Identifying Artifacts from Large Library Docking

Abstract

While large library docking has discovered potent ligands for multiple targets, as the libraries have grown the hit lists can become dominated by rare artifacts that cheat our scoring functions. Here, we investigate rescoring top-ranked docked molecules with orthogonal methods to identify these artifacts, exploring implicit solvent models and absolute binding free energy perturbation as cross-filters. In retrospective studies, this approach deprioritized high-ranking nonbinders for nine targets while leaving true ligands relatively unaffected. We tested the method prospectively against hits from docking against AmpC β-lactamase. We prioritized 128 high-ranking molecules for synthesis and testing, a mixture of 39 molecules flagged as likely cheaters and 89 that were plausible inhibitors. None of the predicted cheating compounds inhibited AmpC detectably, while 57% of the 89 plausible compounds did so. As our libraries continue to grow, deprioritizing docking artifacts by rescoring with orthogonal methods may find wide use.

Graphic Absttact

INTRODUCTION

In the last five years, the number of readily available molecules for ligand discovery has grown from several million to tens of billions. Structure-based docking of those new libraries has revealed new chemotypes with potent affinity for targets ranging from enzymes^1–3 to GPCRS,^4–11 to transporters,^5,12 to kinases.^6,13 As the libraries have grown, however, both simulations and experiments have shown that the very top of the docking-ranked list becomes populated with molecules that “cheat” the scoring function. These false positives can be classified into normal docking failures and cheating artifacts (Figure 1). Normal docking failures, while common, occur through well-known problems of balancing different energy terms in the scoring functions. They often inhabit the same chemical space as true ligands and spread relatively evenly through the docking results. What we call cheating artifacts, conversely, are rare molecules that, through failures of parameterization or adoption of unusual structures that are not properly assessed,¹ rank among the very top molecules, often with scores offset from those of the overall distribution of docked molecules. As our virtual libraries continue to grow,^8,14–19 such cheating molecules may have an ever greater impact on virtual screening hit lists.

Figure 1.

Examples of (A) true inhibitors, (B) normal docking failures, and (C) cheating compounds against AmpC. Ranks are out of 1.7 billion docked.

Since these cheaters arise from holes in a particular scoring function, rescoring high-ranking docked molecules with a second function may identify them as outliers. Rescoring high-ranking docked molecules has been used previously to select molecules more likely to fit well, either because of a consensus agreement of docking scores²⁰ or reranking docked poses at a higher level of theory.^21,22 Here, we seek not so much to pick winners but to eliminate problematic docking hits. We cross-filter high-ranking docked molecules with three different methods: we compare DOCK3.8 implicit solvation energies to those calculated by FACTS (fast analytical continuum treatment of solvation)²³ and to those calculated by GBMV (generalized Born using molecular volume)²⁴ and compare DOCK3.8 rankings to those from an AB-FEP (absolute binding free energy perturbation) calculation on the top-ranking molecules. These calculations are undertaken both retrospectively and prospectively, where 39 cheating artifacts identified by rescoring are synthesized and tested experimentally, as are 89 molecules judged to be likely ligands by the same rescoring strategy. The results of these experiments suggest that a rescoring approach may be useful to deprioritize cheating artifacts that can concentrate among the top-ranking molecules from large library docking.

RESULTS

Retrospective Study on a Target with Hundreds of Experimental Ligands and Decoys.

A good system to retrospectively optimize the cross-filtering approach is one where large libraries of molecules are docked and where hundreds of docking false and true positives are measured by experiment. There are only a handful of these, with among the richest sets being against the enzyme AmpC β-lactamase. In a recent study, 1.71 billion molecules were docked against AmpC, and over 1400 molecules were tested experimentally.²⁵ As in earlier studies, the distribution of docking scores from this campaign suggested that cheating molecules might be present (Figure 2). As the library increased in size, there was a regular improvement in docking scores, until around 300 million molecules began to appear whose very favorable docking scores diverged from the rest of the distribution. As the library grew further, more of these molecules appeared, reaching ever better scores. Both simulation⁵ and experimentation with other systems¹ suggested that these might be cheating molecules. Inspection of some of these suggested that they have unusual partial atomic charge distributions, particularly involving atoms lower in the periodic table (higher atomic numbers, like bromine or even sulfur) and nitriles (Supporting Information Figure 1). The high partial atomic charges lead to unusually favorable scores in AmpC’s highly polarized active site. We note that these particular cheaters did not appear as high-ranking in earlier campaigns where hundreds of molecules had been tested (e.g., the Sigma2 receptor¹); thus far, our observation of cheating molecules has changed with the target.

