Development of a random background to understand ligand optimization

Ligand optimization is central to chemical probe and drug discovery, as initial active molecules rarely have the potency or pharmacokinetic (PK) properties to be viable in vivo^1,2. Between the discovery of an initial hit and a clinical candidate many hundreds, occasionally thousands of optimized analogues might be synthesized^1,5. Several strategies^6,7,8,9,10 have been developed to improve this process, ranging from early empirical approaches such as Topliss Trees¹¹ and quantitative structure–activity relationships¹² to contemporary artificial intelligence (AI)-guided PK¹³ prediction and free energy calculations^14,15. The efficiency of these strategies remains uncertain as there is no random background against which to compare them. Using tenfold improvement between analogue and parent as a benchmark for substantial impact, 10% of analogues meet this standard in the ChEMBL database¹⁶ (Extended Data Fig. 1a). As these ChEMBL results suffer from success bias, sample multiple types of perturbation and suffer from inter-group irreproducibility¹⁷, the public domain offers no sure guidance on what level of improvement one might expect in ligand optimization from random conservative changes.

Random backgrounds have long been used in biology to quantify significance. In genomics, they distinguish between artificial and natural selection¹⁸. In epidemiology, random incidence distinguishes genetic diseases from those driven by environmental factors^19,20. In protein engineering, random backgrounds help to evaluate improvements in enzyme function²¹, and alanine scanning introduces a minimal perturbation to find hot spots for binding and function²². A random background for ligand optimization might quantify progress across chemotypes and targets and compare different optimization strategies.

Like alanine scanning, an unbiased background for ligand optimization might involve conservative perturbations, systematically and comprehensively applied across multiple parents with different targets. This resembles a ‘positional analogue scanning’ approach that substitutes ligand non-polar hydrogens with groups such as CH₃, OH, Cl, F and Br, and aromatic carbon atoms with nitrogen²³. In a retrospective analysis of 110,000 matched molecular pairs in the ChEMBL24 database^3,4, about 30% of analogues improved over threefold in affinity or potency versus their parent. This observation is tempered by design and reporting biases towards compound improvement, by the variability in experimental conditions across datasets that can reduce reproducibility¹⁷, and by the few parent compounds that had multiple conservative changes. Thus, it seemed interesting to create a set of parent ligands that were systematically and comprehensively modified by small perturbations, without design and without a bias towards improvement, using the positional analogue scanning approach^23,24 where the differences between parent and the suite of analogues were quantified by a single laboratory.

To create what we refer to as a random background set for ligand optimization, we chose 18 parent ligands spanning six targets, including three GPCRs (α₂A adrenergic receptor (α₂AAR), μ-opioid receptor (μOR) and cannabinoid receptor type 2 (CB2)), one transporter (the serotonin transporter (SERT)) and two soluble enzymes (AmpC β-lactamase (AmpC) and macrodomain 1 of SARS-CoV-2 (Mac1)). This set was chosen for targets that we had under experimental control and is admittedly far from comprehensive. The 18 parents were chosen for similar pragmatic reasons; they were molecules to which we had ready access. Still, these proteins provide a range of target classes while the 18 parents cover a wide range of physicochemical properties (Extended Data Fig. 2): ranging from 1 nM to 43 µM activity (Table 1) and representing 11 unrelated scaffolds. For each parent, we systematically explored all analogues accessible by a single-atom or methyl modification where such modifications were synthetically feasible and where they did not modify the net charge of the parent at physiological pH. The single-atom modifications involved substituting carbon hydrogens with one of methyl²⁵, hydroxyl²⁶, chloro²⁷, fluoro²⁸ or rarely bromo, or substituting aromatic carbon atoms with nitrogens (Fig. 1). These modifications are not comprehensive but do represent those preferred by medicinal chemists.

These substitutions were systematically and comprehensively applied without design, ensuring an approach unbiased by anything other than synthetic accessibility and price. Overall, 257 single-atom analogues were synthesized and tested for changes in activity on the target, allowing us to measure how often we might expect affinity or potency to improve tenfold or more by minimal perturbation. This begins to provide a background expectation for how often a designed analogue might meaningfully improve affinity or potency. As we show, these undesigned changes had an unexpectedly high success rate, suggesting a systematic strategy for optimization, even though that was not our goal. As affinity or potency maturation is only one criterion for advancing from hit to candidate, we also measured in vitro PK properties for the analogues, including permeability, metabolic stability, plasma protein binding and three other terms, versus the parent compounds. These studies explore trade-offs between PK and target affinity or potency in ligand optimization.

The 257 analogues were tested for activity changes (dissociation constant (K_d), half-maximal inhibitory concentration (IC₅₀) or half-maximal effective concentration (EC₅₀)) versus the parents across the six targets. For μOR, EC₅₀ values were determined in cAMP Glo-sensor assays. α₂A and CB2 ligands were assessed by cAMP assays and radioligand competition binding, while SERT activity was assessed using a transporter uptake assay. AmpC inhibition was measured enzymatically and that of Mac1 was measured by ligand displacement. For every parent, we have only reported changes for one type of activity. For instance, if a parent inhibitory constant (K_i) is reported, then only K_i values are reported for its analogues, and if a parent is an agonist, only EC₅₀ values are reported for its analogues. Each functional or inhibition assay included a literature positive-control molecule (Supplementary Fig. 2) with representative concentration–response curves and Z values for each assay (Supplementary Fig. 3). The effects of receptor expression on agonism were controlled for (Supplementary Figs. 4 and 5). Full concentration–response curves and analyses are shown for every parent–analogue pair, allowing for direct inspection.

Overall, 23% of the analogues affected activity by twofold or less (fold changes rounded to the nearest integer; Supplementary Table 1) and another 30% reduced activity by tenfold or more (Extended Data Fig. 1b). Intriguingly, 29 of the 257 analogues (11.3%) improved activity by tenfold or more (Table 1). We note that four of these 29 had fold changes of between 9.7 and 9.9; we consider these effectively tenfold improvements and used the same convention in analysing the public ChEMBL data (see Table 1 and Supplementary Table 1 for details). Such tenfold or more improvement analogues were found for five of the six targets, and for 10 of the 18 parents (Table 1). The effects and chemical identities of all analogues versus their parents are listed in Supplementary Table 1.

We were interested in which perturbations had the biggest effects, how the rates of affinity or potency improvements compared with the literature, with its admitted biases, and if physical properties of the parents or of the binding sites correlated with greater likelihoods of improvement. Among the analogues, methyl substitutions were the most likely to improve activity tenfold or more, with chlorine substitutions the second most likely; for threefold or more improvements, of which there were 69 among the 257 analogues, both chlorine and methyl substitutions were the most common (Fig. 2a). Given the similar size and similar hydrophobicity of methyls and chlorines, their similar effects seem sensible. Conversely, fluorine substitutions, which are often used to increase metabolic stability, yielded only two tenfold or more and five threefold or more activity improvements after rounding fold changes to the nearest integer. Aromatic carbon-to-nitrogen substitutions also rarely improved affinity or potency, although they often had the greatest effects on PK properties.

Because hydrophobicity increases will increase affinity simply by disfavouring the unbound state, and will have physical property liabilities, it is useful to consider the improved affinity or potency of the analogues in light of ligand efficiency and lipophilic efficiency (LipE). If affinity or potency improvements were largely driven by hydrophobicity, we would expect ligand efficiency to track with hydrophobicity (here, the calculated octanol–water partition coefficient (clog P)) and LipE to remain constant or deteriorate. Instead, analogues that improved tenfold or more had ligand efficiencies that increased much more than hydrophobicity, and had substantially increased LipEs. For instance, an analogue with an added methyl, chloro or other atom that improved in binding or in EC₅₀ by even threefold had a ligand efficiency of 0.65 kcal per atom for the added atom, well above the 0.3 kcal per atom thought to be useful for drug optimization^29,30, whereas for analogues that improved activity tenfold or more, the single-atom ligand efficiency was 1.4 or more, close to the efficiency limit³¹. Similarly, every analogue that improved activity by more than tenfold also improved LipE (Extended Data Fig. 3f). More broadly, among the 220 analogue–parent pairs with complete numerical activity data, ΔclogP showed little quantitative correlation with ΔpActivity (Extended Data Fig. 3a).

We investigated whether there were correlations between likelihood of potency improvement and ligand and binding-site properties. For instance, one might expect that weaker ligands might be easier to optimize, or that smaller binding sites might be better suited to big affinity jumps. We calculated whether the parent molecular weight, binding-site volume, the ligandability of the binding sites (calculated using D-score³²) or the affinity or potency of the starting parent compound was correlated with the likelihood of finding analogues with substantial affinity or potency improvement. Using the fraction of analogues achieving tenfold or more improvement as the outcome, none of these properties was meaningfully correlated (Fig. 2b). If one considers the magnitude of the affinity or potency change overall, and not just whether it improved tenfold or more, correlations did emerge, with weaker versus stronger parents more likely to support improved affinity (R = −0.34, P = 3.4 × 10⁻⁷) and with larger binding sites also tending to do so (R = 0.24, P = 4.0 × 10⁻⁴; Supplementary Fig. 6). Although these trends were echoed among matched pairs in ChEMBL (Fig. 2c and Extended Data Fig. 1), the correlations overall were weak. Together, these results suggest that: (1) as a strategy, systematic small perturbations can be unexpectedly successful at improving ligand affinity or potency^23,24, perhaps even compared with design-heavy approaches³³; (2) this effect is common across diverse parents and binding sites; (3) it can find echoes in the literature; and (4) random expectations for substantial affinity or potency improvement by small perturbation may be as high as 11% (95% CI 7.5–15.2%).

