properties of Orally Bioavailable Drugs Launched before 1983 Compared to Those Launched from 19832002
Table 4. Comparison of the Physicochemical Properties of Orally Bioavailable Drugs Launched Pre-1983 with Those Launched between 1983 and 1992, and 1993 and 2002
oral drugs
oral drugs
Δ mean
all data are
oral drugs
pre-1983
19832002
values
mean values
pre-1983
19831992
19932002
number
864
329
number
864
175
154
MW
331
377
14%
MW
331
374
382
2.1%
clog P
2.27
2.50
10%
clog P
2.27
2.39
2.61
9.2%
% PSA Σ OH + NH
21.1 1.81
21.0 1.77
0% 2%
% PSA Σ OH + NH
21.1 1.81
20.9 1.75
21.2 1.80
1.4% 2.9%
ΣO+N
5.14
6.33
23%
ΣO+N
5.14
6.33
6.32
0.2%
Σ HBA
2.95
3.74
27%
Σ HBA
2.95
3.66
3.82
4.4%
rotatable bonds
4.97
6.42
28%
rotatable bonds
4.97
6.29
6.58
4.6%
rings
2.56
2.88
13%
rings
2.56
2.77
3.02
9.0%
analysis of 2,245 drugs captured in the World Drug Index prior to 1995 that were selected on the basis of consistency with clinical exposure or occurrence in the United States Adopted Names (USAN) or International Nonproprietary Names (INN) databases. This collection was designated as the USAN library, and the physicochemical parameters analyzed were log P, MW, polar surface area (PSA), and the number of H-bond donors (HBDs) and acceptors (HBAs). The analysis related the potential for good oral bioavailability to the physicochemical boundaries summarized in Figure 1, with permeability potentially compromised for compounds that violated 2 or more of the rules. The two notable exceptions recognized by Lipinski were natural products and drugs that are substrates of transporters. Further insights emerged from an analysis of over 1,100 preclinical compounds in the SmithKline Beecham Pharmaceuticals collection that equated a series of physicochemical properties with oral bioavailability in the rat. The results confirmed the Lipinski observations but added an additional factor for consideration, the number of rotatable bonds (RBs) in a molecule.17 A total of e10 rotatable bonds was associated with good oral exposure, and this criterion, when combined with a PSA of e140 Å2, was considered to be sufficient to predict that a compound would exhibit a high probability of showing g20% oral bioavailability in the rat. However, a subsequent analysis of 434 Pharmacia compounds culled from several therapeutic areas indicated that the correlation between oral exposure in the rat and the number of rotatable bonds was less stringent.18 Within some projects, compounds possessing 1520 rotatable bonds were associated with acceptable exposure, providing a cautionary note to the generalization of this property.18 In humans, 13 rotatable bonds have been identified as an upper limit to predict g20% oral bioavailability based on an analysis of 1,014 marketed drugs.19 2.2. Time-Related Differences in the Physical Properties of Oral Drugs. Paul Leeson and colleagues at AstraZeneca have conducted several analyses of the changes in drug properties over time in an effort to identify and understand underlying trends.7,9,10 The initial study focused on comparing the physicochemical properties of 864 orally administered drugs launched prior to 1983 with 329 molecules launched between 1983 and 2002.7 The physicochemical properties that formed the basis of the evaluation were MW, clog P, PSA, the number of HBDs (OH and NH) and HBAs (O and N atoms), the number of RBs, and the number of rings in a molecule. In addition, the analysis was
Δ mean values
extended to a comparison of the properties of drugs across 5 of the major therapeutic areas launched between 1983 and 2002 in order to provide insight into any differences based on the nature of disease targets. The therapeutic areas that formed the basis for this study were categorized as cardiovascular, gastrointestinal and metabolic, infectious diseases, nervous system and respiratory, and inflammation. These analyses revealed several trends in drug properties that differed between those launched before 1983 and those marketed between 1983 and 2002. Mean and median MW, the sum of O and N atoms, HBAs, RBs, and the number of rings all increased, while clog P, % PSA, and the sum of HBDs (OH and NH) were not significantly different (Table 3).7 Drugs launched between 1983 and 2002 were found to be an average of 46 Da larger than those in the pre-1983 data set, with 6.7% of the pre-1983 drugs violating Lipinski’s MW rule of >500 Da, a number that almost doubled to 11.3% of the drugs launched between 1983 and 2002. However, the increase in MW was not accompanied by an increase in mean lipophilicity, an observation that suggested an increase in the incorporation of polar or H-bonding elements in the 19832002 drug set. Indeed, the number of O and N atoms increased in the latter drug set compared to that in pre-1983 drugs, but the number of HBDs was unaltered, attributed to their