Covalent-Allosteric Inhibitors: Do We Get the Best of Both Worlds?

Abstract

Covalent-allosteric inhibitors (CAIs) may achieve the best of both worlds: increased potency, long-lasting effects, and reduced drug resistance typical of covalent ligands, along with enhanced specificity and decreased toxicity inherent in allosteric modulators. Therefore, CAIs can be an effective strategy to transform many undruggable targets into druggable ones. However, CAIs are challenging to design. In this perspective, we analyze the discovery of known CAIs targeting three protein families: protein phosphatases, protein kinases, and GTPases. We also discuss how computational methods and tools can play a role in addressing the practical challenges of rational CAI design.

Introduction

As an expanding therapeutic class, covalent drugs are believed to have several pharmacological advantages over conventional non-covalent drugs,^1–4 such as higher potency and selectivity, prolonged duration of drug action, and good potential to target traditionally undruggable proteins. A covalent drug contains a warhead,^{2, 5} typically a reactive electrophilic group, that can bind reversibly or irreversibly to a reactive nucleophilic amino acid in the protein target. While cysteine remains the predominant amino acid for covalent drug binding, new reactions have been developed to enable covalent ligands for other amino acids like serine, threonine, lysine, tyrosine, and histidine.^{4, 6} Starting from aspirin, nearly 60 covalent drugs have been approved by the United States Food and Drug Administration (FDA),⁷ for indications ranging from inflammation and infection to various cancers. Covalent oncology drugs like Ibrutinib and Osimertinib have been widely used as the first-line treatments.⁸ Nirmatrelvir, a covalent drug that inhibits a key SARS-CoV-2 protease, effectively prevents hospitalization and death.⁹ Many covalent inhibitors have been reported with multi-billion annual sales.⁴ While these examples illustrate the success of covalent drugs, there have been continuous concerns about the off-target effects that trigger toxicity and immune responses. Random binding of covalent drugs and their reactive metabolites have been speculated to form drug-protein adducts,¹⁰ which may be associated with acute tissue damage or adverse immune response.^{1, 4} These concerns cast a significant shadow over covalent drug research and development.

Similar to covalent drugs, allosteric modulators represent another active frontier in small-molecule drug development.¹¹ Unlike orthosteric ligands, allosteric modulators bind to sites distinct from the active or orthosteric pocket and exert their effects by altering the protein’s conformation. The potential advantages of allosteric modulators include enhanced specificity, improved safety profiles, and the ability to fine-tune orthosteric binding events while maintaining the temporal and spatial aspects of orthosteric signaling.¹² Compared to the conserved orthosteric site, targeting the less conserved allosteric sites could provide greater subtype selectivity for covalent ligand binding, which propelled the concept of covalent-allosteric drugs.¹³ The hope is that allosteric binding can transform the target into inactive conformations and eventually alter or block the activity of the target selectively — an effect similar to allosteric post-translational modifications. Targeting allosteric sites may also help overcome drug resistance caused by mutations or other modifications to the sequence.¹⁴ Furthermore, a comparison of over 36,000 ligands from ChEMBL suggested that allosteric ligands generally display weaker binding affinity than orthosteric ones to the same targets.¹⁵ Covalent binding can enhance the affinity of allosteric ligands and prolong the drug-target residence time, which extends the inhibition duration, reduces the required dosage, and potentially lowers toxicity and off-target effects.¹⁶ Therefore, adding covalent binding to an allosteric ligand may enhance their potency and yield desirable drug-like properties, particularly for higher potency, specificity, and safety.

Despite their great potential, small molecules that act as covalent-allosteric ligands are rarely reported, with most focusing on the modulation of receptor proteins such as G protein-coupled receptors.¹⁷ Inspired by recent successes with various protein kinases and protein phosphatases, covalent allosteric inhibitors (CAIs) of enzymes have begun to receive broader attention;¹³ however, only 1–2% of publications on covalent inhibitors or allosteric modulators address CAIs (Figure 1A). Based on a well-accepted model for general covalent ligands,¹⁸ the inhibitory mechanism of CAIs can be interpreted in two steps (Figure 1B): a CAI binds to the enzyme target to create a loose complex (E…I), which then undergoes a chemical reaction and forms the covalent bond, leading to a tight complex (E-I). The association rate constant k_on and the dissociation rate constant k_off describe the reversible binding in the first step, while the rate constant k_inact characterizes the covalent bond formation in the second step (Figure 1B). This model provides valuable insight into CAI potency and selectivity:

Figure 1.

(A) CAI discovery is still in its infancy compared to covalent inhibitors or allosteric modulators. Between 2010 and 2023, there is a clear increase in the number of publications on the topic of “covalent inhibitors” (an average growth of 12% per year) or “allosteric modulators” (an average growth of 4% per year) in the Web of Science core collection. However, the number of publications on “covalent-allosteric inhibitors” or “allosteric-covalent inhibitors” remains low, as shown by the golden bar chart labeled with the actual number of publications each year. (B) A two-step model for the CAI mechanism, with E as the enzyme and I as the inhibitor. Compared with covalent orthosteric inhibitors (COIs), CAIs in general may show lower affinity for the first step and require a greater reaction free energy in the second step to compete with COIs.

