research focus

REVIEWS

Computational methods for the prediction of drug-likeness

David E. Clark and Stephen D. Pickett
Recently, one of the key trends in the pharmaceutical industry has been the integration of what have traditionally been considered development activities into the early phases of drug discovery. The aim of this paradigm shift is the prompt identification and elimination of candidate molecules that are unlikely to survive later stages of discovery and development. In this review, the authors examine the growing role that is being played by computational methods in this filtering process, with a particular focus on the prediction of intestinal absorption and bloodbrain barrier penetration. checked both for their absorption properties using Caco-2 (a human intestinal epithelial cell line derived from a colorectal carcinoma)2 or MDCK (MadinDarby canine kidney)3 cell monolayers and for their susceptibility to metabolic degradation using liver microsomes or hepatocytes4. As valuable as these experimental filters are, they do have some limitations. For example, they require physical samples of compounds for testing and, despite significant technical advances, they remain time-consuming and resource-intensive. Thus, there is currently much interest in the development and application of computational methods for predicting drug-likeness5. Such methods could be applied to virtual compounds or libraries permitting the rapid and costeffective elimination of poor candidates prior to synthesis. The current state-of-the-art in terms of the prediction of drug-likeness will now be reviewed, with particular reference to computational methods for the prediction of: Drug-likeness in a general sense Intestinal absorption Bloodbrain barrier (BBB) penetration. It should be noted that with regard to the latter two categories, only the prediction of absorption by passive mechanisms will be considered. Also, while of much interest and importance, the prediction of both active transport by various carrier systems and the metabolic fate of compounds will not be covered in this review. Prediction of general drug-likeness Initial strategies towards this goal involved the use of computational filters to remove compounds deemed to be chemically unsuitable for screening purposes69. Such filters typically incorporate substructure searches for toxic or reactive groups (such as acid halides) and can include limits on MW and the number of rotatable bonds. These

uch has been written recently concerning the impact of combinatorial chemistry and highthroughput screening (HTS) on drug discovery. However, there is a growing realization that, given the vast size of organic chemical space (possibly 1018 compounds), drug discovery cannot be reduced to a simple synthesize-and-test lottery. Researchers must somehow load the dice to give the compounds they are making the best possible chance of becoming drugs. In other words, there is a need to identify and/or design a subset of drug-like molecules from the vast expanse of what could possibly be synthesized. To this end, a battery of in vitro ADME (absorption, distribution, metabolism, excretion) screens has been implemented in most pharmaceutical companies with the aim of discarding compounds in the discovery phase that are likely to fail further down the line1. The motto in the industry is now fail fast, fail cheap. Thus, candidate molecules are

David E. Clark* and Stephen D. Pickett, Computer-Aided Drug Design, Rhne-Poulenc Rorer Ltd, Dagenham Research Centre, Rainham Road South, Dagenham, Essex, UK RM10 7XS. *tel: 44 020 8919 3353, fax: 44 020 8919 2029, e-mail: [email protected]
DDT Vol. 5, No. 2 February 2000
1359-6446/00/$ see front matter Elsevier Science Ltd. All rights reserved. PII: S1359-6446(99)01451-8

49

REVIEWS
methods provide a very useful initial filtering of any data set and can be used prior to further data analysis. However, by their very nature, instead of addressing the question of what makes a drug, these filters attempt to eliminate definite non-drugs. A more sophisticated approach to predicting drug-likeness is to examine two sets of compounds, one containing known drugs and the other comprising compounds known (or presumed) not to be potential drugs. The properties or characteristics of the former set that differentiate them from the latter can then be elucidated using computer methods. This type of approach has already been applied by several research groups.

research focus
as a drug-like set and the ACD as a non-drug-like set12. In this work, the compounds were characterized using atomtype descriptors originally developed for predicting octanol/water partition coefficients. The resulting scoring scheme succeeded in classifying 77% of the WDI correctly as drugs and 83% of the ACD correctly as non-drugs. The method is very fast, being able to classify approximately 100 compounds per second on a modern workstation. More recently, a different classification technique (recursive partitioning) has been applied to these data sets with similarly successful results [Wagener, M. and van Geerestein, V.J. (1999) Analysing large data sets with decision trees: Discriminating between potential drugs and nondrugs. Abstracts of the 217 th American Chemical Society Meeting. 2125 March, Anaheim, CA, USA].