importance as determinants of oral bioavailability. This presumably reflects a Darwinian-like effect on drug attrition that naturally selects compounds with both good permeability and good overall properties in preclinical species. The number of rings also increased by 13% to 2.88 from the mean of 2.56 noted for pre-1983 drugs. An interesting observation revolved around differences in the distribution of PSA between the two data sets. The % PSA of the 19832002 drugs was found to be narrower than that for the pre-1983 data set, with the 1090th percentile of % PSA spanning 4.539.5% for the pre-1983 drugs. This contrasts with the 19832002 cohort where the distribution was 25% narrower, ranging from 9.9 to 35.9%, an observation thought to be related to increases in drug size and complexity. This notion was supported by the higher number of RBs occurring in the 19832002 drug set compared to that in the pre-1983 collection, a statistic that on the surface would appear to offer some counterbalance to the finding that oral bioavailability in preclinical species decreased with increasing rotatable bond count in the proprietary set of compounds.17,18 However, the number of rotatable bonds in both marketed drug sets is significantly lower, 1423
dx.doi.org/10.1021/tx200211v |Chem. Res. Toxicol. 2011, 24, 1420–1456
Chemical Research in Toxicology
REVIEW
Table 5. Comparison of the Physicochemical Properties of Orally Bioavailable Drugs Launched from 19832002 Arranged by Therapeutic Category all data are mean values
CV (79)
NS (74)
GI and Met (38)
infection (64)
respiratory and inflammation (46)
cancer (14)
other (14)
MW
389
310
378
456
396
313
309
clog P % PSA
3.05 19.8
2.50 16.3
1.90 26.7
1.56 24.6
3.34 20.5
3.02 20.8
1.93 22.9
Σ OH + NH
1.46
1.50
2.71
2.41
1.37
1.00
1.64
ΣO+N
6.73
4.32
6.84
8.78
6.17
4.50
4.29
Σ HBA
3.77
2.12
4.34
5.28
4.24
2.86
2.64
rotatable bonds
8.23
4.70
7.63
6.83
5.52
5.00
4.57
rings
2.84
2.85
2.32
3.45
3.02
2.36
2.36
4.97 in the pre-1983 set and 6.43 in the 19832002 compounds, than the average of 8.19 (upper and lower quartile numbers averaged 6.17 and 10.22) for the GlaxoSmithKline preclinical data set.17 Moreover, the average MW of both marketed drug data sets was below that at which rotatable bond count appeared to exert a significant influence on oral bioavailability in the rat.17 Leeson also examined changes within the 19832002 drug set, dividing the analyzed drugs by decade in an effort to more effectively capture contemporary trends. Although all physicochemical properties exhibited upward trends, none achieved statistical significance (Table 4). However, of particular note, oral drugs launched between 1993 and 2002 contributed to the increases in both clog P and the number of rings observed in the 19832002 data set when compared to those in the pre-1983 drug cohort. The profile of drug properties across therapeutic areas highlighted significant variation, particularly between anti-infective and neuroscience drugs which exhibited the most extreme properties, as summarized in Table 5.7,23,24 Anti-infective drugs possessed the highest mean MW, the lowest mean lipophilicity, and the highest HBA and O and N atom counts.7,23 In contrast, and presumably reflecting the restrictive nature of the blood brain barrier since most drugs included in the data set acted centrally, nervous system drugs showed the lowest mean MW, the lowest mean HBA, O and N count, and the fewest rotatable bonds.7,24 CNS drug space represents a special circumstance that will be discussed in more detail later. Drugs in therapeutic categories beyond anti-infective exhibited a similar distribution of lipophilicity, emphasizing the importance of this property to oral bioavailability irrespective of therapeutic area, with the conclusion that this maybe a more stringent drug-like property than MW. John Proudfoot scrutinized the physicochemical properties of 1791 drugs marketed in the 60 years spanning 19371997, a collection of compounds from which diagnostic, metal-containing drugs, and unmodified natural products were specifically excluded.8 In this data set, median MW increased over time, with drugs launched in the period 19371950 generally 400 Da was more frequently encountered in the drugs launched after 1980 (statistics compiled in Table 6). Only 7% of the complete data set exhibited a MW above the Lipinski guideline of 500 Da. Indeed, the steady increase in MW was the most notable change in properties observed, with just 7 drugs in the 19371951 data set having a MW >500 Da, while 15 drugs in the 19831997 drug set had a MW >500 Da. The median Alog P showed no upward or downward trend over the time frame analyzed, with 8.5% of compounds exhibiting an Alog P of greater than 5 and 5.2% less than 1, which compares with the Lipinski
Table 6. Mean Physicochemical Properties of 1791 Drugs Marketed Between 1937 and 1997 1791