To investigate the structural bases for the affinity or potency changes, we determined the structure of the parent inhibitors in complex with Mac1 and six analogues from the ‘3122 series, 12 from the ‘3453 series and four from the ‘3794 series, each to between 0.97 and 1.02 Å resolution (Extended Data Tables 1–4). The analogues with structures determined represented all six types of single-atom modifications in this study; 15 of the analogues differed in affinity or potency from their parents by threefold or more, two by tenfold or more and among themselves by up to 680-fold.

For some parents, analogues superposed well, facilitating analysis. For the Mac1 ‘3453 (IC₅₀ of 8.9 µM) inhibitor series (Extended Data Fig. 4a), the introduced chloro at the ortho-position of the phenyl ring (‘0676, IC₅₀ of 21 µM) projects into solvent and loses activity versus the parent (Extended Data Fig. 4b). Moving this chloro to the meta-position (‘1304), where it packs with Phe132, Ile131 and Gly48, improves inhibition 11-fold (P < 0.01) to 1.9 µM (Extended Data Fig. 4c). Moving it one atom further over, to the para-position, had little effect (‘6343, IC₅₀ of 2.1 µM), although a methyl at the same position (‘9870, IC₅₀ of 7.6 µM) loses fourfold affinity (P < 0.01), presumably reflecting its poorer packing (Extended Data Fig. 4d).

Many of the effects of the small perturbation analogues could be explained post hoc from their structures; however, fewer were easily anticipated. Even the apparent simplicity of the ‘3453 series is belied by structural accommodations in the enzyme. For instance, the para-chloro of the 2 µM ‘6343 would have clashed with the enzyme without Phe132 and Ile131 rotating away. In doing so, Phe132 adopts a partially eclipsed conformation. Presumably, the resulting strain is overcome by improved ligand packing, but this balance of terms is non-obvious (encouragingly, free energy calculations did correctly predict that the para-chloro and meta-chloro would improve affinity, whereas the ortho-chloro would lose it, although the agreement was more qualitative than quantitative; see below). Conversely, the methyl analogue (‘9249, 53 µM) of ‘0676 (IC₅₀ of 22 µM) is buried from solvent and might be expected to be more potent than its chloro cousin, but instead it loses another 2.5-fold in K_i (P < 0.01; here, the opposite effect was anticipated by the free energy calculation). Meanwhile, in the complex between ‘4905 and the α₂AAR, an added methyl group fits into a pre-existing sub-pocket and qualitatively its improved potency is reasonable (Fig. 5a). Quantitatively, its 52-fold effect on EC₅₀ is at the outer edge of what might be expected, especially as the methyl buries a serine hydroxyl. This is echoed by the free energy calculation, which here too correctly predicted improved activity but by twofold not 52-fold. These are examples of complexes where the analogues superpose on the parent and each other. In complexes in which the single-atom changes led to substantial inhibitor movement or to the adoption of multiple ligand conformations, prediction was harder still (Extended Data Fig. 4f,g). For example, the Mac1 methyl analogue ‘3176 repositions the inhibitors in the pocket (Extended Data Fig. 4g), whereas compounds ‘9249 and ‘3194 bind in two distinct conformations (Extended Data Fig. 4g), a phenomenon that is probably underreported in the Protein Data Bank³⁴ but is readily seen at the ultra-high resolution of these structures. Together, this set of perturbations may provide an interesting and sometimes challenging test for prediction methods.

In optimizing chemical probes and leads, compound PK are as important as molecular target affinity or potency^35,36. We thus explored how the small-perturbation analogues affected PK properties versus those of the parent ligands, and how these changes related to alterations of affinity or potency. We measured the following in vitro PK properties: metabolic stability in microsomes (half-life), plasma stability (half-life), plasma protein binding (fraction unbound), solubility (µM), hERG inhibition (IC₅₀, µM) and membrane permeability (parallel artificial membrane permeability assay (PAMPA); cm s⁻¹), for both the parent ligands and their small-change analogues (Fig. 3 and Supplementary Tables 2–7). These admittedly in vitro measurements are considered predictive of in vivo PK^37,38,39. Some of the small-change analogues showed meaningful improvements relative to the parent ligands for each of these properties (Supplementary Table 8). For example, 12.4% of the analogues had threefold or more improvement in microsomal stability, 20.8% improved threefold or more in plasma stability, 13.8% improved solubility by threefold or more, 6.2% improved threefold or more in permeability, and 13.2% of the analogues improved threefold or more in plasma fraction unbound. Thus, each property can be meaningfully improved by simple changes. Correspondingly, a substantial fraction of the analogues experienced threefold or more decreases in these same properties (Extended Data Fig. 5). Considering what are perhaps the three most impactful in vitro PK measurements, microsomal stability, fraction unbound and permeability, 41.4% of the analogues experienced more than threefold losses in at least one of these, almost all (93.1%) experienced at least some deterioration by the same criterion, and 34.5% deteriorated in all three in vitro PK properties.

Although many of the small-perturbation analogues substantially improved individual PK properties, none of the 29 analogues classified as tenfold or more improved after integer rounding showed improvements across all three key properties: metabolic stability, PAMPA permeability and plasma fraction unbound (Fig. 3a and Supplementary Fig. 7); many were worse in all three properties. Indeed, the changes in in vitro PK were, at best, orthogonal to affinity or potency fold change (Fig. 3b), and most were, if anything, anti-correlated with it, including plasma fraction unbound, microsomal stability, hERG inhibition and, to a minor extent, solubility. This is borne out at the level of individual substitutions: the substituent groups that most often improved affinity or potency substantially, such as methyl and chloro groups (Fig. 2a), were least associated with improved PK properties (Fig. 3c and Supplementary Fig. 8). Meanwhile, those changes least likely to improve affinity or potency substantially, such as aromatic C→N, were most likely to improve microsomal stability or fraction unbound (Supplementary Fig. 8). More broadly, most of the PK properties were largely independent of one another, with no correlations exceeding |ρ | > 0.5 and those that were statistically significant having Spearman ρ values ranging only from |0.16| to |0.38|. What emerges is a frustrated landscape for ligand optimization, with affinity or potency improvements counterbalanced by worsening PK. Although this trade-off has been well noted in medicinal chemistry^40,41, this set can quantify its systematic impact.

The current benchmark only comprises several hundred analogues. If we could predict the results of the small perturbations, rather than having to experimentally test them, it could be much extended. We therefore asked how well modern methods could predict the affinity or potency and PK changes that we observed. In blinded experiments, we used free energy perturbation, arguably the highest level of theory available to the field, with the program FEP+ (Schrödinger) to predict the changes in activity between the parents and analogues (Fig. 4a–c). Similarly, we asked how well we could predict in vitro PK changes between the parents and analogues, using widely accessible tools (Fig. 4d).

There was a strong overall correlation between the FEP+-predicted and experimentally measured binding free energies (ΔG), with 188 of the 219 predictions falling within ±2 kcal mol⁻¹ of the experimental values, and 129 within ±1 kcal mol⁻¹, whereas the set had a mean unsigned error for ΔG of 1.06 kcal mol⁻¹ (Fig. 4a,b and Supplementary Table 9). Of the 18 parents, the FEP+ predictions on 14 had ΔG mean unsigned error below 1.2 kcal mol⁻¹ (Supplementary Table 10). If one considers the correlations qualitatively, of the 35 analogues predicted to lose affinity by 1 kcal mol⁻¹ or more by FEP+ (that is, in the range of confident prediction), 20 did so experimentally. Meanwhile, of the 34 predicted to increase in affinity or potency by more than 1 kcal mol⁻¹ by FEP+, 14 did so experimentally, whereas 12 were experimentally neutral and eight showed decreased potency. Categorization with a 1 kcal mol⁻¹ threshold revealed a significant association between the direction and strength of FEP+ predictions and experimentally observed affinity or potency changes (χ² = 36.4 and P < 0.001; Supplementary Table 11).

Several challenges faced FEP predictions here, which future work may share. The predictive power was probably weakened by our inability to measure activity for weak analogues, reducing activity range and making it impossible to measure affinities predicted to be very low by FEP+. We used an automated workflow to generate initial poses for FEP; on inspection, most of the outliers reflected errors in these poses. FEP accuracy is impacted by the quality of the ligand-complexed structures⁴², and the predictions here began with docked poses, which are less reliable than experimental poses. For the 25 analogue structures that we determined by crystallography (for Mac1) and by cryo-electron microscopy (cryo-EM; for an α₂A analogue, below), where predicted and experimental poses closely superimposed, FEP+ energies were largely consistent with experiment (Fig. 4c).

