Leveraging conformational ensembles in allosteric drug discovery

Abstract

Proteins involved in signaling networks, such as Ras, mTOR, and EGFR, exist as dynamic conformational ensembles in biomolecular condensates. These ensembles play a crucial role in allosteric drug discovery and action. Traditional approaches in drug discovery often trace back to the induced fit model, which viewed proteins as rigid entities with active and inactive states. However, this model’s limitations hindered successful drug development. Advanced molecular dynamics simulations of oncogenic mutants and experiments reveal heterogeneous dynamic ensembles, which can uncover targetable spots like cryptic pockets, and cooperative exosites, that only exist transiently. Here we clarify traditional dogmas and show how recent knowledge improves allosteric drug design by leveraging conformational ensembles, with examples. We further discuss how ensemble-based approaches can advance promising therapeutics, unlocking their potential for more effective future strategies, including in biomolecular condensates.

Protein ensembles in drug discovery

The structure–function paradigm that dominated twentieth-century molecular biology arose from the idea that proteins can exist in only one rigid structure (or perhaps two: active and inactive). This perception, based on early crystal structures, led to the belief that protein structures are static and could serve as the sole drug design targets, and their binding can involve induced fit to accommodate the drug shape [1-3] (Figure 1; Box 1). The dogma persisted, even though the laws of biophysics pointed out that for proteins to function, they must undergo interconversion among many possible structures (conformations) whose likelihood is determined by their energies [4,5]. This biophysics-based perspective recognized that a (non-fibrous) protein can exist in a significantly larger number of flexible structures (i.e., conformational ensembles), thereby substantially expanding the potential range of targets [5,6]. This understanding opened the door to major computational drug discovery advances, permitting allosteric drugging of the ‘undruggable’, transforming biomedical research [7-9]. Key among these advances are the strategies for cryptic pocket detection, which are not observed in native protein structures and whose discovery permits rational drug discovery in precision medicine, and high-throughput screening, which involves docking each candidate drug to every conformation in the ensemble. These state-of-the-art strategies have become cornerstones of allosteric drug discovery protocols [10]. Input data typically include conformations sourced from crystal structures of the native protein, its mutational variants, or complexed states. These are augmented by modeled structures or multiple conformations derived from molecular dynamics (MD) simulations. These approaches follow the premise that a wide range of protein conformations is accessible to drug developers [6,11].

Figure 1. The Ensembles Activity Relationship (EAR).

This schematic diagram encapsulates the ensemble concept and the fundamental role of dynamic ensembles in molecular recognition and function [1]. Here, strategies for drug discovery based on the EAR are suggested. Allosteric drugs are based on ensembles. Without ensembles, there would be no allosteric drugs. The meteoric rise of allosteric drugs and, more recently, their combination with orthosteric drugs is testament to the breakthroughs accomplished by the ensemble nature of proteins. However, some challenges remain. The most formidable challenge is drug resistance, which is characterized by increasing heterogeneity of populations at the cellular and conformational ensemble levels. At the cellular level, there can be broken regulation, such as that emerging at the genomic level resulting in different major target compositions, and at the conformational level, drugs that work against one mutant may or may not work against others. The challenge lies in determining which drug combinations to use and in what order [48,49]. Created with BioRender.

Box 1. Molecular recognition mechanisms: the old, the new, and their convergence.

The old molecular recognition model

The old molecular recognition models proposed by, MJ (Monod and Jacob, 1961 [53]), MWC (Monod, Wyman, and Changeux, 1965 [1]), Koshland (1958 [54]), and KNF (Koshland, Némethy, and Filmer, 1966 [55]) models [3] considered only two discrete states, the active and inactive [11]. In the MJ model, the first to coin the allosteric phenomenon there is an ON/OFF transition between them. The Koshland model suggested that the switch between them is a sequential, induced fit change, initiated by binding at the first site. The essence of the old view has been that (i) there are two distinct conformations; in the absence of a drug (ligand) their relative fractions are controlled by an equilibrium constant; and (ii) that there is a change of shape [3]. This dogma continued for over half a century. It was based on static crystal structures which showed that there are two structures—related to the ligand-bound, and ligand-free, states—and that the shapes of the active sites of the two structures differ. This old dogma has been an all-or-none perception: the protein is either active or inactive [1,55]. It has only two shapes—ON and OFF. This implies that drug-impacted activity cannot be modulated. Further, absent a graded function, a two-state ON/OFF view also implies (iii) that the resulting propagation of the allosteric signal is along a single pathway [1,55] . If only two states—allosteric signal propagating from different drugs will be addicted to that single pathway. Yet, experiments overturned the dogma that switches are controlled by equilibrium, showing that the switch is kinetically controlled [56]. As to pathways, experiments, and multiple computational approaches identified specific allosteric residues in multiple pathways, fitting with Hans Boshard’s elegantly phrased query how fit is the induced fit [57].