Figure 2.

Distribution of docking scores with library growth against AmpC. As the library climbs toward a billion molecules, docking hits with unusually favorable (more negative) scores begin to appear.

We set out to test the cross-filtering approach on the 1440 docked-and-experimentally tested AmpC hits, asking whether it could separate true ligands from nonbinders while minimizing false negatives. We began by calculating FACTS solvation energies for the 1439 compounds, which took an average of 90 s per molecule on a Linux CPU. Plotting the normalized DOCK3.8 and FACTS desolvation energies against each other revealed a bimodal normal distribution (Figure 3A). Clustering those molecules within 3σ of the mean of both distributions reveals an ellipse outside of which 268 of the molecules fall; these molecules are flagged as having unusual solvation energies (i.e., outliers of the solvation energy distribution) by one of the methods. Of these, 262 do not bind AmpC detectably up to 200 μM, while six were decent inhibitors, the best of which had an affinity of 9.7 μM. Thus, insisting that all molecules are within 3σ of both removed more than 22% of the nonbinders while only losing 3.8% of the true ligands, none of which was among the most potent (which ranged down to sub-μM K_d values). While we would prefer not to lose any of the true ligands, we do note that this was a stringent test of the method as it is applied to all the molecules tested in this campaign, not only the thin wedge of unusually well-scoring molecules where we expect the “cheaters” to concentrate.

Figure 3.

FACTS rescoring results for 1440 experimentally validated compounds against AmpC. (A) Joint distribution of normalized DOCK and FACTS solvation free energy contribution. Distribution of (B) DOCK and (C) FACTS solvation free energy contribution. (D) Percentage of ligands and decoys filtered.

Encouraged by these results, we applied the same protocol against the σ2 and D4 receptors, which had the advantage of having had about 500 docking hits synthesized and experimentally tested against them;^1,6 in this sense, they were akin to the AmpC screen, though at a smaller scale. Here too, cross-filtering successfully flagged 22 and 28% nonbinders, respectively (Figure 4), while only losing four and three true ligands, respectively, none of which were among the more potent found in these docking campaigns. Many of the filtered compounds corresponded to molecules that had unusually favorable docking scores and unusual physical features for a hit and might thus be cheating molecules, a point to which we will return in the prospective testing part of this study.

Figure 4.

FACTS rescoring results for (A) D4 and (B) Sigma2. The total number of experimentally validated compounds is 537 and 495, respectively.

Cross-Filtering against Nine Receptors with Different Methods and Parameters.

To further investigate the robustness of this cross-filtering approach, we tested it against other docking targets with binding pocket environments different from those of AmpC, σ2, and D4. With the initial three studies establishing that this approach might be sensible, targeting a wider set of systems also afforded us the chance to try different rescoring methods and to vary the parameters with which we did so (Table 1).

Table 1.

Receptor Data Set

receptor name (abbreviation)	protein class	docking hits rescored	experimentally tested compounds
AmpC β-lactamases (AmpC)	hydrolase	300,000	1440
serotonin transporter (SERT)12	transporter	300,000	40
sigma-2 receptor (Sigma2)1	membrane protein	300,000	495
SARS-CoV-2 macrodomain (Mac1)26	MAR-hydrolase	500,000	78
SARS-CoV-2 main protease (Mpro)27	cysteine hydrolase	500,000	327
melatonin receptor type 1A (MT1)11	GPCR	300,000	38
cannabinoid receptor type 1 (CB1)28	GPCR	300,000	46
dopamine receptor D4 (D4)6	GPCR	300,000	537
α2a adrenergic receptor (alpha2a)9	GPCR	165,000	32

Here again, we sought cases where docked ligands had been tested, admittedly in the 40-molecule range rather than the 500 to 1500 as with σ2, dopamine D4, and AmpC, but nevertheless revealing both true ligands and false positives. There are by now close to 20 of these;^{1,4,6,8,9,11–13,26–30} we focused on 9 to which we had ready access (Table 1). Against each target, hundreds of millions to billions of compounds had been docked; we rescored the several hundred thousand top-ranking poses. For each of the nine targets, we explored rescoring not only with FACTS, as in the initial studies above, but also with a second implicit solvent method, GBMV, and several different calculation parameters (Table S1). As before, we note that this is not the true use case we envision for rescoring as it is applied to all docked-and-tested compounds, not only those very top-ranking ones where we might expect the cheaters to concentrate. Nevertheless, it provides a useful sanity check for the strategy.