1. Unlike reversible inhibitors, where the ligand concentration primarily dictates inhibition, covalent inhibitors show both time and concentration dependency. Although the half-maximal inhibitory concentration (IC50) or the inhibition constant K_i values were often utilized to describe the potency of covalent inhibitors in earlier studies, the second-order rate constant k_inact/K_I is preferred because it characterizes the inhibitor potency in a time-independent manner (Figure 1B).^{19, 20} k_inact/K_I contains the inactivation constant K_I (denoted with a subscript “I”), which is derived from the steady-state approximation as (k_off+k_inact)/k_on. Distinct from K_i, the form of K_I considers both reversible binding and covalent bond formation, which is more suitable for evaluating CAIs. Here, we try to specify the kinetic parameters of covalent inhibition with k_inact/K_I whenever possible, but for many compounds, we have used the IC50 or K_i as a reference according to the original reference, recognizing that IC50 or K_i is not a suitable parameter for evaluating CAIs.

2. The two-step model provides some thermodynamic hints into the binding selectivity, if orthosteric and allosteric sites contain the same reactive amino acid. To design selective CAIs, the reaction free energy in the second step must favor the targeted allosteric site over the orthosteric or other allosteric sites (Figure 1B), which is not trivial to achieve without a deep understanding of the warhead reactions and their mechanisms influenced by various chemical environments.^{13, 21} Currently, too few examples exist to establish such understanding and identify warheads suitable for CAIs. This motivates us to write this perspective to inspire future experimental and computational research, advancing our knowledge of CAI research.

Beyond measuring and comparing the potency, several practical challenges must be addressed for CAI discovery, as allosteric binding sites—some of which are hidden or cryptic—are much less understood and explored than orthosteric sites. Specifically, known three-dimensional (3D) protein structures from methods like X-ray crystallography, nuclear magnetic resonance, and cryogenic electron microscopy (cryoEM) may represent the most commonly populated conformational state, while allosteric pockets may manifest in less populated or transient conformational states.²² Therefore, further efforts are necessary to identify and validate allosteric binding sites for the structure-based design of CAIs. Additionally, there are no general rules for targeting allosteric sites regarding how to connect ligand binding to the impacts on target activity or mechanism. Finally, it remains largely unknown whether the same strategies used to design conventional reversible or covalent inhibitors apply to targeting allosteric sites, particularly concerning warhead reactivity and structure-activity relationships. Consequently, significant efforts are in critical need to enhance current limited knowledge and expedite the discovery of CAIs.

We have analyzed prototypical CAIs targeting protein kinases, protein phosphatases, and GTPases to provide an effective gateway for future CAI discovery. Protein kinases phosphorylate their substrates, whereas protein phosphatases catalyze the reverse process by removing the phosphate group. These enzymes regulate numerous cellular processes, such as proliferation, apoptosis, subcellular translocation, metabolism, and inflammation.²³ Thus, many members in these protein families have been studied as drug targets for cancers, cardiovascular diseases, metabolic diseases, and immune disorders.^{24, 25} Specifically, we present examples of CAI to target two protein tyrosine phosphatases (PTP) — PTP1B and Src-Homology-2-containing PTP 2 (SHP2) — as well as the serine/threonine kinase Akt. In addition, we also analyze the CAIs of the KRAS G12C mutant. After these case studies, we outline the computational tools for CAI design and delve into the current challenges and future pathways for rational CAI discovery.

Covalent-Allosteric Inhibitors of PTP1B

Protein tyrosine phosphatases (PTPs) constitute a large family of signaling enzymes involved in various essential cellular processes. The PTP-catalyzed reaction involves a nucleophilic attack on the phosphorus atom in the substrate by the active site Cys residue.²⁶ Given the importance of the active site cysteine in PTP catalysis, a number of covalent PTP inhibitors targeting the PTP active site Cys have been reported,¹⁹ including alpha-bromobenzylphosphonates²⁷ and aryl vinyl sulfones.²⁸ Unfortunately, while these active site-directed covalent inhibitors can be used for activity-based PTP profiling, they lack the necessary isozyme selectivity due to the conserved nature of the active site pocket. The following section highlights recent developments on covalent PTP inhibitors targeting non-active site Cys residues.

