The molecular chaperone Hsp90 (90 kDa heat-shock protein) is a remarkably versatile protein involved in the stress response and in normal homoeostatic control mechanisms. It interacts with ‘client proteins’, including protein kinases, transcription factors and others, and either facilitates their stabilization and activation or directs them for proteasomal degradation. By this means, Hsp90 displays a multifaceted ability to influence signal transduction, chromatin remodelling and epigenetic regulation, development and morphological evolution. Hsp90 operates as a dimer in a conformational cycle driven by ATP binding and hydrolysis at the N-terminus. The cycle is also regulated by a group of co-chaperones and accessory proteins. Here we review the biology of the Hsp90 molecular chaperone, emphasizing recent progress in our understanding of structure–function relationships and the identification of new client proteins. In addition we describe the exciting progress that has been made in the development of Hsp90 inhibitors, which are now showing promise in the clinic for cancer treatment. We also identify the gaps in our current understanding and highlight important topics for future research.
Hsp90 (90 kDa heat-shock protein) is one of a group of molecular chaperones responsible for managing protein folding and quality control in the crowded environment of the cell . Although Hsp90 is involved in the triage of misfolded proteins under stress, it also plays a key role under normal conditions in regulating the stability and activation state of a range of ‘client’ proteins, many of which are critical for signal transduction . As an extension of its protein-stabilizing role, work in model organisms shows that Hsp90 acts as a buffer or capacitor of genetic variation in morphological evolution . Furthermore, there is strong evidence that Hsp90 plays an important role in disease states, particularly in cancer, where the chaperoning of mutated and overexpressed oncoproteins is critical . This has driven the development of Hsp90 inhibitors for cancer treatment and, potentially, other diseases . Recent systems-biology studies indicate that Hsp90 is a major interaction node, regulating a very diverse set of cellular functions [6,7].
In this review we consider the biology of the Hsp90 molecular chaperone, focusing on recent developments in understanding structure–function relationships, identification of new client proteins and development of Hsp90 inhibitors for therapeutic application. We also identify the deficiencies in our current knowledge and highlight key areas for future investigation.
Hsp90 AND ATP
Structural, biochemical and genetic studies by several groups (reviewed in ) have firmly established Hsp90 as an ATP-dependent system and allowed the rationalization of two classes of natural products, ansamycins and radicicols, as competitive inhibitors of ATP binding (see below). The role that ATP binding and hydrolysis play in the Hsp90-dependent activation of client proteins is still unclear, but considerable progress has been made towards understanding the biochemical mechanism of Hsp90's ATPase activity and of how this is coupled to conformational changes in the chaperone.
Key to this was the observation that ATP-binding engendered a conformation of the chaperone in which the two N-domains (N-terminal ATP-binding domains) in an Hsp90 dimer become closely associated in the ATP-bound state, but not in the apo or ADP-bound state [9,10]. Together with the constitutive dimerization provided by the C-terminal domain, this generates a conformational cycle in which binding and hydrolysis of ATP are directly coupled to a conformational cycle of the chaperone involving transient N-domain dimerization in an ATP-driven ‘molecular clamp’ mechanism . Prevention of this conformational change by ATP-competitive inhibitors blocks activation of Hsp90 client proteins in vivo and highlights the central role of this ATP-driven conformational cycle in Hsp90's biological function.
CO-CHAPERONE REGULATION OF Hsp90
Important insights into Hsp90's ATP-dependent conformational changes have been provided by studies of the interaction of Hsp90 with a range of accessory proteins, or co-chaperones. Hop [Hsp organizing protein; Sti1 (stress-inducible protein 1) is its yeast homologue] is a co-chaperone that binds simultaneously to Hsp90 and Hsp70 (70 kDa heat-shock protein), coupling the two systems [12,13]. Hop/Sti1 is also a potent inhibitor of Hsp90's ATPase activity , binding to the TPR (tetratricopeptide repeat)-binding site at the C-terminus of Hsp90, but stabilizing a conformation that is resistant to ATP-driven N-terminal dimerization. Although the structural basis for recruitment of Hop/Sti1 to the C-terminal MEEVD motif has been elucidated , the mechanism of Hsp90 ‘arrest’ by Hop/Sti1 is not understood.
The ability of Hop/Sti1 to arrest Hsp90's ATPase cycle is shared with a second co-chaperone, Cdc37 (cell-division cycle 37 homologue)/p50 , which recruits protein kinases to the Hsp90 machinery (reviewed in ). The mechanism of ATPase inhibition by Cdc37 was revealed by a crystal structure of the C-terminal part of Cdc37 bound to the N-domain of Hsp90  (Figure 1A) The central domain of Cdc37 binds to the ‘lid’ segment of Hsp90, which closes over the mouth of the ATP-binding pocket when ATP is bound and facilitates N-domain dimerization . By binding to the open conformation of the lid, Cdc37 prevents N-terminal dimerization and coupled ATP hydrolysis.
Hsp90 regulation by co-chaperones
The inherently low ATPase rate of Hsp90 can be substantially increased by Aha1 (activator of Hsp90 ATPase) [19,20], a stress-regulated co-chaperone found in all eukaryotes. The biochemical basis for this activation was revealed by a co-crystal structure of the middle domain of Hsp90 with the N-terminal domain of Aha1 . The structure showed how binding of Aha1 destabilizes an inactive conformation of a loop in the Hsp90 middle domain that carries a catalytically essential arginine residue (Arg380 in yeast), freeing it to interact with bound ATP in the active site of the chaperone  (Figure 1B).
THE ATP-BOUND CONFORMATION OF Hsp90
The first view of the ATP-bound conformation of Hsp90 came from the co-crystal structure of yeast Hsp90 with a non-hydrolysable ATP analogue, p[NH]ppA (adenosine 5′-[β,γ-imido]triphosphate; ‘AMP-PNP’), and with a co-chaperone p23 (mammals)/Sba1 (yeast homologue) . The structure was obtained using an Hsp90 mutant previously shown to activate the ATPase by favouring N-domain association . The A107N mutation within the ‘lid’ segment in the N-domain, does not alter Hsp90 function, and yeast depending on A107N Hsp90 in vivo are viable (P.W. Piper, personal communication). The charged linker connecting the N-domain and middle segment was also substantially removed. This biologically dispensable linker consists of poorly conserved and low-complexity repeats of charged amino acids. Such a substantial highly charged and unstructured segment is inimical to obtaining strongly diffracting crystals, and all crystal structures of eukaryotic Hsp90s, whole or in pieces, have omitted this segment.
