New drugs introduced during the past century, such as antibacterials (penicillin, 1943) anti-inflammatories (cortisol, 1952), antipsychotics (chlorpromazine, 1953), contraceptives (norethindrone, 1960), anxiolytics (diazepam, 1963), immunosuppressant (cyclosporin A, 1983), antidepressants (fluoxetine, 1987), TNFα-inhibitors (infliximab, 1998), and PD-1–PD-L1 inhibitors (pembrolizumab, nivolumab, 2014)—all shown with their first year of US market approval, are responsible for most of the unprecedented medical progress that happened since then and have completely altered the way life is conducted in industrialized nations. However, truly effective antivirals are still lacking, as the recent coronavirus-inflicted COVID-19 pandemic made painfully clear. The search for antivirals has its own particular challenges, as viruses hijack the reproduction machinery of their host organisms, but progress in drug discovery and development as a whole has been frustratingly slow due to a variety of problems (Proudfoot, 2002; Munos, 2009; Paul et al., 2010; Scannell et al., 2012).

Despite enormous increases in research and development (R&D) investments, the number of newly introduced drugs in the United States remained stubbornly stagnant since the 1960s staying around 20–30 per year (Figure 1) and ∼85% of them represented no or only modest improvements (Wolff, 1995) demonstrating a pervasive need for innovation. This is probably best illustrated by the fact that the number of new drugs approved by the United States Food and Drug Administration (FDA) that were developed per $1 billion of R&D spending in the drug industry (inflation-adjusted) has been decreasing exponentially since 1950, being steadily halved about every 9 years (Scannell et al., 2012). This is mainly due to the highly increased regulatory burden, the unrealistic public expectation of no side effects, the need to outperform existing old drugs, and the depletion of effective new targets for traditional drug design approaches (Walters et al., 2011; Bodor and Buchwald, 2012; Scannell et al., 2012). Regarding the last, it is commonly estimated that there are only about 500 to 1,500 human protein targets that are both “druggable” and “disease modifying”, i.e., only about 2–7% of the ∼20,000 canonical (nonmodified) human proteins encoded by individual genes (Hopkins and Groom, 2002; Russ and Lampel, 2005). In general agreement with this, a survey of small-molecule drug targets counted ∼550 human proteins (plus another ∼180 non-human ones) (Santos et al., 2017). Thus, we are probably beginning to run out of traditional protein targets, at least human ones, and quite likely most low-hanging fruits among such targets that can provide therapeutic benefits have already been picked.

Protein-protein interactions (PPIs), the focus of the present review, represent possible additional, alternate targets as evidenced by the increasing number of clinically approved biologics targeting them (Figure 1). For example, one of the latest such successes was the development of cancer immunotherapies targeting immune checkpoint PPIs such as CD80–CTLA4 and PD-1–PD-L1, which has been named Science Breakthrough of the Year in 2013 (Couzin-Frankel, 2013). Unfortunately, PPIs are difficult to modulate with small molecules as the corresponding protein interfaces tend to lack well-defined ligand-binding sites where sufficiently strong interactions can take place to ensure the energy of interaction needed for high affinity binding. Nevertheless, the sheer number of such PPIs, estimated to be in the 300,000 (Zhang et al., 2012; Cheng et al., 2018) to 650,000 (Stumpf et al., 2008) range for humans, implies that a considerable number should still be druggable. Drugs need to be quite potent to be able to compete with naturally present ligands, to be sufficiently specific for their intended target, and to not need unacceptably high doses. Typically, they need to have affinities in the mid-nanomolar range. For example, the median value for all approved drugs has been estimated to be around 20 nM (Overington et al., 2006), which corresponds to a free energy of binding of ΔG ⁰ = –RT·lnK _D = –5.94·log₁₀ K _D [kJ/mol] = 45.7 kJ/mol. To achieve such high energy, small-molecule endogenous agonists and drugs of classic targets such as G-protein coupled receptors (GPCRs) typically bind at binding sites that are fully buried and allow interactions along the entire ligand surface (Figure 2A) (Buchwald, 2019). Since PPI interfaces tend to be relatively large and flat surfaces that lack such well-defined deep pockets, strong binding is difficult to achieve here with small molecules, as interactions are limited to only parts of the total ligand surface. This is illustrated in Figure 2, which compares the 3D structure of a typical fully buried small-molecule agonist at a classic GPCR target (purinergic P2Y12 receptor) with that of a surface-bound small-molecule inhibitor (SMI) of a PPI (venetoclax bound to BCL-2).