We observed similar joint distributions for all nine targets using FACTS (Table 2 and Figure 5). Of the 300,000 to 500,000 high-ranking compounds rescored, between 7 and 19% were outliers by cross-filtering, sitting outside the bounds of the 3σ score distribution. For six of the nine targets, no potent ligands were among these outliers, while experimental nonbinders were captured as outliers, representing 3.85 to 13.51% of the total docking false positives across the screens (Table 2). For three targets, the Alpha2a adrenergic receptor, AmpC, and Sigma2, one potent ligand was categorized as an outlier, representing 12.5, 4.0, and 3.7% of the total ligands found for these targets, respectively; none are among the strongest binders found for their receptors. The best performance was for the CB1 and MT1 receptors and for serotonin transporter (SERT), where none of the true ligands are removed whatsoever. In these studies, we used FACTS with 1000 steps of minimization and ParamChem force-field parameters; the joint distributions for each of the other three parameter setups are plotted in Supporting Information Figures S2, S3, and S4, where similar distributions were observed. This suggests that the approach is robust to different parameters and different implicit solvent models.

Table 2.

Number/Percentage of Compounds Removed by Cross-Filtering

receptor	Ki cutoff (μM)a	potent ligand	decoyb	compoundsc
alpha2a	3.0	1/12.5%	1/10.00%	6239/7.29%
AmpC	10.0	1/4.0%	124/16.56%	55592/18.82%
CB1	10.0	0/0.0%	5/13.51%	34197/11.65%
D4	0.3	0/0.0%	18/14.29%	24411/8.34%
MT1	3.0	0/0.0%	3/11.54%	30934/10.55%
Mac1	300.0	0/0.0%	8/19.51%	26954/9.35%
Mpro	100.0	0/0.0%	6/8.70%	31276/10.83%
SERT	10.0	0/0.0%	1/3.85%	29788/10.17%
sigma2	0.3	1/3.7%	10/9.09%	20734/7.08%

^aThe upper limit of the binding affinity for potent ligands.

^bExperimentally validated nonbinders.

^cAll high-ranking compounds.

Figure 5.

Joint distribution of the solvation free energy for receptor targets (A) Alpha2a, (B) AmpC, (C) CB1, (D) D4, (E) MT1, (F) Mac1, (G) Mpro, (H) SERT, and (I) Sigma2. Relatively potent ligands are colored red, and the rest of the ligands are colored orange. Experimentally tested nonbinders are colored green, and compounds that were not tested are colored blue.

More Compounds Are Flagged as Cheaters among Top-Ranking Results.

The true use case for cross-filtering is to remove molecules that cheat the scoring function, which experiment (Figure 2) and simulation⁵ suggest occur among the very highest-ranking compounds. It is thus interesting to understand how the number of docking hits captured as outliers changes with the docking rank. Overall, cross-filtering highlighted about 10% of the high-ranking docking hits from among the top 300,000 to 500,000 molecules (Table 2). If we plot the number of outliers captured by the cross-filtering across the rank distribution, however, we find higher fractions of outliers among the very top-ranking compounds for six of the nine targets (Supporting Information Figure 5).

Prospective Study—Are Cheaters Experimental Nonbinders?

To test cross-filtering prospectively, we prioritized and synthesized the very top-ranking 128 de novo molecules from the AmpC screen²⁵ (ranked 1 to 415 out of 1.7 billion, DOCK3.8 scores from −151 to −91 kcal/mol). None of these had been previously tested. Of these, solvation cross-filtering identified 39 as cheaters (ranked 1 to 407) falling outside of the 3σ radii of the mean scores of the joint distribution (Figure 6A) and 89 as noncheaters falling within that distribution. Given their good scores, these 89 were considered plausible AmpC inhibitors; their scores overlapped with those of the cheaters (Figure 6B).

Figure 6.

Experimental testing of predicted high-ranking docking cheaters and plausible true actives against AmpC. (A) Bivariate distribution of AmpC top 300 K ranking compounds. The predicted cheaters are highlighted in red. (B) DOCK score distribution for top-ranking putative cheaters and plausible true actives. (C) Percentage inhibition at 200 μM for likely cheaters. (D) Docking hit rate as a function of AmpC apparent K_i. (E) Percentage inhibition at 200 μM for plausible actives; for both panels (C,E), the ZINC numbers of only every fourth compound are shown for clarity.