Genetic algorithm-based approaches A genetic algorithm-based approach has been used10 to calculate weights that describe the differential occurrences of simple, one-dimensional (1-D) molecular properties (MW, the numbers of hydrogen bond donors and acceptors, rotatable bonds and aromatic rings) and a topological shape descriptor in compounds from the World Drug Index (WDI, Derwent Information, http://www.derwent. com), assumed to be drug-like, and compounds from the SPRESI database (Daylight Information Systems, http://www.daylight.com), assumed to be non-drug-like. The resulting biological activity profiles are reasonably effective at discriminating between the two classes, especially considering the simple nature of the descriptors used to characterize the molecules. A method incorporating this work has been used at GlaxoWellcome (Stevenage, Hertfordshire, UK) to filter compounds prior to HTS (Ref. 9). Neural network-based approaches Researchers at Vertex (Cambridge, MA, USA) trained a Bayesian neural network to recognize drug-like molecules using the Comprehensive Medicinal Chemistry (CMC, MDL Information Systems, http://www.mdli.com) and Available Chemicals Directory (ACD, MDL Information Systems) as their sets of drugs and non-drugs, respectively11. Using a combination of simple 1-D descriptors and the ISIS (MDL Information Systems) 2-D substructural keys to characterize the molecules, the neural network classified 90% of the CMC compounds correctly as drugs while incorrectly predicting only 11% of the ACD to be drug-like. The same network also classified nearly 80% of another drug database used as a test set, the MDDR (MDL Drug Data Report, MDL Information Systems), as drug-like. In similarly motivated, but independent, work carried out at BASF AG (Ludwigshafen, Germany), a feed-forward neural network was trained using molecules from the WDI

Limitations Thus, it would seem possible to develop rapid computational filters that can distinguish between compounds that are drug-like and non-drug-like. The lack of validated sets of drugs and non-drugs is a disadvantage for these methods and there are undoubtedly compounds that do not easily fall into either category. However, given the time and expense of conducting such a validation exercise, there is little choice but to use sets like those already mentioned. Another criticism is that the classifiers can only recognize those compounds that resemble existing drugs as drug-like compounds from completely new classes could be misclassified and this could hinder innovation13. In terms of applications, these techniques will probably be most useful in lead discovery where they could be used for prioritizing compounds for HTS or for purchasing from external suppliers. They could also be applied for the design of biased combinatorial libraries. Their use in lead optimization is more restricted, as they can only give a yes/no answer to the question is this a drug-like molecule? At this stage in the drug discovery process, it is necessary to resort to more specific approaches such as those described in the following sections.
Prediction of intestinal absorption For most drugs, the preferred route of administration is by oral ingestion. Researchers have therefore sought to delineate the physicochemical properties that favour intestinal absorption14,15 and develop computational methods for its prediction16.

The rule-of-five Probably the best-known of these methods is the rule-offive17 devised by Lipinski and coworkers at Pfizer

50

DDT Vol. 5, No. 2 February 2000

research focus
(Groton, NJ, USA) from an analysis of 2245 drugs from the WDI believed to have entered Phase II trials. As implemented in the Pfizer registration system, the rule-of-five generates an alert (indicating possible absorption problems) for compounds where any two of the following conditions are satisfied: Molecular weight 500 Number of hydrogen-bond acceptors 10 Number of hydrogen-bond donors 5 Calculated logP 5.0 (if ClogP is used) or MlogP is used).

REVIEWS
Predicting Caco-2 cell permeability Instead of analyzing existing collections of drugs, another approach has been to model specific, absorption-related quantities. For example, the following QSPR (quantitative structureproperty relationship) model was developed to describe the Caco-2 permeabilities of 17 diverse drug compounds20:
logPapp 4.15 (if 0.008 MW 0.043 PSA 5.165 (1)

Here, logPapp is the logarithm of the apparent permeability (cm s 1) through the monolayer and the PSA indicates the polar surface areas of the compounds. The correlation In this rule, any oxygen (O) and nitrogen (N) atoms are coefficient (r) for this equation is 0.833. The equation indidefined as hydrogen-bond acceptors and NH or OH cates that permeability increases with increasing MW and groups are considered as hydrogen-bond donors. LogP decreasing PSA. This data set has also been studied using refers to the octanol/water partition coefficient of a comother computed descriptors and partial least squares (PLS) pound and is used as a measure of lipophilicity. Many regression21. The results of this analysis also indicated the approaches for calculating logP have been developed importance of minimizing properties associated with including ClogP (Daylight Information Systems, Mission hydrogen bonding if good permeability is to be achieved. Viejo, CA, USA) and MlogP (developed by Moriguchi and This conclusion is in accord with the observation that cel18. An analysis of the CMC database by Amgen coworkers) lular permeability is primarily dependent on the desol(Thousand Oaks, CA, USA) researchers19 has reinforced vation potential of the polar functional groups in a molsome of these findings, particularly concerning the preecule and only secondarily dependent on its lipophilicity22. ferred ranges of MW and logP. The concept of using PSA as a predictor of absorption has recently attracted much interest. The PSA is defined as that part of the molecular surface that arises from oxygen (a) (b) or nitrogen atoms, or hydrogen atoms attached to nitrogen or oxygen atoms (see Fig. 1). Some workers also inO clude sulphur (and hydrogen attached to sulphur) in their definitions. Early N work by Palm and coworkers at N Uppsala University23 established a S strong correlation (r2 0.99) between the dynamic PSA (PSAd) and the N O Caco-2 permeabilities of six betablockers (Fig. 2). PSAd is calculated by taking a N H Boltzmann-weighted average of the OH PSA values of all conformers within 2.5 kcal mol 1 of the lowest-energy conformer found during a conformational search (a relatively time-conDrug Discovery Today suming procedure). A similar approach used by other workers showed that Figure 1. The beta-blocker timolol, showing the atoms contributing to its polar surface area in (a) two dimensions (circled atoms) and in (b) three dimensions PSAd correlated well with the Caco-2 (molecular surface coloured purple). permeability data of some beta-blockers and their ester prodrugs24. Lower