90th percentile
mean MW
333
469
mean log P
2.5
4.8
mean H-bond donors
1.5
3
mean H-bond acceptors
5.1
9
# of oral drugs
guideline of a clog P < 5. These results were generally similar to those observed in the Leeson analysis. The median HBD count increased significantly over time but less than 1.1% of the data set incorporated more than 5 HBDs, reflecting the relationship between the number of HBDs and absorption and the observation that HBDs are frequently involved in phase 2 metabolism. It was noted that median HBA count began to increase in the mid1970s, with more substantial increases observed in drugs launched in the 1990s, although the latter comprised a relatively small cohort. Compounds with more than 10 HBAs amounted to 4.8% of the data set, and only 0.6% of the 1791 drugs possessed both a MW over 500 Da and more than 5 HBDs, two of the Lipinski rules. Compounds combining a MW of >500 Da with more than 3 HBDs comprised 2% of the data set, while the combination of a MW of >500 Da and an Alog P of >5 occurred in 2% of the drugs, and less than 5% contained more than 4 H-bond donors. Since the launch of a drug typically occurs a decade or more after its design and discovery, an examination of the physicochemical properties of marketed drugs will not adequately capture contemporary practices. In an attempt to examine this phenomenon in more detail, Leeson and Springthorpe compared the physicochemical properties of 592 oral drugs launched between 1983 and 2007 with those of compounds disclosed in patent applications originating from 4 major pharmaceutical houses that were published between 2001 and 2007.9 Merck, AstraZeneca, Pfizer, and GlaxoSmithKline were selected as substrates for the patent estate evaluation on the basis of their interest in a broad range of therapeutic areas and substantial productivity. In the drug data set, there was a median of 10.5 years between the publication and launch dates, supporting the notion that emerging practices in drug design are more likely to be captured by profiling compounds disclosed in recent patent applications. Of particular interest, a temporal analysis that looked at trends in clog P and MW of compounds disclosed in patent applications published by the individual companies was also included. Two notable basic trends that emerged from the analysis of the 592 approved drugs were an increase of both the median clog P and 1424
dx.doi.org/10.1021/tx200211v |Chem. Res. Toxicol. 2011, 24, 1420–1456
Chemical Research in Toxicology
REVIEW
Table 7. Observations and Trends Captured by an Analysis of Key ADMET and Calculated Descriptors of Preclinical Compounds at GlaxoSmithKline number of function solubility
compounds evaluated
key observations and trends
44,584
As MW increases, solubility decreases. Solubility is dependent on the ionization state with acids generally more soluble than bases, attributed to the ionization state in the pH 7.4 buffer. Solubility correlates negatively with clog P although reducing clog P to neutral > zwitterionic > acidic molecules. CNS penetration increases with clog P, but this correlation is weaker than that for MW.
P-gp efflux
1,975
The P-gp efflux ratio increased with MW. Ionization state played only a minor role on P-gp efflux descending in the order zwitterionic > neutral = basic> acidic molecules. The correlation between clog P and P-gp efflux was weak, but molecules with clog P between 3 and 5 had higher mean ratios than those with clog Ps 5.
plasma protein binding
2,939
Plasma protein binding increased with MW, averaging 72% for MW neutral > zwitterionic > basic molecules. As clog P increased, plasma protein binding increased.
brain tissue binding
986
Brain tissue binding increases as MW increases. Larger molecules bound more tightly to both plasma proteins and brain tissue. Brain tissue binding was not dramatically affected by ionization state, which contrasts with the observations in plasma protein binding. More lipophilic compounds bound more tightly to brain tissue.
in vivo clearance
11,490
No significant relationship between clearance and MW. Acidic molecules showed lower clearance than neutral and zwitterionic compounds, while bases were generally cleared most rapidly. There was a weak but significant correlation between an increase in clog P and increased clearance. 1425
dx.doi.org/10.1021/tx200211v |Chem. Res. Toxicol. 2011, 24, 1420–1456
Chemical Research in Toxicology
REVIEW
Table 7. Continued number of function hERG inhibition
compounds evaluated 35,200
key observations and trends Mean pIC50 values increased as MW increased. Neutral and acidic molecules exhibited weaker hERG inhibition than zwitterions and bases. Mean pIC50 values increased as clog P increased for bases and zwitterions but not for acids and neutral molecules.