All 128 molecules were tested for AmpC inhibition, initially at 200, 100, and 40 μM. None of the 39 outliers inhibited the enzyme substantially at even the top concentration, including compounds ranked 1 to 26 of the 1.7 billion docked; all were classified as nonbinders (Figure 6C). Conversely, of the 89 high-ranking plausible ligands, 51 (57%) inhibited meaningfully at the top concentration (Figure 6D), with inhibition at the lower concentrations consistent with apparent K_i values of 240 μM or better and with 19 inhibiting the enzyme with apparent K_i values between 1.8 and 50 μM (Figure 6E). Concentration–response curves for the eight most potent inhibitors were well-behaved (Figure 7). Their activities place them among the more potent inhibitors discovered for AmpC from docking or high-throughput screens^6,25,31,32 (full compound results are listed in ampc-prospective-result.xlsx), while the gross hit rate of 57% (Figure 6D) is among the highest observed for this enzyme. The experimental results and corresponding SMILES string for each compound are recorded in the Supporting Information file (ampc-prospective-result.xlsx).

Figure 7.

Concentration–response curves for the top eight docking hits. The curves were fit with Hill slopes fixed to 1; had we allowed them to vary, they would have been MUiv −1.0; MUiu −0.82; 8Z5r −0.96; NqT1−0.83; rQg4−0.96; YvzS −0.81; ODhJ-1.05; 86Sb −0.87.

The AmpC assays were run in 0.01% Triton X-100, reducing the likelihood of colloidal aggregation. Nevertheless, for the 10 most potent inhibitors, dynamic light scattering (DLS) and counter-screening against malate dehydrogenase (MDH) were used to investigate colloidal aggregation at concentrations 10-fold higher than their apparent AmpC K_i values (Supporting Information Figures S6 and S7). Only one compound, with an apparent K_i of 8.6 μM, formed colloid-like particles, while none were observed for other top hits, consistent with these latter molecules acting as classic, active-site-directed inhibitors.

Other Methods to Detect Cheaters—AB-FEP and MM/GBMV.

In principle, one should be able to use many orthogonal scoring functions to identify molecules that cheat docking scoring functions. We tried two more here: AB-FEP and molecular mechanics/GBMV (MM/GBMV). Both represent higher levels of theory than implicit solvation methods like FACTS, especially AB-FEP. We note that using these methods differs somewhat from comparing the two solvation energies (those from FACTS and DOCK3.8), which may be thought of as an outlier detection approach, whereas rescoring with MM/GBMV and AB-FEP are orthogonal scoring approaches where we apply a threshold to an energy to predict ligands and decoys. Both methods were applied to the same 128 top-ranking molecules described in the previous section (see Experimental Section).

While GBMV with only minimization had performed relatively well in retrospective calculations (Supporting Information Table 1), the addition of a molecular dynamics component seemed to diminish its ability to prospectively distinguish the 89 plausible inhibitors or the 51 true ones from the 39 cheating artifacts. A challenge here was knowing where to draw the cutoff between plausible inhibitors and likely cheating artifacts since the MM/GBMV energies were much higher in magnitude than the energies inferred from the experimental apparent K_i values (Figure 8a). Accordingly, we investigated different energy cutoffs to distinguish the cheaters from the plausible ligands. Irrespective of where we drew the boundary, however, either very few true cheaters (true negatives) were found or too many true inhibitors were (false negatives). For instance, when we set the cutoff to be worse (greater) than −75 kcal/mol, only 13 of the 39 cheaters were identified, as was one phenolate sulfonamide, a close analogue of true inhibitors and a plausible ligand (Figure 8b). As we raised the stringency of the cutoff, more cheaters were found, but so were more true inhibitors. Increasing the cutoff to −125 kcal/mol identified 24 of the true cheaters but at the cost of adding 13 of the 89 plausible ligands, of which 9 were ultimately shown to be true inhibitors by experiment. Thus, the higher sampling of MM/GBMV led to worse results than the simpler FACTS approach. This may reflect the increased noise on addition of the molecular dynamics and the large magnitude of the MM/GBMV energies. Retrospective calculations in other systems bore this out: against the nonbinders and what turned out to be true ligands from the large dopamine D4 and Sigma2 receptor screens, we observed the same inability to clearly distinguish between nonbinders and true ligands (Supporting Information Figure S8). A caveat here is that many of these nonbinders were what we consider normal-docking failures, though among the very best docking scores, about half of them were likely “cheaters”.