As one of the most central PTPs, PTP1B (encoded by the PTPN1 gene) is known to regulate many pathways implicated in metabolic and oncogenic conditions and thus is extensively studied as a therapeutic target for cancers, diabetes, and obesity.²⁴ PTP1B has an N-terminal catalytic domain (residues 3–277, PDBID: 2BGD), and two structurally disordered domains (residues 278–435).²⁹ The N-terminal domain contains the catalytic cysteine residue (Cys215) and several non-catalytic ones (Cys32, Cys92, Cys121, Cys226, and Cys231). Cys121 has been identified as a potential allosteric site for covalent inhibitors, located just 8 Å from the active site (Figure 2). In 2005, Hansen et al. reported the first CAI of PTP1B, 4-(aminosulfonyl)-7-fluoro-2,1,3-benzoxadiazole (Erlanson-2005-ABDF).³⁰ Erlanson-2005-ABDF was initially identified in a screening campaign to discover covalent probes targeting the orthosteric Cys215. During this screening, however, Erlanson-2005-ABDF demonstrated rapid and specific reactivity with Cys121, reducing enzyme activity with a K_i of 1.3 mM. Subsequently, mass spectrometry (MS) analysis confirmed Cys121 as the binding site for Erlanson-2005-ABDF. The proposed mechanism of Erlanson-2005-ABDF inhibition involves either restriction of the movement of the critical WPD loop, thereby hindering the transition between open and closed conformations during catalysis, or disruption of the hydrogen bond between Tyr124 and His214, which could perturb the active site due to the proximity of Cys215 to Cys121. However, this allosteric inhibition mechanism remains to be confirmed.

Figure 2.

Illustration of CAIs targeting PTP1B allosteric sites. (a) The cartoon of the PTP1B structure highlights the positions of key residues. The active site pocket containing Cys215 is shown by the red surface. (b-d) The chemical structures of known CAIs targeting PTP1B and available crystal structures. Residues within 3 Å of the CAIs are shown as cyan sticks, while the CAIs and their covalently linked residues are shown as yellow sticks.

Following the discovery of Erlanson-2005-ABDF, Punthasee et al. in 2017 found another compound, herein referred to as Gates-2017–5a ²¹, which acts as another CAI to target PTP1B (Figure 2). Gates-2017–5a was derived from an orthosteric inhibitor using an exo affinity labeling strategy, which developed ligands bound to the active site but covalently linked to an amino acid outside the active site.³¹ During the characterization of Gates-2017–5a, the absence of saturation kinetics suggested that it did not operate via the anticipated exo affinity labeling mechanism. However, it still exhibited binding to the active site with an IC50 value of 54 μM. Subsequent MS analysis confirmed that Gates-2017–5a covalently bonded to Cys121 and showed a less intense signal on Cys32. Further mutagenesis studies on Cys32 revealed that this residue is not essential for the inactivation effect of Gates-2017–5a. Based on these observations, a dual-acting mechanism was proposed for Gates-2017–5a: it binds non-covalently to the active site Cys215 and covalently binds to Cys121, resulting in enzyme inactivation through both orthosteric and allosteric interactions.²¹ A non-covalent analog of Gates-2017–5a has been determined to be bound to the active site (PDB ID: 5T19),³² but no crystal structure is available to support the dual-acting mechanism of Gates-2017–5a.

In addition to targeting C121, CAIs have been developed to form covalent bonds with other amino acids in PTP1B. In 2018, Keedy et al. discovered a novel CAI targeting the K197C mutant using the “covalent tethering” method³³: via mutating functionally important K197 to cysteine, a disulfide-capped fragment library can be screened against the site of interest with improved sensitivity because of covalent bond formation. After multiple‐temperature X‐ray crystallography screening and modification of the hit molecule, Fraser-2018–2 was identified as a CAI for the K197C mutant PTP1B (Figure 2c), which exhibited dose-dependent tethering and partial noncompetitive allosteric inhibition of PTP1B K197C at K_i = 7.1 μM. Although it does not affect wildtype PTP1B, Fraser-2018–2 was suggested to be further modified into a non-covalent-allosteric inhibitor for wildtype PTP1B. In 2023, another small molecule, Keedy-2022-z0048, was developed from a room-temperature (RT) X-ray crystallography screening campaign against the wildtype PTP1B.³⁴ It is believed to covalently bind to Lys237 or Lys197, as shown in crystal structures (Figure 2). However, this compound did not exhibit inhibitory effects in vitro due to weak affinity or other reasons, which indicates the challenge of establishing the connection between allosteric binding and target inhibition for potential CAIs.

Covalent-Allosteric Inhibitors of SHP2

Src homology 2-containing protein tyrosine phosphatase-2 (SHP2, encoded by the PTPN11 gene) is another critical member of the PTP superfamily.^{35, 36} SHP2 is overexpressed in cancer and leukemia cells. It holds potential as an anticancer target due to its mutations, which lead to the Noonan and Leopard syndromes, both linked to heightened cancer risks from PTP activity misregulation.^{35, 37} As a central link essential for receptor protein tyrosine kinase-mediated RAS activation, SHP2 represents an attractive anti-cancer target.³⁸ In 1998, Hof and his team determined the first crystal structure of SHP-2 bound to a phosphopeptide, providing insights into how the SH2 domain regulates SHP-2’s catalytic activity and offering successors a 3D structural template of SHP-2 for drug discovery. Structurally, SHP2 consists of two SH2 domains (residues 6–102 and 112–216), a PTP catalytic domain (residues 247–517), and a C-terminus tail (residues 548–571).³⁹ Within its catalytic domain are six cysteine residues (Cys259, Cys318, Cys333, Cys367, Cys459, and Cys486), the active-site residue Cys459 is invariant among PTPs.⁴⁰ Only 9 Å from the catalytic site (Figure 3), the non-conserved Cys333 distinguishes SHP2 from other PTPs except SHP1, which presents an opportunity to discover CAIs selective for SHP1 and SHP2.

