Main

Tau is a microtubule-associated protein expressed throughout the central nervous system (CNS), but predominantly in neuronal axons. Partially phosphorylated tau in the normal adult brain contains sequence motifs that promote association with tubulin, which leads to stabilization of microtubules. However, pathological hyperphosphorylation of these motifs prevents tubulin binding and thereby results in the destabilization of microtubules1 (Fig. 1).

Figure 1: Intracellular neuronal aggregates in tauopathies.
figure 1

a | Microtubules are strong cylindrical polymers composed of α- and β-tubulin that provide structural support to neurons. Tau is the major microtubule-associated protein in neurons and it stabilizes microtubule architecture through tubulin-binding motifs consisting of 18 amino-acid residues arrayed in 3 or 4 imperfect tandem repeats separated by flexible linkers of 13 variable residues. b | Phosphorylation at a number of serine/threonine sites flanking the microtubule-binding repeats by certain affinity-regulating kinases, including GSK3β, CDK5 and ERK2, as well as phosphorylation by other kinases in the repeat region, attenuates tau binding, which results in destabilization of microtubules. c | Abnormally phosphorylated tau protein becomes dissociated from neuronal microtubules and accumulates in paired helical filaments (PHFs). d | Proteolytic processing leads to the formation of tau oligomers and insoluble aggregates called neurofibrillary tangles (NFTs). Tau in these filaments is heavily phosphorylated at several sites and crosslinked by disulphide bonds. The decrease in the association of tau with microtubules and other physiologically relevant proteins, together with the formation of PHFs, are thought to contribute to neuronal dysfunction and eventually cell death. Kinase inhibitors have been shown to prevent tau phosphorylation, so their pharmacological application might result in a reduction of NFTs.

The longest adult tau isoforms contain as many as 79 serine and threonine residues where phosphorylation can potentially occur, although only about 30 of these seem to function as actual phosphorylation sites under normal physiological conditions2. Tau hyperphosphorylation, aggregation and formation of the filamentous form of the tau protein (paired helical filaments or PHF-tau) are all observed in Alzheimer's disease (AD), and a total of 25 sites where abnormal phosphorylation takes place have been identified3. The misfolding of hyperphosphorylated tau leads to the formation of insoluble neurofibrillary tangles (NFTs)4. It has long been believed that NFTs were responsible for the cytotoxic effects of tau, but recent findings suggest that this might not be the case: using transgenic mice expressing a repressible human tau variant it was observed that although neuron numbers stabilized and memory function recovered upon tau suppression, NFTs continued to accumulate5. This raises the possibility that PHF-tau, and perhaps other soluble lower-mass hyperphosphorylated tau aggregates, might induce neurodegeneration owing to loss of the normal function of these tau forms rather than accumulation of toxic NFTs6,7.

Disorders related to tau — collectively referred to as neurodegenerative tauopathies8 — are part of a group of protein misfolding disorders known as foldopathies2. Neurodegenerative diseases that are associated with tauopathy include AD, amyotrophic lateral sclerosis, Pick's disease and progressive supranuclear palsy9,10,11. The neuropathological hallmarks of AD were first reported in 1907 by Alois Alzheimer12. Apart from severe brain atrophy and neuronal loss in the brains of affected individuals, he found dense extracellular deposits and intracellular neuronal aggregates, which were subsequently identified as amyloid (neuritic) plaques and NFTs, respectively4. Amyloid plaques are extracellular accumulations of β-amyloid (Aβ) peptides that are derived from the proteolytic processing of the β-amyloid precursor protein (APP), whereas NFTs are intraneuronal accumulations of insoluble and hyperphosphorylated tau. Tau is intimately associated with β-amyloid-induced neurotoxicity13, as hippocampal neurons cultured from tau knockout mice do not degenerate when challenged with fibrillar β-amyloid14,15.

Although these invariant pathological hallmarks — plaques, tangles and neuronal death — all contribute to the aetiology of AD, the mechanistic relationships between these lesions remain to be unravelled. In this article, we first discuss the evidence indicating that inhibition of tau hyperphosphorylation is a viable therapeutic strategy for AD and other tauopathies, and then summarize progress in the development of inhibitors of the three most relevant kinases — GSK3β (glycogen synthase kinase 3β, also known as tau protein kinase I), CDK5 (cyclin-dependent kinase 5, also known as tau protein kinase II) and ERK2 (extracellular signal-regulated kinase 2) — that could be used to achieve this goal. Although many lead series with activity against these kinases are known, the structural determinants for selectivity remain poorly understood. Specificity in terms of kinase targets and CNS bioavailability will be important design principles for safe and effective new systemic agents for treating tauopathies. We also highlight results from recent medicinal chemistry and structural biology studies that offer important clues about how mono-, oligo- and pan-specific inhibitors for these kinases might be designed.