In the old view it was posited that a protein has one specific crystal form (or two—ON/OFF), it folds quickly into it and even one mutation can distort function [11]. This led to the conviction that only the wild-type shape is pertinent, and that harnessing large computing power, clever strategies and appropriate chemistry, could overcome the immense challenge. The single structure paradigm that dominated tens of years, was doomed by its failure to capture biological functions, since its basic hypothesis did not subscribe to macromolecular dynamics [6]. This failure forced the argument that shapes that do not match, can force—or push—a significant structural change, a view that still echoes in the community.

The new molecular recognition model

In the 1990s, when the prevailing view was of the old model, a view that emerged from rigid crystal structures, we suggested that all dynamic proteins exist in many conformations, and that their population, or the number of molecules in each conformation, depends on the stability, or the energy of that conformation [11]. The more stable–the larger the number of molecules. The conformations are separated by kinetic barriers. Events that change their relative stabilities lead to a shift in their distributions, resulting in function or disease. They jump over low barriers displaying multiple shallow minima of depth (kT), each relating to a conformational state, which can have modified active sites [6]. Allosteric dynamic interconversion can only take place if the protein exists in conformational ensembles [11]. Only conformational ensembles, a purely physical phenomenon, can handle graded, modulated, positive or negative, drug outcomes. Within the ensemble, there is likely to be a state, which could be sparse, high-energy—but already with altered binding site shape that may provide a more complementary surface to the specific drug. Binding will produce a shift of the ensemble, catalyzing re-equilibration to offset the loss of this complementary conformation from solution. Binding is followed by minor backbone and side-chains optimization, making such induced fit dispensable for high affinity interaction [2].

Theories of when the old view of induced fit can take place apart from minor optimization as described for the new model are emerging. Although some studies have suggested it could occur at low drug concentrations [58], the mechanisms involved should be considered thoroughly. Induced fit involves binding practically any conformation (Figure I). Thus, it is a fast event. In contrast, ensemble-based conformational selection is slow. It involves binding a well-matched conformation, followed by a population shift to restore the equilibrium. Thus, population shift as described in the new model takes time. The population needs to overcome multiple kinetic barriers. In experiments and simulations, the faster event is recorded. In molecular dynamics simulations the population shift may take too long to monitor. In experiments, sparse states may not even be seen. However, should the drug exist in very high concentration, an induced fit will likely take place as the drug will bind all (most) free states. In reality, such a scenario is unlikely. Very high concentrations of drugs result in patient toxicity. The aim of drug developers is high affinity drugs which can be then prescribed at as low dosage as possible—avoiding high concentrations. High affinity suggests optimal matching surfaces.

Induced fit and conformational selection are thus not mutually exclusive. We can also interpret the two ON/OFF states as the two observable states in the ensemble. As to conformational changes, we recall that binding does not always incur a conformational change [11]. It can be minimal, mostly expressed in changes in the dynamics.

Exploiting ensemble-based strategies to target cryptic pockets have led to breakthroughs, such as inhibitors of Ras [7,12], and of phosphoinositide 3-kinase (PI3K) (e.g., STX-478, RLY-2608) [13,14]. Moreover, recent studies capturing dynamic transitions of the ensemble via extended MD simulations of oncogenic mutants such as mammalian target of rapamycin (mTOR) [14,15], Abl fusion variants [16,17], and COP9 signalosome (CSN) exosites [18,19] showed how ensemble targeting can be leveraged in allosteric drug discovery. Ensemble-targeting can also involve communication pathways triggered by specific allosteric mutations that propagate to active sites. A promising example of this occurs across the extracellular and intracellular domains of receptor tyrosine kinases (epidermal growth factor receptor, EGFR) [20,21]. Useful criteria for assessing cryptic sites feasibility as drug targets have also been outlined [22].

Here, focusing on allosteric drug discovery efforts targeting mTOR [23], EGFR [23], Abl fusion variants [16,17], and CSN exosites [24,25] as case studies, we leverage our understanding of conformational ensembles in activation and signaling mechanisms to guide allosteric drug design and discovery efforts. Our proposed design strategy prioritizes allosteric drug efficacy over mere affinity. Enhanced efficacy could be accomplished by engineering for stabilizing functional group “anchors”, and for minor electrostatic or steric clashes “drivers” on drug surfaces [26]. A specific “driver” induces a specific allosteric propagation (population shift) that modulates the active site in a specific receptor conformation [26]. Integrating this design approach could broaden therapeutic potential. Current allosteric drug design overlooks the ensemble-biased shift required to modulate, or stabilize, a specific conformation with a specific active site.