Figure 3.

Illustration of CAIs targeting SHP2 allosteric cysteine Cys333. The crystal structure of SHP2 (PDB ID: 8B5Y) with the SH2 domains and the PTP catalytic domain is shown in cartoon, in addition to the chemical structure of the allosteric covalent F1AsH-EDT2 and Bishop-2018–12. The active site pocket is shown by the red surface.

In 2015, aiming to find a cryptic allosteric site within the catalytic domain of SHP2, Bishop et al. identified F1AsH-EDT2, a biarsenical fluorescent chemical probe that reacts with cysteine covalently (Figure 3).⁴⁰ F1AsH-EDT2 demonstrates potent and selective inhibition against SHP2 over other PTPs, with an IC₅₀ value of 74 nM. Supported by site-directed mutagenesis evidence, this specificity might be due to targeting non-conserved Cys333 over other cysteine residues. Following the discovery of the cryptic allosteric site, Bishop et al. in 2018 developed a series of C333-targeted allosteric SHP2 inhibitors,⁴¹ which appear more druglike than F1AsH-EDT2 containing arsenic. In the series, a compound Bishop-2018–12 with an acrylamide warhead (IC₅₀ = 35 μM, k_inact = 0.021 min⁻¹, K_I = 279 μM) displayed modest potency and inhibition kinetics for SHP1 and SHP2 over three other PTPs. LC-MS analysis conducted on the C333P SHP2 mutant revealed a loss of inhibition by Bishop-2018–12, which confirmed its selective binding to Cys333 over the other cysteine. Given the simple structure of Bishop-2018–12, further studies may be needed to understand the structure-activity relationship (SAR) and improve the selectivity of SHP2, which may require an understanding of the difference between SHP1 and SHP2 in the sequence, structure, and even detailed mechanism. After all, as another proof of concept, Bishop-2018–12 suggests the great potential of using CAIs to target a non-conserved allosteric site specific to a difficult protein like SHP2.

Covalent-Allosteric Inhibitors of KRAS G12C

KRAS, NRAS, and HRAS in the RAS family are the most frequently mutated genes across human cancers. Substantial efforts have been devoted to targeting the KRAS mutations, while 80% of the mutation occurs at amino acid position 12 near the GTP/GDP binding site (Figure 4). KRAS was long considered undruggable, as it seemed to be a small protein without a well-defined pocket for small-molecule inhibitors to target. The discovery of a cryptic allosteric pocket in the KRAS G12C mutant and covalent ligands by Shokat et al. in 2013 has shifted the paradigm.^{42, 43} Many covalent and non-covalent inhibitors have been developed to target this allosteric site enclosed by Arg68, Asp69, His95, Tyr96, and other residues (Figure 4).⁴⁴ So far, two KRAS inhibitors, AMG510 (sotorasib) and MRTX840 (adagrasib),⁴⁵ have been approved by the FDA for cancer or tumor treatments. Both molecules are confirmed to covalently bind to the allosteric pocket, which highlights the success of CAIs.

Figure 4.

Illustration of CAIs targeting the KRAS G12C allosteric site (C12 is the mutated form of G in the WT sequence). (a) An overview of the KRAS G12C structure (PDB ID: 4LDJ), highlighting the residue binding to covalent-allosteric inhibitors; (b) The chemical structure of known covalent-allosteric inhibitors: ARS-853, ARS-1620, AMG510 and MRTX849; (c) A detailed view of the binding poses from co-crystal structures of ARS-853, ARS-1620, AMG510 and MRTX849 in complex with KRAS G12C and GDP. Residues binding to the ligand are shown as cyan sticks, while the ligands and covalently linked residues are represented as yellow sticks. The labels for covalently linked residues are highlighted in red sticks. The active site pocket is shown by the red surface.

Prior studies have revealed the details of this allosteric pocket beneath the effector binding switch-II region and its mechanism to favor the inactive GDP state upon inhibitor binding. Instead of directly inhibiting the GTP/GDP binding site, binding to the allosteric site is believed to change the nucleotide preference of KRAS from GTP to GDP, presumably via locking KRAS into certain conformations.⁴² Since the reporting of groundbreaking compounds in 2013, several different scaffolds have been published, and at least five different compounds are in clinical trials or with FDA approval.⁴⁶ In 2016, Liu et al. synthesized the compound ARS-853, among the early examples specifically recognizing KRAS G12C.⁴⁷ An iterative structure-based design strategy was carried out to enhance cellular efficacy, generating candidate compounds for testing with the recombinant KRAS G12C protein and in cell-based assays. These efforts yielded ARS-853 as the most potent compound at a cellular engagement IC₅₀ at 6 hours of 1.6 μmol/L.⁴⁷ In the meantime, Rosen et al. described a plausible mechanism that the drug-bound KRAS G12C is insusceptible to nucleotide exchange factors and thus trapped in its GDP-bound inactive form.⁴⁸ Later, after changing the flexible linker to a rigid quinazoline scaffold, the compound ARS-1620 was reported as a covalent KRAS G12C-specific inhibitor,⁴⁹ which achieves rapid and sustained target occupancy toward tumor regression in vivo. ARS-1620 demonstrated an IC50 value of 120 nM, which represents a 10-fold increase in potency compared to ARS-853. This improvement aligns with G12C occupancy and results in more than a 100-fold selectivity window for the mutant allele in cells. In 2019, a group at Amgen developed a series of inhibitors, including the compound AMG510, which shares a similar binding mode to ARS-1620 (Figure 4).⁵⁰ Preclinical studies showed that AMG510 induces regression of tumors, boosts chemotherapy, and promotes a pro-inflammatory tumor environment in immune-competent mice.⁵⁰ Clinical trials indicated promising anti-tumor activity, positioning AMG510 as a potential transformative therapy for patients without effective treatments.⁵¹ All the 22 mutated cell lines treated with AMG 510 for 2 hours showed that the phosphorylation of downstream proteins was inhibited, with IC₅₀ values ranging from 0.010 to 0.123 μM. In 2021, the FDA granted accelerated approval to AMG510 (sotorasib) for cancer treatment.