Tau hyperphosphorylation as a drug target

Tau is an in vitro substrate for many protein kinases, but the number of those that phosphorylate tau in vivo is probably much smaller and few are actually implicated in abnormal tau hyperphosphorylation16. The physiologically relevant serine/threonine protein kinases fall into two classes: the proline-directed and the non-proline-directed kinases. Of the sites that are phosphorylated in tau, 13 are followed by proline residues and hence are candidate phosphorylation sites for proline-directed kinases, among which GSK3β, CDK5 and ERK1/2 have received the most attention17,18. Relevant non-proline-directed kinases include microtubule affinity-regulating kinase (MARK) and cAMP-dependent protein kinase (PKA). Tyrosine phosphorylation of tau by Src kinases such as FYN has also been suggested to be relevant to neurodegeneration19,20.

Recent data have implicated both GSK3β and CDK5 in aberrant tau phosphorylation and association with microtubules21,22. ERK2 has also been found to be deregulated in post mortem AD brains, and there is evidence that ERK2 is required for neurofibrillary degeneration23. Other stress kinases, including stress-activated C-Jun N-terminal kinase (SAPK/JNK) and p38 mitogen-activated protein kinase (p38 MAPK), have also recently been linked with tau hyperphosphorylation18. Furthermore, a range of kinases, including several from the MAPK family, casein kinase and calcium/calmodulin-dependent kinase II (CaMKII), may also be relevant to tau-related neurodegeneration17. We confine our discussion to CDK5, GSK3β and ERK2, the three kinases that are most strongly linked with tau hyperphosphorylation.

CDK5. All known CDKs require association with a regulatory subunit — usually a cyclin — for activation24. In the case of CDK5, these activators have been identified as the p35 and p39 proteins, which are expressed almost exclusively in postmitotic neurons25. CDK5/p35 knockout mouse studies have shown that CDK5 is essential not only for survival in general, but specifically for neuronal development26,27. Elevated cellular levels of calcium trigger the cleavage of p35 and p39 to the more stable p25 and p29 fragments by the calcium-dependent protease calpain28,29. These truncated forms, which seem to be downstream effectors of Aβ peptides, upregulate CDK5 activity and can therefore be regarded as neurotoxic activators of CDK5 (Refs 3032). Indeed, deregulation of CDK5 is intimately linked with the pathological characteristics of AD33,34,35.

Calpain activation, p25 accumulation, CDK5 co-localization with NFTs and elevation of CDK5 activity have all been observed directly in the AD brain36,37,38,39. Furthermore, the association of CDK5–p25 with pre-tangle neurons in AD brains suggests that CDK5 might be involved in the early stages of NFT formation during AD development40. Recent in vivo studies in rodent models demonstrate that progressive neurodegeneration and NFT formation are triggered by the inducible expression of p25, reinforcing the evidence that the CDK5–p25 pathway is a crucial component of AD pathophysiology41,42,43. It was demonstrated recently that transgenic mice co-expressing both mutant (P301L) human tau and p25 show an accumulation of aggregated and hyperphosphorylated tau, which is associated with GSK3, as well as increased NFTs44. Interestingly, overexpression of p35 in mice that also overexpress tau and CDK5 does not increase tau phosphorylation45, whereas overexpression of p25 does33. Presumably this is due to the fact that p35, the CDK5 activator that predominates under normal conditions, does not co-localize with tau, whereas p25, the main pathological activator, does16. Tau hyperphosphorylation is also closely linked with neuronal cell apoptosis, as the microtubule binding capacity of tau isolated from apoptotic neuronal cells is markedly decreased46,47. Furthermore, the association with tau and the activity of CDK5 are significantly elevated in apoptotic neuronal cells48. The similarity between the tau hyperphosphorylation patterns observed in vitro in apoptotic neuronal cells and those in the AD brain indicate a connection between neuronal apoptosis and AD.

GSK3β. This multifunctional kinase is involved in the regulation of a wide range of cellular processes13,49,50, including metabolism, cell proliferation, cardiac hypertrophy, oncogenesis and apoptosis. Although GSK3β is perhaps best known as a potential drug target for metabolic conditions such as type-2 diabetes and insulin resistance owing to its effects on glycogen metabolism51, it is highly expressed in the brain and is linked to a variety of CNS disease states, including ischaemic stroke, bipolar disorders, Huntington's disease and AD52,53.

The disease association with AD was established when GSK3β was isolated from brain extracts and shown to produce PHF-like epitopes on tau. It is now clear that GSK3β contributes to β-amyloid generation and is involved in brain ageing54. There is strong evidence that GSK3β co-localizes preferentially with NFTs, is active in pre-tangle neurons and contributes to the formation of PHF in the AD brain55. GSK3β has been shown to phosphorylate tau at some of the sites that are hyperphosphorylated in PHFs, both in transfected mammalian neuronal cells and in vivo56,57,58. Furthermore, a potent and specific GSK3 inhibitor, AR-A014418 (Table 1a), was shown to inhibit tau phosphorylation at a GSK3-specific site59. Tau from GSK3β-transfected cells, as well as from mice expressing high levels of GSK3β, displays AD-specific epitopes55. CDK5 has been shown to potentiate tau phosphorylation by priming sites for subsequent phosphorylation by GSK3β60, and CDK5 is known to regulate GSK3β-driven neuronal motility61.