Efficacy is more pivotal than affinity in allosteric drug discovery

Classically, knowledge of the shape of the active site is crucial because it helps evaluate how a drug interacts with the target and the affinity of that interaction [26]. Favorable shape-dependent interactions indicate a tight fit, as do stabilizing hydrophobic interactions. Electrostatic complementarity can indicate specificity but lower affinity. Potency can be measured by the amount of a drug in the biological system. Crucially, an antagonist drug that has high affinity for its receptor, but no efficacy, may not activate the receptor, even at high dosage.

Affinity and allosteric efficacy are major determinants of the drug residence time, making them pivotal components in allosteric drug discovery [27-29]. High affinity is vital in orthosteric and allosteric drugs. However, whereas in orthosteric drugs affinity is the chief factor, for allosteric drugs, effectiveness is dominated by efficacy. Efficacy is the extent to which an allosteric drug stabilizes a specific protein conformation by shifting the population equilibrium of the protein’s different conformations [26]. Productive population shift is triggered by imperfect—albeit specific—mismatches between the drug and the protein pocket, resulting in distinct local conformational changes, which propagate to the active site. Efficacy is particularly crucial, making specific geometric (steric) and electrostatic mismatches key allosteric drivers. Minor differences in the interactions trigger conformational bias [26,30-32]. Here we suggest that allosteric drug protocols tune specific mismatched geometries (or chemistry).

When an allosteric drug binds, it triggers communication pathways to the active site. The resulting conformational changes can optimize the interaction of the orthosteric drug. This efficacy action by the allosteric drug, which can stabilize thus extend the residence time of the orthosteric drug, demonstrates cooperativity, where binding at one site influences another [33]. In an orthosteric+allosteric drug combination regimen, a specific allosteric drug can be more effective when there is a larger difference in the relative stabilities (energies) of the orthosteric-bound conformation versus the orthosteric+allosteric drugs-bound conformation. Since orthosteric drugs may similarly influence allosteric drugs, an orthosteric+allosteric drug combination with bidirectional cooperativity can amplify the outcome. Shortening the distance between orthosteric and allosteric drugs may imply fewer kinetic barriers along the communication pathway.

Below, we present examples of the application of conformational ensembles in allosteric drug discovery.

Leveraging knowledge of mTOR ensemble, activation and signaling mechanism

mTOR serine/threonine kinase, a major drug target, is the central component of two large multiprotein complexes, mTORC1 and mTORC2. Key domains of mTOR include HEAT repeats (N-terminal and middle), FAT (FRAP, ATM, TRRAP) domain, FRB (FKBP-rapamycin binding) domain and kinase domain [15,34,35]. Allosteric mutations A2034V and F2108L in the FKBP12-rapamycin binding (FRB) domain confer resistance to the monovalent allosteric drug everolimus (a derivative of rapamycin, Afinitor, Torpenz, Votubia, Zortress) (Figure 2). The allosteric activating mutation M2327I in the kinase domain promotes resistance to AZD8055, a potent, selective orthosteric inhibitor of mTOR kinase, targeting both mTORC1 (rapamycin-sensitive) and mTORC2 (rapamycin insensitive) complexes [36,37]. Bitopic inhibitors promote affinity and selectivity. Bitopic inhibitors targeting mTORC, which can overcome these resistance mutations, have been developed [36]. For example, Rapalink-1 [36] is a mTOR bitopic inhibitor, consisting of two cooperative components, an allosteric inhibitor rapamycin (sirolimus, Fyarro, Hyftor, Rapamune) and an orthosteric inhibitor MLN0128 (sapanisertib), connected with a PEG-rich linker [15,38,39]. More recently, Revolution Medicines developed a new class of mTORC1 bitopic inhibitors, RMC-5552 and RMC-6272, which build on rapalogs by incorporating elements from active-site inhibitors MLN0128 and XL388 respectively. These highly effective, selective bi-steric inhibitors feature an eight-unit PEG8 linker and demonstrate elevated mTORC1 specificity with potent anti-tumor activity. RMC-5552 is currently in clinical trials (NCT05557292, NCT04774952) for recurrent glioblastoma (GBM) and relapsed/refractory solid tumors.

Figure 2.