Not surprisingly, binding pockets on protein-protein interfaces that are suitable to accommodate small molecules are indeed considerably smaller than those of traditional protein-ligand interactions (Fuller et al., 2009). Typically, existing drugs target a single binding pocket with an average volume of ∼300 Å³, whereas SMIs of PPIs target multiple (3–5) smaller pockets (∼100 Å³) (Fuller et al., 2009). As the achievable maximum energy is limited by the pocket size (Buchwald, 2008), adequate binding affinity at protein interfaces can only be achieved by molecules large enough to reach a sufficient number of such smaller pockets (as illustrated in Figure 2B). The need for larger size can also be seen from the perspective of ligand efficiency, LE, defined as the binding energy per unit size–typically the binding free energy per non-hydrogen atom (N _a ), LE = ΔG ⁰ /N _a (Hopkins et al., 2004). Typical ligand-receptor protein interactions have LE of ∼1.5 kJ/atom (Hopkins et al., 2004; Hajduk, 2006; Reynolds et al., 2007; Buchwald, 2008), which corresponds to an about two-fold increase in affinity (decrease in K _D or IC₅₀) with the addition of each (non-hydrogen) atom. Such high LE is almost impossible to achieve at PPI interfaces where the bound SMI ligand can interact only along part of its surface (Figure 2B). Thus, to achieve the free energy needed for 20 nM binding (45.7 kJ/mol) with an LE of 1 kJ/atom, structures with more than 45 non-hydrogen atoms are needed, which is already larger than desired for typical “druggability”. SMIs of PPIs were indeed found to be larger than classic drugs including receptor ligands, ion channel modulators, and enzyme inhibitors (Neugebauer et al., 2007).

On the other hand, biologics, such as antibodies and fusion proteins, can interact with proteins along a broader surface and a variety of epitopes without having to rely solely on druggable pockets to achieve adequate affinity and specificity. An increasing number of biologics are being used clinically as they can be highly specific (Figure 1); however, they cannot cross cell membranes, thus cannot reach intracellular targets (Verdine and Walensky, 2007; Hughes et al., 2011; Neklesa et al., 2017), and their protein nature also causes solubility, stability, route of administration (i.e., no oral bioavailability), and biodistribution limitations. Further, since they are foreign proteins, they can act as antigens and elicit strong immune responses in some recipients (Suntharalingam et al., 2006; Wadman, 2006; Leader et al., 2008). All these problems are further exacerbated by their typically long elimination half-lives, which makes it difficult to rapidly eliminate unwanted effects when they occur (Huck et al., 2018). Not surprisingly, FDA-approved biologics encountered more post-market safety issues than did small-molecule drugs (Downing et al., 2017). SMIs of PPIs may represent viable alternatives lacking these problems, if the difficulties related to affinity/specificity can be overcome. While such SMIs were not pursued until relatively recently because they were considered unlikely to be successful due to the aforementioned challenges, during the last 2 decades, it has become clear that SMIs can be effective against at least some PPIs. Most small-molecule PPI modulators are SMIs (i.e., antagonists)—our sole focus here, as so far there are only a limited number of identified small-molecule PPI ‘agonists’ (i.e., enhancers or stabilizers) (Thiel et al., 2012; Milroy et al., 2014; Andrei et al., 2017). SMIs of PPIs, as antagonists in general, can be orthosteric, directly interfering with the interface and competing with the protein ligand, or allosteric, binding away from the interface but causing sufficient conformational change to block binding of the protein ligand.