Figure 8.

Rescoring the AmpC hits by MM/GBMV: (A) distribution of energy and (B) number of compounds removed as a function of energy. Rescoring results for AB-FEP: (C) distribution of energy and (D) number of compounds removed as a function of K_d.

AB-FEP performed better. Here too, the range of energies was larger than the experimental values (Figure 8c), reaching affinities in the midfemtomolar, and we considered two energy cutoffs to filter out putative cheating molecules: molecules calculated by AB-FEP to have K_d values worse than 200 μM (essentially the experimental cutoff) and molecules with calculated K_d values worse than 1 μM. At the 200 μM cutoff, AB-FEP removed 30 of the 39 FACTS-predicted cheating artifacts without flagging any false negatives (true inhibitors) (Figure 8d). At the 1 μM cutoff, it found 37 of the 39 cheating artifacts and classified another phenolate sulfonamide, a close analogue of true inhibitors so arguably a plausible ligand, as a cheater, with no experimentally confirmed false negatives found (Figure 8d). As the AB-FEP cutoff K_d became more potent than 1 μM, the number of false negatives began to rise. Detailed calculation results may be found in the Supporting Information (ampc-prospective- result.xlsx). We note that AB-FEP calculates binding affinities (expressed here as K_d values), while inhibition numbers are reported as enzymological K_i values; these two values represent the same thing, and both may be converted into free energies of binding.

DISCUSSION AND CONCLUSIONS

As our virtual libraries grow,^2,8,19 both simulation⁵ and experiment²⁵ suggest that docking results improve. Unfortunately, with this growth have emerged a small group of molecules that cheat our scoring functions and crowd the top scores of docking-ranked libraries.^1,5 While rare, these molecules rise in sheer number as libraries grow; unchecked, they may come to dominate the top-ranking molecules. A key observation from this study is that many of these scoring function “cheaters” may be recognized and deprioritized by an orthogonal scoring function. Perhaps the most compelling evidence for this comes from a prospective study against the model enzyme AmpC β-lactamase. Cross-filtering identified 39 “cheaters” from the very top-ranks of a docking campaign, and on experimental testing none bound substantially. Meanwhile, the method found another 89 compounds, often interspersed among the 39 “cheaters”, that were plausible ligands. On experimental testing, 57% of these 89 docking hits inhibited the enzyme. These results suggest that it is possible to identify molecules that cheat our scoring functions by filtering them out of top-ranking lists to reveal the more plausible and interesting high-ranking molecules.

Several caveats merit mentioning. Although we suspect that molecules that exploit scoring-function holes may be ubiquitous in large-library docking, we have shown that they exist only for DOCK3.7/3.8. Correspondingly, we have focused on physics-based scoring to cross-filter docking results, and while we might expect many methods to perform well in this role, we have not shown that. Finally, we note that our strategy of rescoring the top 300,000 high-ranking molecules seems sensible for billion-molecule libraries, where we expect several hundred cheating artifacts concentrated at the highest ranks. As our libraries grow toward trillions of molecules, however, there may be hundreds of thousands of cheating artifacts, demanding a second criterion to find the region where the cheaters can be separated from the true ligands. One way to do this would be to rescore not so much a fixed number of high-ranking molecules but to consider the region where the docking scores diverge from the distribution for the overall library (Figure 2). Irrespective of library size, such a break in the scoring distribution may identify the occurrence of cheating artifacts.

These caveats should not obscure the key points of this study. We expect ever greater numbers of molecules to find holes in our scoring functions as docking libraries continue to grow and diversify.⁵ Rescoring top-ranking molecules with a second scoring function that, whatever its own holes, is unlikely to share those of the primary scoring function can help to eliminate these molecules, revealing the more interesting and plausible molecules that they rank among. This strategy thus may be generally useful in the field, and multiple scoring functions may be useful for rescoring. The FACTS method used here may be particularly well-suited to physics-based approaches; accordingly, we provide easy-to-use rescoring and analysis scripts in the supplementary files.

EXPERIMENTAL SECTION

Computational Details.

