In the broadest sense, moderately lipophilic drugs cross the BBB by passive diffusion and the hydrogen bonding properties of drugs can significantly influence their CNS uptake profiles. Polar molecules are generally poor CNS agents unless they undergo active transport across the CNS. Size, ionization properties, and molecular flexibility are other factors observed to influence transport of an organic compound across the BBB.² Additionally, an effective drug possesses an appropriate pharmacokinetic profile. One of the major consequences of inadequate pharmacokinetics of both developmental and marketed drugs is failure in advanced development and/or market withdrawal.³

Before the introduction of high-throughput screening (HTS) and combinatorial screening libraries, most screening for new CNS drugs was performed using biochemical methods, followed quickly by descriptive animal studies that could indicate whether or not a compound penetrated the BBB. A lead compound would then be profiled against receptor panels to look for selectivity. For instance, a synthetic derivative ABT-594, a very potent nonopiate analgesic,⁴ derived from the Central American tree frog compound epibatidine,⁵ was profiled against more than 70 receptors and ion channels before clinical investigation. With the introduction of HTS and subsequent huge combinatorial libraries to feed it, whole animal testing has been even further removed from the initial discovery process. Adding in informatics, genomics, and proteomics to the discovery process makes animal testing even more remote from the initial discovery process. In fact, significant resources can be invested in a lead compound and series before even whole cell assays are conducted.⁶

In a global sense, the general requirements for passive and facilitated transport are common. However, depending on the organ, the epithelial cells can have differing specifics for allowing drugs to transfer from the blood to the cells in the organ. The CNS, being exquisitely sensitive to many compounds in the blood and also to drugs, is designed to be very selective in what is allowed in. As such, it is certainly an outlier as is demonstrated by the lack of correlation between intestinal Caco-2 cell and BBB active efflux.⁹

The reason the BBB is an outlier is that BBB epithelial cells form tight junctions that effectively preclude paracellular diffusion. In addition, the cells possess few pinocytotic vesicles and lack fenestration. Therefore, BBB transfer is through transcellular diffusion through the membranes. A drug undergoing transcellular diffusion can be metabolized by a formidable battery of metabolic enzymes. For example, decarboxylation of 3-(3,4-dihydroxyphenyl)-alanine to dopamine occurs during transit. An orally active CNS drug needs not only sufficient metabolic stability to maintain integrity in the intestine and liver but also across the BBB. Alternatively, a drug can be pumped back into the blood by an active transfer process, mainly through p-glycoprotein, using an ATP efflux mechanism.^10–12

On a molecular level, the BBB is not homogenous but consists of a number of partially overlapping zones contained in a highly anisotropic lipid bilayer.¹³ The conformational mobility of the lipid chains is relatively low at or near the water (blood)/lipid interface and increases strongly toward the interface at the center of the bilayer. The lipid-water interface is associated with a layer of perturbed water molecules with significantly different polarization properties. Because of this, the ability of these water molecules to form hydrogen bonds with drug molecules is dramatically reduced and forms part of the desolvation process. In addition, the hydrophilic/lipophilic interface at the blood/membrane boundary consists of perturbed and bound water, charged polar lipid head moieties connected to long lipid chains. As a result, a drug approaching the BBB is confronted with a thick layer that is capable of noncovalent interactions with the drug, similarly to that of receptor but with much looser steric requirements.

An important ADME characteristic is simply the solubility of the drug as only the amount of drug in solution is available for intestinal absorption and blood distribution.¹⁵ Because diffusion across the BBB is a kinetic process and the blood and brain concentrations are in equilibrium, a concentration of drug in the blood high enough to produce a concentration of drug at its receptor in the brain is critical.¹⁶ A drug molecule’s solubility is a function of many of the physical chemical properties that we will discuss further on in relation to BBB penetration. In addition, not only their bulk properties but also the placement of the polar and nonpolar areas of the molecule and how those regions influence the inter- and intramolecular forces in the crystalline state are relevant to solubility and BBB penetration because water has to break these interactions to bring the molecule into solution. General guidelines for solubility are classified as: moderate at 10–60 μg/ml and high over 60 μg/ml. In the blood, drugs reversibly bind to proteins like albumin and α-acidic glycoprotein.¹⁷ The amount of protein binding of CNS drugs is an important consideration. The protein binding of CNS drugs tends to be rather high. A question that needs to be answered is: as these proteins, carrying the drugs molecules, approach the polar surface of the BBB, does the interaction cause the protein to release the bound drug?

