3164
J. Med. Chem. 2005, 48, 3164-3170
A Bioavailability Score Yvonne C. Martin* Advanced Technology Division, Global Pharmaceutical Research and Development, Abbott Laboratories, Abbott Park, IL 60064-6100 Received October 4, 2004
Responding to a demonstrated need for scientists to forecast the permeability and bioavailability (F) properties of compounds before their purchase, synthesis, or advanced testing, we have developed a score that assigns the probability that a compound will have F > 10% in the rat. Neither the rule-of-five, log P, log D, nor the combination of the number of rotatable bonds and polar surface area successfully categorized compounds. Instead, different properties govern the bioavailability of compounds depending on their predominant charge at biological pH. The fraction of anions with >10% F falls from 85% if the polar surface area (PSA) is e 75 Å2, to 56% if 75 < PSA < 150 Å2, to 11% if PSA is g 150 Å2. On the other hand, whereas 55% of the neutral, zwitterionic, or cationic compounds that pass the rule-of-five have >10% F, only 17% of those that fail have > 10% F. This same categorization distinguishes compounds that are poorly permeable from those that are permeable in Caco-2 cells. Further validation is provided with human bioavailability values from the literature. Introduction Recent attention has focused on the need to assess the potential for bioavailability problems of potential drug candidates early in the drug discovery cycle.1 This need is particularly acute for those drug discovery programs that follow a genomic or crystallographic target for which there is no cell- or animal-based assay for potency. One approach to solve the problem has been to use high throughput/low quantity methods such as PAMPA chromatography2 or Caco-2 permeability measurements3 to provide a prediction of whole-cell and animal activity. However, the ideal method would be one that would require no sample, a virtual ADME prediction technique, and so could be used before a compound is synthesized or purchased. The history of the search for such virtual methods dates to the earliest studies of the relationships between the physical chemical and biological properties of molecules.4,5 The search continues to this day,6 with claims of new, improved models appearing frequently.7-12 Others have suggested rules-of-thumb that improve the chances of a compound being well absorbed.13-15 There are examples that consideration of physical properties has been useful in the optimization of the bioavailability or permeability within specific series of compounds.16,17 However, the generalization of such studies to diverse molecules has not been convincingly demonstrated. Indeed, there is no consensus as to the physical properties relevant for bioavailability. For example, although an increase in lipophilicity is generally considered to be related to an increase in bioavailability,18 in at least one study a decrease in cell penetration accompanies an increase in hydrophobicity.19 Others have emphasized the importance of keeping hydrogen-bonding or polar surface area to a * To whom correspondence should be addressed. R47E AP10/2, 100 Abbott Park Rd, Abbott Park IL 60064-6100. e-mail yvonne.c.martin@ abbott.com.
minimum.20-25 In some cases, utilization of a specific transporter, which has structure-activity relationships of its own, is necessary.26 In addition to published models, several vendors offer computer programs for predicting permeability or bioavailability; for example, Cerius2,10,27 Idea,28 Absolv,29-31 QikProp,32 QMPRPlus,33 VolSurf,34,35 and OraSpotter.36 This report describes our investigations that led to a method that provides the probability that a compound will have sufficient rat bioavailability, 10%, such that it will not be dropped for further development because of bioavailability concerns. We envision that the score will be used as a complement to virtual screening, for triaging HTS, and for selecting compounds for purchase. We will show that it outperforms the conventional ruleof-five13 or other simple publicly available algorithms.10,15,37-40 Methods Table 1 provides a summary of the properties of the datasets used. The rat bioavailability data, measured by our standard procedure,41 had been collected over a period of years. It contains compounds that targeted a wide variety of therapeutic targets. This set of 553 compounds contains 99 different rings and 180 different side chains. When clustered at 0.75 similarity, there are an average of 2.27 compounds per cluster. In contrast, the Caco-2 permeability screening data was collected during only one year with the result that this dataset is less structurally diversesit contains an average of 4.41 compounds per cluster. We omitted from consideration only macrolides and a series of compounds that had been synthesized to target the liver. In this dataset 47.74% of the compounds have rat bioavailability less than 10%, 35.80% have rat bioavailability greater than or equal to 10 but less than 50%, 12.66% have bioavailability greater than or equal to 50% but less than 80%, and 3.80% have bioavailability greater than or equal to 80%. We used human bioavailability data curated by others to validate our initial studies. Continuous human bioavailability data was taken from three review articles written by experts in interpreting such data.42-44 A compound was used in the analysis only if the reported bioavailability is a narrow range in all articles in which it was reported; this criterion eliminated