Table 1a Physicochemical properties and predicted BBB permeability of CDK5–GSK3β–ERK2 inhibitors

Transgenic mice with elevated GSK3β expression show increased tau phosphorylation and deficits in spatial learning21,62,63. Apart from its roles in tau phosphorylation, GSK3β is also involved in regulating other AD-related mechanisms13. GSK3β activity can be reduced through phosphorylation of Ser9 by a number of kinases, including AKT/PKB (protein kinase B). Several receptor-coupled signalling systems, such as the anti-apoptotic actions of the AKT pathway, seem to use this method of GSK3β downregulation. Furthermore, amplification mechanisms involving Ser9 phosphorylation and dephosphorylation that affect both inhibition64 and activation65 of GSK3β have been reported. It is therefore possible that GSK3β overexpression might actually result in downregulation of its activity, or that pharmacological GSK3β inhibition may persist after cessation of treatment. However, studies using conditional overexpression of GSK3β in forebrain neurons of transgenic mice have shown that these animals recapitulate aspects of AD neuropathology, including tau hyperphosphorylation, and that GSK3β transgene shut down restores normal GSK3β activity66.

ERK2. ERK2 is highly expressed in neurons. This contrasts with the low number of ERK2 transcripts in sciatic nerves, which contain neuronal axons and glia but lack neuronal bodies and, therefore, neuron-specific mRNA67. ERK2 is also the only known kinase that stoichiometrically phosphorylates the multi-repeat domain of the middle-sized neurofilament (NF-M) in vivo, indicating a role for this enzyme in normal cytoskeletal regulation68. This finding was subsequently corroborated in functional in situ studies69.

ERK2 has an important role in regulating tau functions70 and tau phosphorylation by ERK2 decreases the affinity of tau for microtubles about 10-fold, thereby reducing the ability of tau to stabilize microtubules1. Phosphorylated (activated) ERK2 has been implicated in the early stages of neurofilament and tau phosphorylation in neurons and glial cells in Niemann–Pick type C (NPC), AD and in other tauopathies71. NPC, a disorder of cholesterol metabolism, is characterized by the accumulation of NFTs and progressive neurodegeneration. Using a transgenic mouse model of NPC, it has been demonstrated that increased tau phosphorylation at AD-relevant epitopes coincides with increases in both ERK1 and ERK2 activation72. In this study, no changes were observed in the levels of CDK5 and p35, or in the phosphorylation status of GSK3β. These data suggest that ERK2 activation is relevant to tau phosphorylation under neuropathological conditions.

An immunocytochemical survey of the regulatory state of ERKs in post mortem AD and normal brains showed that activated ERK is a prominent feature in AD, establishing a link between abnormal tau phosphorylation and oxidative stress23. ERK2 appears to be activated in all AD neurons that display tau and neurofilament hyperphosphorylation, suggesting an absolute requirement for ERK2 deregulation in neurofibrillary degeneration. However, unlike CDK5, ERK2 is not associated with apoptosis71.

GSK3β, CDK5 and ERK2 are associated with APP processing. It is established that at least GSK3 is essential for Aβ-induced neurotoxicity73. Exposure of cultured neurons to Aβ activates GSK3, CDK5 and ERK2, leading to tau hyperphosphorylation and, ultimately, to apoptosis74,75,76. Furthermore, the production of Aβ is modulated by APP and an increase in APP phosphorylation at Thr668 by CDK5–p25 has been shown to enhance the production of Aβ77. Further evidence of links between extraneuronal amyloid deposition and tau pathology comes from using a transgenic mouse model of familial AD, in which elevated levels of Aβ and increased exposure of Alzheimer's-type tau phosphoepitopes correlated with deregulation of CDK5, leading to an increase in p25 levels78. Inhibition of CDK5 or calpain activity diminishes Aβ-induced apoptosis of primary cortical neurons31,79, which suggests that increased calcium concentrations attributed to elevated Aβ levels stimulate the cleavage of p35 to p25 in neurons80. However, there is also evidence that CDK5 inhibition might augment Aβ production81. On the other hand, inhibition of GSK3β by lithium has been shown to protect cultured neurons against Aβ-induced neurodegeneration82,83.