Within the Hsp90–p23 complex, the individual Hsp90 protomers have a twisted parallel arrangement, with the constitutively dimerized C-terminal domains at one end and the transiently associated N-domains at the other (Figures 2A and 2B). The C-terminal domain is formed by a three-stranded β-sheet adjacent to a coil of α-helices, which interact with their equivalents in the other protomer to form the core of the dimerization interface. A helical segment projecting from the inner face of the C-domain also interacts with the equivalent in the other protomer, providing a second, though less extensive, dimer interaction. Although the full C-terminus, including the extreme C-terminal MEEVD TPR-domain-binding sequence, is present in the crystallized protein, it is completely disordered.
ATP-bound conformation of Hsp90
As in the isolated structure , the middle segment of Hsp90 consists of three distinct subdomains, a large α–β–α domain connected to a small α–β–α domain via a helical coil segment. These domains have very similar orientations to each other in the full Hsp90 structure, as in the isolated middle segment, but with some flexing at the junctions between them. The small domain connects to the C-terminal domain by a long curved α-helix and packs closely against it. An amphipathic loop, containing residues implicated in client-protein activation in vivo, projects from the inner face of the larger domain in one protomer, across the gap between the two middle segments.
The most significant differences between the isolated domains and the full-length ATP-bound structure occurs in the ATP-binding N-domain. As indicated by previous biochemical studies, the two N-domains in the dimer come into close association, involving a domain-swap of the most N-terminal strand, which detaches from the edge of the main β-sheet in the unassociated N-domain to occupy the equivalent position in the other N-domain in the dimer. This strand-swap brings the first α-helix of the N-domain close to its equivalent in the other protomer, forming a substantial hydrophobic interface. However, these rearrangements can only occur if the lid segment hinges out of the way via a rotation of ∼120°, to lie over the mouth of the ATP-binding pocket. This ‘closed’ conformation of the lid is stabilized by polar interactions between the peptide backbone of a glycine-rich sequence at the C-terminal hinge-point of the lid and the γ-phosphate of ATP, which provides the primary conformational sensor of the phosphorylation state of the bound nucleotide (Figure 2C).
Closure of the lid and movement of the N-terminal strand and α-helix generates a complementary surface for docking of the middle segment. This is accompanied by remodelling of the middle-segment catalytic loop, which interacts with the closed lid segment and inserts Arg380 into the mouth of the ATP-binding pocket. The head group of Arg380 interacts with the γ-phosphate of the ATP and hence promotes hydrolysis by polarizing the β–γ phosphodiester bond and neutralizing the transition state (close-up of ATP interactions; Figure 2D).
ATP hydrolysis by Hsp90 depends on a co-ordinated set of highly interdependent conformational switches that provide access points for regulation by co-chaperones. Thus Cdc37 prevents ATP turnover by binding to the lid in its open conformation and preventing N-domain rearrangement and docking of the middle segment . Aha1 stimulates the ATPase by facilitating remodelling of the catalytic loop in the middle segment towards its active conformation . p23/Sba1, which binds >50-fold more tightly to the ATP-bound state , sits in a depression at the junction of the dimerized N-domains and manipulates both of these control points , binding to the exposed surface of the closed lid and the same residues from the catalytic loop that interact with Aha1 (Figure 2E). Thus p23/Sba1 activates the ATPase in a similar way to, but mutually exclusive with, Aha1, but by stabilizing the closed conformation of the lid, p23/Sba1 slows release of products and full ATP turnover, and consequently behaves as an ATPase inhibitor [24–26].
The involvement of so many interdependent conformational switches in stabilizing the ATP-bound conformation explains the perennial failure to observe a clear structural effect of ATP on the conformation of the isolated N-domain or of the N-domain in larger constructs [27,28]. Compared with the large rearrangements seen in the ATP-bound Hsp90–p23 complex, the small twitches and disordering of the lid observed in other structures are of little biological significance.
OPENING AND CLOSING THE MOLECULAR CLAMP
The ‘closed’ constrained structure of the ATP-bound Hsp90–p23 complex  defines a clear polar state within the ATPase-coupled conformational cycle. It is not clear whether there is also a defined structural state for the opposite phase of the cycle that results from hydrolysis of ATP, an essential step in client activation [29–31]. X-ray-scattering studies  suggest that ATP-free Hsp90 has no defined conformation in solution, but exists in a continuum of ‘open’ conformations in which the N-terminal domains in the dimer are not constrained with respect to each other. Nonetheless, several crystal structures have been determined for multidomain constructs and full-length Hsp90 homologues devoid of ATP, which show distinct conformations [23,27,33,34] (Figure 3). That these significantly different structures were observed shows that they are at least attainable within the conformational flexibility available to ATP-free Hsp90. However, whether they represent biologically relevant states of Hsp90 is difficult to determine. Significantly, in all cases, and unlike the ATP-bound Hsp90–p23/Sba1 structure, there are substantial lattice contacts involving the highly mobile N-terminal domain, so that the observed conformations are intimately coupled to, and stabilized by, their particular crystal lattice. For example, co-crystals of GRP94 (94 kDa glucose-regulated protein, a member of the Hsp90 family) with ADP or with p[NH]ppA are essentially identical , despite clear evidence that GRP94 has a similar ATPase mechanism to cytosolic Hsp90s [34,35]. The existence of fixed conformations of Hsp90 other than the ATP-bound state may depend on interactions with co-chaperones. Biophysical studies indicate that Hsp90 complexes with Hop/Sti1 or with Cdc37 have defined conformations distinct from the flexible structure of free Hsp90 [14,32,36], but, apart from the core interaction of Cdc37 with the N-domain of Hsp90 , these conformations are structurally uncharacterized.