10.1021/jm0492002 CCC: $30.25 © 2005 American Chemical Society Published on Web 04/05/2005
A Bioavailability Score
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3165
Table 1. Properties of the Datasets Analyzed human bioavailability, F dataset
human intestinal absorption
continuous (%)
category (0-4)
variable
rat, %F
Caco-2 permeability category
no. of molecules no. of monocations no. of uncharged no. of monoanions no. of zwitterions
96 36 28 20 9
449 189 137 76 23
519 233 151 83 24
118 43 38 13 11
553 171 150 134 38
617 326 134 114 22
endpoint mean endpoint range
77.8 1.9-100
54.4 0-100
2.61 0-4
-
22.31 0-168
1.11 0-4
pass rule-of-five
96%
94%
94%
86%
79%
72%
log P mean log P range
1.63 -3.65 to 14.36
2.00 -6.66 to 14.36
2.09 -6.66 to 14.36
2.38 -9.8 to 14.36
3.91 -2.2 to 9.5
4.69 -1.5 to 9.5
MW mean MW range
323.9 138.1-1203
335.5 46.1-1203
333.3 46.1-1203
376.8 123-1312
433.6 160-7935
454.4 159-780
Hb donors mean Hb donors range
2.05 0-12
1.83 0-12
1.83 0-12
2.08 0-20
1.46 0-7
2.05 0-6
Hb acceptors mean Hb acceptors range
5.03 1-23
5.56 1-23
5.45 1-23
6.08 1-38
6.19 1-15
6.23 2-15
polar surface area mean polar surface area range
82.7 0-204
83.0 2.91-245
81.7 2.91-245
83.6 5.34-685
86.4 11.5-261
88.7 15.8-234
clusters @0.75 similarity
83
375
244
140
compounds per cluster
1.16
1.38
2.27
4.41
118 compounds, leaving 449 with consistent values. To this set we added additional compounds from a previous model9 and used their criteria to assign the human bioavailability category for compounds from the review articles but not in their model. Because most of the compounds reported in the review articles have good human bioavailability, we sought to add more compounds with low bioavailability. Recognizing that bioavailability cannot exceed intestinal absorption (HIA), we added additional compounds for which HIA is reported8 to be 0-5% and assigned them to the lowest bioavailability category. Note that the majority of the compounds from the literature for human bioavailability pass the rule-of-five.13 We calculated the physical properties that have traditionally been considered to be related to permeability or bioavailability: molecular weight, octanol-water log P,45 log D,46 polar surface area,47 and number of rotatable bonds. For the latter we used a daylight toolkit program to (1) count all single bonds not in a ring, (2) subtract the number of amide bonds not in a ring, and (3) subtract the number of symmetrical terminal groups such as methyl, nitro, nitrile, unusbstituted- or methylalkyne, trifluoromethyl, trichloromethyl, and tribromomethyl. The number of hydrogen bond donors and acceptors was calculated with a simple Daylight toolkit program.48 Except in the case of permanently charged compounds, the properties of the compounds are calculated for the neutral form. Compounds were clustered using the Taylor algorithm.49 We generated the predominant charge state of the molecule by a series of empirical SMARTS targets, derived with the aid of the Biobyte MedChem database.50 The main classes of charged molecules are basic aliphatic nitrogens, acidic phenols, carboxylic acids, sulfonic acids, enolates (all tautomers), basic amidines, phosphates, basic imidazoles and tetrazoles, N-acyl aliphatic sulfonamides, S-aryl sulfonamides, and barbiturates. Independent of this process we calculated the fraction in the various ionic states with ADME Boxes.51 Where the two methods disagreed, we again checked literature sources to decide between them. The standard errors were calculated as:52
standard error ) (observed proportion) × [1 - (observed proportion)] . (sample size)
x
Results Table 2 summarizes the results, which will be discussed with the figures that clarify various points. We first examined properties that have traditionally been considered to be associated with permeability or bioavailability.18 Figure 1 shows that log D seems to be somewhat associated with permeability but not with rat bioavailability. (Not shown is that log D at pH values 4.6, 6.4, and 7.4 perform similarly. Log D at pH 7.4 is used for permeability because that is the pH at which
Figure 1. The dependence of (a) Caco-2 permeability on log D at pH 7.4, the pH of measurement (red, impermeable; yellow, slightly permeable; dark blue, somewhat permeable; light blue, permeable; green, very permeable); and (b) rat bioavailability on log D at pH 6.4.