P450 1A2 inhibition
49,837
Mean pIC50 values decreased as MW increased, reflecting the narrow active site of this enzyme. There was no significant correlation between ionization state and 1A2 inhibition. Lipophilicity exerted a limited effect on pIC50 values.
P450 2C9 inhibition
51,097
There was a parabolic relationship between MW and 2C9 inhibition with compounds with a MW 300700 exhibiting a more potent inhibitory effect. Neutral and acidic molecules exhibited higher affinity for 2C9 than bases and zwitterions. 2C9 inhibition increased with higher clog P values, and the magnitude of this effect was more pronounced for neutral and acidic molecules.
P450 2C19 inhibition
48,464
2C19 inhibition showed minimal dependency on MW and ionization state, but affinity increased with clog P.
P450 2D6 inhibition
50,886
There was a weak parabolic relationship between 2D6 affinity and MW. Basic molecules were more potent inhibitors than zwitterions > neutral > acidic compounds. Molecules with higher clog P values exhibited greater inhibitory activity, an effect that was more pronounced for neutral, basic, and zwitterionic molecules than acids.
P450 3A4 inhibition
42,987
Mean pIC50 values increased with MW. Neutral molecules were more potent 3A4 inhibitors than neutral or zwitterionic compounds which were more potent than acids. An increase in clog P was associated with an increase in the potency of 3A4 inhibition.
MW with time, while the number of drugs approved per year worldwide decreased between 1983 and 2006 as did the proportion of drugs with a MW of less than 350 Da. Trends of increasing clog P and MW compared to the marketed drug data set were observed in compounds abstracted from AstraZeneca, GlaxoSmithKline, Merck, and Pfizer patent applications. The median MW of compounds in the patent applications was 450 Da, and the median clog P was 4.1, figures that compared with a median MW of 432 Da and a median clog P of 3.1 for orally bioavailable drugs launched between 1990 and 2007. Pfizer, however, represented a notable exception from the other 3 pharmaceutical houses, with compounds disclosed in their patent applications generally exhibiting a lower median MW and clog P. This reflects a heightened awareness of the importance of controlling physicochemical properties in the design of drug candidates that presumably has its origin in Lipinski’s analysis of drug oral bioavailability. This phenomenon is reflected in several in depth analyses of the relationship between physicochemical properties and drug developability that have been published by Pfizer scientists over recent years. 2.3. Correlates between Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) Profiles and Calculated Physical Properties. Using a principal components analysis, Gleeson analyzed data from a large and structurally diverse set of preclinical compounds profiled at GlaxoSmithKline seeking a correlation between 15 ADMET assays and 12 calculated physicochemical descriptors.20 The ADMET assays comprised solubility, permeability across an artificial membrane, oral bioavailability, clearance, volume of distribution (Vd) and
Table 8. Odds Ratios for the Appearance of Toxicity in Vivo in a Pfizer Data Set of 245 Compounds Based on TPSA and clog P25 total drug TPSA >75 Å
2
TPSA 75 Å2 TPSA 3
0.41 (38)
2.4 (85)
0.81 (29)
2.59 (61)
CNS penetration in the rat, rat brain tissue and plasma protein binding, P-gp efflux, inhibition of hERG, and inhibition of the P450 isozymes 1A2, 2C9, 2C19, 2D6, and 3A4, while the physicochemical descriptors evaluated were log P, log D, HBAs, HBDs, positive and negative ionization states, molecular flexibility, molar refractivity, MW, PSA, RBs, and HAC. The observations and trends are captured synoptically in Table 7 and provide useful insights into the preferred physicochemical properties that satisfy basic in vitro and in vivo profiling criteria. Ionization state, clog P, and MW were considered to be the most useful predictors of potential ADMET problems with MW and clog P identified as the key drivers, leading to the succinct conclusion that preferable molecules were those with a MW of 75 Å2
promiscuity
TPSA 3
0.44 (13)
6.25 (29)
Table 10. Physicochemical Properties Associated with 2,512 Compounds from a Novartis Data Set Defined As Promiscuous, Moderately Promiscuous, and Selective on the Basis of an Analysis of Activity in Biochemical Assaysa all compounds
a
moderately promiscuous
promiscuous
selective
MW
460
493
472
436
Alog P
3.7
4.4
3.9
3.3
HBA
5.2
5.4
5.2
5.2
HBD
2.0
2.1
2.1
1.9
O count
3.0
2.5
2.8
3.3
N count
4.0
4.6
4.2
3.6
rot bonds
7.0
7.7
7.2
6.6
ring count
4.0
4.6
4.3
3.6
The data presented are mean numbers.