To predict in vitro PK^43,44, we used ADMET-AI, ADMETlab 3.0 and Deep-PK^45,46,47. The predictions of all three correlated with the experimental measurements of plasma fraction unbound and with permeability, whereas microsomal stability, plasma stability and solubility, which are difficult to predict^48,49, were essentially uncorrelated. Correlations in the 0.4–0.5 range for fraction unbound and permeability may often be strong enough to guide compound selection and these open-access tools may have broad impact. Still, even with a correlation of 0.53 in fraction unbound, the highest that we observed, 28 analogues that had substantially increased fraction unbound experimentally were predicted to have lower levels than their parents, and 16 analogues that had higher protein binding than their parents experimentally were predicted to have less. Together, although the FEP and in vitro PK predictions can profitably model the effects of the small-perturbation analogues, their correlations with experiment are not yet strong enough to allow this benchmark to be extended by calculation alone.

Notwithstanding the frustrated landscape for ligand affinity or potency and PK improvement, medicinal chemists can often navigate this multi-parameter landscape to improve the properties of lead molecules. Here too, there were analogues that improved in affinity or potency with only a modest decrease in PK. Among these was an analogue of an α₂A parent, compound ‘4905, where a single methyl group substitution improved potency 52-fold (Fig. 5a,b). In a cryo-EM structure of the ‘4905–α₂A–G protein complex determined to 2.8 Å resolution (Extended Data Table 5), ‘4905 superposed with the docking prediction for the parent ‘3629 and with the experimental structure of a previously described member of this family, ‘9087. This suggests that the methyl perturbation did not change the orientation of the analogue and that this family may be understood in the same structural context. A slight upwards deflection at high concentrations of ‘4905 in the cAMP dose–response curves may reflect secondary G_s coupling, supported by pertussis toxin-treated cAMP assays showing a residual, concentration-dependent increase in cAMP for ‘4905 but not ‘3629 (Supplementary Fig. 9a). Structurally, the added methyl group fills a hydrophobic sub-pocket (Fig. 5a and Supplementary Fig. 9b), helping to explain its improved potency.

Although the improved potency of ‘4905 came at a cost to its in vitro PK, the effects were modest: compared with its parent, ‘3629, the fraction unbound decreased by 57% and the microsomal half-life was reduced by 41% (Supplementary Tables 2–7). The in vivo PK of ‘4905 qualitatively reflected the in vitro results, with cerebrospinal fluid concentrations—a proxy for fraction unbound in the brain⁵⁰—reduced by 36%, whereas half-life was reduced from 196 to 15.2 min (Supplementary Table 12). Accordingly, we compared the in vivo analgesia of ‘4905 to that of its parent ‘3629 and to PS75, the most potent analgesic known for this family of α₂A agonists⁵¹. Despite the deterioration in PK, in mouse nociception assays including tail flick, neuropathic pain (SNI) and hotplate, ‘4905 was far more potent than the parent ‘3629 and at least as potent as PS75 (Fig. 5c–g). If ligand optimization is a frustrated landscape, it remains a navigable one⁴¹. In addition to beginning to provide a random background for such efforts, this study supports a strategy by which that landscape may be reconnoitered^23,24.

Although ligand optimization is at the heart of drug discovery, quantifying the success of optimization strategies has been difficult owing to a lack of a random background. Four key observations emerge from this study. First, systematic and unbiased small perturbations to 18 scaffolds revealed that 11.3% of the analogues improved tenfold or more in affinity or potency versus the parent (Table 1); ‘magic methyls’^25,52 were common in this set. These improvements were observed in five of six targets, across ligand physical properties and across five orders of magnitude of parent activity. This begins to establish a background expectation for affinity or potency optimization against which methods and designs may be compared. It may also suggest a simple, unbiased approach for achieving substantial jumps in affinity or potency, improvements that will not always be obvious from structural analyses or simulations (Figs. 4 and 5 and Extended Data Fig. 4) but may complement them. Second, improved affinity or potency typically came at the cost of ligand PK, with plasma fraction unbound, stability to liver microsomes and permeability, among other terms, declining as analogue affinity or potency increased. The individual in vitro PK terms were orthogonal to each other and several were anti-correlated with affinity or potency improvement (Fig. 3). This suggests a frustrated landscape for ligand optimization. Although the qualitative challenges and trade-offs in ligand optimization are known to medicinal chemists^8,40,41, this study has begun to quantify them across a range of target classes and ligand properties. Third, computational prediction of affinity or potency and PK had some success anticipating the trends that we observed (Fig. 4). Still, their correlations with experiment remained loose enough to preclude replacing systematic exploration and testing of molecules as this background set is expanded, something seen with other potency prediction methods⁵³. Fourth, although potency trades off against PK, it is possible to find analogues where it rises sufficiently to overcome drops in exposure and half-life. Thus, the α₂A analogue ‘4905 is the most potent analgesic in a large series, and is far more potent in vivo than its parent, despite experiencing these trade-offs.

At first glance, this study may recall classical medicinal chemistry where conservative and ligand-based approaches such as Topliss Trees¹¹ dominated the field. However, this would be to misremember the classic age of medicinal chemistry on two counts. First, the systematic small perturbations explored here were rarely used then, partly because the building blocks and reactions to support them were lacking. The ability to systematically explore small perturbations across scaffolds reflects advances in synthetic chemistry and the new building-block approaches that have expanded readily available compounds^54,55. Second, the logic behind Topliss Trees and other quantitative structure–activity approaches^11,12 was anchored in ligand physical chemistry without reference to receptor structure. For instance, in Topliss Trees the success of methyl and chloro derivatives taught opposite lessons because of their different electronic effects on the ligand, whereas the steric and non-polar properties of the two groups will seem similar from the view ‘inside the receptor’^52,56. Meanwhile, the role of a random background would have seemed foreign. Thus, although a strategy of systematic small perturbations may smack of the ‘methyl, ethyl, propyl, butyl, futile’ approach with which classic medicinal chemistry has been tarred^57,58, it has a different basis in theory and reflects new synthetic opportunities.

Certain limitations of this study merit airing. At 257 analogues and six targets, the set remains small and inevitably, if unintentionally, biased. Expanding to additional targets and ligand classes might reveal different trends. We note that the six targets do recognize a range of chemotypes, including large polar co-factors (Mac1), large hydrophobic lipids (CB2), anionic β-lactams (AmpC), cationic neurotransmitters (α₂A and SERT) and peptides (µOR). Experimentally, we assessed activity changes using a transporter uptake assay (SERT), radioligand competition binding assays (α₂A and CB2), second messenger assays (α₂A, CB2 and MOR), and binding and enzyme activity (Mac1 and AmpC) assays. Although ligand competition correlates with K_i, agonist EC₅₀ values are affected by receptor expression. Because we compared the relative activities of analogues and parents, controlled for receptor expression (Supplementary Fig. 4), and measured effects at low expression levels (Supplementary Fig. 5), the effects of receptor expression on relative activity should be modest, although they cannot be completely discounted. For these and related reasons, changes of EC₅₀ between agonist parent and analogues cannot be read as changes in affinity the way that changes in K_i can be, although EC₅₀ changes remain the relevant metric for agonist activity. Overall, the impact of these effects may be inspected in full concentration–response curves for parent–ligand pairs with more than threefold activity improvement (Supplementary Fig. 2). In the area of in vitro PK, our use of liver microsomes rather than hepatocytes to measure metabolic stability meant that some types of metabolism were missed, including glucuronidation. This might make groups such as phenolic hydroxyls seem less labile than they would be in vivo. More broadly, for all but two molecules, we only measured in vitro not in vivo PK. Whereas the latter are widely used in ligand optimization, their prediction of in vivo behaviour is only approximate.

These limitations should not obscure the main observations of this study. Overall, 11.3% of systematic and unbiased small perturbations improved analogue affinity or potency tenfold or more, beginning to establish a background expectation for the likelihood of substantial ligand affinity or potency improvement and a systematic approach to doing so^23,24. Balancing this was a concomitant deterioration in ligand PK, which will lower the exposure and half-life of a ligand in vivo, counteracting the improvements in affinity or potency. Navigating this multi-parameter space is at the heart of medicinal chemistry; this study supports the development of quantitative models to do so.

ChEMBL database molecular pairs

From the PostgreSQL version of CHEMBL34 (ref. ⁴) (https://doi.org/10.6019/CHEMBL.database.34), we filtered the database for compounds with activities reported against a single protein, with either K_i, K_d, IC₅₀ or EC₅₀ as the activity type. To be considered a molecular pair, compounds had to have the same assay ID, the same target ID, the same activity type and the same reference document (publication) ID, in addition to having a parent–analogue relationship as defined in this work, meaning a C–H group replaced by C–OH, C–F, C–Cl, C–Br, C–CH₃ or an aromatic carbon replaced by a nitrogen. This left us with 191,732 parent–analogue pairs.

Analogue enumeration and synthesis

Eighteen parent compounds were selected from previously published literature^51,59,60 or datasets (https://asapdiscovery.org/outputs/molecules/#ASAP-SARS-COV-2-NSP3-MAC1) based on their known binding activity, structural relevance or representation of diverse chemical scaffolds. Starting from a parent compound, we used RDKit (http://www.rdkit.org/) to identify all C–H bonds and iteratively replaced the hydrogen atom with a methyl, hydroxyl, fluoro, chloro or bromo group. We also identified aromatic carbons with two heavy-atom neighbours and replaced them with nitrogen. Every analogue generated was represented as a canonical isomeric SMILES string and added to a set to remove duplicates. For each of the 18 parents, the full set of possible analogues that could be synthesized for less than US$400 for 10 mg was ordered.

