The value of protein allostery in rational anticancer drug design: an update

Abstract

Introduction:

Allosteric drugs are advantageous. However, they still face hurdles, including identification of allosteric sites that will effectively alter the active site. Current strategies largely focus on identifying pockets away from the active sites into which the allosteric ligand will dock, and do not account for exactly how the active site is altered. Favorable allosteric inhibitors dock into sites that are nearby the active sites, and follow nature, mimicking diverse allosteric regulation strategies.

Areas Covered:

The following article underscores the immense significance of allostery in drug design, describes current allosteric strategies, and especially offers a direction going forward. The article concludes with the authors’ expert perspectives on the subject.

Expert opinion:

To select a productive venue in allosteric inhibitor development, we should learn from nature. Currently, useful strategies follow this route. Consider for example the mechanisms exploited in relieving autoinhibition and in harnessing allosteric degraders. Mimicking compensatory, or rescue mutations may also fall into such a thesis, as can molecular glues that capture features of scaffolding proteins. Capturing nature and creatively tailoring its mimicry can continue to innovate allosteric drug discovery.

1. Introduction: The value of protein allostery in rational drug design

Allosteric regulation is a mechanism through which an action in one location of a protein structure affects another, and it does so via propagating a signal communicating between them [1–5]. Allosteric modulators can increase or reduce the protein action [6–8]. While the distance between the sites can be long, short distances, which likely incur fewer transitions, are expected to lead to more productive outcomes. The allosteric event that initiates the signal can be ligand binding or mutation. The outcome can be a conformational change at the active site, promoting, blocking, or modulating substrate binding at the active site and influencing catalytic activity. Exactly how the signal propagates, and the detailed altered active site conformation are of crucial importance. They can help in understanding how cancer-linked allosteric activating mutations work in drug resistance. Crucially, they can also help in targeted drug design which is based on the same principles.

Targeted therapy in cancer uses disease-agnostic drugs. These drugs are based on the cancer’s (or other disease) genetic and molecular features. They disregard the cancer type or the tissue where it was initially discovered. This strategy is rational, and knowledge based. One example is the G12C driver mutation of K-Ras, which has been observed in different cancers. Whereas it is mostly in non-small cell lung cancer (NSCLC), where malignant cells form in the tissues of the lung, it also occurs in colorectal cancer and pancreatic ductal adenocarcinoma [9–11]. Tumor-agnostic treatments harness the same drug to treat all. The drugs can be orthosteric (competitive) or allosteric.

The mechanisms of action of orthosteric and allosteric drugs differ [12]. Orthosteric drugs directly target the active site (Figure 1). The location of the active site is known. Allosteric drugs bind elsewhere, and act allosterically by changing the structure of the protein active site. They do this by shifting the ensemble of the protein conformations to populate disabled active sites, thereby obstructing (or, in cancer repressors like PTEN, promoting) its function. The different mechanisms of the two drug types require different design conceptualizations. With active sites similarity across protein families, an orthosteric drug optimized for one protein may bind to homologous members of the same family. To avoid high dosage and toxicity, selective target-only orthosteric drugs aim to achieve very high binding affinity. Affinity is a cardinal factor for allosteric drugs binding at their corresponding allosteric sites as well. Orthosteric drugs block the active site; allosteric drugs docking into allosteric pockets modify the population of the enzymes’ active state [13]. The conformations with which they interact are largely unchanged. The interaction stabilizes the bound conformation, leading to a redistribution of the ensemble toward this now more stable bound state. Covalent allosteric drugs with irreversible binding, retain this state. Allosteric proteolysis-targeting chimeras (PROTACs) provide one example [14–18]. The role of allostery in the ubiquitin-proteasome system has been documented already over a decade ago [19].

Figure 1.

Overview of drug-protein interaction. Kinase and phosphatase contain an active site for phosphoryl transfer and allosteric sites remote from the active site. For kinases, ATP-competitive orthosteric drugs target the active site, while allosteric drugs bind elsewhere and induce a conformational change of the active site. A bitopic drug contains both orthosteric and allosteric ligands connected by a linker, enhancing drug efficacy. Combinations of orthosteric and allosteric drugs overcome drug resistance mutations and re-sensitize kinases to orthosteric drugs. PROTAC is a heterobifunctional molecule composed of an allosteric ligand for the target protein and a ligand for the E3 ubiquitin ligase, which induces selective intracellular proteolysis.

Allosteric drugs are advantageous [20–22]. Since they bind at sites other than the conserved active sites, they could be less toxic. They do not act by blocking the active sites as an OFF switch, but by shifting the ensemble they permit modulation of the outcome. However, they may still require high residence time at the allosteric site, necessitating high affinity. This requires knowledge of the allosteric sites. Even though there are methods to predict allosteric sites [23–28], including those incorporating advanced physics-based AI and deep learning strategies, as we discuss below, they tend to suffer from inherent limitation. Most account only for the surface of the allosteric pocket, its shape and composition; that is, the amino acids that populate it, side-stepping the mechanistic core of allostery. They do not consider the detailed conformational changes exerted by the drug atoms, ‘pushing’, or ‘pulling’, the receptor protein residues, triggering the allosteric signal, nor its exact potential outcome, thus they are unable to reliably predict the consequences on the active site [29].