2002 and 2006 provided interesting insights into compound survival.2527 The data set comprised 50% basic, 40% neutral, and 10% acidic molecules, and both measured and calculated properties were used for the analysis.25 PK data were available for all of the compounds, and correlates with both the free and total drug concentrations were evaluated, with 10 μM set as the toxicity threshold for total drug and 1 μM for free drug concentrations. The descriptors that emerged as being most closely related to the observation of toxicity were topological polar surface area (TPSA) and clog P. Thresholds were set at a TPSA of 75 Å2 and a clog P of 3 for the analysis of toxicity odds, the ratio of toxic to nontoxic compounds, which were determined for the data set and are captured in Table 8. Compounds with a low clog P and a high TPSA were found to be 2.5-fold less likely to be toxic, while those with a high clog P and a low TPSA were 2.5-fold more likely to be toxic. This afforded an odds ratio of >6 between the two extremes, and the results were similar whether calculations were based on either total or free drug concentration in plasma. The presence of a single risk factor was found to moderately increase the potential for the appearance of toxicity, but compounds with both risk factors showed a significant propensity to manifest problems. As noted by the authors, a low TPSA is associated with deeper tissue penetration, which would increase the chances for toxicity, although, interestingly, there was no apparent correlation between the volume of distribution and toxicity. Biochemical promiscuity was also assessed by analyzing CEREP BioPrint profiling data, available for 108 compounds that had been evaluated in 48 assays, with promiscuous behavior defined as >50% activity at a concentration of 10 μM in 3 or more assays. Biochemical promiscuity showed a similar correlation with TPSA and clog P to the observation of toxicity in vivo, with the strongest effect again seen when both risk factors were present (Table 9). Compounds with a high clog P and a low TPSA were 25-fold less
likely to show specificity than those exhibiting a low clog P and a high TPSA.25 Thus, less polar, more lipophilic compounds were more likely to be problematic, an observation that demonstrated consistency across a broad range of toxicities and chemical structural matter. An interrogation of the biochemical fidelity associated with a set of 3138 compounds at Novartis provided additional insights into correlates with physicochemical properties while also identifying structural elements that are more frequently associated with promiscuity.28 The modeling exercise was based on 2512 compounds selected randomly, with the remaining 626 compounds and 119 marketed drugs used as test sets. The compounds analyzed had been screened in at least 50 of 79 assays, the majority of which were G-protein coupled receptors, and a target hit-rate parameter (THR10) was developed as the index of promiscuity, defined as follows: THR 10 ¼
no: of targets inhibited by 50% at 10 μ M no: of targets tested
A THR10 of g20% was used as the criterion to define a promiscuous compound, while compounds with a THR10 of e5% were considered to be selective and those with a THR in the 520% range viewed as being of moderate promiscuity.28 This analysis allowed categorization of the data set into 604 (24%) promiscuous compounds, 1171 (47%) selective molecules with the remaining 737 (29%) falling into the moderately promiscuous category. MW and Alog P were higher for promiscuous compounds by 10 and 33%, respectively, than for selective compounds, and compounds exhibiting promiscuity incorporated a significantly higher number of rings and rotatable bonds (data summarized in Table 10). Promiscuous compounds also possessed a higher number of N atoms but a lower number of O atoms than selective compounds, possibly reflecting problems with the presence of basic amines, while the prevalence of HBDs and HBAs was not significantly different across the 3 categories.28 A deeper analysis that sought to equate the appearance of specific substructures with compound promiscuity revealed a preponderance of indole, furan, and piperazine heterocycles in the promiscuous compound data set, while carboxylic acids were found to be more frequently associated with the selective class. The latter observation was attributed to the presence of a negative charge potentially leading to unfavorable interactions between a drug candidate and proteins, a tactic well-known to be helpful in avoiding inhibition of the hERG ion channel.29 Compounds containing tetrazole and sulfonamide moieties exhibited little difference in their distribution across the promiscuity landscape, although an indication of the presence of overt acidity in these elements was not provided. The overall conclusion of the synopsis was that small hydrophilic compounds with a carboxylic acid moiety were the most likely to show selectivity while bulky, hydrophobic amines were the most likely to distribute to the promiscuous set. The authors developed a na€ive Bayesian model using a range of descriptor fingerprints that compared the frequency of features between selective and promiscuous sets of compounds. The best models showed lower promiscuity for marketed drugs than those in early development or those that failed during clinical development. Interestingly, the majority of the marketed drugs that were predicted to be promiscuous were compounds that targeted the central nervous system.28 1427