Opening and shutting Hsp90 molecules
DEFINING Hsp90'S CLIENTELE
With the structural and biochemical mechanism underlying Hsp90's ATPase cycle now characterized, it remains almost a complete mystery how it is coupled to the activation and maturation of its client proteins. Results from all of the crystallographic studies carried out so far suggest that the client protein remains, like an elephant in the room, unavoidable but largely ignored. Hsp90 will only be fully understood when its interactions with its clients – and the changes it provokes in them – have been described in atomic detail. Understanding of this central function of Hsp90 is complicated by the broad, yet select, nature of the client proteins themselves. Given the vast structural differences between clients, it is hard to conceive that they have a common mode of interaction with Hsp90 or that they all benefit from that interaction in the same way.
The list of Hsp90 clients is continually expanding, although the literature evidence base implicating them varies considerably in quantity and quality. A catalogue of client proteins is maintained by the Picard laboratory (http://www.picard.ch/downloads/Hsp90interactors.pdf). Recently, a number of genomic and proteomic studies have attempted to better quantify Hsp90 clients and interacting proteins. The first used genome-wide yeast twohybrid and microarray-based chemical genetic screens to identify 198 physical interactions and 451 genetic interactions involving Hsp90 . Notably, two novel Hsp90 co-chaperones were identified, Tah1p (TPR-containing protein associated with Hsp90) and Pih1p (protein interacting with Hsp90), which connect to the chromatin remodelling factor Rvb1p (RuvB-like protein 1)/Rvb2p and provide a clear link from Hsp90 to mechanisms of epigenetic regulation. A similar two-hybrid screen using an ATP-hydrolysisdefective mutant  to stabilize Hsp90 client-protein interactions identified two new clients: the stress-activated mitogen-activated protein kinases Hog1p (high-osmolarity glycerol 1p) and Slt2p, and also the co-chaperone Tah1p  (also identified by the authors of ). The most recent analysis, using a genome-wide chemical-genetic screen in yeast, showed unexpected roles for Hsp90 in the secretory pathway .
Two studies have described changes in expression patterns resulting from exposure of human tumour cells to Hsp90 inhibitors. The first used cDNA microarrays  to survey changes in mRNA profiles in four human colon-cancer cell lines on treatment with 17-AAG [17-(allylamino)-17-demethoxygeldanamycin; tanespimycin; see below]. Hsp90β and Hsc70 (heat-shock cognate 70 kDa protein) mRNA levels increased, consistent with induction of the heat-shock response mediated by HSF1 (heat-shock factor 1). However, with the exception of casein kinase Iγ, mRNAs levels of Hsp90 client proteins were not affected. However, levels of keratin 8, keratin 18 and caveolin-1 mRNAs were altered, consistent with down-regulation of signal transduction through depletion of client proteins. The most recent study combined mRNA and proteomic profiling in a human ovarian-cancer cell line . A heat-shock response profile was again detected after exposure to 17-AAG, including induction of the Hsp90 ATPase-activating protein Aha1, also seen in a previous study . mRNA levels of a group of genes known to be regulated by c-Myc were decreased. Of particular interest were changes in the levels of chromatin-associated proteins, including heterochromatin protein 1, histone acetyltransferase 1 and the PRMT (protein-arginine methyltransferase) PRMT5. The latter was shown to bind to Hsp90 and is a new candidate client protein. These effects further underline the involvement of Hsp90 in epigenetic regulation of gene expression .
The known Hsp90 client proteins can be grouped into three main classes. The largest and most biologically coherent are the protein kinases. The second major class consists of transcription factors, in particular the nuclear receptors for steroid hormones. The third class consists of structurally unrelated clients, but with emerging subsets of viral-replication proteins and a range of intracellular receptors involved in innate immunity. A comprehensive description of Hsp90 clients is beyond the scope of the present review, and the literature on the involvement of Hsp90 with steroid-hormone receptors has been extensively reviewed in depth [41,42]. Consequently, we have restricted our discussion to recent progress towards understanding the mechanism of protein kinase activation by the Hsp90–Cdc37 complex  and to emerging classes of client proteins.
PROTEIN KINASE CLIENTS
Protein kinases represent the largest single group of Hsp90 client proteins with representatives from all branches of the ‘kinome’, including the key oncogenic kinases c-Src, b-Raf (v-raf murine sarcoma viral oncogene homologue B1), PKB (protein kinase B)/Akt1, ErbB2 (v-erb-b2 erythroblastic leukaemia viral oncogene homologue 2) and Cdk4 (cyclin-dependent kinase 4) (reviewed in ). However, the biochemical mechanism of Hsp90-dependent kinase activation is far less well characterized than for steroid receptors, and there is a great need for in vitro model systems. A minimal in vitro model system for the activation of a truncated form of Chk1 (checkpoint kinase 1) has recently been described. In vitro activation required not only Hsp90 and Cdc37, but also Hsp70 and Hsp40, and the protein kinase CK2 (casein kinase II) , which is required for the functionally essential phosphorylation of Cdc37 on Ser13 [44,45]. This system should provide a useful test-bed for mechanistic studies, although, unlike many Hsp90-dependent kinases, such as PKB/Akt1, Raf-1 (v-raf-1 murine leukaemia viral oncogene homologue 1) or b-Raf, Chk1 does not require activation segment phosphorylation to attain catalytic activity.
In the absence of an atomic resolution structure for the complex of a protein kinase client with Hsp90, some insight has been provided by a single-particle electron-microscopy structure determined for Hsp90 bound to Cdc37 and Cdk4 . This structure suggests that one lobe of the kinase interacts with the N-domain of Hsp90, whereas the other lobe associates with the middle domain of Hsp90 (Figure 4). Bivalent interaction of the Cdk4 with Hsp90 provides a means for coupling ATP-dependent changes in the positioning of the N-lobes and middle segments of the chaperone (see above) to changes in the orientation of the N- and C-lobes of the bound kinase. As the orientation of these lobes and the segments that link them is well known to play a significant role in regulating kinase activity, association with Hsp90 might be required to switch the activation segment conformation. Although this idea is speculative, it is consistent with observations that the sequences close to the junction between the lobes play an important role in the selectivity of Hsp90 for its kinase clients and for their Hsp90-dependent activation [47,48]. A bioinformatics analysis of 105 protein kinases suggested that it is the overall surface properties in the vicinity of the N-lobe and hinge region of the kinase that determines selectivity to Hsp90 rather than any simple linear sequence . That study also pointed out that much of this selectivity of Hsp90 is focussed on kinases that integrate multiple regulatory signals and act as informational hubs in signalling networks.