Physiochemical and biological properties are fundamental properties of ADME. Early assessment of the physiochemical properties of potential CNS drugs for their ability to cross the BBB is extremely important. Retrospectively, analysis of experimental data has provided guidelines for physical properties and has been used to derive computational algorithms to predict CNS efficacy. The initial analyses of ADME properties, e.g. anesthetic agents in the late nineteenth century, focused on the partition coefficient (LogP) between water and oil, basically the lipophilicity of the compound. This has served as one of the fundamental principles for drug discovery and design.¹⁸ However, lipophilicity is a two-edged sword. Many other drug parameters are affected by lipophilicity. High lipophilicity frequently leads to compounds with high rapid metabolic turnover¹⁹ and low solubility and poor absorption. As lipophilicity (LogP) increases, there is an increased probability of binding to hydrophobic protein targets other than the desired one, and therefore, there is more potential for toxicity. For example, the active sites for undesirable cytochrome P450 inhibition and human ether-ago-go-related gene (hERG) K-channel blockade are hydrophobic and bind lipophilic substrates.²⁰

Whereas LogP is useful for prediction and optimization, it is clear that other physical-chemical properties are also of potential importance. The next major advance was due to Hansch and colleagues,²¹ who developed regression analyses for these properties and initiated the field of quantitative structure activity relationships (QSAR). Hansch’s initial work was based on a large set of sedative-hypnotic barbituates where he demonstrated biological activity was almost entirely due to their Log P and their rate of metabolism was linearly related to LogP. Furthermore, optimal activity is observed at LogP = 2.²² Other ADME parameters, including cytochrome P450 (CYP450) conversion and drug duration of action, could also be analyzed using QSAR techniques,¹⁸ although it was subsequently acknowledged that, for accuracy, data sets for individual CYPs needed to be analyzed.²³

These early studies underlined the importance of LogP in understanding the activity of CNS drugs and their ability to penetrate the BBB. However, these QSAR equations were designed to both understand the biological activity of drugs as well as guide synthesis of novel analogs. As such, they were limited in scope and the data, and the conclusions derived from were not used to make semiquantitative hypotheses about general rules and molecular requirements for BBB penetration beyond general observations of lipophilicity.²⁹ In an interesting study that proved prescient, Levin determined that rat brain permeability was determined by LogP and had a molecular weight cutoff of 400 or less.³⁰ In another, Young et al.³¹ investigated the effect of dipole moment and LogP in a series of centrally acting H₂-histamine antagonists related to cimetidine. They proposed that, because the molecules were very polar, LogP actually describes a desolvation effect where the molecules in their hydration spheres need to undergo desolvation at the receptor to realize the hydrogen bonding and dipole interactions with the receptor, in general terms, the BBB exterior membrane.

A variation of LogP was introduced by Young et al.,³² who determined that partitioning of a series of H₂-histamine antagonists was best described by a combination of a combination of partition coefficients. The data were explained by the difference between the partition coefficients as measured in octanol/water and cyclohexane water (ΔLogP). Octanol, because of its hydroxyl group, is capable of hydrogen bonding, whereas cyclohexane, being a pure hydrocarbon (lipid) is not. They concluded that the difference between the two measures ΔLogP is a measure of the hydrogen bonding ability of the molecule. However, the generality of this relationship is unclear as a similar study on a series of H₁ antagonists only showed dependence on LogP.³³ Subsequently, Goodwin et al.³⁴ indicated that LogP is predominantly a measure of drug volume, or surface area, plus hydrogen bond acceptor potential. Thus, both hydrogen bonding potential and drug volume contribute to permeability. These general and cited studies, formed the basis for elucidation of the mechanism of BBB penetration by drugs that followed.