Confidence intervals and statistics

When reported, 95% confidence intervals were derived from bootstrap resampling with 10,000 iterations. In cases in which the observed frequency of success is exactly 0, bootstrap will fail to give an upper bound for the interval. In such cases, we used the Clopper–Pearson method to estimate an upper bound based on the sample size⁶¹. Pearson and Spearman correlations with their associated P values were computed using the scipy.stats module from the SciPy package⁶².

FEP simulations

These were conducted using FEP+ within the Schrödinger software suite (v2025-2) with the OPLS4 force field⁶³ and the modified simple point charge water model. The default setting was used for the number of lambda windows selection where it depends on the type of perturbations; charge-changing, core hopping and all other perturbations have 24, 16, and 12 lambda windows, respectively. For alchemical transformations with charge changes, the total charge of the simulation box was kept constant by transmuting a Na⁺ or Cl⁻ ion to water or vice versa. In addition, a 0.15 M concentration of NaCl was added to the simulation box of charge-changing perturbations. For α₂A, CB2 and SERT, the FEP+ membrane protocol was applied where a POPC membrane was added to the system in simulation. All other settings were kept default except that the simulation time was extended from 5 ns to 10 ns.

The default FEP map generation protocol was used with the parent compound selected as the biased node. In preparing proteins and ligands for FEP+, the Schrödinger protein preparation workflow and LigPrep were used. The initial binding poses of parent compounds were from poses generated by DOCK3.8, and analogues were aligned to the parents with severe steric clashes resolved using the FEP+ Pose Builder workflow.

Docking

Whereas no structural information was used in the design of analogues from parent compounds, for FEP+ calculations and for post hoc structural analysis, we generated ligand-bound complexes of the parents and relevant ligands using DOCK.3.8 (refs. ^54,64). Ligands were docked into the receptor-binding site using grids prepared in previous studies^{51,59,60,65,66}.

Assay selection

We used one consistent readout per parent series (no mixing within a series). SERT and AmpC are reported as K_i; Mac1 is reported as IC₅₀ from peptide displacement (a scalable functional hydrolysis assay is unavailable); GPCR agonist series are reported as EC₅₀ and antagonist series as K_i, such that μOR is EC₅₀ only, whereas α₂A and CB2 include EC₅₀ and K_i depending on the parent series. Representative concentration–response curves and the corresponding Z values for each assay are provided in Supplementary Fig. 3.

Transporter assays for SERT K _i

SERT activity was measured using the Neurotransmitter Transporter Uptake Assay Kit from Molecular Devices (#R8174), following the manufacturer’s protocol with slight modifications as previously described⁶⁵. HEK293T cells (ATCC CRL-3216) stably expressing human SERT were plated in poly-L-lysine-coated 384-well black, clear-bottom plates at a density of 15,000 cells in 40 µl per well, using DMEM supplemented with 1% dialysed FBS. Cells were incubated overnight at 37 °C with 5% CO₂ to allow adherence and recovery. The following day, the medium was removed by flicking, and cells were incubated with 25 µl per well of test compound solutions prepared in assay buffer (1× HBSS, 20 mM HEPES, pH 7.4, supplemented with 1 mg ml⁻¹ BSA) for 30 min at 37 °C. After drug treatment, 25 µl per well of dye solution (as provided in the kit) was added directly to the wells, followed by an additional 30-min incubation at 37 °C. Fluoxetine (10 µM) was used as a positive control for SERT inhibition. Fluorescence was measured using the FlexStation II microplate reader with excitation at 440 nm and emission at 520 nm. Relative fluorescence units were exported and analysed using GraphPad Prism 10.0 to calculate IC₅₀ values, from which K_i values were derived using the Cheng–Prusoff equation.

CB2 radioligand-binding assay

CB2 receptor-binding assays were performed using membrane preparations from HEK293 cells (ATCC CRL-1573) stably expressing human CB2, following previously published methods^67,68. Membranes were resuspended in TME buffer containing 0.1% BSA (w/v) and 25 µg of membrane protein was added per well. The assay was conducted using the radioligand [³H]CP-55,940 at a final concentration of 0.75 nM, prepared in assay buffer. Nonspecific binding was defined in the presence of 5 µM unlabelled CP-55,940. Test compounds were applied at increasing concentrations to assess competition. Reactions were incubated at 30 °C for 1 h with gentle shaking. After incubation, samples were transferred to Unifilter GF/B 96-well filter plates and filtered using a Packard Filtermate-196 cell harvester (PerkinElmer). Plates were washed four times with ice-cold wash buffer (50 mM Tris-HCl, 5 mM MgCl₂ and 0.5% BSA, pH 7.4). Radioactivity bound to the filters was quantified via liquid scintillation counting. Specific binding was calculated by subtracting nonspecific binding from total binding. IC₅₀ and K_i values were calculated using nonlinear regression in GraphPad Prism 9 using the Cheng–Prusoff equation.

α₂A receptor-binding assay

α₂AAR binding was performed using membrane preparations from insect cells (Expression Systems, 94-001S) expressing human α₂A receptors, as previously described⁵¹. Membranes were incubated with increasing concentrations of test compounds and 5 nM [³H]rauwolscine in buffer containing 20 mM HEPES (pH 7.5) and 100 mM NaCl, at room temperature for 2 h. After incubation, samples were filtered onto GF/B filter plates, washed with ice-cold buffer and radioactivity was quantified by liquid scintillation counting. IC₅₀ and K_i values were derived using nonlinear regression in GraphPad Prism.

GloSensor cAMP assay for α₂A, CB2 and MOR

The GloSensor cAMP assay was performed following the manufacturer’s instructions (Promega) with slight modifications as previously described^51,69. In brief, wild-type human α₂A, CB2 and μOR were cloned into the pcDNA3.1 vector and co-transfected with the 22F cAMP GloSensor plasmid into HEK293T cells (ATCC CRL-3216) cultured in six-well plates. After 24 h, cells were reseeded into 96-well white plates in CO₂-independent medium and equilibrated with GloSensor cAMP reagent as per the manufacturer’s protocol. Cells were incubated for 1 h at 37 °C followed by 1 h at room temperature. Where applicable, 10 μM forskolin was used to elevate basal cAMP levels for assessing receptor-mediated inhibition. Serially diluted test compounds were added, and luminescence signals were recorded using a PerkinElmer microplate reader. Data were analysed using GraphPad Prism 9.0 to calculate EC₅₀ or IC₅₀ values.

AmpC β-lactamase inhibition assay

The AmpC β-lactamase inhibition assay was performed as previously described⁶⁰. The candidate inhibitors were dissolved in DMSO (20 mM stock) and diluted to maintain a constant 1% DMSO (v/v) in 50 mM sodium cacodylate buffer (pH 6.5). Assays were performed in the presence of 0.01% Triton X-100 to reduce aggregation artefacts. AmpC enzymatic activity was monitored spectrophotometrically using CENTA or nitrocefin as substrates. Initial screening was performed at 200 µM, 100 µM and 40 µM compound concentrations. Substrate concentrations were selected based on known K_m values to achieve defined [S]:K_m ratios: for CENTA ([S] = 50 µM, K_m = 27.6 µM) and nitrocefin ([S] = 100 µM or 28 µM, K_m = 180 µM). Reactions were carried out in 96-well format on a BMG Labtech CLARIOstar plate reader, with substrate and enzyme injected into wells containing the inhibitor, followed by kinetic measurement over 50 s. IC₅₀ values were determined by fitting inhibition curves in GraphPad Prism using a fixed Hill coefficient of 1, and K_i values were calculated using the Cheng–Prusoff equation⁷⁰.

HTRF assay for Mac1

Binding of the compounds to Mac1 was assessed by the displacement of an ADPr-conjugated biotin peptide from His6-tagged protein using a homogeneous time-resolved fluorescence (HTRF)-based assay, as previously described⁷¹. The expression sequences used for SARS-CoV-2 Mac1 are listed below. All proteins were expressed and purified as previously described for SARS-CoV-2 Mac1 (ref. ⁷¹). Compounds were dispensed into ProxiPlate-384 Plus (PerkinElmer) assay plates using an Echo 650 Liquid Handler (Beckman Coulter). Binding assays were conducted in a final volume of 16 μl with 12.5 nM NSP3 Mac1 protein, 200 nM peptide ARTK(Bio)QTARK(Aoa- RADP)S (Cambridge Peptides), 1:20,000 anti-His6-Eu3⁺ cryptate (HTRF donor; PerkinElmer AD0402) and 1:500 streptavidin-XL665 (HTRF acceptor; PerkinElmer 610SAXLB) in assay buffer (25 mM HEPES pH 7.0, 20 mM NaCl, 0.05% bovine serum albumin and 0.05% Tween-20, the latter also to reduce aggregation artefacts). Assay reagents were dispensed manually into plates using an electronic multichannel pipette. Mac1 and peptide were pre-incubated for 30 min at room temperature before HTRF reagents were added. Fluorescence was measured after a 1-h incubation at room temperature using a Perkin Elmer EnVision 2105-0010 Dual Detector Multimode microplate reader with dual emission protocol (A = excitation of 320 nm and emission of 665 nm, and B = excitation of 320 nm and emission of 620 nm). Compounds were tested in triplicate in a 14-point dose response. Raw data were processed to give an HTRF ratio (channel A/B × 10,000), which was used to generate IC₅₀ curves using nonlinear regression using GraphPad Prism v10.0.2 (GraphPad Software).

hERG channel inhibition

hERG channel inhibition was evaluated using a Thallium Flux assay on HEK293 cells (ATCC CRL-1573) stably expressing the hERG potassium channel. Cells were seeded at a density of 8,000 cells per well in 384-well poly-D-lysine-coated plates and incubated for 24 h under standard conditions (37 °C at 5% CO₂). The next day, a thallium-sensitive dye was added to the cells, followed by a 1-h incubation to ensure dye uptake. Test compounds were added to achieve a final concentration of 30 μM in 0.5% DMSO, and the cells were incubated for an additional 30 min at room temperature. Subsequently, a stimulation buffer containing thallium was added, and fluorescence measurements were taken using a FLIPR Tetra system. Data were collected every 3 s for 3 min (excitation of 470–495 nm and emission of 515–575 nm). Fluorescence intensity over time was analysed to calculate the area under the curve from which percentage inhibition was determined versus haloperidol at 100 μM (positive control) and vehicle (DMSO).