Inhibition of mTOR. Molecular structures of drugs targeting mTOR (top panel). Examples are shown for orthosteric inhibitors (MLN0128, AZD8055, and XL388) and allosteric inhibitors (rapamycin and everolimus). Two-dimensional structures of the drugs were obtained from PubChem (https://pubchem.Ncbi.nlm.nih.gov), a public chemical database of the national library of medicine (NLM), for the orthosteric inhibitors and from DrugBank (https://go.drugbank.com/) for the allosteric inhibitors. Bitopic inhibitors for mTOR covalently link these orthosteric and allosteric drugs through a linker (bottom panel). Examples of bitopic inhibitors include Rapalink-1, RMC-5552, and RMC-6272. Three-dimensional structures of bitopic drugs were obtained by optimizing the two-dimensional structures from PubChem (https://pubchem.Ncbi.nlm.nih.gov). These drugs work by binding simultaneously to both an orthosteric site and an allosteric site on a protein, enhancing selectivity and efficacy. The cryo-EM structure of mTORC1 (PDB ID: 8ERA) shows the mTOR kinase domain in complex with RMC-5552 and FKBP12. The MLN0128 analog binds to the active site of the mTOR kinase domain, while the rapamycin analog everolimus binds to the allosteric site in the N-lobe of the kinase domain. A dotted line denotes the linker connecting the two drugs. Everolimus recruits FKBP12, which hinders the substrate from reaching the active site.

Towards understanding the efficacy of these new drugs, a recent study detailed the mechanism of activation of its catalytic core and its oncogenic variant (L2427R) [15]. Simulations of the allosteric mechanism of apo mTOR activation observed an open catalytic cleft [15]. Similarly, in oncogenic variants, the catalytic cleft is highly populated in the open state [15]. A cryo-EM structure of wild-type mTOR (mTOR^WT) bound to RMC-5552 (PDB ID: 8ERA) (Figure 2) has pointed to a closed catalytic cleft. However, the potency of RMC-5552 with mutant mTOR is still unknown. The cryo-EM structure of mTOR^WT/RMC-5552 reveals that the distances between the allosteric and orthosteric inhibitors of RMC-5552 is ~22.7 Å. However, measurements of the distance in the active state of oncogenic variant, captured during the simulations, point to an open cleft state with a longer ~24.7 Å distance [15]. This suggests that the PEG8 linker of RMC-5552 may fall short, resulting in lower affinity for the oncogenic mTORC variant. Long linkers are more flexible and may resolve this lower affinity issue. But longer linker can induce toxicity, as the inhibitor may bind additional targets. Nonetheless, testing tailored linker lengths and optimized nesting of the bitopic inhibitors to mutational variants would be beneficial for best therapeutic intervention.

In a separate study, extensive MD simulations of mTOR^WT and some of its oncogenic variants in the kinase domain revealed their unique conformational ensembles [14]. In the oncogenic variants, the α-helices packing in the kinase domain were ruffled, creating a transient cryptic pocket, whose opening is associated with opening of the catalytic cleft. The cryptic pocket correlated with the allosteric pocket observed in wild-type PI3Kα lipid kinase. This structurally and mechanistically related kinase belongs to the same superfamily, phosphatidylinositol 3-kinase-related kinase (PIKK), and shares catalytic domain similarity.

Learning activation mechanisms can help pharmacology in genomics, homolog- and mutation-guided “basket” trials that test a certain drug or treatment in multiple cancer types that share a specific genetic mutation or biomarker, as defined by the National Cancer Institute (https://www.cancer.gov/publications/dictionaries/cancer-terms/def/basket-trial). In our example, since the PI3K allosteric pocket has been targeted previously by RLY-2608, an allosteric inhibitor [13], docking RLY-2608 into the newly discovered mutant mTOR (I2500F) cryptic pocket showed a good fit [14], clearly demonstrating how detailed understanding of kinase activation mechanisms can inform innovative allosteric inhibitor development in an overlooked pocket likely not observed in a single state crystal structure.

Leveraging understanding of EGFR ensemble, mutational variants, and activation mechanism

We discuss an additional example of the application of conformational ensembles in allosteric drug discovery. However, this example is of a different flavor: it involves the same protein, different cancer types, different mutation environments, and different allosteric drug susceptibilities. EGFR is a crucial node in regulation of multiple pathways, most importantly, MAPK and PI3K/mTOR in cell growth and proliferation. Common examples of EGFR inhibitors include small-molecule tyrosine kinase inhibitors (TKIs) and monoclonal antibodies (see Table 1). Its mutations are often responsible for cancer. EGFR’s complexity has challenged detailed knowledge of the activation mechanisms of the wild type and its mutants. The mutations are asymmetric and tissue-specific and are associated with distinct drug sensitivities [40] (Figure 3). In non-small cell lung cancer (NSCLC), EGFR activating and resistance mutations are largely in its intracellular tyrosine kinase domain encoded by exons 18-24 and respond to type I TKIs, e.g. gefitinib, which binds the active asymmetric kinase domain dimer. In brain GBM, a large proportion of EGFR mutations are in the extracellular ligand-binding ectodomain, especially inframe deletions such as EGFR variant III (EGFRvIII, also known as de2-7EGFR and ΔEGFR) [41]. However, in preclinical models, GBM-activating ectodomain mutations are sensitive to type II EGFR tyrosine kinase inhibitors (e.g., lapatinib) that bind the inactive symmetric kinase domain dimer [20]. The outcome is that both intra- and extracellular mutations exhibit oncogenic ligand-independent activation. Despite the mutational structural heterogeneity, in the intermediate state of GBM mutants, the ectodomain ensemble shifts (untethers), enabling activation of the kinase in a symmetric kinase domain-like organization [20]. The intermediate state features a cryptic antigenic determinant to which the monoclonal antibody, depatuxizumab (mAb806), which was raised against the main EGFRvIII GBM variant that has a large deletion, binds to form an antibody-drug conjugate.