DDT Vol. 5, No. 2 February 2000

51

REVIEWS

research focus

interest in generating predictive models based on (currently quite sparse) human absorption data. The initial work O N N O N O H H H in this area, again by Palm and coworkOH OH OH ers, involved a set of 20 drugs covering a wide range of fractional absorption O HN O values (%FA) in humans (Table 1)27. NH2 O These compounds were carefully selectCH3 Metoprolol Atenolol Practolol ed to ensure they were primarily absorbed by passive processes and to include a wide range of physicochemiN O N O N O H H H cal properties. Using this set, a strong H OH OH O OH N sigmoidal correlation between PSAd and %FA (r2 0.94) was demonstrated, suggesting that compounds with a PSAd Oxprenolol Alprenolol Pindolol 140 2 are likely to show poor Drug Discovery Today absorption characteristics (%FA 10). Figure 2. The beta-blockers showing a good correlation between PSA and Caco-2 28 have applied 23. Norinder and coworkers cell permeability their MolSurf (Qemist AB, Karlskoga, Sweden) descriptors and PLS regression PSAd was proffered as an explanation for the preferential permethods to this data set and arrived at similar conclusions meabilities of the prodrugs. More recent work from the to those reached with their Caco-2 study21. A novel molUppsala group has examined the effect of computing PSAd in ecular surface-based descriptor, known as the molecular different simulated environments (water, vacuum, chlorohashkey, has also been correlated with these data using a form) and concluded that, for most of the nine beta-blockers backpropagation neural network29. studied, the effect of altering the environment on PSAd was An alternative approach using a genetic algorithm with small25. In addition, consideration of the non-PSA has been a neural network scoring function and a larger training shown to be beneficial in modelling the Caco-2 permeabiliset of 76 compounds, resulted in the generation of a ties of an expanded set of 19 oligopeptides26. QSPR model relating %FA to several structure-derived descriptors30. The root-mean square errors (RMSE) predicted Predicting human fractional absorption from this model were quite reasonable: 9.4% for the training set and 16.0% for a test set of ten compounds. While Caco-2 monolayers can provide useful data for predicting intestinal absorption, they are only a model of the However, the data set used in this work is heavily biased human absorption process. There is, therefore, much towards well-absorbed compounds (only 15 of the 86 compounds have %FA values of less than 50), reducing the capacity of the model to Table 1. The 20 drugs studied by Palm and coworkers, together predict %FA values for less well-absorbed 27 with their mean fractional absorption data in humans compounds. Furthermore, the data set conaMean %FA Compound Compound Mean %FA tains compounds that are absorbed by active transport processes as well as passively Metoprolol 102.0 Metolazone 64.0 absorbed compounds. This is problematic Nordiazepam 99.0 Tranexamic acid 55.0 because it would seem unrealistic to expect Diazepam 97.0 Atenolol 54.0 to find a single model that will accurately Oxprenolol 97.0 Sulpiride 36.0 predict such different physical processes. Phenazone 97.0 Mannitol 26.0 Oxazepam 97.0 Foscarnet 17.0 Passive absorption is dependent on the Alprenolol 96.0 Sulfasalazine 12.0 rate of diffusion of a compound, either bePractolol 95.0 Olsalazine 2.3 tween (paracellular) or across (transcelluPindolol 92.0 Lactulose 0.6 lar) the cells comprising the epithelial barCiprofloxacin 69.0 Raffinose 0.3 rier. This diffusion rate will, in turn, a Fractional absorption. depend on the physicochemical properties

52

DDT Vol. 5, No. 2 February 2000

research focus
of the compound. By contrast, each active transport mechanism requires more specific molecular recognition. It has recently been shown how the time-consuming calculation of PSAd can be avoided by using PSA values based on a single, rapidly calculated low-energy conformer31. On application of this method to the 20-compound data set shown in Table 1, an almost identical correlation was obtained (r2 0.94; see Fig. 3). This faster computational protocol enables the rapid screening of large (virtual) data sets, such as combinatorial libraries. The criterion for poor absorption of PSA 140 2 was robust when applied to the set of 76 compounds referred to earlier30, if the actively transported compounds are excluded31.