Acting similarly to AMG510, MRTX849 was reported by a group at Mirati Therapeutics in 2020, using structure-based drug design with a screening approach focusing on the pharmacokinetic properties for oral bioavailability.⁵² MRTX849, featuring a 2-fluoroacrylamide warhead for stability and reduced glutathione conjugation, exhibited potent and selective inhibition of KRAS G12C with high bioavailability and induced rapid and durable tumor regression in preclinical models, including complete responses in mice. The IC₅₀ values of MRTX849 in the NCI-H358 xenograft model and the MIA PaCa-2 xenograft model are 14 and 5 nM, respectively. The k_inact/K_I value for MRTX849 was estimated to be 35 ± 0.3 mM⁻¹s⁻¹ compared with two other KRAS G12C inhibitors, ARS-1620 (1.1 mM⁻¹s⁻¹)⁵³ and AMG510 (9.9 mM⁻¹s⁻¹).⁵⁰ In clinical trials, MRTX849 showed antitumor activity in patients with metastatic colorectal cancer and the KRAS G12C mutant, both as an oral monotherapy and in combination with cetuximab.⁵⁴ In 2024, MRTX849 (adagrasib) with cetuximab was approved by the FDA to treat advanced colorectal cancer, which is the second example of CAI to target KRAS G12C.

The success of CAI in targeting KRAS G12C suggests an exciting mechanism of allosteric modulation through conformational selection, which can be general among other targets. It may be viable to develop CAIs that selectively stabilize the target protein into a desirable state, such as the GDP-bound inactive state in the case of KRAS. Molecular modeling and simulations have helped understand the conformational states and the energetics, identifying the dynamics of the allosteric sites and the impacts of mutations on the protein structure and dynamics.⁵⁵ An expansive focus beyond G12C, targeting other KRAS mutations, is crucial for comprehensive cancer therapeutics; thus, other cysteines (Cys51, Cys80, and Cys118) in KRAS remain to be explored for potential CAIs.

Covalent-Allosteric Inhibitor of Akt

Also known as protein kinase B, Akt is a serine/threonine kinase with three isoforms: Akt1, Akt2, and Akt3 (encoded by AKT1/2/3). While Akt members regulate cellular activities such as cell proliferation, survival, and angiogenesis, dysregulation of Akt expression can lead to various health implications, such as cancers, autoimmune diseases, cardiovascular diseases, and neurological disorders, making it a promising therapeutic target.⁵⁶ Capivasertib (TRUQAP), the first-in-class Akt inhibitor approved by the FDA in 2023, recognizes the ATP-binding pocket of all three isoforms. However, it was suggested that targeting specific Akt isoforms can result in more effective cancer treatments, as these isoforms play differential roles in pathology conditions.⁵⁷ Given the similar ATP-binding pocket among Akt isoforms and other proteins in the same family, developing CAIs for Akt isoforms can be promising for enhanced specificity, effectiveness, and reduced adverse effects.

Some progress in discovering Akt CAIs has been made to target Cys296 or Cys310 at the Akt kinase domain. By modifying a non-covalent inhibitor AKti-1/2, Rauh et al. in 2015 discovered Borussertib and analogs as Akt CAIs that covalently bind to Cys296 or Cys310 (Figure 5).⁵⁸ Selective screening over one hundred protein kinases was carried out, and only the Akt isoforms were significantly inhibited (> 80% inhibition) at over 1 μM Borussertib, which also showed selectivity to Akt1 over the other isoforms. In 2019, the same team performed a sequence and structure comparison of Akt isoforms to reveal differences near Cys296 at the interface between the PH and the kinase domains.⁵⁹ Despite highly conserved sequences, deletions and mutations were found in the loop of residues 259–273 in Akt2 and Akt3, which supports the viability of targeting individual Akt isoforms. This was further supported by two similar compounds Rauh-2019–16a and Rauh-2019–16b (Figure 5), as the former displayed selectivity toward Akt1 (K_i = 59 nM) and the latter toward Akt2 (K_i = 57 nM). More structure-based evidence was found in another publication in 2019 by the same team, ⁶⁰ which further explored the SAR and yielded Rauh-2019–30b with enhanced inhibition (K_i = 6.8 nM) and better pharmacokinetic profile than Borussertib (Figure 5). The MS/MS analysis suggested labeling of Cys296 and Cys310, but the crystal structure revealed a novel binding mode of Rauh-2019–27, only labeling Cys310. Such disagreement may suggest the complex nature of CAIs and the challenge to validate the mechanism of action. A small library of Borussertib analogs was synthesized with click chemistry and tested with various cancer cell lines in 2022, adding more accessible compounds to the class of Akt CAIs.⁶¹ Overall, these Akt CAIs at nanomolar inhibition are promising for future treatment of metastatic breast cancer. Further research is probably required to better understand the mechanism of inhibition at the allosteric pocket, differentiate the labeling of Cys296 and Cys310, and develop next-generation CAIs specifically targeting individual Akt isoforms. The rich structural data of Akt would enable more extensive structure-based design (e.g., covalent docking, molecular simulations, etc.) of compounds with improved druglike properties.