This narrative emphasizes the merits of learning from mother nature and mimicking its strategies. The existence of proteins, and other biomacromolecules as ensembles with conformational distributions, is an inherent physical attribute [30]. Their redistributions, or re-equilibration, following changes, such as non-covalent (e.g., by cofactors), or covalent (e.g., by posttranslational modifications) binding is the essence of functional regulation, that is, allostery.

Above, we considered the value of protein allostery in rational drug design from the standpoint of its mechanism and highlighted the vital importance of identifying ‘good’ allosteric pockets. However, also of vital importance is the choice of proteins to target [31–37] with the allosteric drugs. Normally, the targeted protein is the one harboring the mutation. Yet the likely emergence of drug resistance argues for targeting additional proteins. Two allosteric drugs are not used to co-target the same protein. How then to select the second protein to increase the value of the drug design? Below we first briefly elaborate on the allosteric mechanism of action, allosteric sites, and how allosteric drug design can learn from mechanisms adopted in evolution. Co-selection of protein targets to reduce the chances of resistance was recently discussed [37,38]. Finally, we express our expert opinion as to a successful strategy in protein allostery in rational drug design.

2. Discovery of molecules with mechanisms of action mimicking nature

Drug discovery has been following nature. Since drugs aim to change protein actions, innovative designs can start by querying the mechanisms through which proteins are regulated. Allostery fits the bill, as it is an intrinsic property of all dynamic proteins [39–44]. Only dynamic proteins can be regulated, and only dynamic proteins can catalyze reactions [1]. Dynamics (or flexibility) requires populations that interconvert on various timescales. Allosteric effectors can modulate a protein via stabilization, and destabilization [45–49], acting in inhibition, and activation [50–52]. They redistribute the protein ensembles and can alter the rates of their interconversion between the conformational states. The rates depend on the relative energies of the conformations and the barrier heights separating them. The interconversions, and the amino acid networks through which the allosteric pathways connecting the allosteric and the active sites propagate, clarify how allosteric events cause the specific conformational changes and their extent in the active site. The allosteric effect is caused by frustration [3,53–58]. Binding of the allosteric effector frustrates the residues in the allosteric sites. In turn, accommodating the frustrated atoms frustrate their residue neighbors [59]. The preferred propagation pathways are those where the differences in stabilities of the conformations and the kinetic barriers are the lowest. The lower the barriers, and the higher the minima, the faster the signal. An allosteric effect may cause only very minor, side-chain atoms only, structural alterations at the functional site, and be limited to changes in dynamics, increased flexibility, or rigidity in the global or local structure. Such outcomes are likely to be rarely used in anticancer allosteric drug discovery which seeks to block substrate binding.

Below, we discuss allosteric mechanisms of action exploited by nature that drug discovery can learn. These include direct mechanisms that focus on the protein itself, such as mimicking rescue mutations, allosteric post-translational modifications [38], autoinhibition, and mechanisms focusing on covalently linking the protein to molecules involved in its degradation. These can be allosteric degraders, allosteric PROTACs (Figure 1), noncompetitive molecular glues, and bitopic ligands, with covalent links of allosteric ligands and orthosteric warheads [60–66]. They further include indirect mechanisms, such as rescue pathways and making use of feedback loops, which can also act allosterically via protein-protein interactions.

Broadly, the aim of efficient drugs is to reduce the number of proteins existing in their active, catalysis-ready conformations. Achieving this aim will diminish the strength of the activating allosteric signal that propagates from the protein, blocking proliferation.

2.1. Allosteric drugs: first step is identifying likely sites

One way to capture mechanisms championed by nature is a comprehensive analysis of the allosteric landscape of protein structures, especially allosteric sites, allosteric signaling pathways, and key residues along them. This has recently been done for K-Ras [67], generating atlases of its inhibitory allosteric communication. The analysis quantified the impact of over 26,000 mutations on K-Ras folding and binding to its key six interaction partners and inferred more than 22,000 causal free energy changes. Such analysis can identify the more allosterically conducive hotspots which can serve as allosteric binding sites. Since this innovative approach is based on the powerful energy landscapes, its predictions may well be superior to current prediction approaches. While the potential is immense, it remains to be seen exactly how its findings will come into play in allosteric drug discovery.

Successful discovery of allosteric drugs necessitates identification of allosteric sites [13,68–70] including cryptic pockets [71,72]. As examples of computational schemes to directly identify allosteric sites in the protein structure, in AlloDriver [69], missense mutations are first mapped to the structures of human proteins from the PDB. Mutations falling on allosteric or orthosteric sites are analyzed and scored for their classification, and the sites are designated accordingly. AlloDriver also provides the clinical characterization of the predicted allosteric driver mutations. AlloMAPS [70] is based on the causality and energetics of allosteric communication from structure-based statistical mechanical models of allostery. The Allosteric Signaling Maps (ASMs) were obtained by computational scanning of all (stabilizing and destabilizing) mutations within certain ranges. Computational strategies for predicting allosteric sites also include exploiting machine learning approaches, as comprehensively reviewed in Current Opinion in Structural biology [73]. The review describes Supervised Machine Learning (ML) models for predicting allosteric sites, highlighting their accuracy and merits in potential discovery of new allosteric sites. It further discussed the promise of deep learning in large protein language models (pLMs) for allostery and allosteric mechanisms. The Protein Allosteric Sites Server (PASSer) is another recent addition, providing a fast and accurate prediction of protein allosteric sites with three trained and published machine learning models, an ensemble learning model, an automated machine learning model, and a learning-to-rank model [24]. Additional recent strategies have also been published (e.g., Wu et al. [28]).