Pharmacological target validation. Recent insights into the mechanisms of neurodegeneration suggest that intervention in tau pathologies might be pharmacologically tractable, as genetic suppression of aberrant tau in transgenic models leads to recovery of neurological function6,84. However, until recently, no direct evidence was available to show that suppression of tau hyperphosphorylation with small-molecule agents would be therapeutically useful or even feasible. Several recent reports now provide such evidence. First, intrathecal administration of a calpain inhibitor not only prevented CDK5 activation, but decreased tau hyperphosphorylation and improved neurological function after spinal cord hemisection in rats85. Furthermore, inhibition of GSK3β by lithium ions (which have long been in clinical use for the treatment of bipolar disorder, and which inhibit GSK3β and certain CDK-unrelated kinases by competing for magnesium ions86) has recently been shown to correlate with reduced tauopathy and neurodegeneration in vivo57,87,88. Another study that suggests that pharmacological intervention might not only prevent tau aggregation but actually lead to resolution of preformed aggregates used small-molecule inhibitors that were discovered using an in vitro PHF formation assay89,90. The most direct evidence, however, comes from the report by Le Corre and colleagues7 which shows that small-molecule tau hyperphosphorylation kinase inhibitors (indolocarbazole compounds in Table 1b) can reduce tau hyperphosphorylation and the formation of soluble aggregated tau, and prevent motor deficits on administration to transgenic mice expressing mutant human tau91. The compounds in question are derived from the promiscuous kinase inhibitor staurosporin and are therefore not very selective; the most potent compound SRN-003-556 was reported to inhibit at least ERK2, CDK1, GSK3β, PKA and protein kinase C (PKC) in vitro with Ki values of 0.1–0.6 μM (Ref. 7).

Table 1b Physicochemical properties and predicted BBB permeability of CDK5–GSK3β–ERK2 inhibitors

So, the question remains: which of the tau phosphorylation kinases should be inhibited for optimal therapeutic effects? Although it seems that GSK3β might be the main protagonist in pathological tau hyperphosphorylation, this kinase is potentially a problematic therapeutic target because it is involved in the regulation of so many other physiological processes53. In particular, the role of GSK3β in the Wnt signalling pathway suggests that chronic GSK3β inhibition might be oncogenic92. Many cancers arise as a result of mutation of Wnt pathway components, such as mutations in the β-catenin gene. Normally, proteasomal degradation of β-catenin is promoted by GSK3β phosphorylation, and mutations in β-catenin can produce dominant negative forms that are stabilized and accumulate, leading to the development of cancers in mice.

Similarly, CDK5 and ERK2 belong to kinase families that have multiple roles in many tissues. The main problem with CDK5 inhibitors is the fact that they usually also inhibit some of the 10 other CDKs, which have important roles in regulating the cell cycle and mRNA transcription93. Therefore, the use of these types of inhibitors as CNS drugs might cause undesirable toxicity. Nevertheless, there are indications that selectivity for CDK5 over other CDKs might be achievable. For example, certain compounds described as indolinones have been shown to potently inhibit CDK5, without significantly affecting other CDKs — including CDK2, structurally the closest CDK5 relative — as well as other cell-death-promoting kinases such as JNK and p38 MAPK94,95. Unfortunately, the chemical structures of these indolinones have not been disclosed (presumably these are related to the indirubins discussed below) so it is impossible to draw conclusions that would be helpful in terms of selectivity design.

Partial modulation of more than one of the three prime tau hyperphosphorylation kinases with oligospecific inhibitors, instead of full inhibition of any one of them with mono-specific inhibitors, might result in a better therapeutic margin by maintaining effective suppression of tau hyperphosphorylation while minimizing the potential toxic effects that are associated with the complete suppression of individual non-redundant kinase functions. In the remainder of this article, we discuss the progress so far in the development of such inhibitors with varied selectivity profiles which will allow the rational design of pharmacological tools to probe these questions and provide drug candidates for the treatment of tau-related CNS diseases.

Structure-based kinase inhibitor design

At present, 518 protein kinases that catalyse the transfer of the terminal phosphate group from the cofactor ATP to a macromolecular kinase substrate are known to be encoded in the human genome96. The conservation of structural features within the ATP-binding cleft of kinases initially indicated that specificity for ATP site-directed inhibitors would be difficult to achieve. However, structural elucidation of kinase complexes with ATP and non-hydrolysable ATP analogues has revealed that there are regions at the periphery of, and even within, the binding cleft that ATP does not fully occupy. These regions are structurally diverse between different kinases and provide opportunities for the discovery and design of selective small-molecule ATP-competitive inhibitors97,98.

Structural overview. CDK5, GSK3β and ERK2 are phylogenetically and structurally closely related kinases that share high sequence homology and display similar structural elements that are involved in the regulation of their activity96,99. These three kinases show the typical two-lobed globular kinase fold with a β-strand domain at the N terminus and an α-helical domain at the C terminus100(Fig. 2). ERK2 and GSK3β are both activated by phosphorylation of specific residues that are located in a well-defined region referred to as the activation loop or T loop. Phosphorylation of Thr183 and Tyr185 in the T loop of ERK2 results in a 1,000-fold enhancement of the catalytic activity of this kinase. A corresponding phosphothreonine residue is absent in GSK3β, but Tyr216 in the T loop seems to be an intramolecular autophosphorylation site that is important for GSK3β stability53. The main consequence of T-loop phosphorylation is a structural reconfiguration that allows optimal substrate binding. CDK5 differs from most other kinases, including other CDKs, in this respect: the apoenzyme is in a constitutively active form, and catalytic activity is enhanced following binding of regulatory partners such as p25 and p35 (Ref. 101).