Architecture of an Hsp90–Cdc37–kinase complex
NLR [Nod (NUCLEAR OLIGOMERIZATION DOMAIN)-LIKE RECEPTOR] CLIENTS AND INNATE IMMUNITY
One of the most recently identified members of Hsp90's clientele are the R proteins (resistance proteins), which are key components of plant innate immunity to pathogens. The largest class of R proteins contain a conserved Toll-IL-1 receptor domain or a coiled-coil domain, followed by a nucleotide-binding domain and a LRR (leucine-rich repeat) at the C-terminus . Plant R proteins are structurally and functionally related to the animal NLRs, such as Nod1 and Nod2, which respond to bacterial peptidoglycan fragments and are critical in intracellular immune responses .
The similarity in structure between mammalian NLRs and plant R proteins extends to a common dependence on the Hsp90 chaperone system [52,53]. In plants, Hsp90-dependent activation of R proteins is mediated by Rar1 (required for Mla12 resistance), a CHORD (cysteine- and histidine-rich domain) protein  and Sgt1 [suppressor of G2 allele of Skp1 (S-phase kinase-associated protein 1)], a TPR-domain protein that additionally contains a CS (CHORD/Sgt1) domain related to the Hsp90 p23/Sba1 co-chaperone  (see below). Mammals possess two CHORD proteins, Chp-1 (CHORD-containing protein 1), and the muscle-specific protein melusin, which is involved in the cardiomyocyte hypertrophic response . Unlike Rar1, Chp-1 contains a C-terminal CS domain similar to that of Sgt1, which is necessary, but not sufficient, for interaction with the N-terminal domain of Hsp90 . Furthermore the CS domains of CHORD proteins and Sgt1 bind in a nucleotide-independent manner to Hsp90 [58,59]. In plants the CHORD I domain of Rar1 binds the N-domain of Hsp90 , while the CHORD II domain interacts with the CS domain of Sgt1. Although Chp-1 and RAR proteins appear similar, Chp-1 has not been detected in the active Nod1 complex  and its interaction with Sgt1 has not been demonstrated. How the Sgt1–Rar1 couple mediates Hsp90-dependent R-gene activation is still unknown.
Sgt1 is a highly conserved protein found in plants and mammals. Unlike CHORD proteins, it is also found in yeast . Sgt1 contains an N-terminal TPR repeat domain, an Hsp90-binding CS-domain and a C-terminal SGS (Sgt1-specific) domain. NMR-based interaction mapping and mutational analysis of the plant Sgt1 show that the N-domain of Hsp90 and the CHORD II domain of Rar1 interact with opposite faces of the Sgt1 CS domain and can form a ternary Hsp90–Sgt1–Rar1 complex .
Sgt1 also links Hsp90 to the SCF (Skp1–Cullin 1–F-box) E3 ubiquitin ligase complexes involved in proteasome-mediated targeted destruction of many regulatory proteins [61,63]. The TPR domain of Sgt1 mediates interaction with Skp1, whereas the CS domain binds Hsp90. Thus Sgt1 acts as an adaptor, coupling Hsp90 and SCF complexes , and suggesting a role in regulation of proteasomal degradation of Hsp90 client proteins (see below), although direct evidence for this has yet to be presented. The Sgt1–Skp1 complex is also required for assembly of CBF3 (centromere-binding factor 3, the core kinetochore complex), in both yeast and animal cells via activation of Ctf13 (a subunit of the CBF3 complex). Hsp90 facilitates binding of Sgt1 to Skp1, and in turn the Hsp90–Sgt1 complex stimulates the binding of Skp1 to Ctf13, the F-box core kinetochore protein . In budding yeast, Sgt1 is also involved in interactions with the adenylate cyclase Cdc35p . Interaction with Cdc35p is mediated by the C-terminal SGS region of Sgt1, which is also required for interaction with plant R proteins  and the mammalian Nod1 receptor . In all three cases the SGS region interacts with an LRR domain in the other protein. LRR domains are also common features of the F-box proteins that bind to Skp1 in SCF complexes , but whether the SGS domain of Sgt1 interacts with these has not yet been determined (Figure 5).
CFTR (CYSTIC FIBROSIS TRANSMEMBRANE CONDUCTANCE REGULATOR) IS AN Hsp90 CLIENT
Cystic fibrosis is an inherited disease caused by failure of the CFTR to fold correctly and be delivered to the plasma membrane. Despite the secretory ‘fate’ of CFTR, it is the cytoplasmic Hsp90, rather than its endoplasmic-reticulum homologue GRP94, that is associated with the most common mutant form, namely ΔF508. This association is disrupted on treatment with Hsp90 inhibitors, which promote the proteasomal degradation of CFTR. However, it was subsequently observed that post-translational disruption of an ΔF508 CFTR–Hsp90 complex restores its proteasomal degradation to levels seen for wild-type CFTR . Consistent with this, siRNA (small interfering RNA) silencing of the Hsp90 activator Aha1 rescues delivery of ΔF508 CFTR to the cell surface, suggesting that persistent association of mutant CFTR with the Hsp90 system is responsible for the pathological physiology of cystic fibrosis .
INVOLVEMENT IN VIRAL REPLICATION
Hsp90 was first reported to be involved in virus assembly more than a decade ago, but until relatively recently this was a little explored aspect of Hsp90 biology. Hsp90 is implicated in the replication of many viruses, including HBV (hepatitis B virus) , HCV (hepatitis C virus) , vaccinia , herpes simplex type 1 [73,74], cytomegalovirus , flock house virus , T-cell leukaemia virus type I , influenza A  and a range of negative-strand viruses . Some of these are of considerable importance as they are not only responsible for recurring infectious diseases, but also for emerging infections such as Lassa, Marburg, Hanta and Rift Valley fever. Other viruses, such as HBV, represent major global public-health problems, with over 300 million chronically infected patients worldwide . The involvement of molecular chaperones in viral replication has interesting therapeutic possibilities, as chaperone inhibition offers a means for decreasing the ability of viral genomes to evolve and escape more direct inhibition of their replicative systems .