Figure 5.

Illustration of CAIs targeting Akt1. (a) Insight into the Akt1 protein structure with the orthorsteric and allosteric sites is shown as red and yellow surfaces, respectively. (b) Chemical Structure of the three CAIs. (c) Difference of binding mode for the three covalent-allosteric inhibitors where Borrusertib and Rauh-2019–30b prefer to bind to Cys296 while Rauh-2019–24b prefers to bind to Cys310.

Practical Challenges in Current CAI Discovery

Despite the success of CAI targeting various proteins, most compounds discussed above were developed to inhibit well-established targets and allosteric sites. Experimental screening is still the major approach to discovering new CAIs, as demonstrated by the discovery of Erlanson-2005-ABDF and Keedy-2022-z0048.^{30, 34} Modifying existing orthosteric ligands can sometimes serendipitously yield allosteric modulators, as illustrated by the development of Gates-2017–5a.²¹ In such circumstances, however, a thorough analysis of biological experimental results remains essential for accurately differentiating CAIs and orthosteric ligands. Furthermore, achieving selectivity among various protein subtypes or isoforms continues to be difficult. For example, Erlanson-2005-ABDF not only inhibits PTP1B but also other targets like TC-PTP and LAR³⁰ due to the conserved nature of Cys121, raising concerns about off-target effects. Bishop-2018–12, which aims to target Cys333 present in SHP1 and SHP2, may require further optimization to develop into isoform-specific CAIs. For Akt CAIs, most cysteines such as Cys296 and Cys310 are conserved, complicating the effort to achieve isoform selectivity. Additional efforts may be required to design highly specific CAIs by targeting non-conserved residues and/or detailed mechanistic differences. Finally, while the available structures of KRAS and Akt allow for an extensive structure-based approach to searching for new CAIs or optimizing the hit compounds, protein structural data can be limited for many targets of interest. The lack of structural information hampers the discovery or further development of most CAIs, as seen in the case of the CAI believed bound to PTP1B Cys121.²¹ With advancements in computational approaches and capacity, some of these challenges are likely to yield computational solutions, and more rational design can create new opportunities for CAI discovery. Therefore, we present a critical, forward-looking perspective as follows.

Computer-Aided Design and Optimization of CAIs

Computational approaches can be promising for designing and optimizing CAIs, and thus we summarize available tools and highlight potential computational strategies. Given the lack of protein crystal structures, the cryptic nature of allosteric sites, and the difficulty in drug design and optimization, some of these tools are readily applicable to study CAIs (such as protein structure prediction, molecular docking, etc.), while others require further development, which together may help facilitate CAI discovery at different structure-based discovery stages (Figure 6).

Figure 6.

Summary of CAI design challenges and potential computational solutions.

1. Structure prediction tools such as AlphaFold⁶² and RoseTTAFold⁶³ coupled with simulations can be useful in sampling relevant conformations and providing insight into the mechanisms of action to optimize CAIs. For example, AlphaFold-predicted structures of the human P2X1 and P2X2 proteins were compared with the cryo-EM structure of the zebrafish P2X4 protein.⁶⁴ Combined with mutation studies, this analysis could explain the selectivity of an allosteric modulator. However, uncertainty remains with AlphaFold and similar tools in predicting multiple protein conformational states relevant to the function or mechanism of the same target.⁶⁵ This can restrict the adoption of these tools for structure-based CAI design, as some allosteric pockets may only be observed in less populated or transient conformational states. In such cases, molecular dynamics (MD) simulations and enhanced sampling methods can help sampling protein conformations.^{66, 67} Based on Newtonian laws of motion, an MD simulation can calculate how every molecule in the system moves and interacts with each other, which can sample the protein conformational dynamics at a given condition (i.e., temperature, pressure, salinity, etc.). To choose relevant conformations for docking and other drug design applications, several computational techniques have been developed.⁶⁸ As computational capacity increases and costs decrease, protein structure prediction combined with MD simulation will become a major tool for the structure-based design of CAIs.