Table 1.

Common examples of EGFR inhibitors include small-molecule tyrosine kinase inhibitors (TKIs) and monoclonal antibodies. Type I inhibitors bind to the active, or "DFG-in," conformation of the kinase domain. Type II inhibitors bind to the inactive, or "DFG-out," conformation of the kinase domain. Type I½ inhibitors bind to inactive kinases.

Drug name [Brand name] (Synonyms)	Drug type	PubChem CID	Target disease	Mechanism of action	Route of administration	Reference
Afatinib [Gilotrif] (BIBW-2992)	2nd G, type I, orthosteric, irreversible pan-HER TKI	10184653	Metastatic NSCLC	Covalently binds to the kinase domains of EGFR (ErbB1), HER2 (ErbB2), HER3 (ErbB3), and HER4 (ErbB4), irreversibly inhibits tyrosine kinase autophosphorylation, specifically targeting EGFR with exon 19 deletion and exon 21 L858R mutations	Oral	[59,60]
Dacomitinib [Vizimpro] (PF-00299804)	2nd G, type I, orthosteric, irreversible pan-HER TKI	11511120	Metastatic NSCLC	Covalently binds to the kinase domains of EGFR (ErbB1), HER2 (ErbB2), and HER4 (ErbB4), irreversibly inhibits tyrosine kinase autophosphorylation, specifically targeting EGFR with exon 19 deletion and exon 21 L858R mutations, exhibits more potency against EGFR than the other HER family members	Oral	[61,62]
Erlotinib [Tarceva] (CP-358774)	1st G, type I, orthosteric, reversible TKI	176870	Metastatic NSCLC	Selectively targets EGFR with exon 19 deletion and exon 21 L858R mutations, reversibly inhibits tyrosine kinase autophosphorylation, used in combination with first-line treatment for advanced metastatic pancreatic cancer, but ineffective against T790M resistance mutation	Oral	[63,64]
Gefitinib [Iressa] (ZD-1839)	1st G, type I, orthosteric, reversible TKI	123631	Metastatic NSCLC	Selectively targets EGFR with exon 19 deletion and exon 21 L858R mutations, reversibly inhibits tyrosine kinase autophosphorylation, but ineffective against T790M resistance mutation	Oral	[65,66]
Lapatinib [Tykerb, Tyverb] (GW572016, GSK-572016)	2nd G, type II, orthosteric, reversible dual TKI	208908	Metastatic breast cancer	Selectively targets EGFR and HER2, reversibly inhibits tyrosine kinase autophosphorylation, used to treat HER2-positive breast cancer in combination with trastuzumab targeting extracellular domain	Oral	[67,68]
Neratinib [Nerlynx] (HKI-272)	2nd G, type I, orthosteric, irreversible pan-HER TKI	9915743	Breast cancer	Covalently binds to the kinase domains of EGFR (ErbB1), HER2 (ErbB2), and HER4 (ErbB4), irreversibly inhibits tyrosine kinase autophosphorylation, used to treat HER2-positive breast cancer	Oral	[69,70]
Osimertinib [Tagrisso] (AZD-9291)	3rd G, type I½, orthosteric, irreversible TKI	71496458	Metastatic NSCLC	Covalently binds to the kinase domain of EGFR, irreversibly inhibits tyrosine kinase autophosphorylation, specifically targeting EGFR with exon 19 deletion, exon 21 L858R, and exon 20 T790M mutations	Oral	[71,72]
Vandetanib [Caprelsa] (GNF-PF-2188, ZD-6474)	Early G, type I, orthosteric, reversible TKI	3081361	NSCLC, thyroid cancer	Selectively targets VEGFR, EGFR, and RET, reversibly inhibits tyrosine kinase autophosphorylation	Oral	[73,74]
Cetuximab [Erbitux] (ABP-494, C-225, CDP-1)	IgG1 monoclonal antibody	-	Metastatic colorectal cancer, HNSCC	Binds to the extracellular domain of EGFR, preventing ligand binding, used in combination with encorafenib (B-Raf inhibitor) and mFOLFOX6 (fluorouracil, leucovorin, and oxaliplatin)	IV	[75,76]
Depatuxizumab (ABT-806)	IgG1 monoclonal antibody	-	Glioblastoma	An antibody-drug conjugate, selectively targets cancer cells expressing EGFR, particularly EGFRvIII or overexpressed wild-type EGFR	IV	[77,78]
Necitumumab [Portrazza] (IMC-11F8)	IgG1 monoclonal antibody	-	Metastatic NSCLC, metastatic colorectal cancer, malignant solid neoplasms	Binds to the extracellular domain of EGFR, preventing ligand binding, used in combination with gemcitabine and cisplatin	IV	[79,80]
Matuzumab (EMD-72000)	IgG1 monoclonal antibody	-	Head and neck, colorectal, gastric, cervical, ovarian, pancreatic cancers, and NSCLC	Binds to the extracellular domain of EGFR with high affinity, blocking dimerization and activation, modulates antibody-dependent cellular cytotoxicity	IV	[81,82]
Panitumumab [Vectibix] (E7.6.3)	IgG2 monoclonal antibody	-	Metastatic colorectal cancer	Binds to the extracellular domain of EGFR, preventing ligand binding, used in combination with sotorasib (K-Ras G12C inhibitor)	IV	[83,84]