The mechanistic basis for Hsp90 dependency in viral replication has not been extensively investigated, but appears to centre on the replicative reverse transcriptase (DNA polymerase) or RNA polymerase of the virus [70,74,79,82–84]. In influenza virus, Hsp90 interacts with two of the three RNA polymerase subunits, participates in their nuclear transport and facilitates their assembly into the active heterotrimeric polymerase complex .
Involvement of Hsp90 in assembly of viral replication complexes has echoes of its role in assembly of the eukaryotic telomerase . The role of Hsp90 in facilitating telomerase function is unclear and contentious. Initial studies suggested that Hsp90 and p23 were required for recruitment of the template RNA component hTR (human telomerase RNA) to the reverse-transcriptase component hTERT (human telomerase reverse transcriptase) , whereas recent results suggest that Hsp90 facilitates loading of the assembled telomerase on to telomeric DNA [88,89].
TARGETED DESTRUCTION OF Hsp90 CLIENT PROTEINS
Even before their mode of action had been determined, it was noted that natural-product inhibitors of Hsp90, such as geldanamycin (see below), promoted the ubiquitination and proteasomal degradation of Hsp90 client proteins [90,91], leading to the suggestion that Hsp90 mediates the decision between the activation of a client and its targeted destruction . What remains to be understood is the mechanism by which an Hsp90 client is shifted from the chaperone nexus into the ubiquitin-directed proteasome pathway.
One of the best-known candidates for mediating the chaperone–proteasome ‘shuffle’ is CHIP (C-terminal of Hsp70 interacting protein) [93–95]. CHIP is an active E3 ubiquitin ligase able to interact with E2 ubiquitin-conjugating enzymes via its C-terminal U-box domain [93,96,97] and to bind to the C-terminal EEVD motif of Hsp70 or Hsp90 chaperones via its N-terminal TPR domain. Structural studies of CHIP bound to the C-terminal peptide of Hsp90 and to an E2 ubiquitin-conjugating enzyme revealed the means by which the dimeric CHIP protein can bind exclusively to either Hsp70 or Hsp90, and how the unusual asymmetric dimer structure of CHIP only allows binding of a single E2 enzyme  (Figure 6).
Chaperone-directed E3-ubiquitin ligase CHIP
CHIP-mediated ubiquitylation of a range of proteins has been demonstrated in cell-culture systems, for example Smad (similar to mothers against decapentaplegic) proteins, Tau, ErbB2, Ron receptor tyrosine kinase, DAPK (death-associated protein kinase), Dbl oncoprotein and guanylate cyclase [99–105]. However, these and most other cellular studies utilize overexpression of CHIP, so that the biological relevance of these observations is uncertain. Apparently supportive observations of in vitro ubiquitination by CHIP must also be treated with caution, since the CHIP–UbcH5 E3–E2 couple usually employed is extremely active and non-specifically ubiquitinates essentially any protein with which it is incubated, including controls such as glutathione transferase and BSA (M. Zhang and L. H. Pearl, unpublished work). It is also notable that, with exceptions such as ErbB2, Ron and DAPK, there is only a partial overlap between the proteins identified as targets of CHIP in published studies and the known clientele of Hsp90. Although none of this eliminates CHIP as a biologically significant connection between Hsp90 and ubiquitin-mediated degradation, it looks increasingly less certain that CHIP is universally responsible for the destruction of Hsp90 clients following ATP-competitive inhibition, and other candidates need to be investigated.
ISOFORM FUNCTION AND SPECIFICITY
Eukaryotes have multiple Hsp90 isoforms: zebrafish (Danio rerio) have three cytosolic Hsp90s, budding yeast and mammals have two, while the fruitfly Drosophila and the nematode worm Caenorhabditis elegans have only one. In yeast and humans, one isoform is constitutively expressed (yeast Hsc82 and human Hsp90β), while the other is typically expressed under stress conditions (yeast Hsp82 and human Hsp90α) . In organisms with multiple cytoplasmic Hsp90s, the isoforms are not fully redundant , but the details and molecular basis of the separation of functions are not understood. Expressed in yeast, human Hsp90α or Hsp90β individually confer viability, but show quite different ability to activate some client proteins . In addition, expression of human Hsp90β as the sole isoform renders yeast highly sensitive to Hsp90 inhibitors . Another cytoplasmic variant in mammals, Hsp90N, completely lacks the N-terminal ATPase domain and appears to be membrane associated , but little biology has yet been described for this.
Higher eukaryotes (but not budding yeast) possess a distinct endoplasmic-reticulum isoform, GRP94 , and a mitochondrial isoform, Trap1 (tumour-necrosis-factor-receptor-associated protein 1) . In contrast with the cytoplasmic Hsp90s, for which many clients are known, the GRP94 clientele is far less well described. GRP94 has been found to be associated with various components of the immune system and receptor tyrosine kinases such as p185ErbB2 and EGFRvIII (epidermal-growth-factor receptor variant III) [113,114]. Recent knockout-mouse studies have revealed a role for GRP94 in the secretion of insulin-like growth factors . GRP94 has been widely implicated in the binding and presentation of peptides to MHC class I molecules , but the biochemical basis for this remains poorly understood .
The clientele and biological role of Trap1 is even less clear. Initial reports of an interaction with cytoplasmic and plasmamembrane proteins are difficult to reconcile with the overwhelmingly mitochondrial location of Trap1 [112,118]. More recent work has implicated Trap1 in the regulation of reactive oxygen generation in mitochondria , and Trap1 itself seems to be a regulatory target of the mitochondrion-specific protein kinase PINK1 [PTEN (phosphatase and tensin homologue deleted on chromosome ten)-induced putative kinase 1], which protects cells from oxidative stress . A very recent study using Hsp90 inhibitors has indicated an essential role of Trap1 in maintaining survival of tumour cells, but not normal tissues, by antagonizing mitochondrial collapse induced by cyclophilin D .
