Allosteric regulation, a fundamental mechanism controlling biomolecular activity, occurs when an effector binds to a site distant from the active site, thereby modulating protein function, structure, or flexibility.(p1),(p2),(p3) This mechanism has garnered significant scientific attention in drug discovery, particularly for traditionally ‘undruggable’ targets. Allosteric drugs exhibit several advantages over their orthosteric counterparts, including enhanced selectivity, decreased toxicity, and the ability to modulate protein activity without competing with endogenous ligands.(p4),(p5) Orthosteric ligands typically bind directly to a the active site of a protein, blocking its interactions, whereas allosteric ligands target secondary regulatory sites, enabling fine-tuned control over protein function. This increased selectivity of allosteric ligands can be attributed to the observation that allosteric sites are often less conserved than orthosteric sites, allowing ligands to specially target certain isoforms or conformations of a protein while sparing related proteins, thereby minimizing off-target effects. Furthermore, allosteric ligands preserve baseline signaling and reduce the risk of toxicity associated with complete inhibition or overactivation by modulating protein activity indirectly (e.g. stabilizing active or inactive states). A prominent class of drug targets where this approach is successfully employed is G-protein-coupled receptors (GPCRs).(p6),(p7) Here, positive allosteric modulators (PAMs) amplify receptor activation by enhancing the affinity or efficacy of an orthosteric agonist, negative allosteric modulators (NAMs) suppress pathological overactivity, and neutral allosteric ligands subtly modulate endogenous signaling, collectively enabling fine-tuned therapeutic effects with reduced toxicity (Figure 1a). Similarly, for ion channels, which regulate the ion flow across cell membranes to control processes like neuronal signaling and muscle contraction, allosteric modulators offer precise control over channel activity (Figure 1b). In protein kinases, which are critical in cellular signaling pathways governing processes like proliferation and survival, allosteric inhibitors or activators bind sites distinct from the ATP-binding pocket, offering selectivity over orthosteric inhibitors (Figure 1c). By stabilizing specific kinase conformations, allosteric modulators can fine-tune signaling cascades, avoiding broad suppression or overstimulation that could lead to off-target effects or resistance, as evidenced in cancer therapies.

Despite these advantages, significant challenges remain in identifying allosteric sites because of factors such as limited evolutionary conservation, target conformational flexibility, and transient pockets (i.e. temporary binding sites formed by dynamic conformational changes). Proteins are dynamic entities that transition between multiple conformational states, meaning that allosteric sites might only emerge in specific conformations. This presents considerable difficulties for detecting allosteric sites using static experimental methods like X-ray crystallography and cryogenic electron microscopy (cryo-EM). Additionally, transient pockets formed by dynamic conformational changes further hinder traditional screening methods designed for stable binding pockets. Moreover, although orthosteric sites are frequently conserved across protein families, allosteric regions differ significantly even among structurally similar proteins,(p8),(p9) limiting the application of similarity-based predictions. These challenges render the discovery of allosteric sites resource-intensive, thereby impeding rational allosteric drug design. Despite these limitations, experimental structural biology remains essential for generating high-quality structural data. Moreover, these techniques are the gold standard for experimental validation of computationally predicted allosteric sites, though such validation is a significant challenge because of the transient nature of these pockets and the need for specialized assays.

Although traditional experimental methods face substantial limitations in identifying cryptic allosteric sites (i.e. hidden regulatory binding sites not apparent in unbound protein structures), the field of allosteric drug design is entering a transformative era driven by advancements in ML, MD simulations, and network-based approaches. The remarkable success of AlphaFold2(p10) in predicting protein structures with high accuracy through deep learning (a specialized subset of ML that uses neural networks to model complex patterns) has spurred growing interest in leveraging its capabilities to accelerate drug discovery. Beyond its direct utility in structure prediction, the achievements of AlphaFold2 have also stimulated broader discussions regarding the integration of biophysical principles with deep learning(p11) and its potential applications in exploring allosteric mechanisms or streamlining allosteric drug design.(p12) Meanwhile, MD simulations now capture drug binding to proteins undergoing large conformational changes at the microsecond scale, revealing transient pockets accompanied by conformational shifts that lead to post-drug binding.(p13) In contrast, network-based analyses map communication pathways within proteins, pinpointing residues critical for allosteric signaling and contributing to locating allosteric sites.(p14) Collectively, these developments hold great potential for identifying allosteric sites and designing modulators with tailored effects, heralding a future where allosteric drugs become mainstream therapies for diseases previously deemed undruggable. For example, peripherally restricted cannabinoid receptor (CB1) agonists targeting cryptic allosteric sites demonstrate significant promise for chronic pain, a previously challenging therapeutic target.(p15)

In this review, we focus on the recent advances in computational strategies for allosteric site prediction, including ML, MD, and network-based approaches. We also discuss their respective limitations and highlight the different synergistic combinations of computational methods. Finally, we address current challenges in the computational prediction of allosteric sites and provide future perspectives.