Figure 2: Three-dimensional protein structures of CDK5–p25, GSK3β and ERK2.
figure 2

These kinases, which consist of an N-terminal domain containing mostly β-sheets and a C-terminal α-helical domain, are shown as secondary structure ribbon cartoons (grey). Common structural elements are colour-coded as follows: PSTAIRE/helix αC, yellow; activation loop (T-loop), red; hinge region, green; Gly-rich loop, blue. a | CDK5–p25 complex modelled from PDB (Protein Data Bank) numbers 1UNL (Ref. 102) and 1H4L (Ref. 103). p25 is depicted in cyan. b | GSK3β structure modelled from PDB numbers 1Q4L (Ref. 150), 1Q5K (Ref. 59) and 1I09 (Ref. 152). c | ERK2 structure is based on PDB number 1TVO (Ref. 136). Three-dimensional structure illustrations were prepared using the PyMOL program.

In all three kinases, the ATP-binding pocket is buried deep within the two lobes, the interface of which constitutes the active site cleft. When ATP binds in this cleft the γ-phosphate group projects towards the opening of the pocket, close to the kinase substrate binding region. There are a number of highly conserved amino-acid residues clustered in the active site that are essential for ATP-binding and catalysis (Fig. 3).

Figure 3: The kinase ATP-binding pocket.
figure 3

a | Alignment of the residues that define the ATP-binding cleft in CDK2, CDK5, ERK2 and GSK3β shows that the four kinase sequences differ significantly at the gate-keeper residue and in the specificity surface residues, whereas the sequence is much more conserved at the subsites that interact directly with ATP. b | A surface representation of the ATP-binding region in CDK2 serves as a model for the recognition of ATP by CDK5, GSK3β and CDK5 . Fully activated T-loop-phosphorylated CDK2 (phosphothreonine residue indicated on the surface in yellow), which is cyclin-bound and substrate peptide-complexed, is shown on the left (PDB (Protein Data Bank) number 1QMZ (Ref. 157)) and represents the catalytically competent conformation. The T loop forms a platform for the recognition of the substrate (CPK (Corey–Pauling–Koltun) sticks). In inactive kinase conformations (right) an ATP-bound structure is shown (PDB number 1HCK (Ref. 134)). The opening to the ATP-binding site is wider and the phosphate site is distorted; similar conformations are observed in many inhibitor-complexed structures. Bound ATP (CPK-coloured stick model) interacts with the kinase hinge region (orange) through H-bonds that involve the adenine aminopyrimidine ring. The floor and roof of the ADP-binding site form a hydrophobic clamp (yellow) that accommodates both the adenine purine ring system and the ribose portion of ATP, while the acidic triphosphate group binds in the basic phosphate-binding site (blue). Kinases also require magnesium (purple sphere) as a cofactor for phosphotransfer from ATP to the macromolecular substrate. Two regions of the ATP-binding site that are not involved in direct contacts with ATP but that can be exploited for the design of selective inhibitors are a small hydrophobic pocket delineated by the so-called gate-keeper residue (red) at the base of the ATP-binding site and the opening to the binding cleft, sometimes referred to as the surface specificity site (green). The hinge region residues are also poorly conserved but the side chains of these residues do not protrude into the ATP-binding cleft and the interactions involve the peptide backbone. c | Aligned views of CDK5, ERK2 and GSK3β complexes with the most selective inhibitors are shown (CPK-coloured surfaces for kinases and stick models for inhibitors; PDB numbers 1UNL (Ref. 102), 1UV5 (Ref. 116) and 1WZY (Ref. 111)). The shape and chemical make-up of the ATP-binding pockets in these complexes are significantly different.

Insights into inhibitor design from crystal structures. The deregulation of kinases in many pathological conditions has stimulated an active search for selective inhibitors. The development of such compounds has been aided by detailed knowledge of the ligand interactions with the ATP-binding pocket of the kinases in question gained from X-ray crystallography studies97,98. However, crystal structures provide a static picture of bound ligand–kinase complexes and thus give few clues as to how protein flexibility can be exploited for the purposes of specificity. Comparison of the various published complex structures of CDK5, GSK3b and ERK2 with the inhibitors discussed below shows that such flexibility exists and that even the ATP-binding site is adaptive. Furthermore, some structures may be misleading because the crystallized protein does not fully represent the physiological kinase. A relevant example is CDK5 — when a residue within its ATP-binding cleft was mutated (Asp144Asn) to improve protein expression levels, the resulting recombinant enzyme was inactive102,103.

The design of inhibitors that specifically target inactive kinase conformations, such as the receptor tyrosine kinase (BCR–ABL) inhibitor imatinib, has recently shown much promise104. These so-called type-2 inhibitors — in contrast to type-1 inhibitors that can bind to both active and inactive kinase states — also bind in the ATP site but extend into an allosteric site that arises in some kinases as a result of the DFG (Asp-Phe-Gly) motif of the T-loop folding away from the catalytically competent state. Such inhibitors prevent kinase activation and usually possess potent cellular activity because they recognize and promote kinase conformations that have low affinity for ATP105. Such 'DFG-out' conformations are not usually crystallographically observed in the absence of type-2 inhibitors. For this reason, type-2 leads are difficult to design rationally and none is currently known for the kinases under discussion here. CDK-related kinases are believed to be unable to assume DFG-out inactive conformations, although type-2 inhibitors of CDK-related aurora kinases have recently been reported106.