A conceptual take offered that allosteric drugs have ‘anchors’ and ‘drivers’ atoms [13]. The anchor gets into the pocket. The driver ‘pulls’ and/or ‘pushes’ protein atoms in the pocket. These frustrating actions cause shifts of the protein population to optimize drug binding, triggering propagation of the allosteric signal in the structure. Rescue mutations and autoinhibition mechanisms can point to workable, likely favored sites.

A related strategy involves recruiting a small molecule that stabilizes (destabilizes) an active (inactive) conformation. A drug that acts by altering the stability of the protein is an allosteric drug. Experimentally validating the increased thermodynamic stability can point to a candidate small molecule stability-rescuing agent, serving as starting points for further development. Small molecules that preferentially bind to the folded/active conformation act by shifting the equilibrium towards this state. Experimental verification can also involve Saturation Transfer Difference (STD) NMR [74]. Testing for direct binding can employ STD-NMR, WaterLOGSY, and CPMG [75,76], mutations, and degradation. Additional approaches include FragLites, an efficient approach to druggability assessment and hit generation. Broadly, computational approaches for allosteric and orthosteric drug design are advantageous as they can rapidly generate candidates that can be used as experimental starting points. Experiments are essential for validation. This is especially the case for allosteric candidates. While they may bind as predicted by computations, they may not yield the allosteric inhibition (or activation) effect that the designer sought.

2.2. Rescue mutations are a potentially powerful concept

Activating (or suppressing) mutations are debilitating. But, in rare cases, such as inherited blood and skin disorders, or emerging during treatment, they can reverse symptoms [77]. Rescue mutations are allosteric. They exchange a specific residue by another elsewhere, with this exchange suppressing the mutant phenotype [68]. A drug mimicking the rescue mutation may rescue the oncogenic driver mutant phenotype. Identifying rescue mutations may provide allosteric sites with the conformational changes favorably impacting the active site. The idea is that the drug at that site may act similarly. Mimicking a rescue mutation is however challenging since the detailed conformational changes that the mutation causes may be difficult to imitate. These depend on the details of the atoms, or groups of atoms, ‘pushing’ (or ‘pulling) the neighboring residues. Molecular dynamics simulations may help to identify the positions of such mutations [78]. They may identify unstable regions in the unbound, ligand-free state of the protein, and residues at these sites can be mutated and tested for their active site consequences. In PI3Kα, preferred rescue mutations could allosterically improve the affinity of the ATP-competitive drug. Mutations in the kinase domain’s P-loop could be candidates [68].

Tumor suppressor p53 (encoded by TP53), a highly mutated protein, is a major anticancer drug target [79]. Strategies targeting p53 loss of function include restoration (rescuing) of p53 nonsense mutations [80,81] or stabilization (e.g., eprenetapopt [82]) (see Table 1 in [79]). These also include depletion or degradation of p53 mutants with HSP90 inhibitors (e.g., ganetespib [83]) or with statins, the cholesterol-lowering drugs (e.g., lovastatin, atorvastatin [84]), and exploiting p53 synthetic lethality with G₂ checkpoint inhibitor (e.g., CBP-93872 [85]) or deficiency with Wee1 inhibitors (e.g., adavosertib [86]) or with inhibitors of the DNA damage response, such as ATR (e.g., cisplatin [87]) and CHK1/2 (e.g., UCN-01 [88]) inhibitors (Figure 2). Especially, the 800 most frequent mutations were assessed for their capacity to be rescued by arsenic trioxide (ATO), observing that solvent accessibility and temperature sensitivity were key factors of the rescue potential [89]. This led to a classification for p53 mutations for which ATO may be beneficial. Although ATO, currently in phase I trial in combination with decitabine (Dacogen, MGI Pharma Inc. and SuperGen Inc.) [90], restored the proper folding of a wide range of p53 mutants [91], only a subset of those regained wild-type-like p53 transcriptional activity.

Figure 2.