REVIEWS
groups in a molecule favours passive absorption. The underlying physical explanation for this is that, in an aqueous environment (e.g. the intestine), the polar groups of a molecule will be solvated. Entry to the more lipophilic environment of the intestinal epithelium requires the desolvation of these groups. If this is too energetically expensive, then absorption will be hindered. It would seem from these results, that PSA is in some way describing the degree of solvation and, thus, the energy required for desolvation of a compound33. However, workers at Pharmacia and Upjohn (Kalamazoo, MI, USA) have recently failed to find a good correlation between PSA and the experimentally determined hydrogen-bonding potential for a set of peptides and peptidomimetics34. Thus, there is still much scope for further research in this area, perhaps looking at more sophisticated definitions of the PSA (Ref. 31) and accounting for the relative hydrogen-bond donor and acceptor strengths of molecules35. Prediction of BBB penetration The BBB is a complex cellular system whose purpose is to maintain the homeostasis of the CNS by separating the brain from the systemic blood circulation36. For drugs targeted at the CNS, BBB penetration is a necessity (unless invasive or intranasal delivery routes are being considered). By contrast, for drugs aimed at other sites of action, passage through the BBB might produce unwanted sideeffects. Hence, in drug discovery, the determination of BBB penetration is of great importance, and in vitro37 and in vivo38 techniques have been devised for this purpose. Computational approaches to the prediction of BBB penetration have been recently reviewed39. For the purposes of this review, they can be divided into two classes: those based on observations of CNS activity/inactivity and those seeking to model a specific experimental measure of BBB penetration.

Predicting human effective permeability A different measure of intestinal absorption in humans, effective permeability (Peff), has also been investigated32. Thirteen passively absorbed compounds (Table 2) were considered in the study and their logPeff values correlated with some calculated descriptors using PLS analysis. Notably, the simplest model with reasonable statistics (q2 0.82, where q2 is the cross-validated r2; r2 0.85) included a single-conformer PSA value together with a count of the number of hydrogen-bond donors (HBD):
logPeff 2.546 0.011 PSA 0.278 HBD (2)

As with Equation (1), this model indicates that decreasing the number of polar (hydrogen bonding) functional

100 Fractional absorption (%) 80 60 40 20 0 0 50 100 150 200 250 300

Polar surface area (2)

Drug Discovery Today

Figure 3. Correlation between human fractional absorption and polar surface area calculated using the dynamic (in red)27 and single conformer methods (in green) 31.

Predictions based on CNS activity/inactivity Initial work examining CNS activity40 used a data set of 28 compounds, 16 of which showed high CNS activity (class designated CNS ) with the remainder showing low or no CNS activity (CNS ). Using discriminant analysis, three different models were derived, each using a different combination of topological descriptors together with a hydrogenbonding parameter. Interestingly, although ClogP was available as a variable, it was not incorporated into any of the models during the stepwise selection process. Each of the models performed well, classifying all the CNS compounds correctly, and only one of the CNS set incorrectly. Other workers33 have examined a larger set of

DDT Vol. 5, No. 2 February 2000

53

REVIEWS
Table 2. The 13 passively absorbed compounds and their logPeff values32
Compound
Antipyrine Hydrochlorothiazide Verapamil Ketoprofen Metoprolol Terbutaline Carbamazepine
a

research focus
however, it was demonstrated that this training set was too small to provide a robust predictive model for logBB (Ref. 44). Thus, LogPeff Abraham and coworkers collated a larger set of approximately 60 compounds with logBB 3.36 data that they successfully correlated with 4.70 five solute descriptors45. Their results indi3.62 5.30 cated that increasing molecular volume and 3.08 excess molar refraction favours partitioning 3.54 into the brain, while properties such as dipolarity/polarizability and hydrogen-bond acidity/basicity work against it. One disadvantage of the Abraham approach, as originally described, is that it requires manual dissection of the compound of interest and summation of descriptor values for the resulting fragments. This means it is unsuitable for use in computational screening of large compound collections (virtual HTS). Automation of the calculation of descriptor values has recently been reported46, which should help to alleviate this problem. The Abraham data set has formed the basis for several subsequent studies. Workers at Pfizer developed a QSPR correlating logBB with GoW (the computed free energy of solvation of a compound in water)47. Norinder and coworkers have again used their MolSurf descriptors, in conjunction with PLS analysis, and concluded that partitioning into the brain is favoured by minimizing the number of groups capable of hydrogen bonding and by increasing lipophilicity48. However, both of these approaches require computationally expensive and non-automated calculation procedures. More rapid approaches have recently been reported49,50, which offer the possibility of rapidly screening large (virtual) compound collections for their potential to cross the BBB. In the first of these more rapid approaches49, many topological descriptors are analyzed by a PLS procedure, resulting in an 18-parameter equation that would be difficult to interpret for lead optimization. By contrast, a simple two-variable equation (r 0.887) has been reported, derived from 55 compounds50: logBB 0.0148 PSA 0.152 ClogP 0.139 (3)

aLogP eff

Compound
Desipramine Enalaprilat Fluvastatin Furosemide Naproxen Propanolol

3.35 5.40 3.17 3.08 3.89 4.52 3.37

Effective permeability.