Kinase inhibitors of CDK5, GSK3β and ERK2

A large number of inhibitors that bind to the ATP pocket of various kinases have been described, including many that are more or less selective for CDKs and GSK3. Three prominent CDK5- and GSK3β-inhibitory pharmacophores are the 3′-substituted indolinones107, especially those known as indirubins108, the 2,6,9-trisubstituted purines109 and the aloisines110. In the following sections we describe some of the salient features that determine their potency and selectivity using the known structure–activity relationships (SARs) of these three compound classes.

All of these compounds were originally discovered as CDK inhibitors and they differ in their overall selectivity profile. Apart from CDKs, indirubins generally also inhibit GSK3β, whereas most of the purines also inhibit ERK2. It has been noted that ERK2 is significantly different to most other kinases, apart from its close structural and functional homologue ERK1, and possesses an unusually restrictive ATP-binding site7. As a result, few ERK2 inhibitors are known and none is truly potent or selective. The aloisines are not as well characterized as a compound class as the indirubins and the purines, but they seem to hold promise as a template for the design of oligo-specific inhibitors. Aloisines contain the pyrrolo[2,3-b]pyrazine structure (in the discussion of these compounds we include the structurally related pyrazolo[1,5-a]pyridines) and are the only compound class that contains nanomolar ERK2 inhibitors111.

Indirubins. Indirubins form part of the class of naturally occurring indigo dyes and their pharmacological properties have long been known in traditional Chinese medicine112,113. Planar oxime indirubin derivatives (R3 = NOH in Box 1) have been identified as potent and selective inhibitors of CDKs and GSK3β that seem to be practically inactive towards ERK2 (Refs 108,114116). For example, indirubin-3′-oxime (R1 = R2 = H, R3 = NOH) inhibits GSK3β-mediated tau phosphorylation in vitro (IC50 ≈ 100 nM) and also effectively suppresses cellular tau phosphorylation at AD-specific sites114.

The main interactions that account for the selectivity and potency of indirubins derive from hydrogen bonding to the kinase hinge region, contacts with the gate-keeper residue and apolar van-der-Waals bonds within the ATP-binding cleft102 (described in Box 1). Overall the hydrophobic van-der-Waals energies of interaction seem to account for much of the binding affinity of indirubins for CDK5 and GSK3β115. It is likely that the smaller binding clefts of CDK5 and GSK3β compared with the larger and more flexible ERK2 cleft result in more compact packing around the ligands, as can be observed from complex X-ray crystal structures and modelled interactions. Similar features are probably responsible for the observed selectivity of indirubins for GSK3β over CDK5. As is the case with most CDK5 and GSK3 inhibitors reported in the literature, selectivity beyond these kinases is poorly documented and thus uncertain. Because of the close structural similarity of CDK5 with both CDK1 and CDK2, CDK5 inhibitors usually inhibit CDK2 as well. This seems to be the case with many indirubins117,118,119,120. Furthermore, there are also reports of indirubin derivatives that possess activity against other CDKs, unrelated kinases and even non-kinase targets120,121,122,123. Another unfavourable property that indirubins share with many kinase inhibitors is their poor aqueous solubility, although more soluble derivatives have recently been reported124.

Purines. Substituted purine inhibitors are based on the heterocyclic structural core present in ATP itself. However, despite being similar to ATP in this respect, purine inhibitors do not generally bind in the same manner as ATP and even comparatively insignificant chemical modifications of inhibitors can result in dramatic changes in potency125,126,127,128,129(Box 2). The first synthetic purines to be studied as kinase inhibitors were N6-isopentenyladenine and the 2,6,9-trisubstituted purine olomoucine130,131. Subsequently, many adenine- and guanine-based purines were developed mainly as CDK inhibitors, including roscovitine and purvalanols A and B132,133,134,135. Many of the interactions of these inhibitors with various kinase active sites have been determined from crystal structures of CDK2-bound ligands.

Because of the large number of known purine-based inhibitors, and owing to the fact that the rigid purine scaffold does not permit substituents to bind in overlapping regions of the active site, the kinase inhibition SARs are comparatively well understood for this pharmacophore (described in Box 2).

Pyrrolopyrazines and pyrazolopyridines. Aloisines contain the 6-phenyl-[5H]-pyrrolo[2,3-b]pyrazine scaffold (Box 3) and are selective inhibitors of CDK1, CDK2, CDK5 and GSK3. The most potent compound in this class is aloisine A, with reported IC50 values of 0.15 μM (CDK1), 0.16 μM (CDK5) and 1.5 μM (GSK3β) in functional kinase assays. Aloisine A is less potent against ERK1 (IC50 = 18 μM) and ERK2 (IC50 = 22 μM)110. Other aloisines are less potent towards ERKs, but the known SARs suggest that this pharmacophore might lend itself to the design of inhibitors with differential selectivity profiles (Box 3).