Molecular structures of drugs targeting p53 mutations (top panel). Examples of drugs for p53 reactivation, degradation, and synthetic lethality are shown. Synonyms of the drugs are given in parentheses. Except cisplatin, the three-dimensional drug structures were obtained from PubChem (https://pubchem.ncbi.nlm.nih.gov), a public chemical database of the National Library of Medicine (NLM). The structure of cisplatin was modeled. The PubChem’s compound identifier (PubChem CID) numbers for the drugs are as follows: eprenetapopt (52918385), ganetespib (135564985), lovastatin (53232), atorvastatin (60823), CBP-93872 (3095179), adavosertib (24856436), cisplatin (5702198), UCN-01 (72271). Strategies for direct targeting of mutant p53 (p53^mut) [79–82,90] (lower left panel). Eprenetapopt, a p53 reactivator, targets p53^mut to restore wild-type-like function of p53, leading to transcriptional activity for p53 target genes. Ganetespib inhibits the HSP90 chaperone machinery, leading to degradation of p53^mut by the E3 ubiquitin ligases MDM2 and CHIP. Statins inhibit HSP40 (a.k.a. DNAJA1), resulting in CHIP-mediated p53^mut degradation. Strategies for indirect targeting of p53^mut [83–88] (lower right panel). When DNA is damaged in wild-type p53 (p53^WT) cells, ATM and ATR kinases maintain the replication signaling pathway by activating p53-dependent inhibition of CDK4, which induces G₁ checkpoint arrest, and through CHK1 to inhibit CDC25 phosphatase activation on CDK1, resulting in G₂/M checkpoint arrest. However, upon DNA damage, p53^mut cells lack p53-dependent G₁ checkpoint arrest and rely on the G₂/M checkpoint arrest for survival. Inhibition of the ATR-mediated pathway by cisplatin and CBP-93872 targeting ATR, UCN-01 targeting CHK1, and adavosertib targeting Wee1 may suppress p53^mut cell survival through G₂/M checkpoint arrest. Abbreviations: HSP, heat shock protein; MDM2, mouse double minute 2 homolog; CHIP, C-terminus of HSC70-interacting protein; ATM, ataxia-telangiectasia mutated; ATR, ATM and Rad3-related; CDK, cyclin-dependent kinase; CHK1, checkpoint kinase 1; CDC25, cell division cycle 25.

In another example, the C1156Y mutation in metastatic anaplastic lymphoma kinase (ALK) is resistant to crizotinib (Xalkori, Pfizer), an ATP-competitive drug [68]. Drug resistance also emerged to ATP-competitive lorlatinib (Lorbrena, Pfizer). ALK’s allosteric mutation L1198F also resists lorlatinib. However, L1198F is a rescue mutation which resensitizes ALK to crizotinib.

ATP-competitive drug of BCR-ABL, inhibited mutants in chronic myelogenous leukemia (CML). Serendipitously found allosteric GNF-5 (PubChem CID: 44129660) compound [92] resembling the ALK’s allosteric mutation L1198F, re-sensitized BCR-ABL, essentially mimicking the autoinhibition by SH2 of the kinase domain [93]. The drug docks into the myristoyl pocket of BCR-ABL. The structure clarifies how competitive imatinib, can overcome the T315 drug-resistant hotspot [92,94].

2.3. Autoinhibition

Autoinhibition is an effective regulation mechanism. Its transitions between the protein active and inactive states are governed by allostery [93,95]. Effective regulation requires that it be sensitive, with the difference in stabilities between the states not too large. That is since signaling is most responsive to the cell environment if only a small shift in the equilibrium is required to switch the system from an inactive to an active state. This relatively small difference in the stabilities between the states and the low kinetic barriers separating them, clarify the challenge faced by evolution. Without an autoinhibition mechanisms, the proteins could flip between the inactive and active state, irrespective of the cell needs. Autoinhibition guards against such spurious activation. This emphasizes the difficulties facing pharmacological intervention to stabilize the autoinhibited states. Despite its ubiquity, the mechanisms governing the autoinhibited states and their release differ among the proteins. The autoinhibition mechanism in K-Ras differs from that of the kinases in its network, and that of the B-Raf serine/threonine protein kinase differs from that of the ABL1 tyrosine protein kinase, in turn differing from CDK4/6 serine/protein kinase, and PI3K lipid kinase. All are regulated by ligand, or partner, binding to a minor open state population, stabilizing it, thereby shifting the ensemble toward the open, catalysis ready state. In PI3K, it is binding to a receptor tyrosine kinase and stabilized by the membrane [96,97]. In B-Raf, it is the binding of minor, Ras-free open states, which in the closed state are complexed with the cysteine-rich and Ras binding domains, and stabilized by the 14-3-3 scaffolding protein [98]. Oncogenic mutations commonly stabilize the open state conformation, thereby shifting the equilibrium to the active state.

As an example, in PI3Kα, oncogenic mutations E726K, C901F, T1025A, M1043V/I, H1047R, G1049R, and N1104K, promote membrane interactions in a favorable orientation, pre-organizing PI3K for PIP₂ substrate insertion, facilitating substrate binding at the active site (i.e., acting via the reaction’s km); E542K and E545K, reduces the transition state barrier (ka), releasing autoinhibition by nSH2 [96,97]. Activating mutations most commonly occur in the regulatory-sensitive regions of p110α (Figure 3, top panel), the helical domain (E542K, E545K) or the kinase domain (H1047R, H1047L) [99]. Mutations in p85 that activate the PI3Kα complex are also frequent. To achieve pharmacological success several factors are considered, including the relative differences in the stabilities between the states, the heights of the barrier between them, and critically, the affinity of the drug, thus its residence time. To date, alpelisib (Piqray, Novartis), an orthosteric PI3Kα inhibitor (Figure 3, bottom panel), which targets the wild-type and mutant PI3Kα, is the only FDA-approved PI3Kα-selective inhibitor [99]. It is given in combination with fulvestrant (Faslodex, AstraZeneca), an estrogen receptor antagonist. But no allosteric drugs for PI3K variants are in the clinics. Still, resistance mutations in PI3Kα were identified that alter the inhibitor binding pocket. Recently resistance was observed to be overcome by the allosteric pan-mutant-selective PI3Kα inhibitor, RLY-2608 (Relay Therapeutics), guiding strategies to overcome resistance in PI3Kα-mutated cancers [99–101] (see Figure 1 in [99] for allosteric pocket). RLY-2608 targets the H1047X/E545X mutation in the p110α subunit of PI3Kα. Additional recent allosteric drugs include STX-478 (Scorpion Therapeutics) for the H1047X mutation and LOXO-783 (Eli Lilly) for the H1047R mutation in the p110α subunit [99].