CNS and CNS compounds and concluded that, for brain penetration, MW should be 450 and PSA 90 2. More recently, the Vertex drug/non-drug discrimination approach has been applied to sets of CNS active/inactive compounds to bias the design of combinatorial libraries towards CNS targets [Ajay (1999) Biologically active molecules that cross the BBB: Designing combinatorial libraries with CNS activity. Gordon Research Conference on Quantitative StructureActivity Relationships, 2530 July, Tilton, NH, USA]. This kind of approach can take advantage of the large quantity of data available concerning CNS (in)activity. However, care must be taken to ensure that in the CNS class, activity is caused by the parent compound and not by a metabolite. The CNS class is further complicated because CNS inactivity might not be solely caused by poor BBB penetration, as the compound might undergo extensive first-pass metabolism41 or it might penetrate the brain but not exhibit any receptor binding33. In addition, as discussed in the section on absorption prediction, this method only supplies a binary answer, which is of limited use in lead optimization.

Predicting logBB The second computational approach has been based mainly on the prediction of a brain penetration measure termed logBB, which is the ratio of the steady-state concentrations of the drug molecule between the brain and the blood [i.e. log (Cbrain/Cblood)]. Published values of logBB range from approximately 2.00 to 1.00. Within this range, compounds with logBB 0.3 cross the BBB readily, while those with logBB 1.0 are only poorly distributed to the brain42. Several groups have employed QSPR approaches to develop models for the prediction of logBB. An early attempt at predicting logBB used a small data set of 20 compounds to derive an equation relating logBB to PSA and molecular volume43. As with intestinal absorption, a reduced PSA favoured BBB penetration. Subsequently,

Here, PSA is the single-conformer PSA (Ref. 31) and ClogP is the calculated octanol-water partition coefficient obtained from the Daylight ClogP software mentioned earlier. There are several advantages to this method for logBB prediction. Calculation is very rapid and is fully automated, requiring no human intervention. Furthermore, the quantities that comprise this equation are readily interpretable, making it simple for a medicinal chemist

54

DDT Vol. 5, No. 2 February 2000

research focus

REVIEWS
logPS gives an indication of the permeability of the BBB to a compound52. LogPS is measured using a short-duration vascular perfusion method from which a permeability-surface area product is calculated (hence logPS). LogPS values for a set of 18 compounds have been correlated with Abrahams solute descriptors52. Further logPS data for nine hydroxypyridinones (Fig. 4) have recently been reported53. As more of these data emerge, together with data from in vivo techniques such as microdialysis, it will be interesting to see how they, like logBB, can be used to derive useful, predictive computational models.

O O OH N CH3 N O OH N OH

O OH N

O OH N

-1.89
O OH N

-1.48
O OH N

-0.64
O

-0.38

-0.36
O

OH N N

OH
Drug Discovery Today

-1.03

<-3.00

OH

OH

<-3.00

<-3.00

OH

Applications It was recently predicted that, in less than five years, algorithms for the accurate prediction of passive permeation characteristo alter the molecular structure to change the PSA and/or tics will be developed and available for routine use by ClogP values, thereby changing the predicted logBB of a medicinal chemists [Borchardt, R.T. (1999) Profiling com(hypothetical) compound. Equation (3) has demonstrated pounds for biopharmaceutical properties: Overview. good predictive ability on several test sets50,51. Abstracts of the 217 th American Chemical Society Meeting, 2125 March, Anaheim, CA, USA]. The previous sections Other measures of brain penetration have shown that good progress has been made towards Whereas logBB is a measure of the steady-state distributhis goal in recent years and it seems that our undertion of a compound between the brain and the blood, standing of the molecular determinants of absorption is growing, leading to improved predictive models. Of course, such models are not simply of academic interest 90 they can play a vital role in focusing 80 and accelerating drug discovery. 70 However, it is fair to say that, at this 60 time, most publications in this field 50 40 have been methodological and there 30 are few publications presenting suc20 cesses (or failures) of such methods 10 in an industrial setting. This section 0 will discuss some brief indications of None Poor Moderate High how these types of techniques are Experimental Caco-2 absorption class Drug Discovery Today being applied in drug discovery.
Figure 4. Hydroxypyridinones with their logPS values. Compounds (%) Figure 5. Caco-2 absorption ratings for the two combinatorial libraries, LIB1 (open bars) and LIB2 (shaded bars), showing the improvement in absorption resulting from designing a library subject to polar surface area and rule-of-five constraints. The definitions of the classes are as follows: poor ( 2% absorbed), moderate (220% absorbed) and high ( 20% absorbed). The percentage absorbed figure is calculated by simply taking the peak area for the acceptor chamber and dividing it by the value for the donor chamber. The surface area of the inserts is 0.33 cm2 (Costar HTS 24-well plate). The value shown is the mean taken from two experiments, each run over 3 h.