Abbreviations: G, generation; IV, intravenous; NSCLS, non-small cell lung cancer; RET, rearranged during transfection; HNSCC, head and neck squamous cell carcinoma; VEGFR, vascular endothelial growth factor receptor.

Figure 3.

Inhibition of EGFR. The orthosteric drug strategy targets the intracellular EGFR kinase domain. These ATP-competitive drugs bind to the active site, thereby inhibiting the activity of the EGFR kinase and the downstream signaling pathway. Examples of crystal structures of the EGFR kinase domains in complex with small-molecule tyrosine kinase inhibitors (TKIs) are shown (top panel). The reversible, type I inhibitor gefitinib is loaded into the active site of the active EGFRL858R kinase domain (PDB ID: 2ITZ). The reversible, type II inhibitor lapatinib is loaded into the active site of the inactive EGFR kinase domain (PDB ID: 1XKK). The covalently linked irreversible, type I½ inhibitor osimertinib is loaded into the active site of the active EGFRL858R kinase domain (PDB ID: 6JWL). The orthosteric drug approach is highly effective initially, but it inevitably leads to drug resistance. To overcome this, combination drug therapies are used, such as ortho-allosteric inhibitors or orthosteric drugs in combination with antibodies, to enhance treatment efficacy. Examples are shown for orthosteric drugs and monoclonal antibody treatments targeting EGFR with different mechanisms of action for different cancer types (bottom panel). Orthostatic drugs bind to the intracellular kinase domain, while therapeutic antibodies bind to the ectodomain domain of EGFR. Combining different drug combinations leads to variations in function, cancer specificity of treatment, and the ability to overcome drug resistance. In NSCLC, caner-prone mutations of EGFR mainly occur in the intracellular kinase domain. The most effective approach is to treat with a monoclonal antibody that targets the ectodomain to prevent ligand binding and with orthosteric drugs that inhibit kinase activity in the cytoplasm. In glioblastoma, mutations in the ectodomain of EGFR, such as missense mutations and exon 2-7 deletion, are commonly observed. The antibody-drug conjugate depatuxizumab was previously proposed to target the ectodomain of EGFR, and type II lapatinib was proposed to inhibit the activity of the symmetric-like kinase domain dimer in the intracellular side. However, the efficacy of these glioblastoma drugs is limited due to their inability to cross the blood-brain barrier (BBB). Abbreviations: ECD, ectodomain; JXM, juxtamembrane domain; KD, kinase domain; TMD, transmembrane domain.

A cardinal question is why the mutations occur in different domains of the EGFR gene in NSCLC and GBM tumors and why the differing allosteric drug susceptibilities. We believe that it may reflect the adaptation of the mutations to the environment. The extracellular GBM mutational preference is to facilitate brain cancer, which does not metastasize as the NSCLC cancer does—but aggressively invades neighboring tissue. Invading brain tissue may favor mutations that enhance cell motility and remodeling the extracellular matrix. Mutations in the extracellular domain (ECD) of EGFR in GBM can, however, allosterically activate the kinase domain in a way that leads to different downstream phosphorylation patterns and signaling pathways than those caused by mutations in the kinase domain, which are common in NSCLC. Invasion is influenced by the tumor’s genetic programs and the microenvironment, although how is unclear. Type I TKIs bind to the active conformation of the kinase domain, while type II TKIs bind to the inactive conformation. Inhibiting EGFR effectively is critical, as in GBM, EGFR has the highest enrichment. More importantly, is the crucial pharmacological question of how, in GBM, the allosteric mutations in the ectodomain allosterically propagate to the kinase domain to result in a classical Asp-Phe-Gly (DFG)-out. In the DFG-out state, phenylalanine moves out of the hydrophobic pocket, disrupting the orientation of the Asp of the DFG. This may sterically block the ATP binding site conformation, with Asp no longer able to coordinate magnesium. This activation loop conformation is the target of type II inhibitors. Type I½ inhibitors also bind to an inactive conformation (DFG-in, but αC-helix-out), which is distinct from both type I and type II [35,42]. Understanding exactly how EGFR’s extra- and intra-cellular domains are allosterically linked, identifying their preferred states, revealing their transition intermediates and phosphorylation patterns, and signaling [42,43]—in addition to learning the microenvironment—can advance EGFR allosteric drug discovery, addressing the vital drug resistance.