Because most CNS drugs have either basic (mostly) or acidic properties, or are amphiphilic, molecular charge should be a component of BBB penetration attributes. Because these drugs are weak bases or acids (compared with strong fully ionized ones like sulfuric acid and sodium hydroxide) as measured by their pKa (log of the acid ionization constant), they are in equilibrium with their uncharged neutral species. In measuring LogP, the species in the water layer may be an equilibrium mixture of charged and neutral species. However, the species that is extracted and exists in the organic layer is neutral.³⁵ There are rare exceptions to this, like glycine.³⁶ So, in effect, what is being measured is a distribution ratio. The unraveling of LogP with charged species as a function of pH is complicated. In effect, the amount of uncharged neutral species in water is critical for membrane penetration. Fischer et al.,³⁷ in a study of comparative properties for BBB penetration, estimated pKa limits for penetration between 4 and 10. Similar conclusions were made by Palm et al.³⁸ but using intestinal absorption in Caco-2 cells.

The data set for much of the original research on BBB penetration is based on a set of more than 100 compounds whose LogBBB was determined by radiolabeled drug determinations in blood and brain of anesthetized rats.⁴⁰ This is a slow, labor intensive, and costly process but has the highest predictability. The structures of these compounds are illustrated in a review. Although these LogP values are for the BBB, it should be mentioned that calculation of general LogP values probably has the highest accuracy among computational chemistry programs, whereas calculation of LogBBB is of more limited reliability.⁴¹

Abraham constructed a larger set of molecules in an effort to reliably calculate LogBBB.⁴³ The QSAR equation derived contains terms for excess molecular refraction, solute polarizability, hydrogen bond acidity, and basicity and molecular volume. Although the agreement between calculation and experiment is good and this procedure has been used extensively by Abraham and his group, it fails Rule 2 above as medicinal chemists will find it difficult in using these types of descriptors in designing compounds.

Because the initial PSA research used the low energy conformation of each molecule and most molecules have many accessible conformations at body temperature, Palm et al.⁴⁴ developed a dynamic PSA approach whereby the set of available conformations were used and the contribution of each to the overall PSA was calculated using a Boltzman distribution thereby taking into account conformational flexibility. Based on their results in intestinal Caco-2 cells, drugs with a PSA of 60 Ų or less are completely absorbed, whereas those with at least 140 Ų are not. This approach is computationally intensive and not suitable for analyzing large libraries.

As it turned out, other investigators demonstrated that a single low-energy conformation is sufficient for determining PSA. Kelder estimated LogBBB using a set of 45 compounds, mostly the same set as others used, and compared PSAs for both single lowest energy conformations as well as a Boltzman average of the 100 lowest energy conformations.⁴⁵ There was essentially no difference between the two approaches, indicating the computationally intensive conformational calculations are not required. They further used these compounds as a training set and compared a set of 776 orally active CNS drugs that had entered phase II clinical trials against a set of 1590 non-CNS drugs that had also entered phase II. Using only a single low-energy conformation of each drug for the PSA calculation gave good agreement and separation based on PSA between the CNS and non-CNS drugs. Kelder found that non-CNS drugs transported passively and transcellularly needed a PSA of 120 Ų or less, whereas the drugs can be targeted to the CNS with a PSA less than 60–70 Ų. Similar conclusions were made by van de Waterbeemd based on a study of marketed CNS and non-CNS drugs.⁴⁶ Their cutoff for PSA cutoff for CNS penetration is 90 Ų or below and a molecular weight cutoff of 450.