Tens of PPI-targeting SMIs have reached preclinical development (Arkin and Wells, 2004; Wells and McClendon, 2007; Wilson, 2009; Buchwald, 2010; Arkin et al., 2014; Milroy et al., 2014; Song and Buchwald, 2015; Scott et al., 2016), and three are approved by the FDA for clinical use: lifitegrast (Gadek et al., 2002), venetoclax (Souers et al., 2013), and fostemsavir (Meanwell et al., 2018) (Figure 3). Lifitegrast (SAR 1118) is a LFA-1–ICAM-1 inhibitor developed first at Sunesis (Zhong et al., 2012) from a series originating at Genentech (Gadek et al., 2002) and then clinically by SARcode/Shire; it was approved by the FDA for the treatment of dry eye in 2016 (Xiidra) (Scott et al., 2016). Venetoclax (ABT-199) is part of a small-molecule series developed by Abbott and later AbbVie and designed to target PPIs in the B cell lymphoma 2 (BCL-2) family (Souers et al., 2013). It received FDA approval in 2015 for treatment of chronic lymphocytic leukemia (CLL), small lymphocytic lymphoma (SLL), and later acute myeloid leukemia (AML) (Venclexta, Venclyxto) (Mullard, 2016b). Fostemsavir (BMS-663068) is a water soluble prodrug of temsavir developed by Bristol-Myers Squibb that acts by blocking gp120 binding to CD4 to limit HIV attachment and entry; it was approved by the FDA for clinical use in the US in 2020 as an antiretroviral for adults living with HIV/AIDS (Rukobia) (Meanwell et al., 2018). Finally, maraviroc (Selzentry) is an antiretroviral that can be considered an allosteric SMI of the gp120–CCR5 PPI as it targets CCR5 and stabilizes a conformation no longer recognized by the HIV envelope (Melby and Westby, 2009; Tan et al., 2013). These successes, and particularly that of fostemsavir reemphasize the feasibility of SMIs of PPIs as drug discovery strategy for antivirals. Such SMIs could yield novel therapies that are not only more patient friendly than antibodies (i.e., suitable for oral or inhaled administration), but also less immunogenic, more controllable (shorter half-life/better biodistribution), and possibly even less strain- and mutation-sensitive.

Blocking of PPIs involved in the initiation of cell attachment and entry of CoVs can provide efficient antiviral therapeutics and is the main mechanism of action of biologics such as neutralizing antibodies. SMIs face more challenges to achieve this, as they do for all other PPIs; however, they could lead to new alternative antiviral agents that are suitable for oral administration and act by a different mechanism of action than existing small-molecule antivirals such as protease or viral reproduction inhibitors. Oral bioavailability is highly desirable to achieve widespread usage and compliance (Neklesa et al., 2017), and oral therapeutics are much more suitable for long-term use and/or broadly acceptable preventive use (including for transmission control of viral diseases) than any other routes of administration (Cochrane et al., 1999; Moia et al., 2013). Broadly specific activity is also of considerable interest as it could make possible mutation resistant, multi-strain, or even pan-CoV inhibition. While it is usually difficult if not impossible to achieve with antibodies that tend to be target-specific, it could be more achievable with SMIs. For example, we have shown that while the corresponding antibodies did not cross-react for the human vs. mouse CD40–CD40L PPI, our SMIs did and even maintained similar potencies (Margolles-Clark et al., 2009; Bojadzic and Buchwald, 2019). The impact of SARS-CoV-2 variants on spike and RBD structure and on nAb activity, which could also affect SMIs, has been summarized recently (Nabel et al., 2022). Computational simulations of SARS-CoV-2 spike flexibility and its interactions with other proteins are being carried out and should provide helpful tools for future screening efforts (Pedebos and Khalid, 2022).