Figure 3.

An in silico model structure showing PI3Kα in the inactive state (top left). The crystal structure of PI3Kα (PDB ID: 4OVV) was used to construct the inactive model. The oncogenic mutations E542K and E545K in the helical domain and E726K, C901F, T1025A, M1043V/I, H1047R, and G1049R in the kinase domain are highlighted on the wild-type p110α subunit of PI3Kα (top right). Cryo-EM structure of PI3Kα (PDB ID: 7MYO) bound to alpelisib (PubChem CID: 56649450) and crystal structure of PI3Kα (PDB ID: 8TSD) bound to RLY-2608 (PubChem CID: 166822065) (middle left). Synonyms of the drugs are given in parentheses. Alpelisib is an ATP-competitive orthosteric inhibitor targeting the active site, while RLY-2608 is an allosteric inhibitor targeting the allosteric pocket. The overlay of these two drugs highlights their different binding sites (middle right). Superimposition of the crystal structure of the kinase domain (white carton) of inactive PI3Kα (PDB ID: 4OVV) with the alpelisib-bound cryo-EM structure of the kinase domain (teal cartoon) of PI3Kα (PDB ID: 7MYO) (bottom left) and with the RLY-2608-bound crystal structure of the kinase domain (green cartoon) of PI3Kα (PDB ID: 8TSD) (bottom right). No significant difference between 4OVV and 7MYO with the orthosteric inhibitor alpelisib, but a marked conformational change in the activation loop (A-loop) is observed between 4OVV and 8TSD with the allosteric inhibitor RLY-2608. A small inhibitory helix, which is often observed in the A-loop of inactive protein kinases, can be seen (marked with a black arrow). Abbreviations: ABD, adaptor-binding domain; RBD, Ras-binding domain.

To rationally design the allosteric PI3K inhibitors, tumor-derived fragmented DNA circulating in the bloodstream were serially collected from PIK3CA mutated HR-positive/HER2-negative patients with advanced breast cancer treated with (orthosteric) alpelisib or inavolisib. Analysis pointed to resistance mutations in the catalytic pocket. Modeling indicated mutations impacting Gln859 block binding of the orthosteric PI3Kα inhibitors, but not pan-PI3K inhibitors. On the other hand, mutations of Trp780 stimulate resistance to both PI3Kα and pan-PI3K inhibitors. Allosteric PI3Kα inhibitor RLY-2608 overcomes both. Co-existing AKT1 mutations were also observed, requiring AKT inhibitors [101].

Another example that we noted above in a different context, is BCR-ABL fusion kinase, where the allosteric GNF-5 compound [102] re-sensitized it to imatinib and nilotinib. GNF-5 binds to the C-terminal myristate pocket of BCR-ABL. In the inactive autoinhibited state of ABL1, the SH3 and SH2 domains dock into the kinase domain, with the myristoyl group at the N-terminal serving as the switch, docking into its C-terminal pocket. Its dissociation from the kinase domain releases SH2-SH3, resulting in ABL1 activation [93]. The interaction of SH2 with the N-lobe kinase domain stabilizes the catalysis-primed conformation. Allosteric activating mutations shift the ensemble to the active state, blocking ATP-competitive drugs. In the absence of the myristoylated N-terminal in BCR-ABL, the kinase is in its ON activated state. Allosteric drugs can mimic the myristoyl autoinhibited action, shifting the ensemble to re-favor ATP-competitive drugs. Asciminib (Scemblix, Novartis) is a selective allosteric ABL1 inhibitor that docks into the myristoyl pocket of ABL1 [68], interacting with active site residues like other GNF inhibitors do. On their own, mutations at the myristate site can confer resistance to asciminib, and mutation-induced steric hindrance at the active site can block the ATP-competitive nilotinib. Combined, asciminib and nilotinib [or imatinib or dasatinib (Sprycel, Bristol-Myers Squibb/Otsuka Pharmaceutical)] can successfully kill the cancer cells, including the Y253H and E255V mutations, making it a powerful BCR-ABL inhibition strategy restoring efficacy [94,102,103]. Additional examples of such combinations were also suggested [104,105].

SHP2 allosteric inhibitors directly stabilize the autoinhibited conformation of SHP2, thereby preventing interactions between the catalytic phosphatase domain and SHP2 substrates [106,107]. Allosteric inhibitors in clinical trials include TNO-155 (batoprotafib) [108], JAB-3068 (SHP2-IN-6) [109], RMC-4630 (vociprotafib) [110], JAB-3312 [111], RLY-1971 (migoprotafib or GDC-1971) [112], BBP-398 (IACS-13909) [113] (primarily intended for patients with advanced or metastatic K-Ras^G12C mutated NSCLC), ERAS-601 [114], SH3809 [115], ET-0038 [116], and ICP-189 [117].