Compound and screening set selection One important application of these techniques is in the context of compound and screening set selection. For example, techniques such as those already described can be used

DDT Vol. 5, No. 2 February 2000

55

REVIEWS
to filter a set of compounds from an external supplier prior to purchase. Workers at GlaxoWellcome9 have applied the genetic algorithm-based method described earlier10 to the selection of compounds from the corporate database for generation of a screening set. Profiling using the ruleof-five and PSA criteria can also provide a valuable indicator of the likely absorption characteristics of a combinatorial library or screening set.

research focus

Combinatorial library design In addition to simply profiling libraries, this approach has been taken one stage further and has been applied to the Figure 6. Summary results from molecular property calculations submitted via a Web-based interface in use at RPR. design of combinatorial libraries51. In one example, a chemist had selected reagents for a combinatorial library (LIB1) in an oral drug discovery program to optimize of structural databases. Abstracts of the 217 th American parameters such as MW and ClogP in an approximate manChemical Society Meeting, 2125 March, Anaheim, CA, USA] ner. A follow-up library (LIB2) was designed to optimize and it is likely that this type of system will be commonly PSA and rule-of-five criteria much more rigorously with used in the industry in the future. reagent selection being performed by a Monte Carlo search procedure51. Both libraries were subsequently tested in a Conclusions Caco-2 monolayer absorption system and the designed The rule-of-five was originally derived because of the reallibrary showed much improved absorption (Fig. 5). These ization that HTS was identifying large numbers of hit comresults show the added value of considering quantities such pounds, many of which did not possess drug-like properas PSA in compound (library) design in addition to more ties. It is clear from the growing number of publications in traditional computed descriptors such as ClogP and MW. this area that this message has been taken on board, and methods for predicting drug-likeness are already having a Intranet-based systems major impact on the design and selection of compounds The speed and ease with which many molecular properties in several pharmaceutical companies. The routine use of can be calculated means that such computations can be experimental absorption systems in the pre-screening of readily incorporated into an Intranet system for easy access compounds is providing a wealth of data and should by medicinal chemists and other researchers. In the system allow for the derivation of improved models for drug developed at Rhne-Poulenc Rorer (Dagenham, Essex, UK), absorption. The true worth of such experimental and the user can input a structure using a Java-based sketcher54, theoretical systems will become apparent when a reduca SMILES (simplified molecular input and line entry system) tion is seen in lead optimization times and attrition rates in code55 or, for known compounds, a corporate registry numpreclinical and clinical phases of drug development. ber. In a few seconds, the results are returned including values for simple molecular properties, predictions of absorpNote added in proof tion based upon the rule-of-five and PSA, predicted BBB Readers are directed to the following references that are of penetration, as well as alerts for the presence of potentially relevance to the topic of predicting drug-likeness and that toxic functionality (Fig. 6). Batch calculations using sets of have appeared while this review was in press: SMILES codes are also possible via the Intranet. These computations are now used routinely by bench scientists. Other Kelder, J. et al. (1999) Polar molecular surface area as a workers have reported similar approaches [ODonnell, T.J. dominating determinant for oral absorption and brain and Doman, T. (1999) ChemMart: One-stop Web shopping penetration of drugs. Pharm. Res. 16, 15141519

56

DDT Vol. 5, No. 2 February 2000

research focus
Stenberg, P. et al. (1999) Prediction of intestinal absorption of endothelin receptor antagonists using three theoretical methods of increasing complexity. Pharm. Res. 16, 15201526 Wang, J. and Ramnarayan, K. (1999) Toward designing

REVIEWS
drug-like libraries: A novel computational approach for prediction of drug feasibility of compounds. J. Comb. Chem. 1, 524533 Ajay et al. (1999) Designing libraries with CNS activity. J. Med. Chem. 42, 49424951