Pyrazolo[3,4-c]pyridines, which are structurally related to the aloisines, have recently emerged as potential ERK2-selective inhibitors. The most potent compound to date (shown in Box 3) inhibits ERK2 with an IC50 value of 560 nM (Ref. 136).

General inhibitor design principles. An examination of the key features responsible for the potency and selectivity of known inhibitors of CDK5, GSK3β and ERK2 can be used as a guide in the design of the next generation of inhibitors. Pharmacophore maps of the ATP-binding sites can be derived from these findings (Fig. 4). These maps show that there are two main criteria that need to be met in order to yield compounds that have affinity for the ATP-binding pocket: the compounds need to bind effectively to the hinge region, through two or three H-bond donor or acceptor groups, and there must be sufficient hydrophobic interactions within the adenine-binding domain through two or more aromatic ring systems. Potent compounds feature fused heteroaromatic ring systems decorated with one or two additional cyclic systems with flexible substituents probing residues near the ribose-binding domain and within the surface specificity site. CDK5-selective compounds are generally planar and elongated structures; they often rely on breaking the H-bond between Lys89 and Ile10, either though polar contacts or steric clashes forcing Lys89 away from the opening of the ATP-binding site. Inhibitors that are designed to be selective for GSK3β need to be more compact and have less bulky groups near the gate-keeper residue than either CDK5 or ERK2 inhibitors. Furthermore, GSK3β-selective compounds often require an H-bond acceptor feature close to the ribose-binding domain that can facilitate the interaction with more distal residues though water molecules. ERK2 inhibitors require flexibility between ring systems and H-bond donor groups near the phosphate-binding domain for potency.

Figure 4: Pharmacophore maps of CDK5, GSK3β and ERK2.
figure 4

ac | The main residues lining the ATP-binding pocket of the three kinases are shown as stick models and the prominent pharmacophore features that have yielded potent inhibitors of each kinase are indicated with mesh spheres. Regions that are occupied by hydrophobic and aromatic portions of these inhibitors are depicted as green spheres and those where inhibitors commonly display H-bond donor and acceptor functions are in red and blue, respectively. d | An overlay of the surfaces of these pharmacophore elements for each kinase (CDK5, orange; GSK3β, magenta; ERK2, cyan) shows that potency and selectivity determinants differ significantly.

Other mechanisms for inhibition

The search for kinase inhibitors to date has been dominated by ATP-competitive small-molecule leads; however, alternative strategies also seem to be feasible. The first ATP-non-competitive GSK3β inhibitors that have been described are 2,3-disubstituted thiadiazolidinones and structurally related compounds137. Although the exact mode of kinase inhibition remains unclear, it could involve a binding site that, when occupied, interferes with T-loop conformation and substrate binding, but covalent interaction of these thiol-reactive compounds with a unique active site cysteine (Cys199GSK3β) is the more likely explanation for the observed inhibition kinetics and kinase selectivity138,139. Other compounds such as thienyl and phenyl α-halomethyl ketones that have been reported to inhibit GSK3β non-competitively with respect to ATP presumably work in the same way140.

ATP-non-competitive, but tau-competitive, inhibitors of CDK5 have also recently been reported, suggesting that inhibition of substrate binding might be a a feasible design strategy141. In fact, inhibitors that bind in the ATP cleft but also extend into the substrate-binding domain might offer greater kinase selectivity. Computational modelling studies aimed at the identification of the precise substrate-binding domain in CDK5 have revealed that salt-bridge formation from a Lys residue in peptide substrates to Asp86 and Asp92 of CDK5 is important for peptide binding142. These residues lie within the surface specificity site of CDK5, and ATP-site inhibitors that interact with these residues can be designed.

Because of the unique activation mechanism of CDK5, peptide inhibitors derived from the regulatory subunit of p35 are being pursued143,144. Specifically, a truncated form (residues 154 to 279) of p35 was found to bind with high affinity to CDK5, inhibiting, rather than activating, its kinase activity; this peptide is referred to as CDK5-inhibitory peptide (CIP)145. Transfection of cells with CIP specifically inhibits CDK5-mediated tau phosphorylation146. Furthermore, it was recently demonstrated that neuronal infections with CIP suppress aberrant tau phosphorylation and Aβ-induced apoptosis; surprisingly, this is due to inhibition of abnormally activated CDK5–p25 complexes, as normal CDK5–p35 complexes are not affected147. The reasons for this specificity are not clear, but it is possible that the N-terminal domain of p35, which is absent in p25, might obstruct CIP access to the CDK5 active site. Because of the specificity of CIP, not only for CDK5 versus other kinases, but for the neuropathologically relevant form of CDK5, targeting the CDK5–p25 protein–protein interaction might be especially attractive.