While T790M is a frequent and known mutation, V948R was introduced to facilitate crystallography. It shifted the ensemble to expose the cryptic pocket, which was targeted by small-molecule inhibitor EAI001, in addition to neratinib (Nerlynx) at the ATP site, which also extends into the allosteric pocket. Kannan et al. observed that a small-molecule inhibitor (EAI045) that binds at an allosteric pocket [44], does not compete with ATP. Instead, it favors mutants (L858R, T790M, L858R/T790M). Destabilization of a short helix where L858 is in the wild type leads to the exposure of the cryptic pocket.

Leveraging understanding of ensembles of Abl fusions and activation in leukemia

Oncogenic Abl tyrosine kinase can undergo fusion to form constitutively active Bcr-Abl and Tel-Abl, which promote leukemias [16]. Both fusions are sensitive to orthosteric, ATP-competitive inhibitors, however, their response to the allosteric inhibitor asciminib (Scemblix) differed. While Bcr-Abl was sensitive to the allosteric asciminib, which binds at a site distant from the active site [45-47] and stabilized the inactive state, Tel-Abl was not. Bcr-Abl is predominantly in a dimeric and to a lesser extent tetrameric organization, and Muratcioglu et al. observed that Tel-Abl favors higher-order oligomers, where Abl trans-autophosphorylation blocked the autoinhibited state [16]. We asked why the distinct asciminib outcomes? In the experiments, phosphorylation at Y89 in Tel-Abl was significantly higher than in Bcr-Abl [16]. Y89 lies within the SH3 domain, at the interface between the SH3 domain and the SH2–kinase domain linker and is positioned close to several basic residues (Lys238, Arg239, and Lys241). We believe that phosphorylation at Y89 disrupts the local electrostatic environment, which increases the flexibility of the SH2–kinase domain linker. This would weaken the interaction between SH3 and kinase domain, making it difficult for Abl to adopt its closed conformation. Without this closed conformation, the SH2 domain cannot interact properly with the kinase domain, and the kinked αI’-helix cannot be restored, thereby preventing asciminib and GNF-2, also an allosteric inhibitor, from binding to the allosteric pocket, providing another example of how mechanistic knowledge can help inform pharmacological outcome.

Exosites in COP9 signalosome (CSN): A promising site adjacency organization

Small molecules binding on sites on the target surface are potential regulators, positive or negative. Exosites—outside the immediate ATP site, thus less conserved—can be particularly coveted drug aims. Due to their being adjacent to the active sites they have short allosteric communications. Exosites may be stable and observed in experimental structures [18], or cryptic, discovered by allosteric events that expose them. Communication distances between the allosteric binding site and the target site are a key factor in allosteric efficacy. Short distances, a few angstroms away, can allow for a more direct influence on the active site's conformation upon allosteric ligand binding, and crucially, stronger cooperativity between the ligands (drugs) at the two sites, promoting stability and specificity of both [24]. Exosites are likely to be common and have been prominently pointed out in human kinases and recently, including in the CSN protein complex where targeting them can allosterically modulate CSN activity [19]. CSN, a multi-subunit protein complex, is a key in removing Nedd8 from cullins (deneddylation). Abagyan and his colleagues showed the merit of targeting exosites, which may allosterically alleviate toxicity that stems from ATP-competitive drugs due to off-target effects and have usefully compiled those observed in available structures [25]. Adjacent, cryptic exosites are largely still awaiting.

Concluding remarks and future perspectives

Proteins exist as conformational ensembles. Thus, it is expected that their inhomogeneous populations play cardinal roles in function [6], and that advanced, cutting-edge drug discovery can advantageously harness them. Our ensemble-activity relationship (EAR) concept (Figure 1) is the future. However, as we clarified in this review, their fundamental nature offers much more than has been recognized, including extension of cooperative drug residence times, fusions, as in the case of Abl [16], and favoring active site spatially adjacent cryptic exosites. Biased population shift (efficacy) could be engineered to stabilize cryptic exosite pockets, making them attractive targets. Collectively, we formulate the role of conformational ensembles in productive drug discovery and further modernize the drug discovery paradigm, to go beyond the decades-old induced fit thesis.