Unless otherwise specified, Open Babel³³ was used to generate random ligand conformations, while ParamChem^34,35 was used to prepare ligand topology and parameter files. The CHARMM C36 force fields³⁶ were used, and molecular dynamics was performed in CHARMM.³⁷ The general AMBER force field³⁸ was used for comparison. Detailed molecular dynamics scripts are documented in the Supporting Information.

Rescoring with the Implicit Solvent Model.

We used the implicit solvent model FACTS²³ and the GBMV²⁴ for rescoring. Unless otherwise noted, for each of the docked poses, we performed a short minimization (1000 steps) using FACTS with the CHARMM force field. The protein atoms are fixed to reduce the computational cost. The total energy of the system is recorded (i.e., the sum of E_elec, E_vdw, E_solvation, and the internal energy of the protein and ligand). The solvation energy is explicitly computed by the implicit solvent model. The free energy at the unbound state is computed by first generating 20 random ligand conformations using the Open Babel functionality (obrotamer), followed by minimization. Then, we computed the ensemble average of these 20 trials and considered it as E_{ligand free energy at unbound state} because we are only interested in the same receptor structure. The protein free energy at the bound state is a constant and can be neglected. The detailed CHARMM script for rescoring can be found in Supporting Information.

$$\begin{gathered} \Delta G_{\text{binding}}=E_{\text{elec}}+E_{\text{vdw}} \\ +E_{\text{protein}\text{internal}\text{energy}\text{at}\text{bound}\text{state}} \\ +E_{\text{ligand}\text{internal}\text{energy}\text{at}\text{bound}\text{state}} \\ –E_{\text{protein}\text{free}\text{energy}\text{at}\text{unbound}\text{state}} \\ –E_{\text{ligand}\text{free}\text{energy}\text{at}\text{unbound}\text{state}} \end{gathered}$$ (1)

Cross-Filtering with Bimodal Distribution.

For compounds with different topologies binding to the same receptor binding pocket, it is reasonable to assume that these compounds have similar physiochemical properties so that they could maintain similar key interactions with the binding pocket (i.e., normal distribution). However, the top-ranking artifacts cheat one or multiple energy terms in the scoring function mainly because of (1) incorrect parameterization of the force field and (2) the missing details in the scoring function. Therefore, using another docking method might help separate these artifacts.

As shown in Figure 3, one could draw an ellipse to determine the outliers. The width, height, and center of the ellipse are equivalent to the standard deviation and the mean of each of the normal distributions. While the orientation of the ellipse can be determined by the covariance of the two variants, the size of the ellipse is defined by the user (i.e., the σ cutoff). We determined the ellipse based on the mean and the standard deviation of each of the normal distributions and the covariance of the bivariate normal distribution (eq 2).

$$\begin{gathered} \text{Assume}={\chi }_{\text{joint}}~\left(X_{\text{DOCK}},X_{\text{FACTS}}\right),\text{then} \\ \theta =\text{Cov}\left[X_{\text{DOCK}},X_{\text{FACTS}}\right] \\ \text{Let}c=\text{cos}(\theta );s=\text{sin}(\theta ) \\ \frac{{\left[\text{c}\left(x–{\mu }_{\text{FACTS}}\right)+s\left(y–{\mu }_{\text{DOCK}}\right)\right]}^2}{{\left(3\times {\sigma }_{\text{FACTS}}\right)}^2} \\ +\frac{{\left[s\left(x–{\mu }_{\text{FACTS}}\right)–c\left(y–{\mu }_{\text{DOCK}}\right)\right]}^2}{{\left(3\times {\sigma }_{\text{DOCK}}\right)}^2}=1 \end{gathered}$$ (2)

The detailed python script can be found in Supporting Information.

Calculating Absolute Binding Free Energy.

Absolute binding free energy (AB-FEP) calculations were performed with FEP+ on Schrodinger Web Services. The OPLS4 force field was used along with the Force Field Builder, where missing parameters were calculated. The total molecular dynamics simulation time for each ligand’s AB-FEP calculation was set to 1 ns. The docking poses used as the input for FEP were WScore docking results. All calculations were performed using Schrödinger Software Suite 2023–4.³⁹

The MM/GBMV experiments were run up as follows:⁴⁰ for each compound, we collected the top 5 docking poses. The bound state was minimized in vacuum with a maximum of 2000 steps before running molecular dynamics with the GBMV implicit solvent model. Hydrogen mass repartitioning was applied to reduce the computational cost. The total simulation length was 30 ns, and the last 1 ns of the simulation was used for the MM/GBMV calculation. Each simulation was repeated five times, and the average energy was computed. The lowest energy value among the five docking poses was considered the free energy of the bound state for that compound. The free energy for the unbound ligand state was computed with 25 random conformations generated by Open Babel, followed by the same MM/GBMV protocol.