α₂A receptor purification and structure determination

Wild-type human α₂AAR was cloned into a pVL1392 vector with an N-terminal FLAG tag. The construct was expressed in Spodoptera frugiperda (Sf9) insect cells using the BestBac system. Cells at a density of 4 × 10⁶ cells per millilitre were infected with virus and incubated for 48 h at 27 °C. The receptor was solubilized and purified by FLAG affinity chromatography and size-exclusion chromatography in the presence of 10 μM compound ‘4905. Monomeric peak fractions were concentrated and used for G protein complex formation. G_αoG_β1γ2 heterotrimeric G proteins were expressed in Hi5 insect cells (Expression Systems, 94-002S) and purified using Ni²⁺ affinity following detergent solubilization and dephosphorylation. The final α₂AAR–G_αoG_β1γ2–scFv16 complex⁷² was assembled in the presence of ‘4905 and purified by size-exclusion chromatography. Cryo-EM grids were prepared using UltrAufoil R1.2/1.3 300-mesh grids and vitrified in liquid ethane. Data were collected on a Titan Krios G3 electron microscope equipped with a K3 direct electron detector. Image processing was performed using cryoSPARC, yielding a final reconstruction at approximately 2.8 Å resolution. Model building and refinement were performed with Protein Data Bank (PDB) 7EJ8 as a starting model using ChimeraX⁷³, Phenix⁷⁴ and Coot⁷⁵. The final structures have been deposited in the PDB with accession codes 9PLO and 9PLN. Structural representations of protein–ligand complexes were generated using the PyMOL Molecular Graphics System v2.5.5 (Schrödinger).

Mac1 purification and crystallization

Wild-type Mac1 protein (P4₃ construct, residues 3–169) was expressed in Escherichia coli BL21(DE3) as an N-terminal His₆-tagged construct and purified by Ni²⁺-affinity chromatography⁷¹. The His-tag was cleaved with TEV protease, followed by size-exclusion chromatography (Superdex 75) in 20 mM Tris-HCl (pH 7.5), 150 mM NaCl and 1 mM dithiothreitol. Purified protein was concentrated to 40 mg ml⁻¹ for crystallization and stored at −80 °C.

Crystals were obtained by sitting-drop vapour diffusion in 28% PEG 3000 and 100 mM CHES (pH 9.5). Compounds (100 mM in DMSO) were added to crystal drops using an Echo 650 acoustic dispenser⁷⁶ to a final concentration of 10 mM. Crystals were incubated at room temperature for 2–4 h and vitrified in liquid nitrogen without additional cryoprotection. X-ray diffraction data were collected at 100 K at the Advanced Light Source (beamline 8.3.1) using an X-ray wavelength of 0.88557 Å, and diffraction data were processed using XDS⁷⁷ and Aimless⁷⁸. Structures were determined to resolutions ranging from 0.97 to 1.02 Å. Ligands with low occupancy or conformational disorder were modelled using PanDDA⁷⁹ and Coot⁷⁵ and refined using phenix.refine⁸⁰ as previously described⁸¹. The structure of ‘3184 was determined using a racemic preparation of the compound (‘9037). X-ray data collection and refinement statistics are shown in Extended Data Tables 1–4, as are the 32 PDB IDs. Structural representations of protein–ligand complexes were generated using the PyMOL Molecular Graphics System v2.5.5 (Schrödinger).

Microsomal stability

Microsomal stability of compounds was evaluated using pooled mouse liver microsomes (M3000/lot #2010026, XenoTech) to estimate their metabolic stability and predict hepatic clearance. Each compound was incubated at 2 μM in a reaction mixture containing 0.42 mg ml⁻¹ microsomal protein, phosphate buffer (100 mM, pH 7.4), MgCl₂ (3.3 mM), NADPH (3 mM), glucose-6-phosphate (5.3 mM) and glucose-6-phosphate dehydrogenase (0.67 units per millilitre). Reactions were conducted at 37 °C in 96-well plates with shaking at 100 rpm. Samples were collected at five time points (0, 7, 15, 25 and 40 min), and reactions were quenched by adding five volumes of acetonitrile containing an internal standard. After centrifugation at 5,500 rpm for 5 min, the supernatants were analysed via high-performance liquid chromatography–tandem mass spectrometry (HPLC–MS/MS). The elimination rate constant (kel), half-life (t_1/2) and intrinsic clearance (Clint) were calculated by plotting the natural logarithm of the remaining parent compound versus time. Stability was compared with reference standards such as imipramine and propranolol.

Plasma protein binding

Plasma protein binding was measured using equilibrium dialysis with a 14-kDa molecular weight cut-off membrane in a 96-well HTD96b dialyser. Mouse plasma containing 1 μM test compound (0.005% DMSO and 1% acetonitrile) was placed in one chamber, and phosphate-buffered saline (PBS, pH 7.4) in the opposing chamber. The assembled plates were incubated at 37 °C with 5% CO₂ and approximately 95% humidity, shaking at 250 rpm for 5 h to reach equilibrium. After incubation, aliquots from each chamber were mixed with equal volumes of the blank opposite matrix and processed with acetonitrile containing internal standard. Supernatants obtained after centrifugation were analysed by HPLC–MS/MS. The percentage of compound bound to plasma proteins was calculated using the peak area ratio in buffer to plasma compartments. Recovery and stability standards were included to ensure accuracy and reliability. Verapamil served as a reference control. Most compounds showed moderate binding, typically ranging between 75% and 85%.

Plasma stability

Plasma stability was assessed in non-sterile mouse plasma (Li-heparin treated) at 1 μM concentration (final DMSO content was 0.005%). Incubations were carried out in aliquots of 60 μl each (two per time point) at 37 °C under 5% CO₂ and high humidity (approximately 95%). The reactions were quenched with 240 µl of 90% acetonitrile containing an internal standard at 0, 20, 40, 60 and 120 min, followed by centrifugation at 6,000 rpm for 5 min. Supernatants were analysed via HPLC–MS/MS to determine the percentage of parent compound remaining at each time point. Data were plotted to calculate compound half-lives (t_1/2). Reference compounds, verapamil and propantheline, were used as high and low stability controls, respectively. This assay is critical for identifying compounds susceptible to degradation by plasma esterases or hydrolytic enzymes, helping inform PK optimization strategies during lead selection.

Thermodynamic solubility

Aqueous thermodynamic solubility was determined in PBS (pH 7.4) for 247 compounds using a shake-flask method followed by UV absorbance quantification. Dry powder compounds were dissolved in PBS to a theoretical concentration of 4 mM and incubated in duplicate at 25 °C for 4 h and 24 h with shaking. After incubation, samples were filtered using HTS 96-well filter plates. The filtrates were diluted twofold in acetonitrile with 4% DMSO for UV analysis. The incubation samples for charged molecules were additionally diluted tenfold with 50% acetonitrile–PBS with 2% final DMSO. Calibration curves (0–200 μM) were prepared in 50% acetonitrile–PBS (2% final DMSO). Absorbance was measured between 230 and 550 nm using a SpectraMax Plus microplate reader. Compound-specific absorbance maxima were used to calculate concentration using SoftMax Pro and Excel. The assay dynamic range was approximately 2–400 μM (about 20–4,000 μM for charged molecules), with values near the upper limit treated as semi-quantitative. Ondansetron was used as a reference compound. This method reflects equilibrium solubility under physiologically relevant conditions and helps to rank compounds for formulation feasibility.

PAMPA-BBB

Passive blood–brain barrier (BBB) permeability was estimated using PAMPA (PAMPA-BBB) with a phospholipid-coated membrane simulating the brain endothelium. Test compounds (50 μM in Prisma HT buffer, pH 7.4, with 0.5% DMSO) were added to donor wells, whereas brain sink buffer was added to the acceptor wells. The donor and acceptor chambers were separated by a 0.45-µm filter membrane coated with brain polar lipids. Plates were incubated without agitation at room temperature for 4 h. Post-incubation, samples from both chambers, as well as a standard solution, were diluted with acetonitrile containing an internal standard. Apparent permeability coefficients (log Papp) were calculated based on peak area ratio. Clozapine and chlorpromazine were used as high-permeability controls, whereas ranitidine represented low permeability. The assay provides a high-throughput, non-cell-based method for estimating central nervous system exposure potential.