SMIs of the SARS-CoV-2-S–hACE2 PPI identified so far and summarized above provide proof-of-principle evidence for the feasibility of such a small-molecule approach, but it remains to be seen if they can ultimately lead to clinically usable therapies as specificity and activity profiles still need improvement. While specific goals vary somewhat depending on the specifics of the project, small-molecule drug candidates are generally expected to have, among others: • potency in at least the hundred-nanomolar range (i.e., IC₅₀ < 100 nM meaning pK _i > 7) (the median of existing drugs being ∼20 nM); • adequate selectivity/specificity (>20× versus other targets is a reasonable minimum and >100× is desirable); • good safety profile (TI > 30 and optimally >100 in early studies plus passing of all toxicity studies); • adequate solubility and partition properties (needed to achieve acceptable formulation and desired delivery to the intended target); and • acceptable oral bioavailability and duration of action (somewhat flexible, but oral bioavailability F% > 30% and elimination half-life t _1/2 > 4 h are reasonable goals) (Williams, 2005; Smith and O’Donnell, 2007; Bodor and Buchwald, 2012). Some of these are undoubtedly more difficult to achieve with small molecules targeting PPIs than with those targeting classic drug targets such as GPCRs, ion channels, and enzymes that have pre-formed domains (pockets) to bind their natural ligands with good affinity and specificity. Problems related to lack of good binding pockets and thus a relatively low ligand efficiency (LE) have been reviewed briefly earlier (Small-Molecule Inhibitors of Protein-Protein Interactions; see also illustration in Figure 2). Because of this, SMIs of PPIs tend to be larger structures than classic drugs (Neugebauer et al., 2007), and it is now well-recognized that the chemical space of existing drugs and corresponding screening libraries does not correspond well with that of promising SMIs of PPIs (Pagliaro et al., 2004; Neugebauer et al., 2007; Reynès et al., 2010; Sperandio et al., 2010; Morelli et al., 2011). Fortunately, computational prescreening including exploration of relevant physicochemical properties can provide valuable information (Villoutreix et al., 2012; Trisciuzzi et al., 2019) and there are now databases, such as TIMBAL (Higueruelo et al., 2009), 2P2I (Bourgeas et al., 2010), or iPPI-DB (Labbe et al., 2016; Torchet et al., 2021), that contain an increasing number of 3D structures for protein-protein and protein-inhibitor complexes. These can make computationally enriched library selection much more successful, which has been shown to accelerate hit discovery (Milhas et al., 2016). A chemical library of >10,000 compounds dedicated to PPI inhibition has been developed (Fr-PPIChem) and is freely available upon request for experimental screening against PPIs (Bosc et al., 2020).

Larger structures, often with multiple aromatic rings, are usually better suited for effective PPI inhibition (Che et al., 2006; Fletcher and Hamilton, 2006; Hershberger et al., 2007); however, these tend to violate the widely used “rule-of-five” (Ro5) criteria, which includes MW < 500 (Lipinski et al., 1997; Lipinski, 2004) and has been widely used to guide candidate selection and ensure adequate oral bioavailability and ADME (absorption, distribution, metabolism, and excretion) profile. Nevertheless, an increasing number of new drugs have been launched lately (including venetoclax and fostemsavir discussed earlier) that significantly violate these empirical rules proving that oral bioavailability can be achieved even in the “beyond the rule-of-five” chemical space (DeGoey et al., 2017; Doak and Kihlberg, 2017). Along these lines, it is instructive to highlight that the first promising lead during the development of venetoclax (ABT-199) was ABT-737, which was so far from being suitable for formulation as a drug that one of its developers jokingly described it as having “the biophysical properties of brick dust” (Mullard, 2016b). Similarly, the incredibly tedious process of medicinal chemistry optimization that was required to make the original lead of the series that ultimately led to fostemsavir (BMS-663068) as a clinical product is nicely described in detail in (Meanwell et al., 2018).