The crucial problem facing cancer drug discovery is resistance. The drug discovery challenge involves not only heterogeneous conformational ensembles but also heterogeneous cell populations. These populations, which are a major driver of cell proliferation, result from disproportionate gene over- or underexpression, making the selection of a drug combination regimen a formidable task [48,49]. Innovative biomolecular condensates-dispersing drugs in cancer has been discussed. However, specificity could be a concern.

Conformational ensembles are likely to be constrained in the condensates’ dense, crowded, and complex environment inside membraneless organelles. Dense packing can restrict large-scale rotation of protein domains around a flexible joint. For example, hinge-bending associated cavities are more likely to be constrained due to crowding. As non-equilibrium systems, biomolecular condensates can involve dynamically biased ensembles, making allosteric drugs more effective. While a mostly unexplored area in pharmacology, the non-equilibrium nature can unlock allosteric drugs’ potential for more potent effects than in freely diffusing systems [50]. Tumor conformational heterogeneity could be more limited in biomolecular condensates, primarily because of their functionally specific crowded microenvironment. The condensate environment can also support the powerful proximity-induced pharmacology, including molecular glues and targeted degradation, as well as other ensemble-related biological processes [51].

Protein conformational ensembles have been crucial to significant scientific advancements and have led to widespread applications. This review explores how these ensembles enabled major progress in allosteric drug discovery and proposes novel uses and future directions to realize their full potential. We believe that the emerging and existing innovative allosteric drugs [52] should be advantageous in biomolecular condensates too. Protocol-wise, allosteric drug optimization accounting for the more constrained ensembles, and more nonpolar conditions, and their combinations with condensates-dispersing drugs, could map the future.

Figure I.

Ensembles of protein conformations and drug binding. The classical view of the free energy landscape of protein conformational ensembles and the classical induced fit model of drug binding (left panel). According to the classical view, proteins exist in only two states: active and inactive. In the induced fit model, a protein in a single or dominant conformational state changes its conformation to fit the drug after binding to it. The drug induces a conformational change in the protein for a tighter fit, resulting in lower protein conformational free energy. The current view of the free energy landscape of protein conformational ensembles and the conformational selection model for drug binding (right panel). The current view is of conformational states separated by kinetic barriers, which can be high or low. A conformational state is populated based on its relative stability: the more stable, the more populated. Transitioning over relatively high kinetic barriers require longer time scales, ranging from e.g., μs to ms, and beyond, while transitioning over the low kinetic barriers may require ns. The distributions of protein conformational states are affected by external (allosteric) factors such as ligand binding, posttranslational modifications (e.g., phosphorylation), mutations, changes in the pH). Broadly, binding of drugs to proteins was explained by the classical induced fit model and the conformational selection model. The conformational selection model proposes that proteins exhibit dynamic equilibrium among multiple conformations. This model assumes ensembles of protein conformations. A drug can selectively bind to a pre-existing conformation of the protein, resulting in a shift in the population and redistribution of conformations. In practice, however, protein-drug interactions can involve aspects of both the induced fit and conformational selection mechanisms simultaneously.

Outstanding questions.

• The crucial problem facing drug discovery is resistance. Single drugs can only delay cancer development. Should conformations of common double mutations be considered?

• Can we define oncogenic molecular mechanisms to discover drug soft spots? What properties to look for?

• How to select the ‘optimal’ allosteric drugs in orthosteric/allosteric drug combination, considering their mutual cooperativity?

• Drug residence time is a crucial factor. How to consider it when selecting drug combination?

Highlights.

• Proteins are not rigid structures. They consist of dynamic ensembles. Drugs whose protocols target ensembles are expected to have higher efficacy

• Here we explain the profound roles and merits of conformational ensembles in drug discovery, and why targeting them is expected to have higher efficacy

• Drugs that target ensembles can cooperatively extend residence times and specificity of both orthosteric and allosteric drugs

• Sampling of ensembles can capture cryptic pockets, including exosites. Exosites, which adjoin ATP or active sites, are especially coveted drug targets, with strongest cooperativity

• Learning oncogenic protein mechanisms is essential for successful design protocols. It provides insight and can help in pioneering drug discovery

Glossary

Cryptic pocket

a transient, allosteric binding site for drug binding that targets a protein, which is not present in the static structure obtained through experimentation.

Orthosteric drug

a drug that binds directly to the active site of a protein or enzyme.

Allosteric drug

a drug that binds to a protein at a site distinct from the active site of a protein, known as an allosteric site.

MAPK pathway

mitogen-activated protein kinase pathway, a key pathway regulates a wide variety of cellular processes, including cell proliferation, differentiation, apoptosis, and survival.

PI3K/mTOR pathway

a key pathway regulates a crucial intracellular signaling network, including cell growth, metabolism, survival, and proliferation.