dx.doi.org/10.1021/tx200211v |Chem. Res. Toxicol. 2011, 24, 1420–1456
Chemical Research in Toxicology At Roche, a cohort of 213 compounds originating from 62 projects that had been profiled in a series of biochemical assays between 2004 and 2007 was assessed for promiscuity, defined by counting the number of off-target hits.30 The compounds were divided into 2 categories, those exhibiting g30% and g90% inhibition at a concentration of 10 μM, and the molecular descriptors analyzed for a potential relationship with promiscuity were clog P, clog D , MW, and pKa. The compounds demonstrating g30% inhibition in one assay revealed that pronounced promiscuity was not associated with molecules with a clog P of 4% of targets incorporated a basic element.30 In particular, off-target activity was common among several series of compounds designed as aminergic GPCR ligands or reuptake transporters which together accounted for 49% of the compounds in this category. Over half (55%) of basic compounds from nonaminergic GPCR programs also exhibited promiscuity, confirming that the presence of a basic amine leads to reduced specificity regardless of target class.27,30 A potentially confounding observation revolved around several highly lipophilic compounds for which clog P values exceeded 6 and which were associated with low promiscuity. Most of these compounds were also poorly soluble, 30% inhibition toward a target at a concentration of 10 μM was considered to be promiscuous.9,31 Bases and quaternary bases exhibited higher levels of promiscuity than zwitterions, acids, or neutral compounds, while increased lipophilicity was associated with reduced specificity irrespective of the ionization class. An apparent relationship in which promiscuity increased with the number of rings in a molecule could not be distinguished from lipophilicity. The relationship between MW and specificity for this data set was complicated and failed to reproduce earlier observations with a larger cohort of compounds in which promiscuity declined as MW increased.9,32 An increase in the lipophilicity of drug molecules has also been equated with a higher propensity for inhibition of the hERG cardiac ion channel, an activity associated with arrhythmogenic potential that is manifested in clinical studies as torsades de pointes.29,33 Insight was gleaned from an analysis of 7,685 AstraZeneca compounds for which hERG inhibition data were available in a proprietary whole-cell electrophysiology-based
REVIEW
Table 11. Correlation between hERG Inhibition Data and Lipophilicity Associated with 7,685 Astra-Zeneca Compounds acid
base
neutral
zwitterion
n
350
4302
2598
435
mean hERG pIC50
3.7
5.2
4.5
4.4
mean AZlog D
0.71
2.5
2.9
1.5
mean clog P
3.1
3.6
3.2
4.4
Table 12. Probability of hERG inhibition versus Lipophilicity for Acidic, Basic, Neutral, and Zwitterionic Compounds in an Astra-Zeneca Data Set
Table 13. Upper Limits of Lipophilicity to Avoid hERG Inhibition in Acidic, Basic, Neutral, and Zwitterionic Compounds in an Astra-Zeneca Data Set target upper limits of log D and clog P to predict that >70% of compounds achieve a hERG IC50 >10 μM acids
bases
neutrals
zwitterions
log D
>4
1.4
3.3
2.3
clog P
>9
1.9
4.0
4.4
screening assay.33 Compounds were divided into acidic, basic, neutral, and zwitterionic species, with the distribution of ion class across this data set compiled in Table 11 along with hERG inhibition data and mean lipophilicity profiles. The lipophilicity data are based on experimental log D measured at pH 7.4, available for 1,211 compounds, or calculated as AZlog D using a proprietary algorithm. The data were analyzed and presented as the probability of achieving a compound with a hERG IC50 of >10 μM for the individual ion classes in relationship to their lipophilicity (results summarized in Table 12). Basic compounds exhibited the highest propensity to inhibit hERG, exacerbated by increased AZlog D, while acids, zwitterions, and neutral compounds offered lower potential, although a higher AZlog D correlated with increased inhibitory activity. These data were configured to provide a suggested upper limit of lipophilicity for each ion class that would offer improved odds of reducing the potential to encounter hERG inhibition (Table 13). Lipophilic basic molecules showed the lowest probability to achieve a hERG IC50 of >10 μM, and a clog P of