REFERENCES
01 02 Smith, D.A. and van de Waterbeemd, H. (1999) Pharmacokinetics and metabolism in early drug discovery. Curr. Opin. Chem. Biol. 3, 373378 Delie, F. and Rubas, W.A. (1997) A human colonic cell line sharing similarities with enterocytes as a model to examine oral absorption: Advantages and limitations of the Caco-2 model. Crit. Rev. Ther. Drug Carrier Syst. 14, 221286 Irvine, J. et al. (1999) MDCK (MadinDarby canine kidney) cells: A tool for membrane permeability screening. J. Pharm. Sci. 88, 2833 Eddershaw, P.J. and Dickins, M. (1999) Advances in in vitro drug metabolism screening. Pharm. Sci. Technol. Today 2, 1319 Walters, W.P. et al. (1999) Recognizing molecules with drug-like properties. Curr. Opin. Chem. Biol. 3, 384387 Lewis, R.A. et al. (1997) Similarity measures for rational set selection and analysis of combinatorial libraries: The diverse property-derived (DPD) approach. J. Chem. Inf. Comput. Sci. 37, 599614 Rishton, G.M. (1997) Reactive compounds and in vitro false positives in HTS. Drug Discovery Today 2, 382384 Walters, W.P. et al. (1998) Virtual screening An overview. Drug Discovery Today 3, 160178 Hann, M. et al. (1999) Strategic pooling of compounds for highthroughput screening. J. Chem. Inf. Comput. Sci. 39, 897902 Gillet, V.J. et al. (1998) Identification of biological activity profiles using substructural analysis and genetic algorithms. J. Chem. Inf. Comput. Sci. 38, 165179 Ajay et al. (1998) Can we learn to distinguish between drug-like and non-drug-like molecules? J. Med. Chem. 41, 33143324 Sadowski, J. and Kubinyi, H. (1998) A scoring scheme for discriminating between drugs and nondrugs. J. Med. Chem. 41, 33253329 Polinsky, A. (1999) Combichem and cheminformatics. Curr. Opin. Drug Discovery Dev. 2, 197203 Navia, M.A. and Chaturvedi, P.R. (1996) Design principles for orally bioavailable drugs. Drug Discovery Today 1, 179189 Chan, O.H. and Stewart, B.H. (1996) Physicochemical and drug-delivery considerations for oral drug bioavailability. Drug Discovery Today 1, 461473 Krmer, S.D. (1999) Absorption prediction from physicochemical parameters. Pharm. Sci. Technol. Today 2, 373380; 2, 426 (corrigendum) Lipinski, C.A. et al. (1997) Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings. Adv. Drug Deliv. Rev. 23, 325 Moriguchi, I. et al. (1992) Simple method of calculating octanol/water partition coefficient. Chem. Pharm. Bull. 40, 127130 Ghose, A.K. et al. (1999) A knowledge-based approach in designing combinatorial or medicinal chemistry libraries for drug discovery. 1. A qualitative and quantitative characterization of known drug databases. J. Comb. Chem. 1, 5568 van de Waterbeemd, H. et al. (1996) Estimation of Caco-2 cell permeability using calculated molecular descriptors. Quant. Struct.-Act. Relatsh. 15, 480490 Norinder, U. et al. (1997) Theoretical calculation and prediction of Caco-2 cell permeability using MolSurf parametrization and PLS statistics. Pharm. Res. 14, 17861791 22 Conradi, R.A. et al. (1992) The influence of peptide structure on transport across Caco-2 cells. II. Peptide bond modification which results in improved permeability. Pharm. Res. 9, 435439 Palm, K. et al. (1996) Correlation of drug absorption with molecular surface properties. J. Pharm. Sci. 85, 3239 Krarup, L.H. et al. (1998) Predicting drug absorption from molecular surface properties based on molecular dynamics simulations. Pharm. Res. 15, 972978 Palm, K. et al. (1998) Evaluation of dynamic polar molecular surface area as predictor of drug absorption: Comparison with other computational and experimental predictors. J. Med. Chem. 41, 53825392 Stenberg, P. et al. (1999) Prediction of membrane permeability to peptides from calculated dynamic molecular surface properties. Pharm. Res. 16, 205212 Palm, K. et al. (1997) Polar molecular surface properties predict the intestinal absorption of drugs in humans. Pharm. Res. 14, 568571 Norinder, U. et al. (1999) Theoretical calculation and prediction of intestinal absorption of drugs in humans using MolSurf parametrization and PLS statistics. Eur. J. Pharm. Sci. 8, 4956 Ghuloum, A.M. et al. (1999) Molecular hashkeys: A novel method for molecular characterization and its application for predicting important pharmaceutical properties of molecules. J. Med. Chem. 42, 17391748 Wessel, M.D. et al. (1998) Prediction of human intestinal absorption of drug compounds from molecular structure. J. Chem. Inf. Comput. Sci. 38, 726735 Clark, D.E. (1999) Rapid calculation of polar molecular surface area and its application to the prediction of transport phenomena. 1. Prediction of intestinal absorption. J. Pharm. Sci. 88, 807814 Winiwarter, S. et al. (1998) Correlation of human jejunal permeability (in vivo) of drugs with experimentally and theoretically derived parameters. A multivariate data analysis approach. J. Med. Chem. 41, 49394949 van de Waterbeemd, H. et al. (1998) Estimation of bloodbrain barrier crossing of drugs using molecular size and shape, and H-bonding descriptors. J. Drug Targeting 6, 151165 Goodwin, J.T. et al. (1999) Strategies toward predicting peptide cellular permeability from computed molecular descriptors. J. Pept. Res. 53, 355369 Raevsky, O.A. and Schaper, K-J. (1998) Quantitative estimation of hydrogen bond contribution to permeability and absorption processes of some chemicals and drugs. Eur. J. Med. Chem. 33, 799807 Pardridge, W.M. (1998) CNS drug design based on principles of bloodbrain barrier transport. J. Neurochem. 70, 17811792 Cecchelli, R. et al. (1999) In vitro model for evaluating drug transport across the bloodbrain barrier. Adv. Drug Deliv. Rev. 36, 165178 de Lange, E.C.M. et al. (1999) Microdialysis for pharmacokinetic analysis of drug transport to the brain. Adv. Drug Deliv. Rev. 36, 211227 George, A. (1999) The design and molecular modeling of CNS drugs. Curr. Opin. Drug Discovery Dev. 2, 286292 Basak, S.C. et al. (1996) Predicting bloodbrain transport of drugs: A computational approach. Pharm. Res. 13, 775778 Fischer, H. et al. (1998) Bloodbrain barrier permeation: Molecular parameters governing passive diffusion. J. Membr. Biol. 165, 201211