Alternative inhibitory approaches might also be possible in the case of ERK2. Specific and effective high-affinity substrate recognition by mitogen-activated protein kinases, including ERK2, involves docking sites that are removed from the phosphorylation site148. Although the docking sites that are relevant to tau as a substrate have not yet been characterized, this might represent an interesting line of investigation.

CNS bioavailability

For application in neurodegenerative diseases, kinase inhibitor drugs will have to be able to reach the CNS from systemic circulation by crossing the blood–brain barrier (BBB). The tightness of the epithelia that form the BBB effectively precludes paracellular transport, and transcellular diffusion is rendered difficult by the presence of special metabolic enzymes and efflux pumps in epithelial cells. For these reasons, the physicochemical properties required of a compound to be BBB-permeant — that is, to be capable of crossing cell membranes rapidly and escaping metabolism — are significantly more restrictive than those needed to ensure bioavailability in other tissues.

Because most kinase inhibitors are ATP-competitive, they have to contend with high (mM) cellular ATP concentrations. Kinase inhibitor lead optimization has traditionally been driven by a need for high affinity to ensure adequate dose potency. Potency is also important in terms of kinase selectivity — compounds with high nanomolar and micromolar kinase Ki values are usually promiscuous. Generally, such optimization has been successful and many inhibitors with nanomolar and even picomolar biochemical kinase inhibitory potency have been developed. However, in many cases this has given rise to highly lipophilic compounds with poor aqueous solubility. To counteract this propensity, solubilizing groups are frequently appended as part of the optimization process, so drug candidate molecules with balanced potency and solubility properties tend to be comparatively large molecules. This issue can usually be addressed to ensure the intestinal permeability required for oral bioavailability. However, BBB permeability tends to correlate with reduced molecular mass (Mr is less than 450), rigidity and reduced polar surface area (few H-bond acceptor and especially H-bond donor functions result in a polar surface area of 70 to 90 Å2).

At present, almost nothing is known (at least in the public domain) about the BBB permeability of CDK5–GSK3β–ERK2 inhibitors. Calculated physicochemical properties for some of the compounds discussed above, as well as representatives from some other relevant pharmacophores, are summarized in Table 1. In most cases, comparatively poor BBB permeability is predicted. In fact the only compound with a predicted positive logBB (log [compound]brain/[compound]plasma) value, the acridinyl-thiazolinone derivative, is unlikely to be drug-like owing to excessive lipophilicity. The fact that the Mr criterion for BBB permeability is not violated by most of the compounds is probably a reflection of the fact that most have not yet been potency- and selectivity-optimized, a process that is almost invariably accompanied by increases in mass. Nevertheless, some pharmacophores, such as the aloisines, might hold promise as CNS drug leads.

The study from Le Corre et al. referred to previously disclosed BBB-permeant inhibitors of tau hyperphosphorylation7. The compounds in question are the indolocarbazole analogues K252a and SRN-003-556 shown in Table 1. With SRN-003-556, a maximal concentration of 1.8 μM (corresponding plasma Cmax of 3 μM; that is, logBB = −0.22) was detected in mouse brain 1 hour after oral dosing at 30 mg per kg. This exceeded bioactive in vitro concentrations, and sufficient exposure to allow pharmacological activity was achieved using oral dosing of 10–20 mg per kg twice a day in the transgenic tau mouse model. Apparently SRN-003-556 has better BBB permeability than K252a. This is presumably due to the primary amino function that is present in SRN-003-556 but not in K252a, as compounds that are basic under physiological conditions (7.5 ≤ pKa ≤ 10.5) are more likely to be BBB permeant149. The fact that SRN-003-556 would not have been predicted to have appreciable BBB permeability (Table 1) illustrates the fact that this property is notoriously difficult to predict. Clearly, imparting appropriate physicochemical properties to lead compounds, while maintaining biochemical potency and selectivity, will pose a significant additional hurdle in the discovery of drug candidates.

Conclusions and perspectives

The search for inhibitors of protein kinases has recently culminated in the registration of several such agents for cancer therapies. The challenge now is to develop kinase inhibitors for other therapeutic indications where an urgent need for effective new treatments exists. Neurodegenerative diseases, which are highly debilitating, constitute one such indication. As discussed here, there is good evidence indicating that targeting the kinases responsible for tau protein hyperphosphorylation, one of the causal events in the development of several neurodegenerative disorders, including AD, should be able to arrest — and maybe even reverse — the degeneration that is due to tau protein aggregation. Because the main tau phosphorylation kinases CDK5, ERK2 and especially GSK3β, regulate many other physiological functions apart from tau phosphorylation, it will be important to design kinase inhibitors that minimize the potential toxicity arising from inhibition of such off-target functions, while maximizing effective suppression of tau hyperphosphorylation. On the basis of the current body of knowledge about relevant protein kinase inhibitors, molecules with varying kinase inhibition selectivity profiles and appropriate bioavailability properties might be designed as pharmacological tools and, hopefully, drug candidates for the treatment of neurodegenerative tauopathies.