Hsp90 AS A THERAPEUTIC TARGET IN CANCER
Association with a plethora of signal-transduction and other pathways marks Hsp90 as a target for pharmacological modulation in a number of diseases. However, the major current activity is directed towards the development of inhibitors for treatment of cancer , where Hsp90 is implicated in oncogenesis and malignant progression  and where its increased expression is associated with poor prognosis .
Hsp90 has only recently become accepted as a mainstream drug target, owing to concerns about potential toxicity associated with inhibition of such a ubiquitous chaperone, involved in so many biological functions. However, a key feature of Hsp90 as a cancer drug target is precisely that it is required for the stability and function of so many clients that are bona fide oncoproteins , including kinases such as ErbB2, EGFR, Bcr-Abl tyrosine kinase, Met tyrosine kinase, PKB/Akt, c-Raf and b-Raf, androgen and oestrogen receptors, HIF-1α (hypoxia-inducible factor-1α) and telomerase. Since Hsp90 inhibition leads to the proteasomal degradation of such a large number of oncogenic client proteins, a major benefit is the combinatorial impact on multiple oncogenic pathways and antagonism of all the hallmark traits of cancer, including proliferation, evasion of apoptosis, immortalization, invasion, angiogenesis and metastasis . In addition, this combinatorial action should markedly reduce the opportunities for cancer cells to develop resistance to Hsp90 inhibition.
Therapeutic selectivity for cancer versus healthy cells is based on three main factors . First, cancer cells become ‘addicted’ to the oncogenic processes that drive malignancy . Hence depletion of oncoproteins has a much greater impact on cancer cells than it does on normal cells. Secondly, many oncoproteins are expressed in mutated forms that are much more dependent on Hsp90 for their stability and activity than their normal counterparts, e.g. Bcr-Abl, EGFR and b-Raf [126,127]. Thirdly, cancer cells become dependent on chaperones to manage the cellular stress created by the oncogenic process and the hypoxia, acidosis and nutrient deprivation of the tumour microenvironment . The greater dependence of cancer cells on molecular chaperones may also relate to the observation that, whereas Hsp90 in healthy cells is predominantly uncomplexed, Hsp90 in tumours is involved in large multiprotein complexes that display an enhanced affinity for the natural-product-based inhibitor 17-AAG .
NATURAL-PRODUCT Hsp90 INHIBITORS
Natural-product inhibitors of Hsp90 have played a key role in unravelling the biological function of Hsp90 and in its identification as a molecular target for cancer drugs. Derivatives of one of these, geldanamycin, have been pathfinder molecules in animal models, and early clinical trials, leading to the development of more ‘drug-like’ synthetic small-molecule Hsp90 inhibitors [122,130–132], some of which have recently progressed to the clinic.
The natural products geldanamycin and radicicol (Figure 7A) were initially discovered in cell-based phenotypic screens , and subsequent pioneering studies identified the molecular target of these agents as Hsp90 [134,135]. Of crucial importance were X-ray-crystallographic studies that demonstrated unequivocally that geldanamycin and radicicol are competitive inhibitors, docking into the ATP-binding site in the N-domain of Hsp90  (Figure 7B; geldanamycin and radicicol bound to N-domain). The distinctive properties of the ATP-binding site compared with most other nucleotide-binding proteins  explains the selectivity of the natural products and underpins subsequent inhibitor design. Geldanamycin, radicicol and the newer synthetic inhibitors block the ATPase-coupled chaperone cycle and consign the bound client proteins for ubiquitination and proteasomal degradation [90,91], leading to depletion of oncoproteins and consequent cell-cycle arrest and apoptosis .
While geldanamycin is too toxic for clinical development, a semisynthetic analogue 17-AAG (Figure 7A), with better toxicology and a higher therapeutic index, has displayed promising activity in clinical trials . More soluble derivatives such as 17-DMAG (17-dimethylaminoethylamino-17-demethoxygeldanamycin) (Figure 7A)  and IPI-504 (17-allylamino-17-demethoxygeldanamycin hydroquinone hydrochloride), the hydroquinone form of 17-AAG , have followed 17-AAG into the clinic.
Like geldanamycin, radicicol proved unsuitable for clinical development, owing to its chemical reactivity and instability. Oxime forms showed more promise, but have not progressed. Various analogues have been generated by total synthesis, including cyclopropararadicicol and others [142,143], and detailed analysis of ring and conformational analogues of radicicol has identified synthetically simplified compounds of comparable activity .
Several exotic derivatives of geldanamycin and radicicol have been evaluated, including geldanamycin dimers designed to simultaneously inhibit both N-domain ATP-binding sites in an Hsp90 dimer [145,146] and radicicol–geldanamycin chimaeras . New natural-product scaffolds are being identified, a recent example being the isoflavonoid derrubone from the Indian tree Derris robusta .
SYNTHETIC SMALL-MOLECULE INHIBITORS
The complexity of natural products and their potential for off-target effects prompted a widespread search for synthetic lower-molecular-mass inhibitors of Hsp90. The first of these were based on the purine scaffold of the natural nucleotide ligands and were modelled on the binding of ADP . Co-crystal structures showed the purine ring did bind as in ADP; however, the compounds induced a conformational change in the ATP-binding site, opening a lipophilic pocket in a way that was not predicted . Subsequent studies led to a range of potent and soluble purines with good activity in animal models [151,152]. The optimized purine-based drug BIIB021 (Figure 7A) has recently entered clinical trials.
The next chemical class to receive serious attention was based on the diarylpyrazole Hsp90 inhibitor CCT018159, identified by high-throughput screening against the active yeast Hsp90 [153,154]. Crystallographic studies showed that this compound, which contained the resorcinol unit also present in radicicol, binds deep into the ATP pocket with the resorcinol hydroxy groups and the pyrazole nitrogen atom involved in a network of water-mediated hydrogen-bonding interactions . Structure-based design generated the more potent pyrazole amide CCT0129397/VER-49009  and the even more potent isoxazole CCT0130024/VER-50589, which was active in an animal tumour model . The optimized analogue NVP-AUY922/VER-52296 (Figures 7A and 7B) has now entered clinical trials .