AmpC β-Lactamase Enzymology.

AmpC was purified as described.⁴¹ All candidate inhibitors were dissolved in DMSO at 20 mM, and more dilute DMSO stocks were prepared as necessary so that the concentration of DMSO was held constant at 1% v/v in 50 mM sodium cacodylate buffer, pH 6.5. AmpC activity and inhibition were monitored spectrophotometrically using either CENTA or nitrocefin as substrates. All assays included 0.01% Triton X-100 to reduce compound aggregation artifacts. Active compounds were further investigated for aggregation by DLS and by detergent-dependent inhibition of the counter-screening enzyme MDH.

For initial screening, the docking hits were diluted such that the final concentrations in the reaction buffer were 200, 100, and 40 μM. In these assays, the AmpC substrate nitrocefin¹⁶ was used, with an [S]/Km ratio of 0.56 (Km nitrocefin 180 μM; [S] = 100 μM) and 0.16 ([S] = 28 μM). The colorimetric assay was carried out using a BMG Labtech CLARIOstar for kinetic measurements of 50 s in a 96-well format. IC50 values reflect the percentage inhibition fit to a dose–response equation in GraphPad Prism with a Hill coefficient set to one $\left(f(x)=\text{max}–\frac{\text{max}–\text{min}}{1+\frac{x}{\mathit{\text{IC}}50}}\right)$ K_i was calculated using the Cheng–Prusoff equation $\left(K_i=\frac{\mathit{\text{IC}}50}{1+\frac{[S]}{\mathit{\text{Km}}}}\right).$ For eight of the most potent compounds, based on the initial three concentration-point results, full dose–response curves were measured.

Docking Hits Tested.

All compounds were synthesized by Enamine and used without further purification. Of the 51 active compounds, all but 6 had purities of ≥95%. All of the 6 had purities ≥91%. The identities of all compounds are given in ampc-prospective-result.xlsx of the Supporting Information and are available from Enamine.

MDH Enzyme Inhibition Assay.

Compounds were diluted to 100 μM in 50 mM KPi buffer, pH 7, at a final concentration of 1% DMSO (v/v). Samples were incubated with MDH (Sigma, 442610) for 5 min. The reaction was initiated by adding 200 μM oxaloacetic acid (Sigma, 04126) and 200 μM nicotinamide adenine dinucleotide (NADH) (Sigma-Aldrich,10128023001). The reaction was monitored for 80 s at an absorbance of 340 nm. Sample rates were divided by the DMSO control rates. Compounds that showed less than 35% inhibition were not considered to be inhibitors. Compounds that surpassed the enzyme inhibition threshold were screened as a concentration–response curve from 100 to 0.1 μM. Identified inhibitors were rescreened at 100 μM in the presence of 0.01% (v/v) Triton-X 100 to determine detergent reversibility. All samples were screened in triplicate. Data were analyzed by using GraphPad Prism version 10.2.3 (Boston, MA).

Dynamic Light Scattering.

Compounds were diluted in a filtered 50 mM KPi buffer, pH 7, at a final concentration of 1% DMSO (v/v). All compounds were initially screened at a top concentration of 100 μM using a DynaPro Plate Reader III. Samples that had a scattering intensity >1 × 10⁷ cnts/s scattering in this instrument were considered to form colloid-like particles and were rescreened as a concentration–response curve in eight-point half-log dilutions. Data were separated into two groups: aggregating concentrations (scattering >1 × 10⁷ cnts/s) and nonaggregating concentrations (scattering <1 × 10⁷ cnts/s). A line was generated for each group, and the point of intersection of the two lines serves as the critical aggregation concentration. All samples were screened in triplicate. Data were analyzed using GraphPad Prism.

ABBREVIATIONS USED

AB-FEP

absolute binding free energy perturbation

FACTS

fast analytical continuum treatment of solvation

FEP

free energy perturbation

GBMV

generalized born using molecular volume

MM/GBMV

molecular mechanics/generalized born molecular volume