Clark⁴⁷ also used PSA as a measure of LogBBB using the same small data set using a single low-energy conformation. Interestingly, using only the calculated PSA did not yield the best results. Adding the calculated LogP (ClogP) to the QSAR equation gave more accurate results, essentially adding a lipophilicity term to the hydrogen bonding and polarity terms inherent in PSA. The advantage to Clark’s QSAR equation is that by using a single conformer is fast and easy to calculate LogBBB and thus predict whether a molecule may pass the BBB. This makes the equation suitable for evaluating large screening libraries.

It is reasonable to speculate on the reasons that use of single low-energy conformation of a drug yields a useful estimate of LogBBB and BBB penetration. First, many, but not all, of the drugs contained in the small set do not have a great deal of structural flexibility. This limits the set of low energy conformations to a relatively compact volume. Part of this is the consequence of the optimization process that a medicinal chemist undertakes during discovery where rigidity is incorporated into the molecule to increase biological activity among other factors. Second, the BBB acts as a molecular filter and not a barrier. Some compounds need to be let in and others kept out. There is a molecular recognition event when a drug approaches the external polar face of the BBB. Given the diverse nature of the CNS drugs, this interaction is more of a physical chemical nature than steric, i.e. it does not have the stricter steric requirements that a receptor-ligand pair would have.

Because of the necessity of screening very large real or virtual libraries, on the order of a million compounds or more, approaches were investigated that would significantly speed the evaluation of each compound while maintaining the accuracy. Ertl has developed a topological PSA (TPSA) approach that fits these criteria.⁴⁸ TPSA is based on dissecting the contributions of polar groups in drugs contained in the WDI. Comparison with Clark’s results demonstrated almost no difference between the two approaches. The major advantage of TPSA is that it uses a two-dimensional structure and not the computationally more intensive three-dimensional representation. This enables two to three orders of magnitude increases in throughput.

Because PSA is dependent upon hydrogen bonding and donating atoms, Österberg⁴⁹ investigated simplifying the PSA calculation by incorporating hydrogen bonding descriptors in place of PSA while including ClogP. There was high correlation between the two methods, indicating the importance of hydrogen bonding and lipophilicity for LogBBB. As with TPSA, this simplification allows BBB penetration evaluation of large actual and virtual libraries. This hydrogen bonding approach was later extended and developed into a pair of rather simple rules for predicting BBB penetration:

• Because hydrogen bonding is primarily associated with oxygen and nitrogen moieties in a molecule, then, if the sum of the nitrogen (N) and oxygen (O) atoms in the molecule is five or less, then the molecule has a high probability of entering the CNS.
• If ClogP –(N + O) > 0, then LogBBB is likely to be positive and the compound has a high probability of entering the CNS.

The sum of the (N + O) atoms actually measures the hydrogen bond acceptors associated with nitrogen and oxygen moieties. It should be pointed out that there are other heteroatoms in drugs that can function as hydrogen bond acceptors (HBA) and total HBA, including (N + O) would probably give a better measure.

A molecular recognition event occurs between a drug and the polar end of the phospholipid membrane at the initiation of BBB penetration. This type of interaction was explicitly incorporated in a membrane interaction QSAR (MI-QSAR). In addition to the descriptors used in studies above, Hopfinger and colleagues⁵⁰ used a molecular dynamics trajectory simulation of the drugs through a synthetic membrane. Figure 1 shows acyclovir, a slow penetrator) and diazepam (a fast penetrator) going through an artificial membrane DMPC that corresponds closely with biological membranes. Acyclovir has just passed the polar end while diazepam is almost through half of the bilayer. The illustration demonstrates the path the drugs need to follow from both the side and top views. It is important to realize that the membrane is fluid and especially the lipid side chains have large degrees of freedom to move and interact and thus retard penetration by the drugs. PSA and ClogP were, as before, the most important descriptors but, added to these, were descriptors for the interaction between the drug and the phospholipid polar head and drug molecular flexibility. The polar interaction descriptor recognizes the molecular recognition event at the start of the penetration, whereas the flexibility descriptor corresponds to the ability of a drug to essentially wiggle through the array of long chain lipids in the membrane. Veber et al.⁵¹ have found, however, that increasing flexibility hinders membrane transit. It is likely, however, that the amount of molecular flexibility will follow a parabolic curve similar to LogP where some amount facilitates transit through the membrane, but too much conformational mobility interferes with the fluid lipid side chains.