3166 Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9
Martin
Table 2. Statistical Analysis of the Predictivity of Various Properties for Caco-2 Permeability and Rat Bioavailability Caco-2 cell permeability
criterion
rat bioavailability
number number proportion 95% number number proportion 95% in nonnonstd confidence in rat, rat, std confidence class permeable permeable error limits class F 5
290a 216a 75a
174 87 57
0.600 0.403 0.760
0.029 0.033 0.049
0.058 0.067 0.099
251b 194b 53b
114 101 28
0.454 0.521 0.528
0.031 0.036 0.069
0.063 0.072 0.137
MW < 500 MW > 500
411 206
188 165
0.457 0.801
0.025 0.028
0.049 0.056
390 162
159 105
0.408 0.648
0.025 0.038
0.050 0.075
log D < 2.5 and MW < 500 log D < 2.5 and MW > 500 log D 2.5 to 5.0 and MW < 500 log D 2.5 to 5.0 and MW > 500 log D > 5 and MW < 500 log D > 5 and MW > 500
187 103 168 48 47 28
87 87 58 29 34 23
0.465 0.845 0.345 0.604 0.723 0.821
0.036 0.036 0.037 0.071 0.065 0.072
0.073 0.072 0.074 0.141 0.131 0.145
181 69 124 70 32 21
57 57 63 38 19 9
0.315 0.826 0.508 0.543 0.594 0.429
0.035 0.046 0.045 0.060 0.087 0.108
0.069 0.090 0.089 0.119 0.174 0.216
PSA < 75 Å2 PSA 75-150 Å2 PSA > 150 Å2
219 358 40
91 235 27
0.416 0.656 0.675
0.033 0.025 0.074
0.067 0.050 0.148
240 263 50
94 126 45
0.392 0.479 0.900
0.032 0.031 0.042
0.063 0.062 0.085
pass rule-of-five13 fail rule-of-five
445 172
215 138
0.483 0.802
0.024 0.030
0.047 0.061
436 116
191 73
0.438 0.629
0.024 0.045
0.048 0.090
CLOGP e 5.88 and PSA e 131.6 Å2 10 CLOGP g 5.88 and/or PSA g 131.6 Å2
376 241
179 174
0.476 0.722
0.026 0.029
0.052 0.058
405 147
181 83
0.447 0.565
0.025 0.041
0.049 0.082
PSA < 140 Å2 and number of rotatable bonds < 10 15 PSA g140 Å2 and/or number of rotatable bonds g 10
554
312
0.563
0.021
0.042
493
220
0.446
0.022
0.045
63
41
0.651
0.060
0.120
69
44
0.733
0.057
0.114
ABS ) 0.85 ABS) 0.55 or 0.56 ABS < 0.20 positive charge, pass rule-of-five positive charge, fail rule-of five negative charge, pass rule-of-five negative charge, fail rule-of-five positive charge, PSA < 75 positive charge PSA 75-150 positive charge, PSA >150 negative charge, PSA < 75 negative charge, PSA 75-150 negative charge, PSA > 150
6 455 156 368 114 77 58 213 261 8 6 97 32
1 223 129 169 95 46 43 90 163 1 1 62 26
0.167 0.490 0.827 0.459 0.833 0.597 0.741 0.423 0.625 0.125 0.167 0.639 0.813
0.152 0.023 0.030 0.026 0.035 0.056 0.057 0.034 0.030 0.117 0.152 0.049 0.069
0.304 0.047 0.061 0.052 0.070 0.112 0.115 0.068 0.060 0.234 0.304 0.098 0.138
34 438 80 344 26 93 90 206 160 4 34 103 46
5 191 68 150 23 41 50 89 80 4 5 45 41
0.147 0.436 0.850 0.436 0.885 0.441 0.556 0.432 0.500 1.00 0.147 0.437 0.891
0.061 0.024 0.040 0.027 0.063 0.051 0.052 0.035 0.04
0.121 0.047 0.080 0.053 0.125 0.103 0.105 0.069 0.079
0.061 0.049 0.046
0.121 0.098 0.092
a
pH 7.4.
b
pH 6.4.
the measurements were made; pH 6.4 is used for rat bioavailability to account for the slightly acidic nature of the regions of the GI tract at which most absorption occurs.) For example, 57 of the 75 compounds, 76%, with log D > 5 are not permeable, but only 53% of the compounds with this log D have low rat bioavailability. Figure 2 shows that compounds with a molecular weight greater than 500 are in general less permeable and tend to have lower bioavailability. However, although 80% of the compounds with molecular weight greater than 500 are not permeable, only 65% of such compounds have low rat bioavailability. Figure 3 and Table 2 show that the effect of molecular weight is most pronounced at low log D values in that the 84% of the 103 compounds with log D < 2.5 and MW > 500 are not permeable and 83% have low bioavailability. Presumably the low permeability and bioavailability of these compounds is associated with a large number of polar atoms. Recently the rule-of-five has become a popular way to characterize compounds.13 It was derived by an analysis of the properties of orally active marketed drugs. Figure 4 shows that 80% of the compounds that fail the rule-of-five are not permeable, whereas 48% of those that pass are not permeable. The contrast is less dramatic for rat bioavailability for which the corresponding numbers are 63% and 44%. Thus 43 of the 116
Figure 2. The dependence of (a) Caco-2 permeability (legend as in Figure 1) and (b) rat bioavailability on molecular weight.
A Bioavailability Score
Journal of Medicinal Chemistry, 2005, Vol. 48, No. 9 3167
Figure 4. The predictivity of the rule-of-five for (a) Caco-2 permeability and (b) rat bioavailability. Legend as in Figures 1 and 3.
Figure 3. (a) Caco-2 permeability as a function of log D (pH 7.4) and MW. Note that a log D value could not be calculated for 36 compounds. (b) Rat bioavailability as a function of log D (pH 6.4) and MW (red, 150 Å2, 0.56 if PSA is between 75 and 150 Å2, and 0.85 if PSA is