23 24

03 04 05 06

25

26

07 08 09 10

27 28

29

30

11 12 13 14 15

31

32

33

16

34

17

35

18 19

36 37 38 39 40 41

20

21

DDT Vol. 5, No. 2 February 2000

57

REVIEWS
42 43 44 45 Abraham, M.H. et al. (1997) On the partition of ampholytes: Application to bloodbrain distribution. J. Pharm. Sci. 86, 310315 van de Waterbeemd, H. and Kansy, M. (1992) Hydrogen-bonding capacity and brain penetration. Chimia 46, 299303 Calder, J.A. and Ganellin, C.R. (1994) Predicting the brain-penetrating capability of histaminergic compounds. Drug Des. Discov. 11, 259268 Abraham, M.H. et al. (1994) Hydrogen bonding. 33. Factors that influence the distribution of solutes between blood and brain. J. Pharm. Sci. 83, 12571268 Platts, J.A. et al. (1999) Estimation of molecular free energy relation descriptors using a group contribution approach. J. Chem. Inf. Comput. Sci. 39, 835845 Lombardo, F. et al. (1996) Computation of brainblood partitioning of organic solutes via free energy calculations. J. Med. Chem. 39, 47504755 Norinder, U. et al. (1998) Theoretical calculation and prediction of brainblood partitioning of organic solutes using MolSurf parametrization and PLS statistics. J. Pharm. Sci. 87, 952959 Luco, J.M. (1999) Prediction of the brainblood distribution of a large set

research focus
of drugs from structurally derived descriptors using partial least-squares (PLS) modeling. J. Chem. Inf. Comput. Sci. 39, 396404 Clark, D.E. (1999) Rapid calculation of polar molecular surface area and its application to the prediction of transport phenomena. 2. Prediction of bloodbrain barrier penetration. J. Pharm. Sci. 88, 815821 Pickett, S.D. et al. Enhancing the hit-to-lead properties of lead optimization libraries. J. Chem. Inf. Comput. Sci. (in press) Gratton, J.A. et al. (1997) Molecular factors influencing drug transfer across the bloodbrain barrier. J. Pharm. Pharmacol. 49, 12111216 Habgood, M.D. et al. (1999) Investigation into the correlation between the structure of hydroxypyridinones and bloodbrain barrier permeability. Biochem. Pharmacol. 57, 13051310 Ertl, P. (1998) World Wide Web-based system for the calculation of substituent parameters and substituent similarity searches. J. Mol. Graph. Model. 16, 1113 Weininger, D. (1988) SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 28, 3136

50

51 52 53

46

47 48

54

55

49

Collaboration
Cambridge Drug Discovery (Cambridge, UK) has engaged in a collaborative agreement with KuDOS Pharmaceuticals (Cambridge, UK) for high-throughput screening (HTS) to identify lead compounds as cancer therapies. As part of the agreement, both companies will work together to identify novel inhibitors of enzymes involved in DNA repair, while KuDOS will gain access to Cambridge Drug Discoverys fully automated screening facilities. Barry Kenny, the Research Director of Cambridge Drug Discovery, said, Integration of assay development into a quality HTS environment is a key activity for us and we are confident of providing exciting leads on this programme.

Contributions to Drug Discovery Today

Drug Discovery Today publishes topical information on all aspects of drug discovery molecular targets, lead identification, lead optimization and associated technologies together with overviews of the current status of compound classes, approaches in specific therapeutic areas or disease states and novel strategies, such as gene therapy. Areas of pharmaceutical development that relate to the potential and viability of drug candidates are also included, as are those relating to the strategic, organizational and logistic issues underlying pharmaceutical R&D. Authors should aim for topicality rather than comprehensive coverage. Ultimately, articles should improve the readers understanding of the field addressed and should therefore assist in the increasingly important decisionmaking processes for which drug discovery scientists are responsible. Most articles appearing in Drug Discovery Today are commissioned. However, suggestions and proposals for full reviews or shorter items for the Editorial, Monitor or Update sections are welcomed; in the first instance, a tentative title and brief outline of the proposed article should be supplied. Typically, full reviews will extend to 4000 words with up to 60 references. Update and Monitor items (news and views, reports of new technological advances, conferences, experimental methods, and critical assessment of important new literature and other media) do not usually exceed 1000 words, and one or two figures plus up to ten references may be included. The Editorial represents a personal perspective on contemporary issues and controversies affecting R&D and the pharmaceutical industry. If you would like to contribute to Drug Discovery Today in future, please submit your proposal to: Dr Debbie Tranter, Editor, Drug Discovery Today, Elsevier Science London, 84 Theobalds Road, London, UK WC1X 8RR.

58

DDT Vol. 5, No. 2 February 2000