2. With the structural information of the target protein, it is essential to identify the allosteric pocket—a critical task given the often-elusive nature of these pockets. Without experimentally validated binding sites, identifying allosteric binding sites with target amino acids (such as cysteine) is a prerequisite for structure-based CAI discovery. Several modern computational methods have been previously reviewed based on sequence, structural, dynamic, or energetic data.⁶⁹ Specifically, tools like CavityPlus⁷⁰ and SiteMap⁷¹ (Schrödinger, Inc.) are available to predict druggable pockets. These tools are simple and fast, as they only need a protein conformation as the input and can seamlessly connect with a molecular docking program (e.g., from SiteMap to Glide⁷²). They mainly utilize advanced algorithms to cluster points from a three-dimensional grid, identify the pockets from large clusters, and rank these pockets when considering volume and protein surface interactions.⁷¹ For example, SiteMap has been used to identify a tunnel-like allosteric site in SHP2, which was reported to stabilize SHP2 in an auto-inhibited conformation upon ligand binding.⁷³ Moreover, pK_a prediction (especially for cysteine tarting) with empirical⁷⁴ or physical approaches⁷⁵ can also identify potential allosteric pockets for CAIs. While the active-site cysteine is often acidic, allosteric cysteine residues may span a wide range of pK_a values (and reactivities with diverse warheads),^{67, 76} which helps understand CAI selectivity or choose the allosteric site for CAI design. Beyond these tools, new approaches based on AI/ML^77–79 and MD simulations⁸⁰ have been developed, which will find more general applications for CAI discovery. For instance, PocketMiner,⁷⁷ a graph neural network trained with 2,400 structure ensembles that contain examples of pocket opening events from MD simulations, can successfully identify the allosteric binding site on the GDP-bound KRAS^G12C structure. Based on different theories and algorithms, these tools can complement each other in identifying potential CAI-binding sites, which may enable de novo CAI design in the future.

3. With the knowledge of the allosteric sites, covalent docking is a popular technique to screen small molecules. However, current covalent docking approaches, built on conventional molecular docking developed for non-covalent binding, primarily focus on binding poses and affinity prediction. For example, CovDock⁸¹ was built upon two Schrödinger programs, the Glide docking program⁷² and the Prime protein structure refinement program,⁸² which has recently been applied to identify CAIs for the transcriptional enhanced associated domain and other targets.⁸³ CovDock samples the orientation of the target cysteine side chain and the conformations of the ligand in tandem, under the user-chosen reactions. Once the reaction type is chosen, the warhead reactivity is assumed unaffected by the remaining moiety and stays the same across the docking library. Therefore, little or no consideration is given to the warhead reactivity in CovDock, which can affect accurate docking when electron-withdrawing or donating groups modify the warhead reactivity. In our practice, the CovDock score is typically low (e.g., below −5 kcal/mol) for validated covalent ligands, which could make it hard to distinguish many covalent ligands with the same or different warheads. Other popular docking programs like Autodock⁸⁴ and GOLD⁸⁵ also provide covalent docking applications. More systematic testing and large-scale applications will be useful to demonstrate the accuracy of these approaches. As an alternative to docking, AI/ML-powered generative ligand design techniques, such as those based on generative adversarial networks and diffusion models, have emerged.⁸⁶ Despite current limitations to generating highly viable compounds,⁸⁶ generative CAI design will become available in the foreseeable future.

In addition to the abovementioned computational approaches, several others may be complementary or add new capacities. For CAI hit compounds, quantum mechanics (QM) can evaluate warhead reactivity ⁸⁷ and identify warheads with modest electrophilic activity to reduce off-target toxicity.⁸⁸ Some typical warheads have been investigated,⁸⁹ while more general studies are needed, given the high computational cost of QM calculations. Also, there are several available tools to optimize the CAI hits from virtual or experimental screening, including the water dynamics approach (like WaterMap⁹⁰) and alchemical free energy methods like free-energy perturbation (FEP) or thermodynamic integration (TI).⁹¹ For instance, WaterMap calculates the free energy of water molecules on the protein surface from MD simulations and recently was used to design covalent inhibitors of KRAS mutants with conserved water.⁹² It is suggested that a compound can be optimized by maintaining the conserved water molecule or replacing it without compromising the binding affinity, which can also be helpful for CAI design.⁹² Furthermore, with improvements in force field accuracy, sampling efficiency, and employed algorithms, alchemical free energy approaches have been widely used to estimate relative binding free energies to targeted proteins in drug discovery.^{91, 93} Among these approaches, FEP is a popular one that calculates the free energy difference between two ligands by simulating the transition from a reference ligand to another. This alchemical transformation acts as a “perturbation” to the protein-ligand complex, providing an accurate measure of the energy difference between the two ligands. Generally, the hit compound is used as the reference, and the energy difference indicates the relative binding energy of the new ligand compared to the reference. A negative relative binding energy implies that the new ligand might bind more effectively than the reference. With increasing applications to covalent ligand design,⁹⁴ FEP and related approaches can be very useful for CAI optimization. Finally, AI/ML techniques have proven powerful through the pipeline of CAI discovery. However, the performance of AI/ML models depends heavily on the quality and quantity of data, which poses a significant challenge, especially for allosteric modulators⁷⁸ and covalent ligands⁹⁵ for traditionally undruggable proteins, where data are relatively limited. In this case, combining multiple structural databases and methodologies could offer a possible solution.