The tractability of the Hsp90 N-domain for crystallography is being successfully exploited to identify a range of new synthetic small-molecule inhibitors, with compounds based on aminopyrimidine and quinazolinone leads identified by structure-based fragment-binding and virtual-screening respectively [158,159].
The bacterial gyrase inhibitor novabiocin, a member of the coumeromycin family of antibiotics, is an interesting Hsp90 antagonist, since it does not bind to the N-terminal ATP site, but has been suggested to interact with a putative second ATP binding site at the N-terminus . However, no such second site is evident in the crystal structures for this region of Hsp90 that are now available [23,33,34] and the biochemical basis for inhibition of Hsp90 by novabiocin and other compounds such as celestrol  is unclear.
NEW APPROACHES TO Hsp90 INHIBITION
Although inhibition of ATPase activity offers the most direct route to therapeutic manipulation of Hsp90, it is inherently non-selective for the Hsp90-dependent clients whose activation is disrupted by that inhibition. A more challenging, but potentially far more specific, approach is to develop agents that block the interaction between Hsp90 and its client or accessory proteins. Although blocking such protein–protein interactions is technically more difficult than inhibiting ATP binding, it is facilitated by the availability of crystal structures from which lead compounds might be identified using fragment-binding and virtual-screening approaches. Attractive targets would include Aha1, which interacts with the middle segment of Hsp90 , and Cdc37, which is required for loading of kinase clients on to Hsp90 and has been linked to transformation ; an inhibitor of the Cdc37–Hsp90 interaction would clearly have therapeutic potential in kinase-addicted cancers.
TPR-domain co-chaperones, such as Hop, which couples Hsp90 and Hsp70, the Hsp90-targeted protein phosphatase PP5 or the E3 ubiquitin ligase CHIP, are also very attractive targets for interaction inhibition, as they interact with Hsp90 (and/or Hsp70) via a short peptide motif at the C-terminus of the chaperone, binding into a cleft in the TPR domain [15,98]. This type of protein–peptide interaction, comparable with substrate binding to proteases, is eminently ‘druggable’ by small-molecule compounds, which are less able to selectively block the flat hydrophobic surfaces involved in many protein–protein interactions.
Inhibition of client-protein interaction with Hsp90 offers the ultimate in selectivity, but little is known about the molecular basis for these interactions, either whether they are specific or generic or even where the interaction occurs. Nonetheless, attempts have been made to specifically block binding of particular clients. The peptidomimetic shepherdin was designed to specifically inhibit the interaction between Hsp90 and the anti-apoptotic client protein survivin . Although shepherdin did exhibit anti-leukaemic activity , its apparent interaction with the ATPase pocket of Hsp90 and effect on a range of Hsp90 clients suggests that it may have a different mode of action.
Several studies have shown that Hsp90 is associated with HDACs (histone deacetylases), particularly HDAC6 , and that broad-range acetyltransferase inhibitors, such as hydroxamic acid derivatives, elicit very similar profiles of client-protein degradation as Hsp90 ATPase inhibitors . The effect of these compounds is to enhance the level of acetylation of Hsp90, particularly on Lys294, which reduces Hsp90 functionality in yeast in vivo . It is not known whether the acetylation of Hsp90 that occurs when deacetylases are inhibited is a normal regulatory modification or what acetyltransferase is responsible for it. Nor is the mechanistic basis for modulation of Hsp90 activity by this acetylation clear, as the equivalent to Lys294 in the crystal structure of yeast Hsp90 is surface-exposed and distant from the catalytic machinery of the chaperone . Nonetheless, that Hsp90's activity can be modulated by post-translational modifications opens up an extra avenue of indirect intervention via inhibition of the enzymes responsible. Observations of inhibition of Hsp90 function by disruption of phosphatase activity  and the well-documented requirement of Cdc37 phosphorylation  open this approach to a wider range of indirect targets.
NEWS FROM THE BEDSIDE – INHIBITING Hsp90 IN PATIENTS
Although it is still early days, reports from the front line of early clinical trials indicate promising initial results for Hsp90 ATPase inhibitors, which are proving to be well tolerated and show clear signs of therapeutic activity. A detailed discussion is outside the scope of this review. Readers are referred to clinical reviews [168,169], very recent results [168–171] and proceedings of recent meetings of the American Association for Cancer Research, the American Society of Clinical Oncology and the American Society of Haematology.
Importantly, Phase I studies with 17-AAG showed clear evidence of Hsp90 inhibition, as revealed by the molecular signature of depletion of client proteins and up-regulation of Hsp70 . Early signs of therapeutic activity were seen in melanoma, potentially via depletion of c-Raf/b-Raf [126,127]; in breast cancer, most likely through the depletion of ErbB2; in prostate cancer, potentially via effects on the androgen receptor and phosphoinositide 3-kinase pathway clients such as Akt; and in multiple myeloma, probably involving the unfolded protein response [168–170]. These results show that the arguments for cancer selectivity, outlined above, may well hold true in cancer patients.
CONCLUSIONS AND PERSPECTIVE
The remarkable properties of the versatile Hsp90 molecular chaperone continue to surprise us. Though its involvment in the stress response is undiminished, the range of its activities under ‘normal’ conditions continues to expand. Our understanding of the extent of its participation in all aspects of the life and death of cells is being extended by high-throughput systems-biology approaches, as well as by studies in model organisms. It is clear that Hsp90 acts a master regulator, controlling critical hubs in homoeostatic signal transduction and regulating chromatin structure and gene expression, development and morphological evolution. Unexpected roles in processes such as the secretory pathway have been revealed.
In the present review we have focused especially on recent exciting developments in our understanding of Hsp90's structure–function relationships, the discovery of new Hsp90 clients and the development of Hsp90 inhibitors for cancer treatment.