Hits obtained so far for this PPI (Figure 5) reemphasize that our approach relying on the chemical space of organic dyes as a starting point when screening for SMIs of PPIs makes sense. For example, Evans blue which was the best hit identified from a HTS of >3,000 compounds selected after in silico prescreening of ∼60,000 structures (Day et al., 2021), came up as a hit from our screening of a much smaller library of <100 dyes (Bojadzic et al., 2021b). For obvious reasons, organic dyes have good affinity for proteins (Hunger, 2003), and they contain privileged structures for protein binding (Che et al., 2006; Fletcher and Hamilton, 2006; Hershberger et al., 2007). Thus, contrary to commonly available drug-like libraries, they are a good starting point to identify SMIs of PPIs. We have even found organic dyes that are promiscuous PPI inhibitors (Ganesan et al., 2011; Ganesan and Buchwald, 2013). Of course, organic dyes are not particularly suitable for therapeutic development because of their strong color (plus, for azo dyes, their quick metabolic degradation (Levine, 1991; Feng et al., 2012)). Nevertheless, we have shown in at least one case (the CD40–CD40L PPI, a member of the TNF superfamily) that new drug-like SMIs can be developed by first using this chemical space to identify the molecular scaffold required for activity and then removing the color-causing chromophore(s) while retaining PPI inhibitory activity (Chen et al., 2017; Bojadzic et al., 2018). Specificity can be an issue with dyes, and indeed most dyes found here as promising SMIs of the SARS-CoV-2-S–hACE PPI seem to be quite non-specific as the specificity plot shown in Figure 6 illustrates (data from (Bojadzic et al., 2021b)). Also, many azo-containing dyes are likely PAINS (pan-assay interference compounds) that can show up as false positives in screening assays (Baell and Walters, 2014; Aldrich et al., 2017); thus, they need to be treated carefully to ensure, for example, that the PPI inhibitory activity seen is not due to aggregation/polymolecular conglomeration. Nevertheless, medicinal chemistry optimization for specificity should still be feasible, and it is worth remembering that modern medicinal chemistry emerged in the early 20th century from the synthetic dye industry of the late 19th century (mostly in Germany at that time) (Paterson, 1984). Following the discovery of the first synthetic dye in 1856, Paul Ehrlich (1854–1915) (Drews, 2004; Bosch and Rosich, 2008), who acquired his medical doctor’s degree with a thesis on “the theory and practice of histological staining”, laid the foundations of chemotherapy. The analogy between the azo –N=N– bond common in many of these dyes and the arsenic bond –As=As– led to his search of arsenicals and ultimately to the discovery of arsphenamine (Salvarsan, no. 606), the first modern antimicrobial, in 1909. A few years later, the testing of thousands of azo dye related compounds and the contributions of Gerhard Domagk (1895–1964) led to the discovery of Prontosil (1932), the first effective sulfonamide antibacterial.

None of the SMIs of SARS-CoV-2-S–hACE2 identified so far and discussed here (Figure 5) have reached clinical development (Zhao et al., 2022); in fact, none seem to have even been evaluated in existing preclinical animal models for SARS-CoV-2 (Takayama, 2020; Saravanan et al., 2022). Methylene blue, a phenothiazine dye we have identified as such an SMI (Bojadzic et al., 2021a), is, in fact, included in the WHO List of Essential Medicines and is orally bioactive; thus it might have some potential for repositioning for COVID-19 prevention and treatment especially as its low micromolar anti-CoV activity, possibly due to multiple mechanisms of action, has been confirmed by several other groups (Chuang et al., 2022). Overall, results summarized here provide proof-of-principle evidence for the feasibility of such SMI approaches toward antivirals that inhibit CoV attachment and entry, and they serve as a first guide of the chemical space needed to achieve this.