In sum, we anticipate a bright future for CAIs, which harness the combined advantages of allosteric modulators and covalent inhibitors. Many proteins were previously labeled as “undruggable” because they lacked appropriate pockets or had highly similar active sites to other proteins. However, the success of CAIs has demonstrated that they can be effectively converted into druggable targets. Furthermore, computational tools, particularly with the advance of AI/ML techniques, are poised to significantly impact CAI design and optimization, heralding an exciting era for CAIs in drug discovery ahead.

Significance

• CAIs represent an emerging yet underdeveloped field. With recent success in cancer therapeutics, CAIs present numerous challenges and opportunities for drug discovery.

• This perspective analyzes the discovery of known CAIs from a structure and mechanism-based view, providing critical insights into the discovery of future CAIs.

• Computational tools play a growing role in identifying allosteric sites, recognizing hits, and optimizing hit/lead compounds, potentially paving the way for rational CAI design.

Biographies

Hui Tao received his B.S. degree in Chemistry at Michigan State University in 2023. He is now a Ph.D. student at Purdue University under the co-guidance of Prof. Jianing Li and Prof. V. Jo Davisson. His research focuses on drug discovery, specifically developing potent small molecules to combat highly proliferative cells.

Bo Yang received his B.S. degree in Pharmaceutical Engineering from the East China University of Science and Technology in 2021 and his M.S. degree in Pharmaceutical Science from the University of Pittsburgh in 2023. He is a Ph.D. student under the supervision of Dr. Jianing Li at Purdue University. His current research focuses on developing and applying novel algorithms to accelerate hit identification and lead optimization processes of small molecules for various targets.

Atena Farhangian was a postdoctoral researcher at Purdue University, specializing in computational drug design. She has a B.S. degree in Material Science from Shiraz University (2010) and an M.S. degree in Biomedical Engineering from the University of Tehran (2014). She also received a Ph.D. degree in Bioengineering at the University of Vermont, exploring biomaterials and computational physics through multiscale modeling of lipid bilayers. Now, she is a postdoctoral associate in the Department of Medicine, Division of Pulmonary and Critical Care Medicine at the University of Vermont.

Ke Xu received her B.S. degree in chemistry (2017) and her Ph. D. in organic chemistry (2023) at Sun Yat-Sen University. She is a joint postdoctoral scholar with Prof. Severin Schneebeli and Prof. Jianing Li at Purdue University. Her research includes the design and synthesis of sequence-defined covalent cages, small-molecule drugs, and mimicking larger peptide drugs.

Tongtong Li is a postdoctoral researcher in the Borch Department of Medicinal Chemistry and Molecular Pharmacology at Purdue University. She received her B.S. degree in Chemistry from Qingdao University in 2014, an M.S. degree in Materials Science from the University of Chinese Academic of Science in 2017, and a Ph.D. degree in Chemistry from the University of New Mexico in 2024. Her current work uses computational approaches to drug discovery, building on her doctoral research in protein-ligand interactions and protein dynamics-function relationships. She is particularly interested in designing novel small-molecule inhibitors for target proteins.

Zhong-Yin Zhang obtained his Ph.D. in Chemistry from Purdue University in 1990. He completed his postdoctoral training at the University of Michigan from 1991–1994 in the laboratory of Dr. Jack Dixon. He established his laboratory in 1994 at Albert Einstein College of Medicine and moved to Indiana University School of Medicine in 2005. In 2016, he moved back to Purdue as Distinguished Professor of Medicinal Chemistry, Robert C. and Charlotte P. Anderson Chair in Pharmacology, Head of the Borch Department of Medicinal Chemistry and Molecular Pharmacology, and Director of the Institute for Drug Discovery. His research spans the disciplines of chemistry and biology with an emphasis on the structure/function and therapeutic targeting of protein tyrosine phosphatases (PTPs). His team pioneers the development of potent, selective, and bioavailable PTP inhibitors (orthosteric, allosteric, covalent, and small molecule degraders) for functional PTP target interrogation and therapeutic translation.

Jianing Li is an Associate Professor in the Borch Department of Medicinal Chemistry and Molecular Pharmacology at Purdue University. She received her B.S. degree in Chemical Physics at the University of Science and Technology of China in 2006 and Ph.D. degree in Chemical Physics from Columbia University in 2011. After three-year postdoctoral training at the University of Chicago, she joined the University of Vermont and rose to Associate Professor in Chemistry with tenure. She joined Purdue in 2022 and is a member of the Purdue Institute for Drug Discovery. Her current research focuses on multiscale modeling and AI/ML for drug discovery, which allows accurate design of small molecules, biologics, and nanoparticles as potential therapeutics and for drug delivery. She has received several awards, including the 2019 OpenEye Junior Faculty Award in Computational Chemistry, the 2020 NSF CAREER award, and the 2023 AnalytiXIN Fellow.