Behavioural analyses

Male C57BL/6J mice 7–8 weeks of age were obtained from The Jackson Laboratory (JAX strain #000664) and housed five per cage under a standard 12–12-h light–dark cycle at 20.7 °C (69.3 °F) and 58% humidity. All animal procedures were approved by the University of California, San Francisco Institutional Animal Care and Use Committee (protocol #AN208219). Animals were randomly assigned to treatment and control groups. For behavioural experiments, mice were initially placed together in a cage and allowed to move freely for a few minutes; each mouse was then randomly selected, injected with compound or vehicle, and placed in a separate cylinder before testing. The experimenter assessing behaviour was different from the experimenter administering injections and placing mice into the cylinders; experimenters were therefore blinded to treatment. Behavioural experiments were conducted in two independent cohorts. Mice were habituated individually in Plexiglas enclosures for 1 h before testing. Compounds were administered subcutaneously 30 min before behavioural assessment, and, where applicable, the α₂AAR antagonist atipamezole (2 mg kg⁻¹, intraperitoneally) was given 15 min before compound injection. Tail-flick latency was measured by immersing the distal third of the tail in a 50 °C water bath and recording the withdrawal time. For the neuropathic pain model, SNI was performed under isoflurane anaesthesia by ligating and transecting two of the three branches of the sciatic nerve, sparing the sural nerve. Mechanical thresholds were assessed 7–14 days post-surgery using von Frey filaments and the up–down method, and values were normalized to the baseline of each animal. Thermal nociception was evaluated using a 55 °C hotplate, and the latency to nocifensive behaviour (paw lick or jump) was recorded with a cut-off of 45 s to prevent tissue damage. Behavioural data are summarized as group size (n), mean and s.d. (Supplementary Table 13). Statistical analyses were performed in GraphPad Prism. For the SNI experiment (Fig. 5f), we used two-way ANOVA followed by Tukey’s multiple-comparisons test. The hotplate assay (Fig. 5g) was analysed using the Friedman test followed by Dunn’s multiple-comparisons test. For the tail-flick assays, the 4905 dose–response (Fig. 5c) was analysed by two-way ANOVA followed by Dunnett’s multiple-comparisons test against vehicle (control), whereas the PS75 and 3629 dose groups (Fig. 5d,e) were analysed using the Kruskal–Wallis test followed by Dunn’s multiple-comparisons test against vehicle. The Kruskal–Wallis and Friedman tests were nonparametric; Prism calculated two-sided, multiplicity-adjusted P values for the Dunn comparisons (Supplementary Table 13). For sample size, we did not perform formal a priori power calculations. Instead, group sizes were guided by our previous experience with these assays and by precedent in the literature. This approach may limit sensitivity to small effects. Raw data for the animal assays underlying Fig. 5c–g are provided in the Source Data file.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

All primary data are available through this article and its Extended Data and Supplementary Information. The cryo-EM structures and maps have been deposited in the PDB and Electron Microscopy Data Bank, respectively, under the accession codes 9PLO and EMD-71719 (the ‘4905-bound α₂AAR complex with G proteins and scFv16), and PDB 9PLN and EMD-71718 (the ‘4905-bound locally refined α₂AAR structure). The Mac1 crystal structures and supporting structure factors have been deposited in the PDB under the accession codes 14AB, 7IIW, 7IIX, 7IIY, 7IIZ, 7IJ0, 7IJ1, 7IJ2, 7IJ3, 7IJ4, 7IJ5, 7IJ6, 7IJ7, 7IJ8, 7IJ9, 7IJA, 7IJB, 7IJC, 7IJD, 7IJE, 7IJF, 7IJG, 7IJH, 7IJI, 7IJJ, 7IJK, 7IJL, 14AM, 14AN, 14AO, 14AP and 7IJM. Source data are provided with this paper.

1. Hughes, J. P., Rees, S., Kalindjian, S. B. & Philpott, K. L. Principles of early drug discovery. Br. J. Pharmacol. 162, 1239–1249 https://doi.org/10.1111/j.1476-5381.2010.01127.x (2011).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

2. Bleicher, K. H., Bohm, H. J., Muller, K. & Alanine, A. I. Hit and lead generation: beyond high-throughput screening. Nat. Rev. Drug Discov. 2, 369–378 https://doi.org/10.1038/nrd1086 (2003).

   Article  CAS  PubMed  Google Scholar 

3. Gaulton, A. et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 40, D1100–D1107 https://doi.org/10.1093/nar/gkr777 (2012).

   Article  CAS  PubMed  Google Scholar 

4. Zdrazil, B. et al. The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Res. 52, D1180–D1192 https://doi.org/10.1093/nar/gkad1004 (2024).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

5. Paul, S. M. et al. How to improve R&D productivity: the pharmaceutical industry’s grand challenge. Nat. Rev. Drug Discov. 9, 203–214 https://doi.org/10.1038/nrd3078 (2010).

   Article  CAS  PubMed  Google Scholar 

6. Lipinski, C. A. Lead- and drug-like compounds: the rule-of-five revolution. Drug Discov. Today Technol. 1, 337–341 https://doi.org/10.1016/j.ddtec.2004.11.007 (2004).

   Article  CAS  PubMed  Google Scholar 

7. Stumpfe, D., Hu, H. & Bajorath, J. Evolving concept of activity cliffs. ACS Omega 4, 14360–14368 https://doi.org/10.1021/acsomega.9b02221 (2019).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

8. Leeson, P. D. & Springthorpe, B. The influence of drug-like concepts on decision-making in medicinal chemistry. Nat. Rev. Drug Discov. 6, 881–890 https://doi.org/10.1038/nrd2445 (2007).

   Article  CAS  PubMed  Google Scholar 

9. Wunberg, T. et al. Improving the hit-to-lead process: data-driven assessment of drug-like and lead-like screening hits. Drug Discov. Today 11, 175–180 https://doi.org/10.1016/S1359-6446(05)03700-1 (2006).

   Article  CAS  PubMed  Google Scholar 

10. Jorgensen, W. L. Efficient drug lead discovery and optimization. Acc. Chem. Res. 42, 724–733 https://doi.org/10.1021/ar800236t (2009).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

11. Topliss, J. G. Utilization of operational schemes for analog synthesis in drug design. J. Med. Chem. 15, 1006–1011 https://doi.org/10.1021/jm00280a002 (1972).

    Article  CAS  PubMed  Google Scholar 

12. Hansch, C. & Fujita, T. ρ-σ-π Analysis: a method for the correlation of biological activity and chemical structure. J. Am. Chem. Soc. 86, 1616–1626 (1964).

13. Dong, J. et al. ADMETlab: a platform for systematic ADMET evaluation based on a comprehensively collected ADMET database. J. Cheminform. 10, 29 https://doi.org/10.1186/s13321-018-0283-x (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

14. Wang, L. et al. Accurate and reliable prediction of relative ligand binding potency in prospective drug discovery by way of a modern free-energy calculation protocol and force field. J. Am. Chem. Soc. 137, 2695–2703 https://doi.org/10.1021/ja512751q (2015).

    Article  ADS  CAS  PubMed  Google Scholar 

15. Klimovich, P. V., Shirts, M. R. & Mobley, D. L. Guidelines for the analysis of free energy calculations. J. Comput. Aided Mol. Des. 29, 397–411 https://doi.org/10.1007/s10822-015-9840-9 (2015).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

16. Hajduk, P. J. & Sauer, D. R. Statistical analysis of the effects of common chemical substituents on ligand potency. J. Med. Chem. 51, 553–564 https://doi.org/10.1021/jm070838y (2008).

    Article  CAS  PubMed  Google Scholar 

17. Kramer, C., Kalliokoski, T., Gedeck, P. & Vulpetti, A. The experimental uncertainty of heterogeneous public data. J. Med. Chem. 55, 5165–5173 https://doi.org/10.1021/jm300131x (2012).

    Article  CAS  PubMed  Google Scholar 

18. Booker, T. R., Jackson, B. C. & Keightley, P. D. Detecting positive selection in the genome. BMC Biol. 15, 98 https://doi.org/10.1186/s12915-017-0434-y (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

19. Pritchard, J. K. & Cox, N. J. The allelic architecture of human disease genes: common disease—common variant… or not? Hum. Mol. Genet. 11, 2417–2423 https://doi.org/10.1093/hmg/11.20.2417 (2002).

    Article  CAS  PubMed  Google Scholar 

20. Manolio, T. A. et al. Finding the missing heritability of complex diseases. Nature 461, 747–753 https://doi.org/10.1038/nature08494 (2009).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

21. Fowler, D. M. & Fields, S. Deep mutational scanning: a new style of protein science. Nat. Methods 11, 801–807 https://doi.org/10.1038/Nmeth.3027 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

22. Wells, J. A. Systematic mutational analyses of protein protein interfaces. Method Enzymol. 202, 390–411 (1991).

    Article  CAS  Google Scholar 

23. Pennington, L. D., Aquila, B. M., Choi, Y., Valiulin, R. A. & Muegge, I. Positional analogue scanning: an effective strategy for multiparameter optimization in drug design. J. Med. Chem. 63, 8956–8976 https://doi.org/10.1021/acs.jmedchem.9b02092 (2020).