New clients that we have highlighted here include Tah1p, Pih1p and PRMT5, further strengthening the link between Hsp90 and chromatin remodelling and epigenetic regulation, and multiple innate immunity and ubiquitin ligase systems linked to Hsp90 via Sgt1. Major advances have been made in the last couple of years in our understanding of the structure–function relationships for Hsp90 and its co-chaperones. The molecular-clamp model underpinning the chaperone cycle, driven by ATP binding and hydrolysis and involving N-terminal dimerization and conformational changes across the chaperone, has now been fully vindicated by structural data. The importance of this clear understanding of the ‘opening and closing’ mechanism of Hsp90 cannot be underestimated.
Also of note has been the emerging structural and biochemical understanding of the roles of co-chaperones and accessory proteins in the complex regulation of the chaperone cycle. The molecular basis for regulation by p23, Aha1 and Cdc37 has been defined, but that for many others remains unknown, and a picture of the detailed roles of many of these co-chaperones in the biology of particular client-protein classes is only just starting to appear. Electron-microscopic studies have provided the first view of the interaction of Hsp90 with a client protein, Cdk4, but the level of detail is still very low. The major challenge ahead is to comprehend the structural basis for the molecular recognition of diverse client proteins at atomic resolution, and how this determines their subsequent activation or degradation. Although focus has been on CHIP, it seems likely that additional E3 ligases will turn out to be involved in the ubiquitination of clients prior to proteasomal breakdown.
Another area in which we are fairly ignorant concerns the biological functions of the different isoforms of Hsp90, and this requires greater attention. It is increasingly clear that the isoforms can play different roles. Of recent interest was the observation that extracellular Hsp90α is involved in invasion through the chaperoning of the matrix metalloprotease II . The role of Trap1 in the regulation of mitochondrial membrane permeability and apoptosis has been described .
One of the most notable aspects of Hsp90 research in recent years has been the mutually beneficial symbiosis between basic and translational studies. The natural products that helped to elucidate the structure and function of Hsp90 have paved the way for drugs that are now beginning to show early promise in cancer patents. It is becoming clear that malignant cells are much more dependent on Hsp90 than are their healthy counterparts. Encouraging clinical results with 17-AAG show that the Hsp90 target can be inhibited without causing unacceptable toxicity. Moreover, there are clear indications of therapeutic activity and the latest results with 17-AAG in breast cancer are especially encouraging .
A key challenge is to gain an understanding of the major mechanisms governing the responses of particular tumour types to Hsp90 inhibitors, and especially to figure out which client proteins are critical. New pharmacodynamic biomarkers for use in proof of mechanism studies continue to be identified on the basis of expression profiling [38,39] and convenient and quantitative serum-based ELISA markers are beginning to emerge . Biomarkers that are predictive of a response are now needed. NQO1 [NAD(P)H dehydrogenase, quinone 1] levels are predictive of sensitivity to 17-AAG, but not other chemical classes of Hsp90 inhibitors .
Many new small-molecule ATPase inhibitors are emerging for clinical evaluation, and the value of isoform-selective compared with pan-inhibitors needs to be explored, potentially learning from our experience with kinase drugs. Although there is optimism for the success of Hsp90 N-terminal ATPase inhibitors, an exciting approach that we have highlighted for the future is to attack interactions between the chaperone and its co-chaperone and accessory proteins. An advantage of the ATPase inhibitors is that they block all chaperone functions, depleting all Hsp90 clients and producing strong combinatorial effects on all of the cancer hallmark traits. However, it is possible that more sophisticated blockade of specific Hsp90 functions can be obtained by targeting its protein interactions. In the present review we have indicated the potential of attacking Hsp90's interactions with its activating protein Aha1, the kinase-selective co-chaperone Cdc37, the phosphatase PP5, the E3 ligase CHIP and, for ultimate selectivity, the client proteins themselves. Meeting the technical challenges will be facilitated by the increasingly sophisticated understanding of the structural biochemistry of Hsp90 ‘super-chaperone’ complexes.
The initial success with Hsp90 inhibitors in cancer opens up the potential for their use in other diseases. In view of our increasing understanding of the activity of Hsp90 in pathologically relevant processes such as immunity and viral infection, as well as its potential role in cystic fibrosis, research into the therapeutic applications is likely to accelerate. Other potential disease areas include inflammation and neurodegenerative disorders. In addition, the early promise of Hsp90 inhibitors is stimulating interest in additional chaperone targets, such as Hsp70 which may limit response to Hsp90 inhibition , as well as other targets in protein quality control.
We predict an exciting few years ahead in basic and translational research on Hsp90.
Hsp90 research at the Institute of Cancer Research Chester Beatty Laboratories is supported by a Wellcome Trust Programme Grant to L. H. P. P. W. is supported by Cancer Research UK Programme grant number C309/A8274. P. W. is a Cancer Research UK Life Fellow. We thank our many colleagues and collaborators for valuable discussions. L. H. P., C. P. and P. W. were involved in a research collaboration with Vernalis to develop Hsp90 inhibitors that have been licensed to Novartis. P. W. is a consultant to Novartis. L. H. P. is a consultant to Biotica.
activator of Hsp90 ATPase
v-raf murine sarcoma viral oncogene homologue B1
centromere-binding factor 3
cell-division cycle 37 homologue
cyclin-dependent kinase 4
cystic fibrosis transmembrane conductance regulator
C-terminal of Hsp70 interacting protein
checkpoint kinase 1
cysteine- and histidine-rich domain
CHORD-containing protein 1
- CS domain
death-associated protein kinase
v-erb-b2 erythroblastic leukaemia viral oncogene homologue 2
94 kDa glucose-regulated protein
hepatitis B virus
heat-shock protein organizing protein
heat-shock 70 kDa cognate protein
heat-shock factor 1
70 kDa heat-shock protein
90 kDa heat-shock protein
N-terminal ATP-binding domains
nuclear oligomerization domain
protein kinase B
- R protein
required for Mla12 resistance
RuvB-like protein 1
suppressor of G2 allele of Skp1 (S-phase kinase-associated protein 1)
stress-inducible protein 1 (the yeast homologue of Hop)
tetratricopeptide-repeat-containing protein associated with Hsp90
tumour-necrosis-factor-receptor-associated protein 1.