Adenot extracted a large CNS library from the WDI containing approximately 1700 compounds segregated according to positive and negative BBB penetration and whether the molecule was a Pgp substrate.⁷² Discrimination analysis yielded an initially surprising result. Most of the non-BBB-penetrating compounds have a number of heteroatoms larger than 8, whereas most of the BBB permeable compounds have a number of heteroatoms lower than 9. The reason that the number of heteroatoms classified most of the compounds is due to its correlation to the polar surface area. In this study, heteroatoms included oxygen and nitrogen as well as phosphorus, sulfur, and halogens, explaining the difference with Österberg’s studies.⁴⁹ This approach, however, compares two sets of drugs looking for differences. As a result, the similarities are factored out.

Neural networks are being used as part of an effort to qualify compounds based on their “drug-likeness” for inclusion in large screening libraries. Sadowski and Kubinyi⁷³ encoded and analyzed over 200,000 molecules from WDI and ACD using topological descriptors. Neural network analysis qualified drug-likeness on non-drug-likeness success percentage of approximately 80%. A similar study by Ajay⁷⁴ used a Bayesian neural network to discriminate between drug-likeness and non-drug-likeness. Subsequently, Ajay and Murcko⁷⁵ published a study on designing libraries with CNS activity. They reported the development of a neural network trained on different sets of about 9000 compounds from CMC and MDDR drug databases separated into CNS-active and CNS-inactive drugs. They used an automated classification scheme based on therapeutic use. After training, classification of test sets of about 13,000 CNS-active and 53,000 CNS-inactive compounds from both databases was performed with approximately 80% accuracy. These neural network-based classifications appear to be practical tools for in silico screening of large compound libraries for identifying potential leads and early ADME profiling but not for high-resolution physical chemical understanding of BBB penetration.

Although the previous comparative analyses are useful for large libraries, Lipinski et al.⁷⁶ at Pfizer were looking for a set of general rules that would govern drug-like properties aimed primarily at solubility and membrane penetration that could be used by medicinal chemists. Lipinski et al. noted that with the advent of HTS and combinatorial chemistry starting in the late 1980s, the trend for lead molecules identified by HTS was toward increasing molecular weight and increasing lipophilicity. Both of these contributed toward decreased solubility and membrane penetration. A database of about 2500 orally active drugs that had entered at least phase II, based on United States Adopted Name (USAN) or International Non-proprietary Name (INN) names, was used to analyze the physical chemical properties of these drugs. From this analysis, the “Rule of Five” was developed. The “Rule of Five” is so named because all the essential physical properties are parameters of five. According to this rule, a good absorption and permeability is likely if:

• Molecular weight is ≤500
• Oil/water distribution coefficient (LogP) is ≤5
• Hydrogen bond donors ≤5 (expressed as the sum of OHs and NHs)
• Hydrogen bond acceptor ≤10 (expressed as the sum of Ns and Os)

A fifth rule was added later:

• Number of rotatable bonds ≤10

This was originally based on rat studies³⁴ but is supported by the observation that drugs discontinued during the clinical development process have more rotatable bonds than those that are successfully marketed.⁷⁷ This effect may very well be a reflection of a conformationally mobile compound trying to permeate through mobile lipid chains in the bilayer.

These rules were explicitly stated as not applying to actively transported substrates. If any two of the rules were exceeded, the compounds were likely to possess poor permeability and solubility properties.⁷⁸ Because, however, these properties are all correlated in individual molecules, exceptional strength in one could counteract a substandard score in another.

Because Lipinski’s data set contained drugs that entered clinical trials, the probability existed that many of these failed on the path to approval. Wenlock⁷⁷ evaluated a set of marketed drugs and those that failed in clinical trials. The results are shown in Table 1. The agreement between Lipinski’s Rules (clinical candidates set) and Wenlock’s marketed drug data set is excellent. Vieth⁷⁹ also studied a set of 1193 marketed drugs with results essentially confirming those of Wenlock and Lipinski.⁸⁰

Lipinski⁸⁰ also laid down a set of rules derived from the set of 1500 drugs that were filtered from USAN or INN names for good CNS penetration. CNS penetration is likely if:

• Molecular weight ≤400
• Log p ≤ 5
• Hydrogen bond donor ≤3
• Hydrogen bond acceptor ≤7

This study indicated that, for CNS penetration, the physical properties in general have a smaller range than general therapeutics. These “Rules” certainly confirmed the observations made independently by earlier workers. Lipinski’s contribution is that the study was based on clinically studied drugs and looked at their overall requirements not against a set of control drugs where common factors were filtered out. Lipinski’s Rule of Five has thus provided medicinal chemists with a simple mnemonic for identifying compounds with medicinally relevant physical chemical properties. As such, medicinal chemists have enthusiastically adopted the Rule of Five.⁸¹

Two studies have contrasted physical chemical properties of CNS-marketed drugs to either non-CNS drugs in general or to other classes of non-CNS drugs. Mahar Doan⁸² conducted both biochemical and computational experiments on 48 CNS and 45 non-CNS drugs. The results (Table 2) did not differentiate between the two classes in molecular weight, volume, hydrogen bond acceptors, and number of aromatic rings. Significant differences were observed with hydrogen bond donors, ClogP, PSA, and molecular flexibility (rotatable bonds). In general, although there are some differences, perhaps due to the smaller data set, the results are in agreement with both laboratory and computational results.

An interesting study was published by Leeson and Davis⁵³ that used the year of market introduction of drugs to determine not only their physical chemical properties but also how they may have changed as a function of time. The time course was grouped into two parts, launched before 1983 (864 drugs) and launched between 1983 and 2002 (329 drugs). The pre-1983 was taken from Vieth’s⁷⁹ compilation and from 1983 from the last chapter in Annual Reports in Medicinal Chemistry. In contrast to all the other studies, Leeson’s was subdivided into six therapeutic areas, i.e. CNS, cardiovascular, anti-infective, etc. This study demonstrated that cardiovascular and CNS drugs were both outliers compared to the other classes (Table 3). For the 74 CNS drugs, the results show, compared with all drugs, that these drugs have statistically different (ρ = 0.001 − 0.01) means and medians in some physical chemical aspects and do not differ in others. They have significantly reduced molecular weight, polar properties %PSA [(O+N), total hydrogen bond acceptors], (OH + NH) hydrogen bond donors), and rotatable bonds (molecular flexibility). Thus, indicating that CNS drugs are smaller with a more compact and less flexible structure, and that the surface has fewer polar groups able to function as hydrogen bond donors and acceptors and that compared to the total surface area, the PSA is reduced.

To function as a drug, a molecule has to meet a set of physical chemical requirements in addition to the steric and energetic requirements at its receptor. Many of these requirements are the same for both CNS and non-CNS drugs. For instance, whereas lipophilicity is important for CNS penetration, the ClogP average values for CNS and cardiovascular drugs are the same. Based on the data in the tables, CNS drugs tend to be more lipophilic, be less polar, have less flexibility, and have lower molecular weights than drugs used for other therapeutic indications. In addition, the molecular volume of CNS drugs appears to be smaller than other drugs.

In terms of design of new drugs or analogs, besides the physical chemical attributes listed above, the fact that these attributes are correlated is important. Thus, changing a molecule to adjust one attribute will change others as well. This was clearly shown by a simple QSAR equation where LogP is clearly defined by molecular weight and hydrogen bond donors and acceptors. In designing new drugs, there is a balancing act required to address the physical chemical requirements and making the best compromises in properties from those available and the actuator requirements at the receptor. At the end, attributes of a successful CNS drug are shown in Table 4 and can be used as a guide to design CNS therapeutic agents with better drug-like properties.