    Article  CAS  PubMed  Google Scholar 

24. Muegge, I. & Hu, Y. In silico positional analogue scanning with Amber GPU-TI. J. Chem. Inf. Model. 62, 4448–4459 https://doi.org/10.1021/acs.jcim.2c00860 (2022).

    Article  CAS  PubMed  Google Scholar 

25. Schönherr, H. & Cernak, T. Profound methyl effects in drug discovery and a call for new C–H methylation reactions. Angew. Chem. Int. Ed. 52, 12256–12267 (2013).

    Article  Google Scholar 

26. Cramer, J., Sager, C. P. & Ernst, B. Hydroxyl groups in synthetic and natural-product-derived therapeutics: a perspective on a common functional group. J. Med. Chem. 62, 8915–8930 (2019).

    Article  CAS  PubMed  Google Scholar 

27. Chiodi, D. & Ishihara, Y. “Magic chloro”: profound effects of the chlorine atom in drug discovery. J. Med. Chem. 66, 5305–5331 (2023).

    Article  CAS  PubMed  Google Scholar 

28. Gillis, E. P., Eastman, K. J., Hill, M. D., Donnelly, D. J. & Meanwell, N. A. Applications of fluorine in medicinal chemistry. J. Med. Chem. 58, 8315–8359 (2015).

    Article  CAS  PubMed  Google Scholar 

29. Hopkins, A. L., Groom, C. R. & Alex, A. Ligand efficiency: a useful metric for lead selection. Drug Discov. Today 9, 430–431 https://doi.org/10.1016/S1359-6446(04)03069-7 (2004).

    Article  PubMed  Google Scholar 

30. Zhu, T. et al. Hit identification and optimization in virtual screening: practical recommendations based on a critical literature analysis. J. Med. Chem. 56, 6560–6572 https://doi.org/10.1021/jm301916b (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

31. Kuntz, I. D., Chen, K., Sharp, K. A. & Kollman, P. A. The maximal affinity of ligands. Proc. Natl Acad. Sci. USA 96, 9997–10002 https://doi.org/10.1073/pnas.96.18.9997 (1999).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

32. Halgren, T. A. Identifying and characterizing binding sites and assessing druggability. J. Chem. Inf. Model. 49, 377–389 https://doi.org/10.1021/ci800324m (2009).

    Article  CAS  PubMed  Google Scholar 

33. Thomas, M., Bender, A. & de Graaf, C. Integrating structure-based approaches in generative molecular design. Curr. Opin. Struct. Biol. 79, 102559 https://doi.org/10.1016/j.sbi.2023.102559 (2023).

    Article  CAS  PubMed  Google Scholar 

34. Flowers, J. et al. Expanding automated multiconformer ligand modeling to macrocycles and fragments. eLife 14, RP103797 https://doi.org/10.7554/eLife.103797 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

35. Di, L., Kerns, E. H. & Carter, G. T. Drug-like property concepts in pharmaceutical design. Curr. Pharm. Des. 15, 2184–2194 https://doi.org/10.2174/138161209788682479 (2009).

    Article  CAS  PubMed  Google Scholar 

36. Morgan, P. et al. Can the flow of medicines be improved? Fundamental pharmacokinetic and pharmacological principles toward improving phase II survival. Drug Discov. Today 17, 419–424 https://doi.org/10.1016/j.drudis.2011.12.020 (2012).

    Article  CAS  PubMed  Google Scholar 

37. Houston, J. B. Utility of in vitro drug metabolism data in predicting in vivo metabolic clearance. Biochem. Pharmacol. 47, 1469–1479 https://doi.org/10.1016/0006-2952(94)90520-7 (1994).

    Article  CAS  PubMed  Google Scholar 

38. Iwatsubo, T. et al. Prediction of in vivo drug metabolism in the human liver from in vitro metabolism data. Pharmacol. Ther. 73, 147–171 https://doi.org/10.1016/s0163-7258(96)00184-2 (1997).

    Article  CAS  PubMed  Google Scholar 

39. Amidon, G. L., Lennernas, H., Shah, V. P. & Crison, J. R. A theoretical basis for a biopharmaceutic drug classification: the correlation of in vitro drug product dissolution and in vivo bioavailability. Pharm. Res. 12, 413–420 https://doi.org/10.1023/a:1016212804288 (1995).

    Article  CAS  PubMed  Google Scholar 

40. Hopkins, A. L., Keseru, G. M., Leeson, P. D., Rees, D. C. & Reynolds, C. H. The role of ligand efficiency metrics in drug discovery. Nat. Rev. Drug Discov. 13, 105–121 https://doi.org/10.1038/nrd4163 (2014).

    Article  CAS  PubMed  Google Scholar 

41. Young, R. J. & Leeson, P. D. Mapping the efficiency and physicochemical trajectories of successful optimizations. J. Med. Chem. 61, 6421–6467 https://doi.org/10.1021/acs.jmedchem.8b00180 (2018).

    Article  CAS  PubMed  Google Scholar 

42. Rocklin, G. J., Mobley, D. L., Dill, K. A. & Hunenberger, P. H. Calculating the binding free energies of charged species based on explicit-solvent simulations employing lattice-sum methods: an accurate correction scheme for electrostatic finite-size effects. J. Chem. Phys. 139, 184103 https://doi.org/10.1063/1.4826261 (2013).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

43. Kim, H. et al. Artificial intelligence in drug discovery: a comprehensive review of data-driven and machine learning approaches. Biotechnol. Bioprocess Eng. 25, 895–930 https://doi.org/10.1007/s12257-020-0049-y (2020).

    Article  CAS  PubMed  Google Scholar 

44. Kumar, A., Kini, S. G. & Rathi, E. A recent appraisal of artificial intelligence and in silico ADMET prediction in the early stages of drug discovery. Mini Rev. Med. Chem. 21, 2788–2800 https://doi.org/10.2174/1389557521666210401091147 (2021).

    Article  CAS  PubMed  Google Scholar 

45. Fu, L. et al. ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support. Nucleic Acids Res. 52, W422–W431 https://doi.org/10.1093/nar/gkae236 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

46. Myung, Y., de Sa, A. G. C. & Ascher, D. B. Deep-PK: deep learning for small molecule pharmacokinetic and toxicity prediction. Nucleic Acids Res. 52, W469–W475 https://doi.org/10.1093/nar/gkae254 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

47. Swanson, K. et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics https://doi.org/10.1093/bioinformatics/btae416 (2024).

48. Bergstrom, C. A., Wassvik, C. M., Johansson, K. & Hubatsch, I. Poorly soluble marketed drugs display solvation limited solubility. J. Med. Chem. 50, 5858–5862 https://doi.org/10.1021/jm0706416 (2007).

    Article  CAS  PubMed  Google Scholar 

49. Obach, R. S. Prediction of human clearance of twenty-nine drugs from hepatic microsomal intrinsic clearance data: an examination of in vitro half-life approach and nonspecific binding to microsomes. Drug Metab. Dispos. 27, 1350–1359 (1999).

    Article  CAS  PubMed  Google Scholar 

50. Friden, M. et al. Structure-brain exposure relationships in rat and human using a novel data set of unbound drug concentrations in brain interstitial and cerebrospinal fluids. J. Med. Chem. 52, 6233–6243 https://doi.org/10.1021/jm901036q (2009).

    Article  CAS  PubMed  Google Scholar 

51. Fink, E. A. et al. Structure-based discovery of nonopioid analgesics acting through the α_2A-adrenergic receptor. Science 377, eabn7065 https://doi.org/10.1126/science.abn7065 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

52. Leung, C. S., Leung, S. S., Tirado-Rives, J. & Jorgensen, W. L. Methyl effects on protein-ligand binding. J. Med. Chem. 55, 4489–4500 https://doi.org/10.1021/jm3003697 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

53. Sindt, F., Bret, G. & Rognan, D. On the difficulty to rescore hits from ultralarge docking screens. J. Chem. Inf. Model. 65, 5553–5566 https://doi.org/10.1021/acs.jcim.5c00730 (2025).

    Article  CAS  PubMed  Google Scholar 

54. Lyu, J. et al. Ultra-large library docking for discovering new chemotypes. Nature 566, 224–229 https://doi.org/10.1038/s41586-019-0917-9 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

55. Moroz, Y., Chuprina, A. & Mykytenko, D. Enamine REAL DataBase – an instrumental and practical vehicle for charting new regions of the relevant drug discovery chemical space. In 251st American Chemical Society National Meeting & Exposition (ACS, 2016).

56. Verteramo, M. L. et al. Interplay of halogen bonding and solvation in protein-ligand binding. iScience 27, 109636 https://doi.org/10.1016/j.isci.2024.109636 (2024).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

57. Krimmer, S. G., Betz, M., Heine, A. & Klebe, G. Methyl, ethyl, propyl, butyl: futile but not for water, as the correlation of structure and thermodynamic signature shows in a congeneric series of thermolysin inhibitors. ChemMedChem. 9, 833–846 https://doi.org/10.1002/cmdc.201400013 (2014).