Finally, recently, allosteric and selective CDK2 inhibitors with negative cooperativity to cyclin binding were developed in an underexplored mechanism for CDK2 inhibition [118].

2.4. Feedback loops, crosstalk, and network rewiring

Feedback control aims to adapt the behavior of the system to maintain homeostasis [119,120]. Homeostasis relates to the system’s ability to recover after being changed. Feedback control is a requirement for normal cell function. It is handled via feedback loops, crosstalk, and broadly, network (re)wiring. Since the network, including loops and crosstalk, consists of protein-protein interactions, the influence is inherently allosteric. The network is complex. Drug treatment influences network wiring, thus signaling responses [121]. Drug resistance has been attributed to negative and positive feedback loops. On their own, feedback loops are incapable of completely reactivating steady-state physiologic signaling; they can only transiently restore it. Activation and inhibition events can however accomplish this e.g., by connecting an upstream protein to a downstream output, activated either by oncogenic mutations or overexpression. One example involves the PI3K/AKT pathway. Autopsy from patients with metastatic PIK3CA-mutant HR-positive/HER2-negative breast cancer treated with PI3Kα inhibitors, observed PIK3CA mutations in 38% of the cases, and in 13% PTEN and AKT1 mutations [101]. Resistance was overcome by the allosteric pan-mutant-selective PI3Kα-inhibitor RLY-2608. Crosstalk-wise, the network is rewired by a crosstalk between the PI3K/AKT and MEK/ERK pathways [122], resulting in signaling in one pathway promoting or quelling the other, with growth factors released by ERK1/2 signaling spurring other signaling pathways, such as PI3K/AKT (Figure 4). Thus, AKT blocked by PI3K inhibition can be reactivated by ERK2 in tumor cells with constitutive K-Ras activity [123]. Drug resistance includes network adaptations to crosstalk, feedback activation, and bypass signaling [121]. One such example of extensive rewiring of protein-protein-interaction networks is the EGFR network in colorectal cancer cells expressing transforming levels of K-Ras^G13D [124].

Figure 4.

Overview of orthosteric and allosteric inhibitors in signaling pathways. Examples are shown for SHP2, EGFR, and BCR-ABL in the MAPK pathways; ALK fusion protein in the JAK/STAT pathways; and P13K and AKT in the PI3K/AKT pathway. In tumor cells with AKT blocked by PI3K inhibition, ERK can lead to mTORC1 activation through ERK-mediated phosphorylation of TSC1/2. Abbreviations: mTORC1, mammalian target of rapamycin complex 1; TSC1/2, tuberous sclerosis complex 1/2.

As another example, Notch1 receptor and YAP1 signaling can regulate breast cancer metastasis. YAP1 expression is positively correlated with Notch1, and the YAP1-Notch1 positive feedback loop promotes lung metastasis of breast cancer by modulating self-renewal and inhibiting the BMP4–SMAD1/5 signaling [125].

To date, no allosteric drug has been derived from this strategy. As such, this approach has not yet proven successful.

2.5. Allosteric degraders, PROTACs, molecular glues, and bitopic ligands

Targeted protein degradation is a highly innovative and promising approach to targeted therapy in cancers. For several breast and multiple myeloma cancers, heterobifunctional PROTACs have advanced quickly to currently already being in clinical trials [126].

PROTACs work by reining in the ubiquitin-proteasome system, inherent to the cellular proteolytic machinery, to degrade unwanted proteins (Figure 1). Unlike targeted small molecule drugs, PROTACs tether their targets, and because the targets are then eliminated, they may offer a way to alleviate resistance and toxicity. Their cycling and re-use permits low dosage. Only a subset of the degraders forms a covalent bond with the protein targeted for degradation. Those that are covalently linked do not depend as much on the occupancy at the protein pocket, although for covalent bond formation the residues on the protein surface around the linkage site have to provide a favorable chemically conducive environment, yielding good affinity, with a long-enough residence time. PROTACs consist of two ligands joined by a linker, which is engineered to precisely coordinate the distance and orientation between the target and the recruited E3. One ligand, the warhead, binds a target protein and the other an E3 ubiquitin ligase, which induces ubiquitylation of the target, with its subsequent degradation. Well-designed linkers, and good binding affinity of the small molecule ligands for the E3 ligases are vital for successful PROTAC development [127]. In heterobifunctional allosteric PROTAC constructs, one end binds at an allosteric site, such as the myristoyl pocket of ABL or BCR-ABL, while the other end recruits the E3 ligase. With productively oriented and placed ligands, allosteric PROTACs, noncompetitive molecular glues, and broadly allosteric bitopic ligands, increase the effective local concentration of the target protein and its degrader. Clinical data of extensively pre-treated populations are promising, pointing to their favorable potential for difficult to target proteins [128], whether due to broad and shallow pockets or surfaces that offer few candidate sites for a small molecule. Noncompetitive, or allosteric, molecule glues are small molecules that interact with the surfaces of two proteins to allosterically enhance the stability of their complexed structure. Bitopic, or bivalent ligands can act similarly.

The allosteric GNF-5-linked PROTAC, targeting BCR-ABL myristoylation site discussed above [129], has been the first allosteric example [16,130]. It can collaborate with orthosteric drugs, like imatinib, to degrade ABL mutants such as those harboring the T315I mutation [92]. GNF-5–PROTAC can act on BCR-ABL lacking mutations for therapeutic stem cell applications [131]. GNF-5 binds to the myristoyl-binding pocket of ABL. Its other end is tethered to the linker which is covalently linked to another small molecule that here binds the von Hippel-Lindau (VHL) Cullin RING E3 ligase. E3 interacts with E2, which transfers ubiquitins to the BCR-ABL, resulting in BCR-ABL degradation.

AKT, a vastly important kinase target in cancer therapeutics provides another example [132]. AKT ATP-competitive inhibitors GSK690693 (PubChem CID: 16725726) [133], GDC-0068 (ipatasertib, Genentech) [134], and AZD5363 (Truqap, AstraZeneca) [135] and allosteric inhibitors MK-2206 (PubChem CID: 24964624) [136–138], ARQ-092 (miransertib, ArQule, Inc.) [139], and TAS-117 (pifusertib, Taiho Pharmaceutical) [140], were developed [141–145], but in clinical trials [146], showed severe side-effects (for competitive drugs), or limited efficacy (for allosteric drugs). Several AKT PROTACs derived from competitive drugs were developed but were ineffective in K-Ras/B-Raf mutated cells. AKT allosteric inhibitor based PROTACs were also developed [132], including degrader 62 (MS15, PubChem CID: 165368938), achieving potent and selective AKT degradation and antiproliferative activity in mutated K-Ras/B-Ras and PI3K/PTEN cancer cells. Another example concerns successful selective targeting of EGFR^L858R/T790M cancer mutant through an allosteric degrader, DDC-01–163 (PubChem CID: 132020463) in Ba/F3 cells [15]. DDC-01–163 is also effective against osimertinib (Tagrisso, AstraZeneca) resistant cells with L858R/T790M/C797S and L858R/T790M/L718Q mutations. Co-targeted with osimertinib, a competitive inhibitor, anti-proliferative activity of DDC-01–163 against EGFR^L858R/T790M mutant Ba/F3 cells increased.

Altogether, targeted protein degradation is a promising drug development approach, to date with under 2% of the hundreds of E3 ligases in the human genome recruited in PROTAC targeted protein degradation [147]. Allosteric degraders harness the allosteric drug advantages. In particular, allosteric degraders can combine with competitive drugs, promising a pharmacological breakthrough in future cancer treatments.

3. Allosteric inhibitor rational design

Allosteric drugs have been designed for protein-protein interfaces, nuclear hormone receptor modulators, kinases, GPCRs, ion channels and more. Innovative chemical modalities, such as targeted protein degraders, cyclopeptides and macrocycles, covalent inhibitors and antibodies are also being designed (for a comprehensive review with details see [21]). Allosteric drugs are often designed to fit into known allosteric pockets, as in the case of Abl’s N-terminal region, to which the myristoyl moiety is linked [148], molecular chaperones, as in the case of small molecules that inhibit (or promote) inter-chaperone protein-protein interactions [149], with applications to mTOR allosteric inhibitors, and dual binding site, orthosteric-allosteric inhibitors, according to their sites of action [150]. For mTOR RMC-5552 (RMC-6272) bitopic inhibitors, we showed that they may benefit from optimizing the (PEG8) linker length which bridges the distance between the allosteric and orthosteric inhibitors when targeting certain mTOR variants [52]. Liu et al. showed how discovering the mTOR kinase allosteric activation mechanism can potentially increase the affinity of inhibitors targeting mutants. Also harnessing the essence of allostery, the perturbative nature and reversibility of allosteric communication were proposed as a basis for computational design of allosteric effectors [151].

4. Examples of computational/experimental allosteric inhibitor rational design

Examples of successful rational design and the methods that were used include M2 muscarinic acetylcholine receptor (mAChR) and complement component 5a receptor (C5aR). The methods include homology modeling, MD simulations, clustering, ensemble docking, Glide structure-based drug design [152–154], and virtual screening [155]. These led respectively to the discoveries that positive allosteric modulators (PAMs) and negative allosteric modulators (NAMs) act via two different allosteric mechanisms and DF2593A, a potent and orally active C5a noncompetitive allosteric molecule. The benzodiazepine lorazepam that targets ovarian cancer G-protein coupled receptor 1 (OGR1) is a cancer-related computational/experimental example. It was known to be a PAM, but structural details were missing. The receptor was modeled [156], the ligand docked to the generated ensemble, scored, ranked, optimized, and tested experimentally. For comprehensive descriptions, including additional receptors, lipoxygenases, viral, and heat shock proteins see Chatzigoulas and Cournia [157]. PDK1 is a kinase example, whose PDK1-interacting fragment (PIF) pocket is established as allosteric site. Its structure was modeled, and the AGC kinase subfamily containing this allosteric site, to which it belongs, was used to define a pharmacophore model, followed by screening [158]. Binding of the top candidates shifted the PDK1 ensemble toward the active state. In a subsequent allosteric PDK1 targeting of the PIF pocket [159], an ensemble of six PDK1 with different PIF pocket conformations was generated, optimized, and used for docking a constructed compound library. The results were filtered, and manually inspected. The strongest inhibitor was crystallized. EGFR was also targeted computationally and experimentally [160]. In none of these cases the predicted compound reached the clinics.

In the case of the allosteric pan-mutant-selective PI3Kα inhibitor, RLY-2608 [100], MD simulations played a crucial role. They probed the effects of activating H1047R mutation, W1051 and F954, and the disengaged C-terminal tail conformation. The simulations suggested that H1047R weakens the interactions with the p85α regulatory subunit and favors a disengaged tail. Constraints imposed on DNA-encoded library (DEL) screening included validated full-length constructs, blocked orthosteric site with an inhibitor for detection of allosteric ligands, and enrichment for H1047R, E542K, and E545K. The obtained candidate compounds were synthesized and tested, and X-ray crystallography identified the mutant-selective cryptic pocket. The top compound showed conformation stabilizing features versus the wild type. This example shows that conformational dynamics-based design can drive mutant selectivity, emphasizing the crucial role of simulations in successful allosteric drug design.

Common tools/techniques in allosteric inhibitor rational design include DEL screening technology. Imposing constraints on the selection can bias it toward potential allosteric compounds, as we discussed above for RLY-2608. The recently developed eDESIGNER increases the diversity and properties space by comprehensively generating all possible library designs, then sorting and evaluating them to design ligand with desired attributes [161]. Another approach is high-throughput screening (HTS) [162] in the presence of orthosteric ligands. Recent developments in structure-based and ligand-based computational methods contributed to rational bioactive allosteric ligands [157].

Thus, computational strategies harnessing conformational dynamics are vital for rational allosteric drug designs. In addition to capturing preferred mutant-specific states, computational screening often employs simulations seeking ‘hidden’ pockets. These can also be small structural variations of a large pocket already known from the crystal structures of complexes. None of the examples above addresses unstable proteins. In such cases, allosteric small molecules that stabilize the native protein state can act as allosteric drugs.

5. Expert Opinion:

Here we ask: have we harnessed allostery to the fullest? Protein allostery has proven powerful in drug discovery. Drug development has usefully embraced it, and it keeps giving. New drugs that overcome resistance are being developed, including the recent kinome-wide selective EAI-432 [163], an allosteric EGFR inhibitor for NSCLC [164] for targeting mutant-EGFR driven NSCLC, including drug resistant C797S considered ‘undruggable’ [165]. This drug binds outside the ATP pocket. Co-binding with Osimertinib, it targets the allosteric L858R mutation and its variants, L858R/T790M, L858R/C797S, and L858R/T790M/C797S. The new allosteric drug is another successful recent targeting following on RLY-2608 for PI3Kα^H1047R mutant [99]. Still, have we harnessed allostery-based innovations to the fullest?

Our premise is that we should keep learning from nature, however, do so creatively, aiming to innovate, and base new ventures on solid mechanistic principles. Here we recounted rescue mutations, autoinhibition, catalysis-ready conformations and ensembles, allosteric degraders, noncompetitive molecular glues, and bitopic ligands, with covalent links of allosteric ligands and orthosteric warheads. Molecular glues can enhance protein-protein interactions, create connections between proteins, and allosterically influence the strength and kinetics of binding for interacting proteins. Indirect mechanisms such as via feedback loops and crosstalk are also inherently allosteric. All are used for regulation by nature and can be mimicked by allosteric drugs too.

Function is determined by the extent that a macromolecule populates its active conformation. Allosteric drugs affect this foundational principle. The unified view of allostery [7] underscored that (i) preexisting communication pathways couple the active and the allosteric sites and (ii) the relative population of a specific state is determined by the specific interactions that it harbors. In our case, the interactions of anchor atoms, or chemical groups, of the allosteric drugs with those of the protein allosteric site. The interactions can stabilize the active conformation and/or destabilize the inactive conformation. These principles explain how similar modulators can bind at the same allosteric site, with distinct outcomes. They also imply that rational allosteric inhibitor design should consider not only anchor atoms, which get into the pocket, and stabilize the interaction [13], but also the ‘drivers’ atoms, which ‘pull’ and/or ‘push’ protein atoms in the pocket, triggering the communication signal, and the consequent conformational change of the active site.

Recent remarkable advances [106] provide an optimistic outlook for creative and effective drug discovery progress. Still, it will take time, and drugs targeting rare diseases are lagging. As more allosteric drugs enter clinical trials, imaginative strategies will continue to revolutionize drug discovery and therapeutic strategies. We note however that oncogenic proteins drive proliferation not only via mutations involving residue substitutions; proliferation is also driven by over- (or under-) expressed proteins, e.g., via gene duplication (or deletions for repressors), gene fusions, alteration of epigenetic markers, and super-enhancers. Both residue substitutions and overexpression act through higher population of the active, catalysis-ready protein states. Despite potential emerging toxicity, drugs should consider broadly the active state.

Article highlights Box.

• Allosteric regulation is a mechanism through which an action in one location of a protein structure affects another.

• Allosteric drugs are advantageous because they bind at sites other than the conserved active sites, they could be less toxic.

• Favorable allosteric inhibitors dock into sites that are nearby the active sites, and follow nature, mimicking diverse allosteric regulation strategies.

• Examples of mechanisms adopted by nature include relieving autoinhibition and allosteric degradation. Mechanisms adopted by nature also include molecular glues which mimic scaffolding proteins

• Mimicking nature can powerfully innovate and enhance allosteric drug discovery