    Article  CAS  PubMed  Google Scholar 

58. Wermuth, C. G. The Practice of Medicinal Chemistry (Elsevier Science & Technology, 2008).

59. Gahbauer, S. et al. Iterative computational design and crystallographic screening identifies potent inhibitors targeting the Nsp3 macrodomain of SARS-CoV-2. Proc. Natl Acad. Sci. USA 120, e2212931120 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

60. Liu, F. et al. The impact of library size and scale of testing on virtual screening. Nat. Chem. Biol. 21, 1039–1045 https://doi.org/10.1038/s41589-024-01797-w (2025).

61. Clopper, C. J. & Pearson, E. S. The use of confidence or fiducial limits illustrated in the case of the binomial. Biometrika 26, 404–413 (1934).

    Article  ADS  Google Scholar 

62. Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat. Methods 17, 261–272 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

63. Lu, C. et al. OPLS4: improving force field accuracy on challenging regimes of chemical space. J. Chem. Theory Comput. 17, 4291–4300 https://doi.org/10.1021/acs.jctc.1c00302 (2021).

    Article  CAS  PubMed  Google Scholar 

64. Balius, T. E. et al. Testing inhomogeneous solvation theory in structure-based ligand discovery. Proc. Natl Acad. Sci. USA 114, E6839–E6846 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

65. Singh, I. et al. Structure-based discovery of conformationally selective inhibitors of the serotonin transporter. Cell 186, 2160–2175.e17 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

66. Vigneron, S. F. et al. Docking 14 million virtual isoquinuclidines against the μ and κ opioid receptors reveals dual antagonists–inverse agonists with reduced withdrawal effects. ACS Central Sci. 11, 770–790 (2025).

    Article  CAS  Google Scholar 

67. Hua, T. et al. Activation and signaling mechanism revealed by cannabinoid receptor-Gi complex structures. Cell 180, 655–665.e18 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

68. Hua, T. et al. Crystal structures of agonist-bound human cannabinoid receptor CB1. Nature 547, 468–471 (2017).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

69. Manglik, A. et al. Structure-based discovery of opioid analgesics with reduced side effects. Nature 537, 185–190 (2016).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

70. Yung-Chi, C. & Prusoff, W. H. Relationship between the inhibition constant (Ki) and the concentration of inhibitor which causes 50 per cent inhibition (I50) of an enzymatic reaction. Biochem. Pharmacol. 22, 3099–3108 (1973).

    Article  Google Scholar 

71. Schuller, M. et al. Fragment binding to the Nsp3 macrodomain of SARS-CoV-2 identified through crystallographic screening and computational docking. Sci. Adv. 7, eabf8711 https://doi.org/10.1126/sciadv.abf8711 (2021).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

72. Maeda, S. et al. Development of an antibody fragment that stabilizes GPCR/G-protein complexes. Nat. Commun. 9, 3712 https://doi.org/10.1038/s41467-018-06002-w (2018).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

73. Goddard, T. D. et al. UCSF ChimeraX: meeting modern challenges in visualization and analysis. Protein Sci. 27, 14–25 https://doi.org/10.1002/pro.3235 (2018).

    Article  CAS  PubMed  Google Scholar 

74. Liebschner, D. et al. Macromolecular structure determination using X-rays, neutrons and electrons: recent developments in Phenix. Acta Crystallogr. D Struct. Biol. 75, 861–877 https://doi.org/10.1107/S2059798319011471 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

75. Emsley, P., Lohkamp, B., Scott, W. G. & Cowtan, K. Features and development of Coot. Acta Crystallogr. D Biol. Crystallogr. 66, 486–501 https://doi.org/10.1107/S0907444910007493 (2010).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

76. Collins, P. M. et al. Gentle, fast and effective crystal soaking by acoustic dispensing. Acta Crystallogr. D Struct. Biol. 73, 246–255 https://doi.org/10.1107/S205979831700331X (2017).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

77. Kabsch, W. XDS. Acta Crystallogr. D Biol. Crystallogr. 66, 125–132 https://doi.org/10.1107/S0907444909047337 (2010).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

78. Evans, P. R. & Murshudov, G. N. How good are my data and what is the resolution? Acta Crystallogr. D Biol. Crystallogr. 69, 1204–1214 https://doi.org/10.1107/S0907444913000061 (2013).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

79. Pearce, N. M. et al. A multi-crystal method for extracting obscured crystallographic states from conventionally uninterpretable electron density. Nat. Commun. 8, 15123 https://doi.org/10.1038/ncomms15123 (2017).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

80. Afonine, P. V. et al. Towards automated crystallographic structure refinement with phenix.refine. Acta Crystallogr. D Biol. Crystallogr. 68, 352–367 https://doi.org/10.1107/S0907444912001308 (2012).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

81. Correy, G. J. et al. Exploration of structure-activity relationships for the SARS-CoV-2 macrodomain from shape-based fragment linking and active learning. Sci. Adv. 11, eads7187 https://doi.org/10.1126/sciadv.ads7187 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

Download references

Author notes

1. These authors contributed equally: Xinyu Xu, Olivier Mailhot, Galen J. Correy, Xi-Ping Huang

Authors and Affiliations

1. Department of Pharmaceutical Chemistry, University of California San Francisco, San Francisco, CA, USA

   Xinyu Xu, Olivier Mailhot, Karthik Srinivasan, Moira M. Rachman, Fangyu Liu, Katie L. Holland, Yujin Wu, Aashish Manglik & Brian K. Shoichet

2. Department of Bioengineering and Therapeutic Sciences, University of California San Francisco, San Francisco, CA, USA

   Galen J. Correy, Kara Zielinski & James S. Fraser

3. Department of Pharmacology, NIMH Psychoactive Drug Screening Program, School of Medicine, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA

   Xi-Ping Huang, Jing Wang & Bryan L. Roth

4. Department of Anatomy, University of California San Francisco, San Francisco, CA, USA

   Joao M. Braz & Allan I. Basbaum

5. Schrödinger Inc., San Diego, CA, USA

   Da Shi & Yuqi Zhang

6. Enamine Ltd., Kyiv, Ukraine

   Yuliia Holota, Yuliia Kuziv & Yurii S. Moroz

7. Center for Drug Discovery, Department of Pharmaceutical Sciences, Northeastern University, Boston, MA, USA

   Christos Iliopoulos-Tsoutsouvas & Alexandros Makriyannis

8. Department of Medicinal Chemistry, College of Pharmacy, University of Utah, Salt Lake City, UT, USA

   Nathan D. Levinzon

9. Helen Diller Family Comprehensive Cancer Center, University of California, San Francisco, CA, USA

   Yagmur U. Doruk, Morgan E. Diolaiti, Maisie G. V. Stevens & Alan Ashworth

10. Department of Chemistry and Pharmacy, Medicinal Chemistry, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany

    Harald Hübner & Peter Gmeiner

11. Chemspace LLC, Kyiv, Ukraine

    Yurii S. Moroz

12. Taras Shevchenko National University of Kyiv, Kyiv, Ukraine

    Yurii S. Moroz

13. Schrödinger Inc., New York, NY, USA

    Robert Abel

Contributions

B.K.S., X.X. and O.M. designed the project. X.X. and O.M. designed the analogues with help from N.D.L. α₂ Receptor studies were performed by X.X. and H.H., with guidance from P.G. α₂A-related behavioural analyses were conducted by J.M.B., with guidance from A.I.B. α₂ Structural studies were performed by K.S. under the supervision of A. Manglik. SERT uptake assays were carried out by X.-P.H. and J.W., with guidance from B.L.R. and ligand choice from Y.W. Mac1 biochemical assays were conducted by Y.U.D., M.G.V.S., M.E.D. and K.Z., and were supervised by A.A. Mac1 crystallography was performed by G.J.C. with guidance from J.S.F. CB2 receptor assays were performed by X.X. and C.I.-T., with guidance from A. Makriyannis and ligand advice from M.M.R. AmpC assays were conducted by X.X., F.L. and K.L.H. Y.S.M. coordinated and supervised small-molecule synthesis. All in vitro ADME and safety assays were supported by Y.H. and Y.K. through Bienta. FEP calculations and analyses were performed by D.S., guided by Y.Z. and R.A. O.M. performed statistical analyses for both ChEMBL and experimental data. X.X., O.M. and B.K.S. prepared the manuscript. B.K.S. supervised the project. All authors reviewed and approved the final manuscript.

Corresponding authors

Correspondence to Bryan L. Roth, James S. Fraser or Brian K. Shoichet.

Competing interests

B.K.S. is co-founder of Epiodyne, BlueDolphin and Deep Apple Therapeutics; serves on the Scientific Advisory Board (SAB) for Schrödinger, Vilya Therapeutics and Frontier Discovery; and is on the Scientific Resource Board (SRB) of Genentech. B.L.R. is founder of Onsero Therapeutics. J.S.F. is a consultant to and a shareholder of Vilya Therapeutics and Relay Therapeutics. D.S., Y.Z. and R.A. are employed by Schrödinger Inc. Y.H., Y.K. and Y.S.M. are employed by Enamine Ltd. Y.S.M. is employed by Chemspace LLC. The remaining authors declare no competing interests.

Peer review information

Nature thanks Derek Lowe, Celine Valant and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions