Toward a Random Background for Ligand Optimization

40 Ligand optimization is central to drug discovery as hundreds of analogs might be 41 designed and synthesized between an initial hit and a therapeutic candidate. The efficiency of 42 this process is unclear, at least partly because there is no random background for optimization 43 against which to compare. Such a random bac kground might emerge from synthetically 44 accessible but otherwise systematic random small substitutions across starting ligands, 45 measuring likelihood of achieving a substantial improvement in affinity/potency or other property 46 by any single perturbation. Recent literature and ligand-affinity/potency databases suggest that 47 perhaps 10% of analogs with minor modifications improve upon a parent’s potency substantially 48 (by >10-fold), but this number is clouded by reporting bias, intentional improvement, and inter-49 group reproducibility. To begin to establish a background expectation for ligand optimization, 50 we comprehensively and systematically modified 18 lead molecules across six targets with 51 single atom changes; 257 compounds were synthesized . Unexpectedly, 1 1.2% of these 52 random small perturbation analogs improved potency by >10-fold over their parents. 53 Conversely, these more potent analogs typically had worse in vitro pharmacokinetics (e.g. 54 reduced metabolic stability, lower plasma free fraction). While it was possible to find analogs 55 where the potency increase compensated for inferior exposure and half -life, resulting in more 56 potent compounds in vivo, overall a frustrated landscape for ligand optimization is revealed. 57 This study begins to establi sh a background expectation for ligand potency optimization and 58 offers a simple strategy to do so. It also begins to quantify the challenges confronting the field 59 in moving beyond in vitro potency. 60 61 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 3

62 Ligand optimization is central to chemical probe and drug discovery, as initial active 63 molecules rarely have the potency or pharmacokinetic properties to be viable in vivo 1,2. 64 Accordingly, between the discovery of an initial hit and a clinical candidate many hundreds, 65 occasionally thousands of optimized analogs might be synthesized1,3. Several strategies4-6 7,8 66 have been developed to improve this process , ranging from early empirical approaches such 67 as Topliss Trees9 and Hansch QSAR10 to contemporary AI-guided ADMET11 prediction and free 68 energy calculations 12,13. How efficient these strategies are remains uncertain, as there is no 69 random background against which to compare them. If we use a ten-fold improvement between 70 analog and parent as a benchmark for substantial impact, then only about 10% of analogs meet 71 this standard in the ChEMBL ligand -protein activity database (Extended Data Figure 1 a)14. 72 Since these ChEMBL results suffer from success bias, sample multiple types of perturbation, 73 and struggle with inter -group irreproducibility15, the public domain offers no sure guidance on 74 what level of improvement one might expect in ligand optimization if one were making random 75 conservative changes. 76 Random backgrounds have long been used in biology to help quantify significance. In 77 genomics, they help to distinguish between artificial and natural selection 16. In epidemiology, 78 random incidence rates help distinguish genetic diseases fr om those driven by environmental 79 factors17,18. In protein engineering, random backgrounds are used to evaluate improvements in 80 enzyme function 19, and alanine scanning has been used to introduce a minimally -biased 81 perturbation to find hot spots for ligand binding and protein func tion20. A random background 82 for ligand optimization, conceivably, could quantify progress across chemotypes and targets 83 and compare different optimization strategies. 84 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 4 Akin to alanine scanning, an unbiased background for ligand optimization might involve 85 conservative perturbations, systematica lly and comprehensively applied across multiple 86 unrelated parents across several different targets. Muegge and colleagues have described a 87 “positional analog scanning” approach, which substitutes ligand non-polar hydrogens (C—Hs) 88 with groups like CH ₃, OH, Cl, F, and Br, and aromatic carbon atoms with nitrogen 21. In 89 retrospective analyses of the ChEMBL24 database22,23, effects on affinity, functional activity, 90 solubility, clearance, and permeability were studied. Among over 110,000 matched molecular 91 pairs (MMPs), about 30% of the time of one analog improved over 3-fold in affinity /potency 92 versus the other. Our understanding of this observation is tempered by design and reporting 93 biases toward compound improvement , by the variability in experimental conditions across 94 datasets that can reduce reproducibility15, and by the few parent compounds that had multiple 95 conservative changes. Thus, it seemed interesting to create a set of parent ligands that were 96 systematically and comprehensively modified by small perturbations, without design and 97 without a bias toward improvement, using the positional analog scanning approach 21,24 where 98 the differences between parent and the suite of analogs were quantified by a single lab. 99 100 To create what we will refer to as a random background set for ligand optimization, we 101 chose 18 parent ligan ds spanning six targets, including three GPCRs (alpha2A adrenergic 102 receptor (2a), μ-opioid receptor (MOR) and cannabinoid receptor type 2 (CB2)), one 103 transporter (the serotonin transporter, SERT) and two soluble enzymes (AmpC β-lactamase 104 (AmpC), and Macrodomain 1 of SARS-CoV-2 (Mac1)). This set was chosen for targets that we 105 had under experimental control and is admittedly far from comprehensive: we explore only class 106 A GPCRs, no ion channels nor kinases are represented, only one transporter, and only two of 107 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 5 almost innumerable soluble enzymes. The 18 parent compounds were chosen for similar 108 pragmatic reasons—they were molecules to which we had ready access. Still, these proteins 109 provide a range of target classes, while the 18 parents cover a wide range of physico-chemical 110 properties (Extended Data Table 2) and values, ranging from 1 nM to 43 µM (Table 1), and 111 most were topologically unrelated . For each of the 18 parents we systematically explored all 112 analogs accessible by a single atom modification where such modifications were synthetically 113 feasible—typically costing no more than $400 to acquire—and where they did not modify the 114 net charge of the parent at physiological pH. The single atom modifications involved 115 substituting carbon hydrogens with one of methyl 25, hydroxyl 26, chloro 27, fluoro 28, or rarely 116 bromo, or substituting aromatic carbon atoms with nitrogen s (Figure 1). These changes are 117 not comprehensive, but they do represent changes preferred by medicinal chemists while 118 avoiding those that would introduce oxidation liabilities (e.g., adding a thiol or aromatic sulfur) 119 or changes of valence (e.g., nitrogen to oxygen or sulfur) ; moreover, they were made without 120 design and as systematically as pragmatism would allow. 121 122 These substitutions were systema tically and comprehensively applied without design, 123 ensuring an approach unbiased by anything other than synthetic accessibility and price (which 124 we admit are meaningful constraints, as we make clear below) . Overall, 257 single-atom 125 analogs were synthesize d and tested for changes in activity on the target, allowing us to 126 measure how often we might expect affinity/potency to improve >10-fold by minimal 127 perturbation. This begins to provide a background expectation for how often a designed analog 128 might meaningfully improve affinity/potency. As we will show, these undesigned changes had 129 an unexpectedly high success rate, suggesting also a systematic strategy for optimization, even 130 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 6 though that was not our primary goal. Since affinity/potency maturation is only one criterion for 131 advancing from hit to candidate, w e also measured in vitro pharmacokinetic properties for the 132 analogs, including permeability, metabolic stability, plasma protein binding, and three other 133 terms, versus the parent compounds. These pharmacokinetic studies explore what are perhaps 134 inevitable trade-offs between pharmacokinetics and target affinity/potency in ligand optimization. 135 136 Fig.1: Experiment and Analysis Workflow for One Heavy Atom Analogs. a, Parent ligands from 137 six drug targets (the aminergic GPCR Alpha2a-AR, the lipid GPCR CB2R, the peptide GPCR µOR, the 138 serotonin SLC transporter SERT, and the soluble enzymes AmpC and Mac1) were modified by one-at-139 a-time, one atom substitutions: HCH3, HCl, HOH, HF, and ring CN. These modifications 140 were made systematically and comprehensively; only compounds that were too expensive to 141 synthesize, or that changed the charge state of the parent were left out. The analogs were tested for 142 changes in affinity/potency, solubility, plasma protein binding, plasma stability, hepatic microsomal 143 stability, permeability, and hERG inhibition, versus their parent ligands. 144 145 146

147 We began with 18 parent ligands, most of which are unrelated to one another (ECFP4 148 Tanimoto coefficients, Tcs, <0.35), with affinities ranging from 1 nM to 43 µM and molecular 149 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 7 weights (MW) ranging from 200 to 400 amu (Extended Data Figure 2). Each of the 18 parents 150 was modified, one atom at a time, with five types of single atom substitutions (Figure 1a, central 151 panel), subject only to the expense and charge modification constraints (above). About 7% of 152 the possible molecules met these criteria and of those ~80% were successfully synthesized, 153 resulting in 257 analogs. In the analyses that follow, we pool all 257 analogs for analysis, which 154 reduces the noise from a single scaffold or single target and improves statistical power 155 (Supplementary Figure 1). 156 157 These 257 analogs were tested for activity changes (Kd, IC50, or EC50) versus the parent 158 ligands across the six targets. For the μ-opioid receptor, EC50 values were determined in cAMP 159 Glo-sensor assays. 2A and CB2 ligands were assessed by cAMP assays and by radioligand 160 competition binding with 3H-rauwolscine or [³H]CP-55940, respectively . Serotonin transporter 161 interactions were analyzed using [3H]-citalopram binding assays. AmpC β -lactamase (AmpC) 162 inhibition was measured enzymatically while that of Mac1 was measured by ligand 163 displacement. For every parent, and for most receptors overall, we only report changes for one 164 type of activity, i.e., if a parent K i is reported than only K i values are reported for its analogs, 165 and if a parent is an agonist, only EC 50s are reported for its analogs. Each functional or 166 inhibition assay is done in reference to a literature positive control molecule ( Supplementary 167 Figure 2). Representative concentration–response curves and the corresponding Z values for 168 each assay are prov ided in (Supplementary Figure 3). Especially for agonists, the effect of 169 receptor expression is controlled for (Supplementary Figure 4, 5). Because changes in 170 agonist efficacy can sometimes affect interpretation of EC 50, we inspect efficacy values to 171 ensure that they are similar between parent and analog; full concentration-response curves and 172 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 8 analyses are shown for every parent-analog pair, allowing for direct inspection of these results. 173 174 Overall, 23% of the analogs affected activity by 10-fold (Extended Data Figure 1 b). Intriguingly, of the 257 analogs tested, 29, or 177 11.3%, improved activity by >10-fold rounding to the nearest integer (Table 1). We note that 4 178 of these 29 had fold -changes of between 9.7 and 9.9 to 1 decimal place; we consider these 179 effectively 10 -fold improvements and used the same convention in analyzing the public 180 ChEMBL data (see Table 1 and Supplementary Table S1 for details) . Such >10-fold 181 improvement analogs were found for five of the six targets, and for 10 the 18 parents (Table 1). 182 The effects and chemical identities of all analogs relative to their parents are listed in 183 Supplementary Table 1. 184 185 Table 1. Analogs from all parents improving affinity/potency by >3x or >10x 186 Protein Assay readout Parent Ligand Parent Affinity/potency Total Analogs Analogs with ≥3X Affinity/potency improvement Analogs with ≥10X Affinity/potency improvement MOR EC50 Z4407498716 9.2 nM 16 2 2 alpha2 EC50 Z2750653629 1.4 nM 31 7 4a alpha2 Ki Z3034773248 22.1 µM 18 3 1 alpha2 EC50 Z4376630014 8.3 nM 10 2 0 alpha2 Ki Z8727395870 13 µM 18 1 0 AmpC Ki Z2610488449 24.2 µM 10 2 0 cb2 Ki Z52076138 2.3 µM 7 7 6b cb2 Ki Z6969215903 0.4 µM 6 0 0 cb2 EC50 Z8184698918 0.1 µM 20 2 2 Mac1 IC50 Z5398393122 2.2 µM 5 0 0 Mac1 IC50 Z1039063794 0.7 µM 12 0 0 Mac1 IC50 Z7534253453 8.9 µM 22 8 2 SERT Ki Z2009978218 1.1 µM 14 4 0 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 9 SERT Ki Z2573292480 0.08 µM 11 1 0 SERT Ki Z2573292509 0.2 µM 9 2 1 SERT Ki Z6971277399 13.4 µM 15 10 4 SERT Ki Z8727393896 42.7 µM 14 8 3c SERT Ki Z8731642686 1.8 µM 19 10 4 Total All Parents 257 69 29 *Fold-changes are displayed as nearest -integer values; exact (unrounded) fold -changes are provided in 187 Supplementary Table S1. Using exact fold -changes, 62 analogs improve ≥3-fold and 25 improve ≥10-188 fold. a. One of these had a fold change of 9.7. b. Two of these had fold-changes of 9.7 and 9.8. c. One 189 of these had a fold-change of 9.9, see Supplementary Table S1. 190 191 We were interested in which perturbations had the biggest effects, how the rates of 192 affinity/potency improvements compared to the literature, with its admitted biases, and if 193 physical properties of the parent compounds or of the binding sites correlated with greater 194 likelihoods of affinity/potency improvements. Among the analogs, methyl substitutions were the 195 most likely to improve activity >10-fold, with chlorine substitutions the second most likely; for >3-196 fold improvements, of which there were 69 among the 257 analogs, both chlorine and methyl 197 substitutions were the most common (Figure 2a ). Given the similar size and similar 198 hydrophobicity of methyls and chlorines, their similar effects may be rationalized. Conversely, 199 fluorine substitutions, which are often used to increas e metabolic stability, yielded no >10-fold 200 and few >3-fold improvements in activity. Aromatic carbon to nitrogen substitutions also rarely 201 improved affinity/potency, though as we will see they often had the most favorable impacts on 202 pharmacokinetic properties. 203 204 Because increases in hydrophobicity will increase affinity/potency simply by disfavoring 205 the unbound state, and have physical property liabilities, it is us eful to consider the improved 206 affinity/potency of the analogs in light of ligand efficiency (LE) and liphophilic efficiency (LiPe) , 207 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 10 which control for these effects . If affinity/potency improvements were largely driven by 208 hydrophobicity we would expect to see both terms deteriorate for the analogs. Instead, for 209 analogs that improved >10-fold both ligand efficiency and lipophilic efficiency much improved. 210 For instance, an analog with an added methyl, chloro or other atom that improved in binding or 211 in EC50 by even 3-fold enjoyed a ligand efficiency of 0.65 kcal/atom for the added atom, well 212 above the 0.3 kcal/atom that is thought to be useful for drug optimization 29,30. Meanwhile, 213 analogs that improved activity >10-fold had ligand efficiency of 1.4 for a single added atom (Ext 214 Data Figure 4 ), close to the efficiency limit 31 (several even exceeded single atom ligand 215 efficiencies of 2.0). Similarly, every analog that improved activity by >10-fold also improved 216 lipophilic efficiency (Ext. Data Figure 4). More broadly, while there was little correlation 217 between clopgP and pKi or logEC50 over the the 257 analogs (Ext Data Figure 4b), there 218 was a strong linear correlation between both ligand efficiency and activity change and between 219 lipophilic ligand efficiency and activity change —in both cases, substantially improved activity 220 was accompanied by improved ligand and lipophilic efficiency. For analogs with much improved 221 activity—especially those improved by >10 -fold but also for those improved by >3 -fold—the 222 improvements cannot be simply laid at the door of increased hydrophobicity. 223 224 We investigated whether there were correlations between likelihood of potency 225 improvement and ligand and binding site properties. One might expect that certain ligands or 226 sites might better lend themselves to affinity/potency improvement, which would affect the 227 domain of applicability of this study . For instance, weaker ligands might be easier to optimize, 228 or smaller binding sites might be better suited to big affinity/potency jumps. We calculated if 229 the parent molecular weight, binding site volume, the ligandability of the binding sites 230 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 11 (calculated using Dscore 32), or the affinity/potency of the starting parent compound was 231 correlated with the likelihood of finding analogs with substantial affinity/potency improvement. 232 Using the fraction of analogs achieving ≥10 -fold improvement as the outcome, none of these 233 properties showed a statistically meaningful correlation (Figure 2b ). If one considers the 234 magnitude of the affinity/potency change overall, and not just whether it improved >10-fold, 235 correlations did emerge, with weaker vs stronger parents more likely to support improved affinity 236 (R = −0.34, p = 3.3×10 −7) and with larger bi nding sites also tending to do so (R = 0.23, p = 237 4.3×10−4) (parent size and site ligandability continued to show no association with affinity 238 improvement). While these trends were echoed among matched pairs in ChEMBL (Extended 239 Data Figure 1), the correlations overall were weak (even an R of -0.34 explains only 10% of 240 set variance), and as activity changes drop below a 3-fold effect our confidence in the 241 experimental values begins to drop. Taken together, these results suggest that : (1) as a 242 strategy, systematic small perturbations can be unexpectedly successful at improving ligand 243 affinity/potency21,24, perhaps even compared to modern design -heavy approaches33; (2) this 244 effect is common across diverse parents and binding sites, (3) finds echo in the literature, and 245 (4) random expectations for substantial affinity/potency improvement may be as high as 11% 246 (95% CI 7.5%-15.2%). 247 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 12 248 Fig.2 Property correlations of affinity/potency improvement and comparison with literature 249 values. a. The fraction of compounds improving by >3- or >10-fold in activity by type of single 250 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 13 modification (CH₃, F, Cl, Br, N, OH). b. Correlation heatmap between physicochemical properties of 251 binding pockets and parent compounds with log₁₀ of affinity/potency fold-change. c. The fraction of 252 single atom modifications (CH ₃, F, Cl, N, OH) that improve a ctivity, comparing the experimental 253 observations in this study with results from the ChEMBL database. Shaded regions indicate 95% 254 confidence intervals, estimated from 10,000 iterations of bootstrap resampling . The ChEMBL ligands 255 are a composite of 54,021 parents and 79,984 small change analogs; for only 4.7% of these were there 256 more than three analogs per parent. 257 258 To investigate the structural bases for the affinity/potency changes, we determined the 259 structure of the parent inhibitors in complex with Mac1 and six analogs from the '3122, 12 260 analogs from the '3453 series, and four analogs from the ‘3794 series. These structures were 261 determined to between 0.97 and 1.02 Å resolution (Supplementary Table 2 ), supporting 262 atomic resolution analysis. The analogs with structures determined represented all six types of 263 single-atom modifications in this study; 15 of them differed in affinity/potency from their parents 264 by >3x, two by over 10-fold, and among themselves by up to 680-fold. 265 266 For some parents, analogs superposed well facilitating analysis of the effects. This was 267 often true for the Mac1 ‘3453 (IC50 8.9 µM) inhibitor series, for instance (Figure 3a). Displayed 268 at the ortho-position of the phenyl ring the chloro of ‘0676 (IC50 21 µM) projects into solvent and 269 loses activity versus the parent ( Figure 3b). Moving this chloro to the meta -position ('1304), 270 where it packs with Phe132, Ile131, and Gly48 , improves inhibition 11 -fold (p-value <0.01) to 271 1.9 µM (Figure 3c). Moving this substitution one atom further over, to the para-position, had 272 little effect ('6343, IC50 2.1 µM) though a methyl at the same position ('9870, IC50 7.6 µM) loses 273 4-fold affinity (p-value <0.01), presumably reflecting its poorer packing versus the chloro analog. 274 275 If many of the effects of the small perturbation analogs could be explained post hoc from 276 their structures, however, fewer were easily anticipated. Even the apparent simplicity of the 277 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 14 ‘3453 series is belied by structural accommodations in the enzyme site. For instance, the para-278 chloro of the 2µM ‘6343 would have clashed with the enzyme without Phe132 and Ile131 279 rotating away. In doing so, Phe132 adopts a partially eclipsed conformation. Presumably the 280 resulting strain is overcome by the improved packing of the buried para-chloro, but the outcome 281 of this balance of terms seems difficult to intuit (encouragingly, free energy calculations did 282 correctly predict that the para -and meta-chloro ‘6343 and ‘1304, respectively, would improve 283 affinity while the ortho -chloro ‘0676 would lose it, though the agreement was more qualitative 284 than quantitative; see below). Conversely, the methyl analog (‘9249 , 53 µM) of '0676 (IC50 22 285 µM) is buried from solvent, not exposed, and might be expected to be more potent than its 286 chloro cousin, but instead it loses another 2.5-fold in Ki (p-value <0.01; here, the opposite effect 287 was anticipated by the free energy calculation, below). Meanwhile, in the complex between 288 ‘4905 and the 2a adrenergic receptor, an added methyl group fit s into a pre -existing sub-289 pocket and qualitatively its improved potency is reasonable. Quantitatively, its 52-fold effect on 290 EC50 is at the outer edge of what might be expected by ligand -efficiency, especially since the 291 methyl buries a serine hydroxyl. This is echoed by the free energy calculation, which here too 292 correctly predicts improved activity, but only by 2 - not 52 -fold. Th ese are examples of 293 complexes where the analogs largely superpose on the parent and each other. In complexes 294 where the single-atom changes led to substantial inhibitor movement , or to the adoption of 295 multiple ligand conformations, prediction was harder still (Extended Data Fig 5). For example, 296 the Mac1 methyl analog ‘3176 repositions the inhibitors in the pocket (Extended Data Fig. 5b), 297 while compounds ‘9249 and ‘3194 adopt two distinct conformations in the site (Extended Data 298 Fig. 5b), a phenomenon that i s likely underreported in the PDB 34 but is readily seen at ultra -299 high resolution. Taken together, this set of perturbations may provide an interesting and 300 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 15 sometimes challenging test for prediction methods, and su ggests that there may be merit in 301 systematic small perturbations as a strategy, though showing that was not a goal of this study. 302 303 Figure 3. Structural basis of potency changes induced by single -atom modifications for Mac1 304 inhibitors. a. Chemical structure of the ’3453 parent and overlay of X -ray crystal structures of 12 305 analogs. b. Alignment of ‘0676 (21 μM) and ‘9249 (53 μM) with the ‘3453 parent showing the difference 306 in binding pose for the ortho -chloro and ortho-methyl substituted analogs. c. Comparison of ‘6343 (2.1 307 μM), ‘1304 (1.9 μM) and ‘0676 (21 μM) reveals that repositioning the aryl chloride disrupts packing with 308 I131 and F132, reducing potency. d. Para-methyl (‘9870, 7.6 μM), para -chloro (‘6343, 2.1 μM) 309 substitutions have different effects despite similar sterics. Fluoro analog ‘6404 (5.5 μM) shows weaker 310 binding than the ‘chloro (‘6343), likely reflecting weaker non -covalent interactions. e. Nitrogen 311 substitutions (C→N) in ‘3169, ‘6627, and ‘8675 all pay increased desolvation penalties. Compared to 312 their shared parent (8.8 μM), ‘8675 (meta -N) improved potency nearly 2-fold (4.5 μM), ‘6627 (8.7 μM) 313 showed little improvement, and ‘3169 (10 μM) is slightly worsened. 314 315 In optimizing chemical probes and leads , compound pharmacokinetics (PK) is as 316 important as molecular target affinity/potency 35,36. We thus explored how the small perturbation 317 analogs affected the PK properties versus those of the parent ligands, and how these changes 318 related alterations of affinity/potency. We measured the following in vitro PK properties: 319 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 16 metabolic stability in microsomes (half-life), plasma stability (half-life), plasma protein binding 320 (fraction unbound), solubility (uM), hERG inhibition (IC50, uM) , and membrane permeability 321 (PAMPA assay) (cm/s) of both the parent ligands and their small change analogs (Figure 4, 322 Supplementary Table 3-8). While these are all in vitro measurements, they are proxies that 323 are considered predictive of in vivo PK37-39. Some of the small change analogs showed 324 meaningful improvements relative to the parent ligands for each of these properties (Extended 325 Data Table 2). For example, 12.4% of the analogs had ≥3-fold improvement in microsomal 326 stability, while 20.8% of them had a >3-fold improvement in plasma stability, versus the parents. 327 For 13.8% of the analogs, solubility improved by ≥3-fold, while 6.2% of them improved >3-fold 328 in permeability. Examining the fraction unbound in plasma, 13.2% of the analogs improved ≥3-329 fold. These findings illustrate that each property is amenable to meaningful optimization even 330 by simple changes. We also observed a substantial fraction of analogs with ≥3 -fold decreases 331 in these same properties (Extended Data Figure 6). Considering what are perhaps the three 332 most impactful in vitro PK measurements, microsomal stability, fraction unbound, and 333 permeability, 36% of the analogs suffered >3 -fold losses in at least one of these , almost all 334 (96%) suffered at least some deterioration by the same criterion, and 37% deteriorated in all 335 three in vitro PK properties. 336 337 If many of the small perturbation analogs improved substantially in individual PK 338 properties, none of those with >10-fold increased potency improved or even maintained more 339 than one of what we might consider the three most important ones: metabolic stability, PAMPA 340 permeability, and fraction unbound in the plasma (Figure 4.a, Extended Data Figure 7); most 341 were worse by all three properties. Indeed, the changes in in vitro PK were, at best, orthogonal 342 to affinity/potency fold-change (Figure 4b), and several properties were anti-correlated with it. 343 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 17 This is born out on an atom -by-atom level: those atoms that most often improved 344 affinity/potency substantially, like methyls and chloros (Figure 2a), were least associated with 345 improved pharmacokinetic properties, such as microsomal stability or fraction unbound (Figure 346 4c, Extended Data Figure 8). Meanwhile, those atoms least likely to improve affinity/potency 347 substantially, like aromatic C N, were most likely to improve microsomal stability or fraction 348 unbound. Density and cumulative distribution plots illustrate this divergence in effect (Extended 349 Data Figure 8), with distinct shifts in property distributions across modifications. More broadly, 350 most of the pharmacokinetic properties were largely independent of one another, with no 351 correlations exceeding |ρ| > 0.5 and those that were statistically significant having Spearman ρ 352 values ranging only from |0.16| to |0.38|. What emerges is a frustrated landscape for ligand 353 optimization, with affinity/potency improvements counterbalanced by often worsening 354 pharmacokinetics. While this trade-off has been noted in medicinal chemistry, 40,41 this set 355 quantifies its systematic impact. 356 357 Fig.4 A frustrated landscape for ligand optimization. a, Fold-changes in pharmacokinetic 358 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 18 properties for the 29 analogs that were >10-fold improved in affinity/potency versus a parent. b, 359 Correlations between affinity/potency fold changes and those of the other PK properties. c, The 360 fraction of single atom changes that were >3x fold improved in each parameter. Error bars represent 361 95% confidence intervals, estimated from 10,000 iterations of bootstrap resampling. 362 363 The current benchmark only comprises several hundred analogs. If we could predict the 364

of the small perturbations, rather than having to experimentally test them, the set could 365 be much extended. We therefore asked how well modern methods could predict the relative 366 affinity/potency and pharmacokinetic changes we observed. In blinded experiments, we used 367 free energy perturbation , arguably the highest level of theory available to the field, with the 368 program FEP+ (Schrodinger, New York) to predict the changes in activity between the parents 369 and analogs (Figure 5 a-c). Similarly, we asked how well we could predict in vitro 370 pharmacokinetic changes between the parents and analogs, using widely accessible tools 371 (Figure 5d). 372 373 There was a strong overall correlation between the FEP+-predicted and experimentally 374 measured binding free energies (ΔG), with most predictions (190) falling within ±2 kcal/mol of 375 the experimental values, and 131 within ±1 kcal/mol, while the set had a mean unsigned error 376 (MUE) to ΔG as 1.07 kcal/mol (Figure 5a,b). Of the 18 parents, the FEP+ predictions on 14 had 377 ΔG MUE below 1.2 kcal/mol (Supplementary Table 9). If one considers the correlations 378 qualitatively, of the 35 analogs predicted to lose affinity by 1 kcal/mol or more by FEP+ (i.e., in 379 the range of confident prediction), 20 did lose affinity/potency by experiment. Meanwhile, of the 380 34 predicted to increase in affinity/potency by more than 1 kcal/mol by FEP +, 14 did so 381 experimentally while 12 were experimentally neutral and 8 instead showed decreased potency. 382 Categorization with a 1 kcal/mol threshold reveals a significant association between the 383 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 19 direction and strength of FEP+ predictions and experimentally observed affinity/potency 384 changes (² of 36.4 and a p-value < 0.001) (Supplementary Table10). 385 386 It is useful to note several of the challenges facing accurate FEP predictions for this 387 series, which future work may share. The predictive power was likely weakened by our inability 388 to measure affinity/potency for weak analogs, diminishing the overall range of affinities and 389 making it impossible to measure affinities predicted to be very low by FEP +. We used an 390 automated workflow to generate input poses for the calculations; on visual inspection, most of 391 the outliers were caused by suboptimal initial poses of analogs. The accuracy of FEP 392 calculations is strongly impacted by the quality of the ligand -complexed structures42, and the 393 FEP+ predictions here began with docking poses, which are less reliable than experimental 394 structures. For the 25 analog structures that we determine d by crystallography (for Mac1 395 analogs) and by cryoEM (for an 2a analog, below), there were many cases where the 396 predicted and experimental poses closely superimposed, and for these the FEP+ energies were 397 largely consistent with the experimental measurements (Figure 5C). 398 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 20 399 Figure 5: Accuracy of computational predictions of analog relative affinity and in vitro 400 pharmacokinetics. a. Comparison of FEP-predicted and experimentally measured binding free 401 energies (ΔG). Dashed lines indicate ±1 and ±2 kcal/mol error margins. b. Correlation between 402 FEP-predicted and experimental measured relative binding free energy changes (ΔΔG). c. An 403 example of good structural alignment between the predicted (cyan) and experimental (grey) 404 poses, accompanied by a small error in the predicted ΔΔG. d. Heatmap of Spearman 405 correlations between AI-predicted and experimentally measured fold changes across six in vitro 406 PK properties, evaluated using three models. Color intensity indicates correlation strength; 407 significance: p < 0.05 (*), < 0.01 (**), < 0.001 (***). 408 409 Recently, AI-based platforms have been developed to facilitate ligand optimization by 410 predicting ADMET properties with increasing accuracy and scale, guiding experiment43,44. We 411 compared the ADMET predictions from ADMET -AI, ADMETLab 3.0, and Deep -PK45-47 to the 412 experimentally measured values. The predictions of all three were significantly correlated with 413 experimental measurements of plasma fraction unbound and with permeability, while 414 microsomal stability , plasma stability and solubility, which are widely considered difficult to 415 predict48,49, were essentially uncorrelated. Correlations in the 0. 4 to 0.5 range, as they often 416 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 21 were for fraction unbound and permeability, are strong enough to help guide compound 417 selection in many circumstances, and these open -access tools may have broad impact. Still, 418 even with a correlation of 0. 53 in fraction unbound, the highest we observed, 28 analogs that 419 had substantially increased fraction unbound experimentally were predicted to have lower 420 levels than their parents, and 16 analogs that had higher protein binding than their parents 421 experimentally were predicted to have less. Taken together, while the FEP and ADMET 422 predictions can profitably model the effects of the small perturbation analogs, their correlations 423 with experiment are not yet strong enough to allow this set to be extended by calculation alone. 424 425 Multi-parameter optimization for an improved analgesic. Notwithstanding the 426 frustrated landscape for ligand affinity /potency and pharmacokinetic improvement, medicinal 427 chemists can often navigate the multi-parameter optimization to overall improve the properties 428 of lead molecules. In this study, too, there were analogs that improved in affinity/potency with 429 only modest sacrifice in pharmacokinetics (PK). Among these was an analog of an 2a parent, 430 compound ‘4905, where a single methyl group substitution improved potency 52-fold (Figure 431 6a, b). To understand this improvement at atomic resolution, we determined the structure of 432 ‘4905/alpha2a/G protein complex by single particle cryoEM to 2.8 Å resolution. Encouragingly, 433 the experimental structure of ‘4905 superposed closely with the docking prediction for the 434 parent ‘3629 and with the experimental structure of a previously described lead member of this 435 family of alpha2a agonists, ‘ 9087, suggesting that the methyl perturbation did not change the 436 orientation of the analog, and that this family may be understood in the same structural context. 437 From the cAMP dose –response curves, we obser ved a slight upward deflection at high concentrations 438 for ‘4905. Consistent with this, PTX -treated cAMP assays showed a residual, concentration -dependent 439 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 22 increase in cAMP for ‘4905 but not 3629 (Supplementary Figure 6), suggesting secondary Gs coupling 440 by the ligand. Structurally, the extra methyl in 4905 fills a hydrophobic subpocket and shifts Y409 6.55 441 (implicated in Gs activation), which may bias the receptor toward conformations permissive for Gs 442 engagement. Improved steric complementarity with residues in TM5 and TM3, and a closer 443 distance between the partly cationic bridging nitrogen with the crucial Asp128 on TM3, helps to 444 explain the improved potency of ‘4905 over its parent. 445 446 While the improved potency of ‘4905 came at a cost to its in vitro pharmacokinetics, the 447 effects were relatively modest: Compared to its parent, ‘3629, the fraction unbound decreased 448 by 57% and the microsomal half -life was reduced by 41%. (Supplementary Table 3-8). The 449 in vivo PK of ‘4905 qualitatively reflected the in vitro results, with CSF concentrations—a proxy 450 for fraction unbound in the brain50—reduced by about 36%, while the half-life was reduced from 451 196 to 15.2 minutes (Supplementary Table 3-8). These size of these effects suggested that 452 they might be outweighed by the analog’s great increase in activity. Accordingly, we compared 453 the in vivo analgesia conferred by ‘4905 to its parent ‘3629 and to PS75, the most potent 454 analgesic known for this family of alpha2a agonists 51. Despite the deterioration in 455 pharmacokinetic properties, in mouse nociception assays including tail flick, neuropathic pain 456 (SNI), and hot-plate the new analog ‘ 4905 was far more potent than the parent ‘3629 and at 457 least as potent as PS75 (Figure 6 c-e). If ligand optimization is a frustrated landscape, it 458 remains a navigable one as is well-known to medicinal chemistry 41 . In addition to beginning to 459 provide a random background for such efforts, this study supports a strategy by which that 460 landscape may be reconnoitered21,24. 461 462 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 23 463 Figure 6. A single-atom addition to an alpha2a receptor agonist creates a more potent 464 analgesic in vivo. a. Docking pose of parent ‘3629 superposed on the cryoEM structure of analog 465 ‘4905, which differs from ‘3629 by the addition of a single methyl group. 466 b. ‘4905 is 50-fold more potent than the parent ‘3629 in receptor activation in vitro. We note that the 467 apparent reduced efficacy through Gi coupling at concentrations greater than 1 nM likely reflects the 468 activation of Gs. c-g. ‘4905 is more potent than the parent ‘3629 in c-e. Reflex pain as measured in 469 tail flick response to heat; f. Reversal of mechanical allodynia in a neuropathic pain model (SNI). g. 470 Efficacy in reducing thermal acute pain assessed via the hotplate test. 471 472 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 24

473 While ligand optimization is at the heart of drug discovery, quantifying the success of 474 strategies to modify ligands has been difficult owing to a lack of a random background 475 comparison set. Four key observations emerge from this study. First, systematic and unbiased 476 small perturbations to 18 different scaffolds revealed that 11.2% of the analogs improved >10-477 fold in affinity /potency versus the parent (Table 1); “magic methyls” 25,52 were common in this 478 set. These improvements were observed against five of six receptors targeted, across physical 479 properties of target and ligand, and across five orders of magnitude of parent affinity/potency. 480 This begins to establish a background expectation for affinity /potency optimization against 481 which methods and design s may be compared . It may also suggest a simple, unbiased 482 approach for achieving substantial jumps in affinity/potency, improvements that will not always 483 be obvious from structural analyses or even high -level simulations (Figures 2 and 5) but may 484 complement them. Second, improved affinity/potency typically came at the cost of ligand 485 pharmacokinetics (PK), with plasma fraction unbound, stability to liver microsomes, and 486 permeability—among other terms —declining as analog affinity /potency increased. The 487 individual in vitro pharmacokinetic terms were orthogonal to each other, and several were anti-488 correlated with affinity/potency improvement (Figure 4). This suggests a frustrated landscape 489 for ligand optimization. Although the qualitative challenges and trade-offs in ligand optimization 490 are known to medicinal chemists 6,40,41, this study begins to quantify them at scale and across 491 a range of target classes and ligand properties . Third, computational prediction of 492 affinity/potency and PK had some success anticipating the trends we observed ( Figure 5) 493 implying that they may even tually guide compound selection . Still, their correlations with 494 experiment remained loose enough to preclude replacing systematic exploration and testing of 495 molecules as this background set is further expanded , something seen with other potency 496 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 25 prediction methods53. Fourth, while potency trades-off against PK, it is possible to find analogs 497 where it rises sufficiently to overcome drops in exposure and half-life. A testament to this is the 498 in vivo potency of the 2a analog ‘4905, which despite suffering these trade-offs nevertheless 499 is the most potent analgesic in vivo in a large series, and is certainly far more potent in vivo 500 than its parent. It may be interesting to model the implications of the orthogonal and anti -501 correlated terms here to understand how best to navigate this multi -parameter optimization 502 problem. 503 504 At first glance, this study is reminiscent of learnings from classical medicinal chemistry 505 where conservative and ligand-based approaches like Topliss Trees9 dominated the field. But 506 this would be to misremember the classic age of medicinal chemistry on two counts. First, the 507 systematic small perturbations explored here were rarely used then, partly because the building 508 blocks and reactions to support them were lacking. The ability to systematically explore small 509 perturbations, across starting scaffolds and targets, reflects the advances in synthetic chemistry 510 over the last 50 years and also the new building-block approaches that have expanded readily 511 available compounds 54,55. Second, the logic behind Topliss Trees and other quantitative 512 structure-activity approaches9,10 were anchored in ligand physical chemistry without reference 513 to receptor structure. For instance, in Topliss Trees the success of methyl and chloro 514 derivatives were thought to teach opposite lessons because of their different el ectronic effects 515 on the ligand, whereas the steric and non-polar properties of the two groups will seem similar 516 from the view “inside the receptor”52,56.Indeed, this is what we find in the similar affinity/potency 517 effects of the two substitutions. Meanwhile, the advantage of a random background, so integral 518 to molecular biology, would have seemed foreign to classical medicinal chemistry. Thus, while 519 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 26 a strategy of systematic small perturbations may smack of the “methyl, ethyl, propyl, butyl, futile” 520 approach with which classic medicinal chemistry has been tarred 57,58, it has a different basis in 521 theory and reflects new synthetic opportunities. 522 523 Certain limitations of this study merit airing. At 257 analogs and six targets, our set 524 remains small and inevitably, if unintentionally, biased. We do not pretend that three class A 525 GPCRs, one transporter, and two soluble enzymes adequately represent pharmacologically 526 relevant targets, nor that the 18 parent ligands can adequately represent ligand space. 527 Expanding the analysis to additional targets and ligand classes might reveal different trends. 528 We note that the six targets do recognize a range of chemotypes, including large polar co -529 factors (Mac1) , large hydrophobic lipids (CB2), anionic -lactams ( AmpC), cationic 530 neurotransmitters (2a and SERT), and peptides (µOR). Experimentally, we assessed activity 531 changes using radioligand competition binding assays (SERT, 2a), second messenger assays 532 (2a, CB2, and MOR), and binding and enzyme activity (Mac1 and AmpC) assays. While ligand 533 competition correlates with Ki, agonist EC50s are affected by receptor expression. Because we 534 compared the relative activities of analogs versus their parents , controlled for receptor 535 expression (Supplementary Figure 4), and measured effects in a low expression level domain 536 (Supplementary Figure 5), the effects of receptor expression on relative activity should be 537 modest, though they cannot be completely discounted. For these and related reasons, changes 538 of EC50 between agonist parent and analogs cannot be read as changes in affinity the way that 539 changes in Ki can be, though EC 50 changes remain the relevant metric for agonist activity. 540 Overall, the impact of these effects may be inspected case by case in full co ncentration-541 response for any parent ligand pair with >3 -fold improvement in activity (Supplementary 542 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 27 Figure 2). In the area of in vitro pharmacokinetics, our use of liver microsomes rather than 543 hepatocytes to measure metabolic stability meant that some types of metabolism were missed, 544 including glucuronidation. This might make groups like phenolic hydroxyls seem less labile 545 than they would be in vivo. More broadly, for all but two molecules we only measured in vitro 546 not in vivo pharmacokinetics. Whereas the in vitro measurements qualitatively anticipated the 547 in vivo effects where we did measure them, and are widely used in ligand optimization, their 548 quantitative prediction of in vivo behavior is only approximate. 549 550 These limitations should not obs cure the main observations of this study. Over 11.2% 551 of systematic and unbiased small perturbations improved analog affinity/potency >10-fold, 552 beginning to establish a background expectation for the likelihood of substantial ligand 553 affinity/potency improvement and a systematic approach to doing so 21,24. Balancing this was 554 a concomitant deterioration in ligand pharmacokinetics, which will lower the exposure and half-555 life of a ligand in vivo, counteracting the improvements in affinity/potency. Navigating this multi-556 parameter space is at the heart of medicinal chemistry; this study supports to the development 557 of quantitative models to do so. 558 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 28

559 1 Hughes, J. P., Rees, S., Kalindjian, S. B. & Philpott, K. L. Principles of early drug discovery. 560 Brit J Pharmacol 162, 1239-1249, doi:10.1111/j.1476-5381.2010.01127.x (2011). 561 2 Bleicher, K. H., Bohm, H. J., Muller, K. & Alanine, A. I. Hit and lead generation: beyond 562 high-throughput screening. Nat Rev Drug Discov 2, 369-378, doi:10.1038/nrd1086 (2003). 563 3 Paul, S. M. et al. How to improve R&D productivity: the pharmaceutical industry's grand 564 challenge. Nature Reviews Drug Discovery 9, 203-214, doi:10.1038/nrd3078 (2010). 565 4 Lipinski, C. A. Lead - and drug -like compounds: the rule -of-five revolution. Drug Disc ov 566 Today Technol 1, 337-341, doi:10.1016/j.ddtec.2004.11.007 (2004). 567 5 Stumpfe, D., Hu, H. & Bajorath, J. Evolving Concept of Activity Cliffs. ACS Omega 4, 14360-568 14368, doi:10.1021/acsomega.9b02221 (2019). 569 6 Leeson, P. D. & Springthorpe, B. The influence o f drug-like concepts on decision-making 570 in medicinal chemistry. Nat Rev Drug Discov 6, 881-890, doi:10.1038/nrd2445 (2007). 571 7 Wunberg, T. et al. Improving the hit-to-lead process: data-driven assessment of drug -like 572 and lead -like screening hits. Drug Disco v Today 11, 175 -180, doi:10.1016/S1359 -573 6446(05)03700-1 (2006). 574 8 Jorgensen, W. L. Efficient drug lead discovery and optimization. Acc Chem Res 42, 724-575 733, doi:10.1021/ar800236t (2009). 576 9 Topliss, J. G. Utilization of Operational Schemes for Analog Synthes is in Drug Design. J 577 Med Chem 15, 1006-&, doi:DOI 10.1021/jm00280a002 (1972). 578 10 Hansch, C. Citation Classic - Rho-Sigma-Pi-Analysis - a Method for the Correlation of 579 Biological-Activity and Chemical-Structure. Cc/Life Sci, 18-18 (1982). 580 11 Dong, J. et al. ADMETlab: a platform for systematic ADMET evaluation based on a 581 comprehensively collected ADMET database. J Cheminformatics 10, doi:ARTN 29 582 10.1186/s13321-018-0283-x (2018). 583 12 Wang, L. et al. Accurate and Reliable Prediction of Relative Ligand Binding Po tency in 584 Prospective Drug Discovery by Way of a Modern Free -Energy Calculation Protocol and 585 Force Field. J Am Chem Soc 137, 2695-2703, doi:10.1021/ja512751q (2015). 586 13 Klimovich, P. V., Shirts, M. R. & Mobley, D. L. Guidelines for the analysis of free energy 587 calculations. J Comput Aided Mol Des 29, 397 -411, doi:10.1007/s10822 -015-9840-9 588 (2015). 589 14 Hajduk, P. J. & Sauer, D. R. Statistical analysis of the effects of common chemical 590 substituents on ligand potency. J Med Chem 51, 553-564, doi:10.1021/jm070838y (2008). 591 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 29 15 Kramer, C., Kalliokoski, T., Gedeck, P. & Vulpetti, A. The Experimental Uncertainty of 592 Heterogeneous Public Data. J Med Chem 55, 5165-5173, doi:10.1021/jm300131x (2012). 593 16 Booker, T. R., Jackson, B. C. & Keightley, P. D. Detecting positive selection in the genome. 594 Bmc Biol 15, doi:ARTN 98 10.1186/s12915-017-0434-y (2017). 595 17 Pritchard, J. K. & Cox, N. J. The allelic architecture of human disease genes: common 596 disease - common variant ... or not? Hum Mol Genet 11, 2417 -2423, doi:DOI 597 10.1093/hmg/11.20.2417 (2002). 598 18 Manolio, T. A. et al. Finding the missing heritability of complex diseases. Nature 461, 747-599 753, doi:10.1038/nature08494 (2009). 600 19 Fowler, D. M. & Fields, S. Deep mutational scanning: a new style of protein science. Nat 601

11, 801-807, doi:10.1038/Nmeth.3027 (2014). 602 20 Wells, J. A. Systematic Mutational Analyses of Protein Protein Interfaces. Method Enzymol 603 202, 390-411 (1991). 604 21 Pennington, L. D., Aquila, B. M., Choi, Y., Valiulin, R. A. & Muegge, I. Positional Analogue 605 Scanning: An Effective Strategy for Multiparameter Optimization in Drug Design. J Med 606 Chem 63, 8956-8976, doi:10.1021/acs.jmedchem.9b02092 (2020). 607 22 Gaulton, A. et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic 608 Acids Res 40, D1100-1107, doi:10.1093/nar/gkr777 (2012). 609 23 Zdrazil, B. et al. The ChEMBL Database in 2023: a drug dis covery platform spanning 610 multiple bioactivity data types and time periods. Nucleic Acids Res 52, D1180 -D1192, 611 doi:10.1093/nar/gkad1004 (2024). 612 24 Muegge, I. & Hu, Y. In Silico Positional Analogue Scanning with Amber GPU -TI. J Chem 613 Inf Model 62, 4448-4459, doi:10.1021/acs.jcim.2c00860 (2022). 614 25 Schönherr, H. & Cernak, T. Profound Methyl Effects in Drug Discovery and a Call for New 615 C H Methylation Reactions. Angewandte Chemie International Edition 52, 12256-12267 616 (2013). 617 26 Cramer, J., Sager, C. P. & Ernst, B. Hydroxyl groups in synthetic and natural -product-618 derived therapeutics: a perspective on a common functional group. J Med Chem 62, 8915-619 8930 (2019). 620 27 Chiodi, D. & Ishihara, Y. “Magic chloro”: profound effects of the chlorine atom in drug 621 discovery. J Med Chem 66, 5305-5331 (2023). 622 28 Gillis, E. P., Eastman, K. J., Hill, M. D., Donnelly, D. J. & Meanwell, N. A. Applications of 623 fluorine in medicinal chemistry. J Med Chem 58, 8315-8359 (2015). 624 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 30 29 Hopkins, A. L., Groom, C. R. & Alex, A. Ligand efficiency: a useful metric for lead selection. 625 Drug Discov Today 9, 430-431, doi:10.1016/S1359-6446(04)03069-7 (2004). 626 30 Zhu, T. et al. Hit identification and optimization in virtual screening: practical 627 recommendations based on a critical literature analysis. J Med Chem 56, 6560 -6572, 628 doi:10.1021/jm301916b (2013). 629 31 Kuntz, I. D., Chen, K., Sharp, K. A. & Kollman, P. A. The maximal affinity of ligands. Proc 630 Natl Acad Sci U S A 96, 9997-10002, doi:10.1073/pnas.96.18.9997 (1999). 631 32 Halgren, T. A. Identifying and Char acterizing Binding Sites and Assessing Druggability. J 632 Chem Inf Model 49, 377-389, doi:10.1021/ci800324m (2009). 633 33 Thomas, M., Bender, A. & de Graaf, C. Integrating structure -based approaches in 634 generative molecular design. Curr Opin Struct Biol 79, 10255 9, 635 doi:10.1016/j.sbi.2023.102559 (2023). 636 34 Flowers, J. et al. Expanding automated multiconformer ligand modeling to macrocycles and 637 fragments. Elife 14, doi:10.7554/eLife.103797 (2025). 638 35 Di, L., Kerns, E. H. & Carter, G. T. Drug -like property concepts in pharmaceutical design. 639 Curr Pharm Des 15, 2184-2194, doi:10.2174/138161209788682479 (2009). 640 36 Morgan, P. et al. Can the flow of medicines be improved? Fundamental pharmacokinetic 641 and pharmacological principles toward improving Phase II survival. Drug Discov Today 17, 642 419-424, doi:10.1016/j.drudis.2011.12.020 (2012). 643 37 Houston, J. B. Utility of in vitro drug metabolism data in predicting in vivo metabolic 644 clearance. Biochem Pharmacol 47, 1469-1479, doi:10.1016/0006-2952(94)90520-7 (1994). 645 38 Iwatsubo, T. et al. Prediction of in vivo drug metabolism in the human liver from in vitro 646 metabolism data. Pharmacol Ther 73, 147 -171, doi:10.1016/s0163 -7258(96)00184-2 647 (1997). 648 39 Amidon, G. L., Lennernas, H., Shah, V. P. & Crison, J. R. A theoretical basis for a 649 biopharmaceutic drug classification: the correlation of in vitro drug product dissolution and 650 in vivo bioavailability. Pharm Res 12, 413-420, doi:10.1023/a:1016212804288 (1995). 651 40 Hopkins, A. L., Keseru, G. M., Leeson, P. D., Rees, D. C. & Reynolds, C. H. The role of 652 ligand efficiency metrics in drug discovery. Nat Rev Drug Discov 13, 105 -121, 653 doi:10.1038/nrd4163 (2014). 654 41 Young, R. J. & Leeson, P. D. Mapping the Efficiency and Physicochemical Trajectories of 655 Successful Optimizations. J Med Chem 61, 6421 -6467, 656 doi:10.1021/acs.jmedchem.8b00180 (2018). 657 42 Rocklin, G. J., Mobley, D. L., Dill, K. A. & Hunenberger, P. H. Calculat ing the binding free 658 energies of charged species based on explicit -solvent simulations employing lattice -sum 659 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 31

an accurate correction scheme for electrostatic finite -size effects. J Chem Phys 660 139, 184103, doi:10.1063/1.4826261 (2013). 661 43 Kim, H. et al. Artificial Intelligence in Drug Discovery: A Comprehensive Review of Data -662 driven and Machine Learning Approaches. Biotechnol Bioprocess Eng 25, 895 -930, 663 doi:10.1007/s12257-020-0049-y (2020). 664 44 Kumar, A., Kini, S. G. & Rathi, E. A Recent Appraisal of A rtificial Intelligence and In Silico 665 ADMET Prediction in the Early Stages of Drug Discovery. Mini Rev Med Chem 21, 2788-666 2800, doi:10.2174/1389557521666210401091147 (2021). 667 45 Fu, L. et al. ADMETlab 3.0: an updated comprehensive online ADMET prediction platform 668 enhanced with broader coverage, improved performance, API functionality and decision 669 support. Nucleic Acids Res 52, W422-W431, doi:10.1093/nar/gkae236 (2024). 670 46 Myung, Y., de Sa, A. G. C. & Ascher, D. B. Deep -PK: deep learning for small molecule 671 pharmacokinetic and toxicity prediction. Nucleic Acids Res 52, W469 -W475, 672 doi:10.1093/nar/gkae254 (2024). 673 47 Swanson, K. et al. ADMET-AI: a machine learning ADMET platform for evaluation of large-674 scale chemical libraries. Bioinformatics 40, doi:10.1093/bioinformatics/btae416 (2024). 675 48 Bergstrom, C. A., Wassvik, C. M., Johansson, K. & Hubatsch, I. Poorly soluble marketed 676 drugs display solvation limited solubility. J Med Chem 50, 5858 -5862, 677 doi:10.1021/jm0706416 (2007). 678 49 Obach, R. S. Prediction of human clearance of twenty-nine drugs from hepatic microsomal 679 intrinsic clearance data: An examination of in vitro half -life approach and nonspecific 680 binding to microsomes. Drug Metab Dispos 27, 1350-1359 (1999). 681 50 Friden, M. et al. Structure-brain exposure relationships in rat and human using a novel data 682 set of unbound drug concentrations in brain interstitial and cerebrospinal fluids. J Med 683 Chem 52, 6233-6243, doi:10.1021/jm901036q (2009). 684 51 Fink, E. A. et al. Structure-based discovery of nonopioid analgesics actin g through the 685 alpha(2A)-adrenergic receptor. Science 377, eabn7065, doi:10.1126/science.abn7065 686 (2022). 687 52 Leung, C. S., Leung, S. S., Tirado-Rives, J. & Jorgensen, W. L. Methyl effects on protein -688 ligand binding. J Med Chem 55, 4489-4500, doi:10.1021/jm3003697 (2012). 689 53 Sindt, F., Bret, G. & Rognan, D. On the Difficulty to Rescore Hits from Ultralarge Docking 690 Screens. J Chem Inf Model 65, 5553-5566, doi:10.1021/acs.jcim.5c00730 (2025). 691 54 Lyu, J. et al. Ultra-large library docking for discovering new chemotypes. Nature 566, 224-692 +, doi:10.1038/s41586-019-0917-9 (2019). 693 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 32 55 Moroz, Y., Chuprina, A. & Mykytenko, D. Enamine REAL DataBase - an instrumental and 694 practical vehicle for charting new regions of the relevant drug discovery chemical space . 695 Abstr Pap Am Chem S 251 (2016). 696 56 Verteramo, M. L. et al. Interplay of halogen bonding and solvation in protein-ligand binding. 697 iScience 27, 109636, doi:10.1016/j.isci.2024.109636 (2024). 698 57 Krimmer, S. G., Betz, M., Heine, A. & Klebe, G. Methyl, ethyl, propyl, butyl: futile but not for 699 water, as the correlation of structure and thermodynamic signature shows in a congeneric 700 series of thermolysin inhibitors. ChemMedChem 9, 833-846, doi:10.1002/cmdc.201400013 701 (2014). 702 58 Wermuth, C. G. The Practice of Medic inal Chemistry. (Elsevier Science & Technology, 703 2008). 704 59 Gahbauer, S. et al. Iterative computational design and crystallographic screening identifies 705 potent inhibitors targeting the Nsp3 macrodomain of SARS -CoV-2. Proceedings of the 706 National Academy of Sciences 120, e2212931120 (2023). 707 60 Liu, F. et al. The impact of library size and scale of testing on virtual screening. Nature 708 Chemical Biology, 1-7 (2025). 709 61 Clopper, C. J. & Pearson, E. S. The use of confidence or fiducial limits illustrated in the 710 case of the binomial. Biometrika 26, 404-413 (1934). 711 62 Virtanen, P. et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nat 712

17, 261-272 (2020). 713 63 Lu, C. et al. OPLS4: Improving Force Field Accuracy on Challenging Regimes of Chemical 714 Space. J Chem Theory Comput 17, 4291-4300, doi:10.1021/acs.jctc.1c00302 (2021). 715 64 Balius, T. E. et al. Testing inhomogeneous solvation theory in structure -based ligand 716 discovery. Proceedings of the National Academy of Sciences 114, E6839-E6846 (2017). 717 65 Singh, I. et al. Structure-based discovery of conformationally selective inhibitors of the 718 serotonin transporter. Cell 186, 2160-2175. e2117 (2023). 719 66 Vigneron, S. F. et al. Docking 14 Million Virtual Isoquinuclidines against the μ and κ Opioid 720 Receptors Reveals Dual Antagonists–Inverse Agonists with Reduced Withdrawal Effects. 721 ACS Central Science (2025). 722 67 Hua, T. et al. Activation and signaling mechanism revealed by cannabinoid receptor -Gi 723 complex structures. Cell 180, 655-665. e618 (2020). 724 68 Hua, T. et al. Crystal structures of agonist-bound human cannabinoid receptor CB1. Nature 725 547, 468-471 (2017). 726 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 33 69 Manglik, A. et al. Structure-based discovery of opioid analgesics with reduced side effects. 727 Nature 537, 185-190 (2016). 728 70 Yung-Chi, C. & Prusoff, W. H. Relationship between the inhibition constant (KI) and the 729 concentration of inhibitor which causes 50 per cent inhibition (I50) of an enzymatic reaction. 730 Biochemical pharmacology 22, 3099-3108 (1973). 731 71 Schuller, M. et al. Fragment binding to the Nsp3 macrodomain of SARS -CoV-2 identified 732 through crystallographic screening and computational docking. Science advances 7, 733 eabf8711 (2021). 734 72 Maeda, S. et al. Development of an antibody fragment that stabilizes GPCR/G -protein 735 complexes. Nat Commun 9, 3712, doi:10.1038/s41467-018-06002-w (2018). 736 73 Goddard, T. D. et al. UCSF ChimeraX: Meeting modern challenges in visualization and 737 analysis. Protein Sci 27, 14-25, doi:10.1002/pro.3235 (2018). 738 74 Liebschner, D. et al. Macromolecular structure determination using X -rays, neutrons and 739 electrons: recent developments in Phenix. Acta Crystallogr D Struct Biol 75, 861 -877, 740 doi:10.1107/S2059798319011471 (2019). 741 75 Emsley, P., Lohkamp, B., Scott, W. G. & Cowtan, K. Features and development of Coo t. 742 Acta Crystallogr D Biol Crystallogr 66, 486-501, doi:10.1107/S0907444910007493 (2010). 743 76 Schuller, M. et al. Fragment binding to the Nsp3 macrodomain of SARS -CoV-2 identified 744 through crystallographic screening and computational docking. Sci Adv 7, 745 doi:10.1126/sciadv.abf8711 (2021). 746 77 Correy, G. J. et al. Exploration of structure -activity relationships for the SARS -CoV-2 747 macrodomain from shape -based fragment linking and active learning. Sci Adv 11, 748 eads7187, doi:10.1126/sciadv.ads7187 (2025). 749 750 751 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 34

752 ChEMBL database molecular pairs. From the PostgreSQL version of CHEMBL3423 753 (https://doi.org/10.6019/CHEMBL.database.34), we filtered the database for compounds with 754 activities reported against a single protein, with either Ki, Kd, IC50 or EC50 as the activity type. 755 To be considered a molecular pair, compounds had to have the same assay ID, the same 756 target ID, the same activity type, and the same reference document (publication) ID, in 757 addition to having a parent-analog relationship as defined in this work, meaning a C-H group 758 replaced by C-OH, C-F, C-Cl, C-Br, C-CH3 or an aromatic carbon replaced by a nitrogen. This 759 left us with 191,732 parent-analog pairs. 760 761 Analog enumeration and synthesis. Eighteen parent compounds were selected from 762 previously published literature51,59,60 or datasets 763 (https://asapdiscovery.org/outputs/molecules/#ASAP-SARS-COV-2-NSP3-MAC1) based on 764 their known binding activity, structural relevance, or representation of diverse chemical 765 scaffolds. Starting from a parent compound, we used RDKit (www.rdkit.org) to identify all C-H 766 bonds and iteratively replace the hydrogen atom with a methyl, hydroxyl, fluoro, chloro or 767 bromo group. We also identified aromatic carbons with two heavy-atom neighbors and 768 replaced them with nitrogen. Every analog generated was represented as a canonical 769 isomeric SMILES string and added to a set to remove duplicates. For each of the 18 parents, 770 the full set of possible analogs that could be synthesized for < $400 for 10mg were ordered. 771 772 Confidence intervals and statistics. When reported, 95% confidence intervals were 773 derived from bootstrap resampling with 10,000 iterations. In cases where the observed 774 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 35 frequency of success is exactly 0, bootstrap will fail to give an upper bound for the interval. In 775 such cases, we used the Clopper-Pearson method to estimate an upper bound based on the 776 sample size.61 Pearson and Spearman correlations with their associated p-values were 777 computed using the scipy.stats module from the SciPy package62. 778 779 FEP simulations. These were conducted using FEP+ within the Schrödinger software 780 suite (versions 2025-2) with the OPLS4 force field63 and the modified SPC water model. The 781 default setting was used for the number of lambda windows selection where it depends on the 782 type of perturbations; charge-changing, core hopping, and all other perturbations have 24, 16, 783 and 12 lambda windows respectively. For alchemical transformations with charge changes, 784 the total charge of the simulation box was kept constant by transmuting a Na+ or Cl- ion to 785 water or vice versa. In addition, a 0.15 M concentration of NaCl was added to the simulation 786 box of charge-changing perturbations. For α2A, CB2, and SERT, the FEP+ membrane 787 protocol was applied where a POPC membrane was added to the system in simulation. All 788 other settings were kept default except the simulation time was extended from 5 ns to 10 ns. 789 790 The default FEP map generation protocol was used with the parent compound selected 791 as the biased node. In preparing proteins and ligands for FEP+, the Schrödinger protein 792 preparation workflow and LigPrep were used. The initial binding poses of parent compo unds 793 were from poses generated by DOCK3.8, and analogs were aligned to the parents with severe 794 steric clashes resolved using the FEP+ Pose Builder workflow. 795 796 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 36 Docking: Whereas no structural information was used in the design of analogs from 797 parent compounds, for FEP+ calculations and for post hoc structural analysis, we generated 798 ligand bound complexes of the parents and relevant ligands using DOCK.3.854,64. Ligands were 799 docked into the receptor binding site using grids prepared in previous studies51,59,60,65,66. 800 801 Assay selection. We use one consistent readout per parent series (no mixing within a 802 series). SERT and AmpC are reported as Ki; Mac1 is reported as IC50 from peptide 803 displacement (a scalable functional hydrolysis assay is unavailable); GPCR agonist series are 804 reported as EC50 and antagonist series as Ki, such that MOR is EC50 -only whereas α2A and 805 CB2 include EC50 and Ki depending on the parent series. Representative concentration –806 response curves and the corresponding Z values for each assay are provided in 807 Supplementary Figure 3. 808 809 Transporter assays for SERT Ki. SERT activity was measured using the 810 Neurotransmitter Transporter Uptake Assay Kit from Molecular Devices (Catalog #R8174), 811 following the manufacturer’s protocol with slight modifications as described previously65. 812 HEK293 cells stably expressing human SERT were plated in poly-L-lysine (PLL)-coated 384-813 well black, clear-bottom plates at a density of 15,000 cells in 40 µL per well, using DMEM 814 supplemented with 1% dialyzed FBS (dFBS). Cells were incubated overnight at 37°C with 5% 815 CO₂ to allow adherence and recovery. The following day, the medium was carefully aspirated, 816 and cells were incubated with 25 µL per well of test compound solutions prepared in assay 817 buffer (1× HBSS, 20 mM HEPES, pH 7.4, supplemented with 1 mg/mL BSA) for 30 minutes at 818 37°C. After drug treatment, 25 µL per well of dye solution (as provided in the kit) was added 819 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 37 directly to the wells, followed by an additional 30-minute incubation at 37°C. Fluoxetine (10 820 µM) was used as a positive control for SERT inhibition. Fluorescence was measured using 821 the FlexStation II microplate reader with excitation at 440 nm and emission at 520 nm. 822 Relative fluorescence units (RFUs) were exported and analyzed using GraphPad Prism 10.0 823 to calculate IC₅₀ values, from which Ki values were derived using the Cheng-Prusoff equation. 824 825 CB2 radioligand binding assay. CB2 receptor binding assays were performed using 826 membrane preparations from HEK293 cells stably expressing human CB2, following 827 previously published methods67,68. Membranes were resuspended in TME buffer containing 828 0.1% BSA (w/v) and 25 µg of membrane protein was added per well. The assay was 829 conducted using the radioligand [³H]CP-55,940 at a final concentration of 0.75 nM, prepared 830 in assay buffer. Nonspecific binding was defined in the presence of 5 µM unlabeled CP-831 55,940. Test compounds were applied at increasing concentrations to assess competition. 832 Reactions were incubated at 30 °C for 1 hour with gentle shaking. After incubation, samples 833 were transferred to Unifilter GF/B 96-well filter plates and filtered using a Packard Filtermate-834 196 cell harvester (PerkinElmer). Plates were washed four times with ice-cold wash buffer (50 835 mM Tris-HCl, 5 mM MgCl₂, 0.5% BSA, pH 7.4). Radioactivity bound to the filters was 836 quantified via liquid scintillation counting. Specific binding was calculated by subtracting 837 nonspecific from total binding. IC₅₀ and Ki values were calculated using nonlinear regression 838 in GraphPad Prism 9 using the Cheng-Prusoff equation. 839 840 α2A Receptor Binding Assay. α2A adrenergic receptor binding was performed using 841 membrane preparations from insect cells expressing human α2A receptors, as previously 842 described51. Membranes were incubated with increasing concentrations of test compounds 843 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 38 and 5 nM [³H]-rauwolscine in buffer containing 20 mM HEPES (pH 7.5) and 100 mM NaCl, at 844 room temperature for 2 hours. After incubation, samples were filtered onto GF/B filter plates, 845 washed with ice-cold buffer, and radioactivity was quantified by liquid scintillation counting. 846 IC₅₀ and Ki values were derived using nonlinear regression in GraphPad Prism. 847 848 849 GloSensor cAMP assay for α2A, CB2, and MOR. The GloSensor cAMP assay was 850 performed following the manufacturer’s instructions (Promega) with slight modifications as 851 previously described51,69. Briefly, wild-type (WT) human α2A, CB2, and MOR were cloned into 852 the pcDNA3.1 vector and co-transfected with the 22F cAMP GloSensor plasmid into 853 HEK293T cells cultured in 6-well plates. After 24 hours, cells were reseeded into 96-well white 854 plates in CO₂-independent medium and equilibrated with GloSensor cAMP reagent as per the 855 manufacturer's protocol. Cells were incubated for 1 hour at 37 °C followed by 1 hour at room 856 temperature. Where applicable, 10 μM forskolin was used to elevate basal cAMP levels for 857 assessing receptor-mediated inhibition. Serially diluted test compounds were added, and 858 luminescence signals were recorded using a PerkinElmer microplate reader. Data were 859 analyzed using GraphPad Prism 9.0 to calculate EC₅₀ or IC₅₀ values. 860 861 AmpC β-lactamase Inhibition Assay. The AmpC β-lactamase inhibition assay was 862 performed as previously described60. The candidate inhibitors were dissolved in DMSO (20 mM stock) 863 and diluted to maintain a constant 1% DMSO (v/v) in 50 mM sodium cacodylate buffer (pH 6.5). 864 Assays were performed in the presence of 0.01% Triton X-100 to reduce aggregation artifacts. AmpC 865 enzymatic activity was monitored spectrophotometrically using CENTA or nitrocefin as substrates. 866 Initial screening was performed at 200 µM, 100 µM, and 40 µM compound concentrations. Substrate 867 concentrations were selected based on known Km values to achieve defined [S]/Km ratios: for CENTA 868 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 39 ([S] = 50 µM, Km = 27.6 µM) and nitrocefin ([S] = 100 µM or 28 µM, Km = 180 µM). Reactions were 869 carried out in 96-well format on a BMG Labtech CLARIOstar plate reader, with substrate and enzyme 870 injected into wells containing the inhibitor, followed by kinetic measurement over 50 seconds. IC₅₀ 871 values were determined by fitting inhibition curves in GraphPad Prism using a fixed Hill coefficient of 1, 872 and Ki values were calculated using the Cheng-Prusoff equation70. 873 874 875 HTRF assay for Mac1. Binding of the compounds to Mac1 was assessed by the 876 displacement of an ADPr conjugated biotin peptide from His6-tagged protein using a HTRF- 877 based assay, as previously described71.The expression sequences used for SARS-CoV-2 878 Mac1 are listed below. All proteins were expressed and purified as described previously for 879 SARS-CoV-2 Mac171. Compounds were dispensed into ProxiPlate-384 Plus (PerkinElmer) 880 assay plates using an Echo 650 Liquid Handler (Beckman Coulter). Binding assays were 881 conducted in a final volume of 16 μl with 12.5 nM NSP3 Mac1 protein, 200 nM peptide 882 ARTK(Bio)QTARK(Aoa- RADP)S (Cambridge Peptides), 1:20000 Anti-His6-Eu3+ cryptate 883 (HTRF donor, PerkinElmer AD0402) and 1:500 Streptavidin-XL665 (HTRF acceptor, 884 PerkinElmer 610SAXLB) in assay buffer (25 mM 4-(2-hydroxyethyl)-1-piperazi‐ 885 neethanesulfonic acid (HEPES) pH 7.0, 20 mM NaCl, 0.05% bovine serum albumin and 886 0.05% Tween-20, the latter also to reduce aggregation artifacts). Assay reagents were 887 dispensed manually into plates using an electronic multichannel pipette. Mac1 and peptide 888 were preincubated for 30 min at room temperature before HTRF reagents were added. 889 Fluorescence was measured after a 1 hour incubation at room temperature using a Perkin 890 Elmer EnVision 2105-0010 Dual Detector Multimode microplate reader with dual emission 891 protocol (A = excitation of 320 nm, emission of 665 nm, and B = excitation of 320 nm, 892 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 40 emission of 620 nm). Compounds were tested in triplicate in a 14-point dose response. Raw 893 data were processed to give an HTRF ratio (channel A/B × 10,000), which was used to 894 generate IC50 curves using nonlinear regression using GraphPad Prism v.10.0.2 (GraphPad 895 Software, CA, USA). 896 897 hERG channel inhibition. hERG channel inhibition was evaluated using a Thallium 898 Flux assay on HEK293 cells stably expressing the human Ether-à-go-go Related Gene 899 (hERG) potassium channel. Cells were seeded at a density of 8000 cells per well in 384-well 900 poly-D-lysine–coated plates and incubated for 24 hours under standard conditions (37 °C, 5% 901 CO₂). The next day, a thallium-sensitive dye was added to the cells, followed by a 1-hour 902 incubation to ensure dye uptake. Test compounds were added to achieve a final 903 concentration of 30 μM in 0.5% DMSO, and the cells were incubated for an additional 30 904 minutes at room temperature. Subsequently, a stimulation buffer containing thallium was 905 added and fluorescence measurements were taken using a FLIPR Tetra system. Data were 906 collected every 3 seconds for 3 minutes (excitation: 470–495 nm, emission: 515–575 nm). 907 Fluorescence intensity over time was analyzed to calculate the area under the curve (AUC) 908 from which percentage inhibition was determined versus haloperidol at 100μM (positive 909 control) and vehicle (DMSO). 910 911 α2A Receptor Purification and Structure Determination. Wild-type human α2A 912 adrenergic receptor (α2AAR) was cloned into a pVL1392 vector with an N -terminal FLAG tag. 913 The construct was expressed in Spodoptera frugiperda (Sf9) insect cells using the BestBac 914 system. Cells at a density of 4 × 10⁶ cells/mL were infected with virus and incubated for 48 hours 915 at 27 °C. The receptor was solubilized and purified by FLAG affinity chromatography and size-916 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 41 exclusion chromatography in the presence of 10 μM compound ‘4905. Monomeric peak 917 fractions were concentrated used for G protein complex formation. GαoGβ1γ2 heterotrimeric G 918 proteins were expressed in Hi5 insect cells and purified using Ni² ⁺-affinity following detergent 919 solubilization and dephosphorylation. The final α 2AAR–GαoGβ1γ2–scFv1672 complex was 920 assembled in the presence of ‘4905 and purified by size -exclusion chromatography. Cryo-EM 921 grids were prepared using UltrAufoil R1.2/1.3 300 mesh grids and vitrified in liquid ethane. Data 922 were collected on a Titan Krios G3 electron microscope equipped with a K3 direct electron 923 detector. Image processing was performed using cryoSPARC, yielding a final reconstruction at 924 ~2.8 Å resolution. Model building and refinement was performed using PDB 7EJ8 as a starting 925 model using ChimeraX73, Phenix74, and Coot75. The final structures have been deposited in the 926 PDB with accession codes 9PLO and 9PLN. 927 928 Mac1 Purification and Crystallization. Wild-type Mac1 protein (P43 construct, 929 residues 3–169) was expressed in E. coli BL21(DE3) as an N-terminal His₆-tagged construct 930 and purified by Ni²⁺ affinity chromatography 76. The His-tag was cleaved with TEV protease, 931 followed by size-exclusion chromatography (Superdex 75) in 20 mM Tris-HCl (pH 7.5), 932 150 mM NaCl, and 1 mM DTT. Purified protein was concentrated to 40 mg/ml for 933 crystallization and stored at -80°C. 934 935 Crystals were obtained by sitting-drop vapor diffusion in 28% PEG 3000 and 100 mM 936 CHES (pH 9.5). Compounds (100 mM in DMSO) were added to crystal drops using an Echo 937 650 acoustic dispenser to a final concentration of 10 mM. Crystals were incubated at room 938 temperature for 2–4 hours and vitrified in liquid nitrogen without additional cryoprotection. X-939 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 42 ray diffraction data were collected at ALS (beamline 8.3.1) and processed using XDS and 940 Aimless. Structures were determined to resolutions ranging from 0.97 to 1.02 Å. Ligands with 941 low occupancy or conformational disorder were modeled using PanDDA and refined using 942 Phenix 77.PanDDA event maps are shown in Supplementary Figure 7. The structure of ‘3184 943 was determined using a racemic preparation of the compound (‘9037). X-ray data collection 944 and refinement statistics are summarized in SI Table 2, as are the 32 PDB IDs. 945 946 Microsomal Stability. Microsomal stability of compounds was evaluated using pooled 947 mouse liver microsomes (XenoTech, M3000/lot #2010026) to estimate their metabolic stability 948 and predict hepatic clearance. Each compound was incubated at 2 μM in a reaction mixture 949 containing 0.42 mg/mL microsomal protein, phosphate buffer (100 mM, pH 7.4), MgCl₂ (3.3 950 mM), NADPH (3 mM), glucose-6-phosphate (5.3 mM), and glucose-6-phosphate 951 dehydrogenase (0.67 units/mL). Reactions were conducted at 37 °C in 96-well plates with 952 shaking at 100 rpm. Samples were collected at five time points (0, 7, 15, 25, and 40 minutes), 953 and reactions were quenched by adding five volumes of acetonitrile containing an internal 954 standard. After centrifugation at 5500 rpm for 5 minutes, the supernatants were analyzed via 955 HPLC-MS/MS. The elimination rate constant (kel), half-life (t₁/₂), and intrinsic clearance (Clint) 956 were calculated by plotting the natural logarithm of the remaining parent compound versus 957 time. Stability was compared with reference standards such as imipramine and propranolol. 958 959 Plasma Protein Binding (PPB). Plasma protein binding (PPB) was measured using 960 equilibrium dialysis with a 14 kDa molecular weight cut-off membrane in a 96-well HTD96b 961 dialyzer. Mouse plasma containing 1 μM test compound (0.005% DMSO, 1% acetonitrile) was 962 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 43 placed in one chamber, and phosphate-buffered saline (PBS, pH 7.4) in the opposing 963 chamber. The assembled plates were incubated at 37 °C with 5% CO₂ and ∼95% humidity, 964 shaking at 250 rpm for 5 hours to reach equilibrium. After incubation, aliquots from each 965 chamber were mixed with equal volumes of the blank opposite matrix and processed with 966 acetonitrile containing internal standard. Supernatants obtained after centrifugation were 967 analyzed by HPLC-MS/MS. The percentage of compound bound to plasma proteins was 968 calculated using the peak area ratio in buffer to plasma compartments. Recovery and stability 969 standards were included to ensure accuracy and reliability. Verapamil served as a reference 970 control. Most compounds showed moderate binding, typically ranging between 75–85%. 971 972 Plasma stability. Plasma stability was assessed in non-sterile mouse plasma (Li-973 heparin treated) at 1 μM concentration (final DMSO content was 0.005%). Incubations were 974 carried out in aliquots of 60 μL each (two per time point) at 37 °C under 5% CO₂ and high 975 humidity (∼95%). The reactions were quenched with 240 µL of 90% acetonitrile containing an 976 internal standard at 0, 20, 40, 60, and 120 minutes, followed by centrifugation at 6000 rpm for 977 5 minutes. Supernatants were analyzed via HPLC-MS/MS to determine the percentage of 978 parent compound remaining at each time point. Data were plotted to calculate compound half-979 lives (t₁/₂). Reference compounds, verapamil and propantheline, were used as high and low 980 stability controls, respectively. This assay is critical for identifying compounds susceptible to 981 degradation by plasma esterases or hydrolytic enzymes, helping inform pharmacokinetic 982 optimization strategies during lead selection. 983 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 44 984 Thermodynamic Solubility. Aqueous thermodynamic solubility was determined in 985 PBS (pH 7.4) for 247 compounds using a shake-flask method followed by UV absorbance 986 quantification. Dry powder compounds were dissolved in PBS to a theoretical concentration of 987 4 mM and incubated in duplicates at 25 °C for 4 and 24 hours with shaking. After incubation, 988 samples were filtered using HTS 96-well filter plates. The filtrates were diluted 2-fold in 989 acetonitrile with 4% DMSO for UV analysis. The incubation samples for charged molecules 990 were additionally diluted 10-fold with 50% acetonitrile/PBS with 2% final DMSO. Calibration 991 curves (0–200 μM) were prepared in 50% acetonitrile/PBS (2% final DMSO). Absorbance was 992 measured between 230–550 nm using a SpectraMax Plus microplate reader. Compound-993 specific absorbance maxima were used to calculate concentration using SoftMax Pro and 994 Excel. The assay dynamic range is ~2–400 μM (~20–4000 μM for charged molecules), with 995 values near the upper limit treated as semi-quantitative. Ondansetron was used as a 996

compound. This method reflects equilibrium solubility under physiologically relevant 997 conditions and helps rank compounds for formulation feasibility. 998 999 PAMPA-BBB. Passive blood-brain barrier (BBB) permeability was estimated using a 1000 Parallel Artificial Membrane Permeability Assay (PAMPA-BBB) with a phospholipid-coated 1001 membrane simulating the brain endothelium. Test compounds (50 μM in Prisma HT buffer, pH 1002 7.4, with 0.5% DMSO) were added to donor wells, while brain sink buffer was added to the 1003 acceptor wells. The donor and acceptor chambers were separated by a 0.45 µm filter 1004 membrane coated with brain polar lipids. Plates were incubated without agitation at room 1005 temperature for 4 hours. Post-incubation, samples from both chambers, as well as a standard 1006 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 45 solution, were diluted with acetonitrile containing an internal standard. Apparent permeability 1007 coefficients (log Papp) were calculated based on peak area ratio. Clozapine and 1008 chlorpromazine were used as high-permeability controls, while ranitidine represented low 1009 permeability. The assay provides a high-throughput, non-cell-based method for estimating 1010 CNS exposure potential. 1011 1012 Behavioral analyses. All animal behavior experiments were conducted with the 1013 experimenter blinded to treatment. Mice were habituated individually in Plexiglas enclosures 1014 for 1 hour prior to testing. Compounds were administered subcutaneously 30 minutes before 1015 behavioral assessment, and where applicable the α2A adrenergic receptor antagonist 1016 atipamezole (2 mg/kg, intraperitoneally) was given 15 minutes before compound injection. Tail 1017 flick latency was measured by immersing the distal third of the tail in a 50°C water bath and 1018 recording the withdrawal time. For the neuropathic pain model, spared nerve injury (SNI) was 1019 performed under isoflurane anesthesia by ligating and transecting two of the three branches 1020 of the sciatic nerve, sparing the sural nerve. Mechanical thresholds were assessed 7 to 14 1021 days post-surgery using von Frey filaments and the up-down method, and values were 1022 normalized to each animal’s baseline. Thermal nociception was evaluated using a 55°C hot 1023 plate, and the latency to nocifensive behavior (paw lick or jump) was recorded with a cutoff of 1024 45 seconds to prevent tissue damage. Behavioral data are summarized as group size (n), 1025 mean, and SD (Supplementary Table 12). For the SNI experiment, we used two-way 1026 ANOVA followed by Tukey’s post-hoc multiple comparisons. For the hotplate assay, we used 1027 one-way repeated-measures ANOVA with Friedman post-hoc testing. For tail-flick, the 4905 1028 dose–response was analyzed by two-way ANOVA with Dunnett’s post-hoc comparisons to 1029 vehicle/control, whereas PS75 and 3629 dose groups were analyzed by one-way ANOVA 1030 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 46 with Kruskal–Wallis post-hoc testing (Supplementary Table 12).For sample size，we did not 1031 perform formal a priori power calculations. Instead, group sizes were guided by our prior 1032 experience with these assays and by precedent in the literature. This approach may limit 1033 sensitivity to small effects. Raw data for the animal assays are provided in Supplementary 1034 Table 13. 1035 1036 Data availability. Most primary data is available through this manuscript and its extended 1037 and supplementary materials. Crystal and cryoEM structures and supporting electron density 1038 are being made available via the Protein Data Bank, including PDB 9PLN, the structure of the 1039 2a/’4905 complex and PDB 14AB, 7IIW, 7IIX, 7IIY, 7IIZ, 7IJ0, 7IJ1, 7IJ2, 7IJ3, 7IJ4, 7IJ5, 7IJ6, 1040 7IJ7, 7IJ8, 7IJ9, 7IJA, 7IJB, 7IJC, 7IJD, 7IJE, 7IJF, 7IJG, 7IJH, 7IJI, 7IJJ, 7IJK, 7IJL, 14AM, 14AN, 1041 14AO, 14AP, and 7IJM, structures of Mac1 in complex with inhibitors.. 1042 1043 Code availability. ZINC tools used in the selection of the small perturbation analogs are 1044 openly available and zinc22.docking.org. 1045 1046

Supported by US DARPA grant HR0011-19-2-0020 (to B .K.S., A .M., 1047 A.I.B., and B.L.R.), by US NIH R35GM122481 (to B.K.S.) by US ARPA-H grant 1AY1AX000035 1048 (PI J.S.F.), O.M. partially supported by US NIH postdoctoral fellowship F32GM154469. We 1049 thank Yeyue Xiong for help with the FEP studies. We thank the UCSF Cryo-EM facility staff for 1050 training and technical assistance. UCSF Cryo-EM equipment is partially supported by NIH 1051 grants S10OD020054, S10OD021741 , and S10OD026881 , and by the Howard Hughes 1052 Medical Institute. 1053 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 47 1054 Author contributions: B.K.S., X.X., and O.M. designed the project. X.X. and O.M. designed 1055 the analogs with help from N.L. Alpha₂ receptor studies were performed by X.X. and H.H., with 1056 guidance from P.G. Alpha2A -related behavioral analyses were conducted by J.B., with 1057 guidance from A.I.B. Alpha₂ structural studies were performed by K.S., under the supervision 1058 of A.M. SERT binding studies were carried out by X.P.H. and J.W., with guidance from B.L.R. 1059 and ligand choice from Y.W. Mac1 biochemical assays were conducted by Y.U.D. M.S and M.D. 1060 supervised by A.A., and Mac1 crystallography was performed by G.C. with guidance from J.S.F. 1061 CB2 receptor assays were performed by X.X. and C.T., with ligand advice from M.R. AmpC 1062 assays were conducted by X.X. and F.L. All in vitro ADME and safety assay s were supported 1063 by Y.H. and Y.K. through Bienta. FEP calculations and analysis were performed by DS guided 1064 by GZ and RA . O.M. performed statistical analyses for both ChEMBL and experimental data. 1065 X.X., O.M., and B.K.S. prepared the manuscript. B.K.S. supervised the project. All authors 1066 reviewed and approved the final manuscript. 1067 1068 Competing interests B.K.S. is co-founder of Epiodyne, BlueDolphin, and Deep Apple 1069 Therapeutics, and serves on SAB for Schrodinger LLC, Vilya Therapeutics, Frontier Discovery 1070 Ltd, and on the SRB of Genentech . B.L.R. is founder of Onsero Therapeutics. J.S.F. is a 1071 consultant to and a shareholder of Vilya Therapeutics and Relay Therapeutics. The remaining 1072 authors declare no competing interest. 1073 1074 1075 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 48 Extended Data Fig. 1 1076 1077 Extended Data Fig. 1 Frequency of activity changes from single-atom substitutions in 1078 ChEMBL and in this study. a, Cumulative frequency of single non-hydrogen atom 1079 substitutions (CH3, F, Cl, Br, N, OH) in ChEMBL that improve activity, stratified by that of the 1080 parent compound. “Potent” denotes parents with activity ≤32 nM, “mid” 32 nM–1 μM, and 1081 “weak” 1 μM–1 mM. b, For each target in this study, the percentage of analogs that improve 1082 or decrease activity by ≥3-fold or ≥10-fold versus their parent. Fold changes were rounded to 1083 the nearest integer prior to thresholding (e.g., 9.6–9.9 counted as 10-fold). Error bars denote 1084 95% bootstrap confidence intervals (percentile method; 20,000 resamples). Compounds 1085 yielding no measurable curves were counted as ≥10-fold decreases. c, Parent-level summary 1086 of analog effects in this study (rounded fold-change). 1087 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 49 Extended Data Fig. 2 1088 1089 1090 Extended Data Fig. 2 Overview of Parent Compounds' Structures and Properties. 1091 a. 2D structures of the parent compounds. b. Molecular Weight (MW) and Affinity/potency 1092 values for the parent compounds. c. Structural Similarity matrix showing the relationships 1093 between parent compounds based on molecular fingerprint comparisons. 1094 1095 1096 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 50 Extended Data Fig. 31097 1098 Extended Data Fig. 3 Correlation between parent compound or binding pocket 1099 properties and affinity/potency fold change. 1100 a. Correlation between parent molecular weight (MW) and log₁₀ affinity/potency fold 1101 change 1102 b. Correlation between parent pKi and log₁₀ affinity/potency fold change. 1103 c. Correlation between binding pocket volume and log₁₀ affinity/potency fold change. 1104 d. Correlation between SiteMap Dscore and log₁₀ affinity/potency fold change. 1105 1106 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 51 Extended Data Fig. 4 1107 1108 Extended Data Fig. 4 Correlations between physicochemical descriptors and 1109 ΔpKi/pEC50 (log M) . a. Scatter plot of ΔpKi or pEC50 (log M) versus cLogP (RDKit). Each 1110 point represents one analog; the solid line indicates a linear fit. Pearson’s R and Spearman’s ρ 1111 are shown. b. ΔpKi or pEC50 (analog − parent) against ΔcLogP for all analog–parent pairs with 1112 ≥3× rounded potency/affinity improvement; ≥10× improvements are highlighted in red. c. 1113 Scatter plot of ΔpKi or pEC50 (log M) versus ΔLE (analog − parent), where LE = 1.37 × pKi or 1114 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 52 pEC50 / Nheavy. Pearson’s R and Spearman’s ρ are shown. d. Scatter plot of ΔpKi or pEC50 1115 (log M) versus ΔLipE (analog − parent), where LipE = pKi or pEC50 − cLogP. Pearson’s R and 1116 Spearman’s ρ are shown. e. Binding free -energy change versus heavy -atom change for 1117 improved analogs. The y -axis shows ΔΔG (kcal/mol), computed from the affinity metric 1118 (Ki/IC50/EC50) as ΔG = RT ln(K) (298 K) and ΔΔG = ΔG_analog − ΔG_parent; negative ΔΔG 1119 indicates improved binding. The x -axis is ΔN (a nalog − parent; heavy -atom count difference). 1120 Points with rounded fold ≥10 are highlighted in red; rounded 3 –9 are shown in blue. f. 1121 Affinity/potency gains versus changes in lipophilic efficiency. Scatter plot of ΔpKi or pEC50 1122 (analog − parent) against ΔLipE (analog − parent) for all analog–parent pairs with ≥3× rounded 1123 potency/affinity improvement; ≥10× improvements are highlighted in red. 1124 1125 1126 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 53 Extended Data Fig. 5 1127 1128 Extended Data Fig. 5 Structural variability caused by single-atom modifications 1129 complicates prediction of affinity/potency changes. a. Overlay of the X-ray crystal 1130 structures for 12 ligand-bound complexes from the ‘3453 series showing that single-atom 1131 modifications can result in substantial ligand movement or induce multiple binding poses, 1132 complicating structure-based analysis and predictions. b. Example of a binding pose shift: 1133 compounds ‘9249, ‘3176 and ‘3194 adopt markedly different pose compared to the ‘3454 1134 parent. Two conformations were identified for ‘3194. 1135 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 54 Extended Data Fig. 6 1136 1137 Extended Data Fig. 6 Single-atom substitutions frequently worsen PK properties a.Fraction of 1138 single-atom substitutions with ≥3-fold decreases in PK properties. Error bars represent 95% confidence 1139 intervals, estimated from 10,000 iterations of bootstrap resampling.b. Counts of analogs with measured 1140 PK parameters and >3-fold or >10-fold losses. 1141 1142 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 55 1143 Extended Data Fig. 7 1144 1145 Extended Data Fig. 7 Pharmacokinetic profiles of 29 individual analogs with >10-fold 1146 affinity/potency improvement. Fold-changes in in vitro pharmacokinetic properties, 1147 including microsomal stability, plasma fraction unbound, PAMPA permeability, plasma 1148 stability, solubility, and hERG inhibition, for 29 analogs showing >10-fold affinity/potency 1149 enhancements. 1150 1151 1152 1153 1154 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 56 Extended Data Fig. 8 1155 1156 Extended Data Fig. 8 Distribution and cumulative effects of atom modifications on 1157 molecular properties. a. Density plots showing the distribution of log ₁₀ fold improvement for 1158 different properties after one -atom modifications (Chlorine, Fluorine, Hydroxyl, Methyl, and 1159 Nitrogen). Each row represents a property (e.g., Affinity/potency, Microsome stability, Solubility), 1160 and each column corresponds to a specifi c modification. Vertical dashed lines indicate 0 (no 1161 change) and ±0.5 log ₁₀ fold change (approximately 3 -fold improvement or reduction). b. 1162 Cumulative distribution plots of the same data as in (a), showing the proportion of analogs 1163 achieving various levels of improvement or reduction. This visualization allows for comparison 1164 of the shift in property distributions across different modifications. 1165 1166 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint Extended Data Table1 FEP+ predictions and experimental activity readouts for analogs. 1167 Name modification protein affinity/potency- fold Affinity/potenc y_uM exp_G Parent exp G parent name exp_G FEP_G fep_unc fep_G fep_G err_abs fep_G err_abs Z8771520648 C MOR 13.39 0.00 -12.50 -10.96 Z4407498716 -1.54 -11.85 0.41 -0.89 0.65 0.65 Z8771520652 C MOR 0.12 0.08 -9.69 -10.96 Z4407498716 1.27 -8.95 0.59 2.01 0.74 0.74 Z8771520655 C MOR 0.53 0.02 -10.59 -10.96 Z4407498716 0.37 -10.98 0.40 -0.02 0.39 0.39 Z8771520657 C MOR 0.13 0.07 -9.76 -10.96 Z4407498716 1.20 -11.81 0.41 -0.85 2.05 2.04 Z8771520658 C MOR 0.48 0.02 -10.53 -10.96 Z4407498716 0.44 -11.14 0.41 -0.18 0.61 0.61 Z8771520665 O MOR 0.07 0.14 -9.36 -10.96 Z4407498716 1.60 -12.00 0.50 -1.04 2.64 2.64 Z8771520670 O MOR 1.24 0.01 -11.09 -10.96 Z4407498716 -0.13 -11.19 0.50 -0.23 0.10 0.10 Z8771520680 F MOR 0.73 0.01 -10.77 -10.96 Z4407498716 0.19 -9.96 0.40 1.00 0.81 0.81 Z8771520687 F MOR 0.67 0.01 -10.72 -10.96 Z4407498716 0.24 -10.77 0.41 0.19 0.05 0.05 Z8771520688 F MOR 0.26 0.04 -10.17 -10.96 Z4407498716 0.80 -11.38 0.41 -0.42 1.21 1.21 Z8774357908 O MOR 0.77 0.01 -10.81 -10.96 Z4407498716 0.16 -9.40 0.45 1.56 1.41 1.41 Z8793812850 C MOR 0.93 0.01 -10.92 -10.96 Z4407498716 0.04 -8.62 0.75 2.34 2.30 2.30 Z8793812851 C MOR 0.82 0.01 -10.84 -10.96 Z4407498716 0.12 -10.85 0.54 0.11 0.01 0.00 Z8798881729 C MOR 10.80 0.00 -12.37 -10.96 Z4407498716 -1.41 -11.39 0.75 -0.43 0.98 0.98 Z8812737822 Cl MOR 0.04 0.26 -8.98 -10.96 Z4407498716 1.98 -10.86 0.41 0.10 1.88 1.88 Z8812742460 Cl MOR 1.67 0.01 -11.27 -10.96 Z4407498716 -0.30 -11.89 0.54 -0.93 0.62 0.62 Z1800559918 Cl alpha2 0.00 6.84 -7.05 -12.09 Z2750653629 5.04 -13.03 1.14 -0.94 5.98 5.98 Z2750652484 N alpha2 0.01 0.16 -9.28 -12.09 Z2750653629 2.81 -8.88 1.52 3.21 0.40 0.40 Z4767467064 F alpha2 18.39 0.00 -13.81 -12.09 Z2750653629 -1.73 -12.37 1.28 -0.28 1.44 1.44 Z4767467067 N alpha2 0.01 0.09 -9.59 -12.09 Z2750653629 2.50 -11.33 1.98 0.76 1.74 1.73 Z4767467070 N alpha2 0.01 0.14 -9.36 -12.09 Z2750653629 2.73 -11.26 1.37 0.83 1.90 1.90 Z4767467072 N alpha2 0.01 0.10 -9.53 -12.09 Z2750653629 2.56 -12.64 2.78 -0.55 3.11 3.11 Z4954471231 C alpha2 5.48 0.00 -13.10 -12.09 Z2750653629 -1.01 -12.58 1.92 -0.49 0.52 0.52 Z5295862251 N alpha2 0.09 0.01 -10.69 -12.09 Z2750653629 1.40 -13.12 1.85 -1.03 2.43 2.43 Z5874039302 N alpha2 0.01 0.14 -9.34 -12.09 Z2750653629 2.75 -11.09 2.73 1.00 1.75 1.75 Z8598084898 N alpha2 0.01 0.16 -9.29 -12.09 Z2750653629 2.80 -12.08 2.00 0.01 2.79 2.79 Z8598084899 C alpha2 0.01 0.20 -9.13 -12.09 Z2750653629 2.96 -12.34 2.13 -0.25 3.21 3.21 Z8598084900 C alpha2 0.36 0.00 -11.49 -12.09 Z2750653629 0.60 -12.09 1.70 0.00 0.60 0.60 Z8598084901 C alpha2 9.72 0.00 -13.44 -12.09 Z2750653629 -1.35 -11.71 2.51 0.38 1.73 1.73 Z8598084905 C alpha2 52.17 0.00 -14.43 -12.09 Z2750653629 -2.34 -12.60 1.46 -0.51 1.83 1.83 Z8598084906 C alpha2 10.68 0.00 -13.49 -12.09 Z2750653629 -1.40 -12.37 1.86 -0.28 1.12 1.12 Z8598084908 F alpha2 0.16 0.01 -11.02 -12.09 Z2750653629 1.07 -12.62 1.57 -0.53 1.60 1.60 Z8598084909 F alpha2 2.47 0.00 -12.63 -12.09 Z2750653629 -0.54 -13.60 1.86 -1.51 0.97 0.97 Z8598084911 F alpha2 1.70 0.00 -12.40 -12.09 Z2750653629 -0.31 -11.76 1.85 0.33 0.64 0.64 Z8598084912 Cl alpha2 0.10 0.01 -10.71 -12.09 Z2750653629 1.37 -12.93 1.52 -0.84 2.21 2.21 Z8598084913 Cl alpha2 1.56 0.00 -12.35 -12.09 Z2750653629 -0.26 -12.77 2.04 -0.67 0.41 0.41 Z8598084914 Cl alpha2 0.08 0.02 -10.61 -12.09 Z2750653629 1.48 -13.39 1.82 -1.30 2.77 2.77 Z8598084918 Cl alpha2 2.38 0.00 -12.60 -12.09 Z2750653629 -0.51 -11.61 1.86 0.48 0.99 0.99 Z8598084922 O alpha2 0.67 0.00 -11.85 -12.09 Z2750653629 0.24 -13.02 1.90 -0.93 1.17 1.17 Z8598084925 O alpha2 0.18 0.01 -11.07 -12.09 Z2750653629 1.02 -8.13 1.45 3.96 2.95 2.95 Z8701729806 O alpha2 4.80 0.00 -13.02 -12.09 Z2750653629 -0.93 -12.59 2.37 -0.50 0.43 0.43 Z8731865364 O alpha2 0.17 0.01 -11.06 -12.09 Z2750653629 1.03 -12.14 2.06 -0.05 1.09 1.09 Z8747421921 O alpha2 1.51 0.00 -12.33 -12.09 Z2750653629 -0.24 -12.16 2.30 -0.07 0.17 0.18 Z8755839162 O alpha2 2.82 0.00 -12.70 -12.09 Z2750653629 -0.61 -12.60 2.42 -0.51 0.10 0.10 Z8825803704 C alpha2 0.58 0.00 -11.77 -12.09 Z2750653629 0.32 -11.65 2.44 0.44 0.11 0.11 Z8825803705 Cl alpha2 0.00 1.94 -7.79 -12.09 Z2750653629 4.30 -8.73 1.28 3.36 0.93 0.93 Z3034773353 C alpha2 0.31 72.29 -5.65 -6.35 Z3034773248 0.70 -5.49 1.33 0.86 0.16 0.16 Z3034773362 C alpha2 0.00 8270.00 -2.84 -6.35 Z3034773248 3.51 -5.73 1.31 0.62 2.89 2.89 Z3034773379 Cl alpha2 0.98 22.58 -6.34 -6.35 Z3034773248 0.01 -6.08 1.51 0.27 0.26 0.26 Z3034773381 Cl alpha2 0.49 45.57 -5.92 -6.35 Z3034773248 0.43 -6.25 1.40 0.10 0.33 0.33 Z3034773663 N alpha2 6.95 3.18 -7.50 -6.35 Z3034773248 -1.15 -5.82 1.77 0.53 1.68 1.68 Z3305295413 F alpha2 0.71 31.07 -6.15 -6.35 Z3034773248 0.20 -6.26 1.21 0.09 0.11 0.11 Z8727395864 O alpha2 2.38 9.29 -6.86 -6.35 Z3034773248 -0.51 -6.03 2.42 0.32 0.83 0.83 Z8836317802 O alpha2 0.35 63.92 -5.72 -6.35 Z3034773248 0.63 -5.77 1.88 0.59 0.04 0.04 Z8904328913 F alpha2 1.69 13.12 -6.66 -6.35 Z3034773248 -0.31 -6.07 1.42 0.28 0.59 0.59 Z8904332023 O alpha2 1.55 14.31 -6.61 -6.35 Z3034773248 -0.26 -4.34 1.69 2.02 2.27 2.27 Z8908003437 O alpha2 8.45 2.62 -7.61 -6.35 Z3034773248 -1.26 -6.38 1.72 -0.02 1.24 1.24 Z3034773358 C alpha2 33.63 0.66 -8.43 -6.35 Z3034773248 -2.08 -7.08 1.55 -0.73 1.36 1.36 Z6071720064 C alpha2 7.06 0.00 -12.18 -11.02 Z4376630014 -1.16 -11.03 1.21 -0.00 1.15 1.15 Z6071720076 Cl alpha2 0.15 0.06 -9.88 -11.02 Z4376630014 1.14 -9.98 1.15 1.04 0.10 0.10 Z6071720077 F alpha2 0.07 0.12 -9.45 -11.02 Z4376630014 1.58 -11.48 0.90 -0.46 2.03 2.03 Z6071720079 F alpha2 2.70 0.00 -11.61 -11.02 Z4376630014 -0.59 -11.51 1.33 -0.49 0.10 0.10 Z6071720081 Cl alpha2 0.02 0.38 -8.75 -11.02 Z4376630014 2.27 -11.62 0.89 -0.59 2.87 2.87 Z8081978921 Cl alpha2 0.00 28.99 -6.19 -11.02 Z4376630014 4.83 -12.87 0.94 -1.85 6.68 6.68 Z8598084957 C alpha2 0.04 0.21 -9.11 -11.02 Z4376630014 1.91 -10.84 1.25 0.18 1.73 1.73 Z8598084962 Cl alpha2 0.00 17.51 -6.49 -11.02 Z4376630014 4.53 -13.05 1.15 -2.03 6.57 6.56 Z3034773369 F alpha2 2.28 5.85 -7.14 -6.65 Z8727395870 -0.49 -6.18 1.23 0.47 0.96 0.96 Z3034773371 F alpha2 0.78 17.08 -6.50 -6.65 Z8727395870 0.15 -6.41 1.28 0.24 0.09 0.09 Z8727394867 Cl alpha2 1.09 12.23 -6.70 -6.65 Z8727395870 -0.05 -5.95 1.27 0.70 0.75 0.75 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 58 Name modification protein affinity/potency- fold Affinity/potenc y_uM exp_G Parent exp G parent name exp_G FEP_G fep_unc fep_G fep_G err_abs fep_G err_abs Z8731642722 F alpha2 1.32 10.10 -6.81 -6.65 Z8727395870 -0.16 -5.51 1.64 1.14 1.30 1.30 Z8731733054 C alpha2 1.81 7.36 -7.00 -6.65 Z8727395870 -0.35 -5.67 1.39 0.98 1.33 1.33 Z8735760397 C alpha2 0.95 13.97 -6.62 -6.65 Z8727395870 0.03 -6.27 1.50 0.38 0.35 0.35 Z8836317766 C alpha2 0.90 14.78 -6.59 -6.65 Z8727395870 0.06 -6.71 1.36 -0.06 0.12 0.12 Z8836317772 O alpha2 0.40 33.53 -6.10 -6.65 Z8727395870 0.55 -6.41 2.40 0.24 0.31 0.31 Z8904328914 C alpha2 0.66 20.12 -6.41 -6.65 Z8727395870 0.24 -6.06 1.46 0.59 0.34 0.35 Z8904332000 Cl alpha2 3.68 3.62 -7.42 -6.65 Z8727395870 -0.77 -6.81 1.57 -0.16 0.61 0.61 Z8904332024 O alpha2 0.00 24700.00 -2.19 -6.65 Z8727395870 4.46 -6.77 2.74 -0.12 4.58 4.57 Z8904332025 O alpha2 0.85 15.66 -6.55 -6.65 Z8727395870 0.10 -7.01 2.66 -0.36 0.45 0.45 Z8908003436 O alpha2 0.65 20.46 -6.40 -6.65 Z8727395870 0.25 -6.07 2.55 0.58 0.33 0.33 Z9089437746 Cl alpha2 0.72 18.62 -6.45 -6.65 Z8727395870 0.20 -7.08 1.27 -0.43 0.63 0.63 Z2610512538 Cl ampc 1.37 17.60 -6.49 -6.30 Z2610488449 -0.19 -6.31 0.43 0.18 0.18 0.37 Z2610513554 F ampc 0.95 25.46 -6.27 -6.30 Z2610488449 0.03 -7.03 0.52 -0.54 0.76 0.57 Z2610513975 F ampc 0.74 32.65 -6.12 -6.30 Z2610488449 0.18 -6.50 0.50 -0.01 0.38 0.19 Z8584748090 F ampc 1.41 17.13 -6.50 -6.30 Z2610488449 -0.20 -6.49 0.41 0.00 0.01 0.20 Z8584748099 N ampc 8.38 2.89 -7.56 -6.30 Z2610488449 -1.26 -6.61 0.41 -0.12 0.95 1.14 Z8584748244 F ampc 1.68 14.42 -6.60 -6.30 Z2610488449 -0.31 -6.39 0.41 0.10 0.21 0.41 Z8584748322 Cl ampc 1.22 19.77 -6.42 -6.30 Z2610488449 -0.12 -6.99 0.43 -0.50 0.57 0.38 Z8584748346 Cl ampc 2.72 8.88 -6.89 -6.30 Z2610488449 -0.59 -6.83 0.41 -0.34 0.06 0.25 Z1355328651 Cl cb2 36.17 0.06 -9.81 -7.69 Z52076138 -2.13 -9.45 0.60 -1.85 0.36 0.27 Z1511587581 C cb2 2.66 0.87 -8.26 -7.69 Z52076138 -0.58 -8.29 0.52 -0.69 0.03 0.11 Z1745610848 C cb2 21.10 0.11 -9.49 -7.69 Z52076138 -1.81 -7.94 0.41 -0.34 1.55 1.46 Z1820897159 F cb2 9.74 0.24 -9.03 -7.69 Z52076138 -1.35 -7.91 0.41 -0.31 1.12 1.04 Z185279188 C cb2 9.83 0.24 -9.04 -7.69 Z52076138 -1.35 -8.22 0.49 -0.62 0.82 0.73 Z6755334561 Cl cb2 17.44 0.13 -9.38 -7.69 Z52076138 -1.69 -8.91 0.49 -1.31 0.47 0.38 Z8383080706 C cb2 19.26 0.12 -9.44 -7.69 Z52076138 -1.75 -9.50 0.45 -1.90 0.06 0.15 Z8701681958 C cb2 0.04 11.24 -6.75 -8.72 Z6969215903 1.97 -6.62 0.91 1.33 0.13 0.65 Z8701681960 C cb2 0.07 5.53 -7.17 -8.72 Z6969215903 1.55 -6.83 0.84 1.12 0.34 0.44 Z8701681963 C cb2 0.49 0.83 -8.30 -8.72 Z6969215903 0.43 -8.15 0.43 -0.20 0.15 0.63 Z8701681979 F cb2 0.06 7.23 -7.01 -8.72 Z6969215903 1.71 -8.18 0.83 -0.23 1.17 1.95 Z8701681983 Cl cb2 1.09 0.37 -8.77 -8.72 Z6969215903 -0.05 -8.29 0.43 -0.34 0.48 0.30 Z8701681984 Cl cb2 0.11 3.74 -7.40 -8.72 Z6969215903 1.32 -7.51 0.91 0.44 0.11 0.89 Z8701681647 N cb2 0.14 0.90 -8.24 -9.42 Z8184698918 1.18 -10.25 1.15 -0.84 2.01 2.02 Z8701681649 C cb2 1.80 0.07 -9.77 -9.42 Z8184698918 -0.35 -7.41 0.47 2.00 2.36 2.35 Z8701681656 C cb2 29.61 0.00 -11.43 -9.42 Z8184698918 -2.01 -11.20 0.98 -1.79 0.23 0.22 Z8701681673 O cb2 0.20 0.62 -8.46 -9.42 Z8184698918 0.96 -4.54 1.25 4.87 3.92 3.91 Z8701681677 F cb2 1.57 0.08 -9.69 -9.42 Z8184698918 -0.27 -8.12 0.47 1.29 1.57 1.56 Z8701681695 Cl cb2 58.46 0.00 -11.83 -9.42 Z8184698918 -2.41 -11.01 1.25 -1.60 0.82 0.81 Z8701681697 Cl cb2 1.40 0.09 -9.62 -9.42 Z8184698918 -0.20 -7.71 0.46 1.70 1.91 1.90 Z9187344338 N mac1 0.09 7.96 -6.96 -8.40 Z1039063794 1.44 -7.10 0.43 1.30 0.14 0.14 Z9187344350 N mac1 0.00 935.70 -4.13 -8.40 Z1039063794 4.26 -5.56 0.60 2.83 1.43 1.43 Z9187344358 N mac1 0.46 1.54 -7.93 -8.40 Z1039063794 0.47 -8.99 0.38 -0.59 1.06 1.06 Z1359487111 C mac1 0.00 745.80 -4.27 -8.40 Z1039063794 4.13 -6.41 0.43 1.99 2.14 2.14 Z8756044544 C mac1 0.10 6.94 -7.04 -8.40 Z1039063794 1.36 -7.82 0.48 0.57 0.78 0.78 Z8756044539 C mac1 0.16 4.30 -7.32 -8.40 Z1039063794 1.08 -8.69 0.44 -0.29 1.37 1.37 Z237580338 N mac1 0.06 11.59 -6.73 -8.40 Z1039063794 1.66 -8.24 0.76 0.15 1.51 1.51 Z9043825450 O mac1 0.04 19.29 -6.43 -8.40 Z1039063794 1.96 -3.50 0.39 4.90 2.93 2.93 Z4223060659 N mac1 0.04 18.18 -6.47 -8.40 Z1039063794 1.93 -9.73 0.34 -1.33 3.26 3.26 Z8207543137 Cl mac1 0.00 220.00 -4.99 -8.40 Z1039063794 3.41 -6.79 0.46 1.60 1.80 1.80 Z8207599372 Cl mac1 0.06 12.22 -6.70 -8.40 Z1039063794 1.69 -9.59 0.45 -1.19 2.88 2.88 Z8207316857 Cl mac1 0.05 13.37 -6.65 -8.40 Z1039063794 1.75 -8.70 0.36 -0.30 2.05 2.05 Z8598075280 N mac1 0.16 15.18 -6.57 -7.72 Z5398393122 1.14 -7.94 0.52 -0.23 1.37 1.37 Z7691912366 C mac1 0.77 3.23 -7.49 -7.72 Z5398393122 0.23 -7.31 0.62 0.41 0.18 0.18 Z8598075283 C mac1 0.47 5.29 -7.20 -7.72 Z5398393122 0.52 -8.78 0.56 -1.06 1.58 1.58 Z8598075292 Cl mac1 2.24 1.12 -8.12 -7.72 Z5398393122 -0.40 -7.38 0.56 0.34 0.74 0.74 Z8598075284 C mac1 0.40 6.25 -7.10 -7.72 Z5398393122 0.62 -7.65 0.62 0.07 0.55 0.55 Z8928826043 C mac1 0.36 27.75 -6.22 -6.89 Z7534253453 0.68 -6.49 0.48 0.40 0.28 0.28 Z8727396582 C mac1 0.70 14.15 -6.61 -6.89 Z7534253453 0.28 -7.02 0.41 -0.13 0.41 0.41 Z8727396638 C mac1 4.13 2.40 -7.67 -6.89 Z7534253453 -0.77 -6.54 0.61 0.35 1.12 1.12 Z8929429249 C mac1 0.19 52.56 -5.84 -6.89 Z7534253453 1.05 -8.03 0.50 -1.14 2.19 2.19 Z8990523176 C mac1 0.46 21.65 -6.36 -6.89 Z7534253453 0.53 -7.28 0.55 -0.39 0.92 0.92 Z8990523178 O mac1 39.68 0.25 -9.01 -6.89 Z7534253453 -2.11 -7.14 0.65 -0.25 1.86 1.86 Z8990919048 O mac1 0.32 30.69 -6.16 -6.89 Z7534253453 0.74 -7.45 0.64 -0.56 1.30 1.30 Z8598075302 O mac1 0.70 14.27 -6.61 -6.89 Z7534253453 0.28 -6.04 0.60 0.85 0.57 0.57 Z8990523184 O mac1 19.84 0.50 -8.60 -6.89 Z7534253453 -1.70 -8.49 0.55 -1.60 0.10 0.10 Z7692056404 F mac1 1.79 5.53 -7.17 -6.89 Z7534253453 -0.28 -7.45 0.42 -0.56 0.28 0.28 Z8928826033 F mac1 2.82 3.51 -7.44 -6.89 Z7534253453 -0.55 -8.56 0.49 -1.67 1.12 1.12 Z8990523169 N mac1 0.99 10.04 -6.82 -6.89 Z7534253453 0.07 -7.01 0.61 -0.12 0.19 0.19 Z8990527720 F mac1 0.94 10.51 -6.79 -6.89 Z7534253453 0.10 -6.49 0.47 0.40 0.30 0.30 Z7692046343 Cl mac1 4.75 2.09 -7.75 -6.89 Z7534253453 -0.86 -7.62 0.56 -0.73 0.13 0.13 Z8768700676 Cl mac1 0.46 21.48 -6.37 -6.89 Z7534253453 0.52 -5.15 0.79 1.74 1.22 1.22 Z8990523194 Cl mac1 6.04 1.64 -7.89 -6.89 Z7534253453 -1.00 -7.27 0.69 -0.38 0.62 0.62 Z8727401304 Cl mac1 5.17 1.92 -7.80 -6.89 Z7534253453 -0.91 -8.74 0.52 -1.85 0.94 0.94 Z7692056627 N mac1 1.14 8.67 -6.91 -6.89 Z7534253453 -0.01 -6.90 0.72 -0.01 0.01 0.01 Z8929428675 N mac1 2.22 4.46 -7.30 -6.89 Z7534253453 -0.41 -7.99 0.52 -1.10 0.69 0.69 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 59 Name modification protein affinity/potency- fold Affinity/potenc y_uM exp_G Parent exp G parent name exp_G FEP_G fep_unc fep_G fep_G err_abs fep_G err_abs Z7691912473 N mac1 0.30 33.10 -6.11 -6.89 Z7534253453 0.78 -6.03 0.50 0.86 0.08 0.08 Z7140729870 C mac1 1.30 7.61 -6.98 -6.89 Z7534253453 -0.09 -6.12 0.41 0.77 0.86 0.86 Z3033185506 N sert 0.07 16.52 -6.52 -8.13 Z2009978218 1.61 -7.24 2.99 0.89 0.72 0.72 Z3991412496 F sert 0.93 1.18 -8.09 -8.13 Z2009978218 0.04 -8.86 2.18 -0.73 0.77 0.78 Z4224292450 O sert 0.20 5.53 -7.17 -8.13 Z2009978218 0.96 -5.66 2.48 2.47 1.51 1.51 Z4767467074 Cl sert 9.48 0.12 -9.46 -8.13 Z2009978218 -1.33 -9.28 2.39 -1.16 0.18 0.18 Z4767467093 C sert 1.22 0.89 -8.25 -8.13 Z2009978218 -0.12 -8.71 2.60 -0.59 0.46 0.47 Z4955958409 N sert 0.12 9.07 -6.88 -8.13 Z2009978218 1.25 -4.95 2.39 3.17 1.92 1.92 Z4957966902 N sert 0.02 60.37 -5.76 -8.13 Z2009978218 2.38 -6.02 2.68 2.11 0.26 0.27 Z5022589075 C sert 4.16 0.26 -8.98 -8.13 Z2009978218 -0.84 -8.81 2.46 -0.68 0.17 0.16 Z5425693717 Cl sert 0.44 2.49 -7.64 -8.13 Z2009978218 0.49 -7.60 2.42 0.53 0.04 0.04 Z5766294474 F sert 1.04 1.05 -8.16 -8.13 Z2009978218 -0.02 -8.04 2.44 0.08 0.11 0.11 Z8646465742 Cl sert 5.05 0.22 -9.09 -8.13 Z2009978218 -0.96 -9.41 2.57 -1.29 0.32 0.33 Z8827560439 C sert 2.92 0.37 -8.77 -8.13 Z2009978218 -0.63 -9.59 2.68 -1.47 0.83 0.83 Z8827560440 C sert 0.75 1.46 -7.96 -8.13 Z2009978218 0.17 -8.73 2.30 -0.60 0.76 0.77 Z8827560451 O sert 0.14 7.56 -6.99 -8.13 Z2009978218 1.15 -5.28 2.99 2.84 1.70 1.70 Z2573292509 C sert 0.35 0.24 -9.04 -9.67 Z2573292480 0.63 -12.05 1.66 -2.38 3.01 3.01 Z3498174416 N sert 0.03 2.64 -7.61 -9.67 Z2573292480 2.06 -7.12 1.18 2.56 0.49 0.50 Z4067645979 N sert 0.03 2.98 -7.54 -9.67 Z2573292480 2.13 -8.58 1.19 1.09 1.04 1.04 Z4224285938 N sert 0.14 0.58 -8.50 -9.67 Z2573292480 1.16 -9.97 1.47 -0.30 1.47 1.46 Z4765367699 C sert 0.69 0.12 -9.45 -9.67 Z2573292480 0.22 -10.84 1.59 -1.17 1.39 1.38 Z5458682006 C sert 3.17 0.03 -10.35 -9.67 Z2573292480 -0.68 -9.29 1.18 0.39 1.06 1.07 Z8055284773 N sert 0.50 0.16 -9.25 -9.67 Z2573292480 0.41 -10.21 1.26 -0.54 0.96 0.95 Z8501235596 C sert 1.27 0.06 -9.81 -9.67 Z2573292480 -0.14 -9.79 1.91 -0.11 0.02 0.03 Z8647047828 C sert 1.54 0.05 -9.92 -9.67 Z2573292480 -0.26 -7.84 1.13 1.83 2.09 2.09 Z8827560409 O sert 0.24 0.35 -8.81 -9.67 Z2573292480 0.85 -9.20 1.55 0.47 0.39 0.38 Z8827560428 Cl sert 1.56 0.05 -9.93 -9.67 Z2573292480 -0.26 -10.67 1.27 -0.99 0.74 0.73 Z3950610525 N sert 0.26 0.91 -8.24 -9.04 Z2573292509 0.80 -9.77 1.42 -0.70 1.53 1.49 Z4067646017 N sert 0.02 10.81 -6.77 -9.04 Z2573292509 2.26 -8.04 1.61 1.03 1.27 1.23 Z4765367717 C sert 2.28 0.10 -9.53 -9.04 Z2573292509 -0.49 -10.88 1.45 -1.81 1.35 1.32 Z5106507813 N sert 0.01 16.52 -6.52 -9.04 Z2573292509 2.52 -7.13 1.56 1.94 0.61 0.57 Z5458682044 C sert 16.14 0.01 -10.69 -9.04 Z2573292509 -1.65 -9.21 1.48 -0.13 1.47 1.51 Z8647047832 C sert 1.79 0.13 -9.39 -9.04 Z2573292509 -0.35 -7.97 1.44 1.11 1.42 1.46 Z8781161134 N sert 0.07 3.57 -7.43 -9.04 Z2573292509 1.61 -7.91 1.74 1.17 0.48 0.44 Z8827560496 O sert 0.63 0.37 -8.77 -9.04 Z2573292509 0.27 -9.10 1.55 -0.02 0.33 0.29 Z8827560515 Cl sert 2.76 0.09 -9.64 -9.04 Z2573292509 -0.60 -9.58 1.85 -0.50 0.06 0.10 Z6971277405 N sert 0.55 24.20 -6.30 -6.65 Z6971277399 0.35 -5.25 1.28 1.23 1.04 0.88 Z6971278430 C sert 24.81 0.54 -8.55 -6.65 Z6971277399 -1.90 -6.44 1.28 0.03 2.11 1.94 Z6971278455 C sert 4.08 3.28 -7.48 -6.65 Z6971277399 -0.83 -6.21 1.35 0.27 1.27 1.10 Z6971278464 C sert 3.99 3.36 -7.47 -6.65 Z6971277399 -0.82 -6.23 1.36 0.24 1.23 1.06 Z6971278779 F sert 0.42 31.49 -6.14 -6.65 Z6971277399 0.51 -6.54 1.43 -0.06 0.39 0.56 Z6971279383 Cl sert 5.09 2.62 -7.61 -6.65 Z6971277399 -0.96 -6.97 1.28 -0.49 0.64 0.47 Z6971279394 Cl sert 19.99 0.67 -8.42 -6.65 Z6971277399 -1.77 -7.37 1.29 -0.89 1.05 0.89 Z6971279399 Cl sert 3.57 3.74 -7.40 -6.65 Z6971277399 -0.75 -7.00 1.25 -0.52 0.40 0.23 Z6971279682 Br sert 11.93 1.12 -8.12 -6.65 Z6971277399 -1.47 -7.71 1.44 -1.23 0.40 0.24 Z6971279689 Br sert 7.58 1.76 -7.85 -6.65 Z6971277399 -1.20 -7.15 1.59 -0.68 0.69 0.52 Z6971279702 Br sert 16.09 0.83 -8.29 -6.65 Z6971277399 -1.65 -7.25 1.59 -0.77 1.04 0.88 Z8172508405 O sert 4.18 3.20 -7.50 -6.65 Z6971277399 -0.85 -6.87 1.48 -0.39 0.63 0.46 Z3034773248 Cl sert 9.24 4.62 -7.28 -5.96 Z8727393896 -1.32 -6.68 0.99 -1.00 0.60 0.32 Z8024305237 Br sert 9.88 4.32 -7.32 -5.96 Z8727393896 -1.36 -6.48 1.17 -0.79 0.84 0.56 Z8727394028 O sert 5.45 7.84 -6.96 -5.96 Z8727393896 -1.00 -6.79 1.32 -1.11 0.17 0.11 Z8727394030 Br sert 7.56 5.65 -7.16 -5.96 Z8727393896 -1.20 -6.76 0.96 -1.08 0.40 0.12 Z8727395866 F sert 2.44 17.51 -6.49 -5.96 Z8727393896 -0.53 -6.26 1.02 -0.57 0.23 0.05 Z8727395870 Cl sert 5.79 7.37 -7.00 -5.96 Z8727393896 -1.04 -6.20 1.00 -0.52 0.80 0.52 Z8727396158 Br sert 10.50 4.07 -7.35 -5.96 Z8727393896 -1.39 -6.74 1.07 -1.05 0.62 0.34 Z8731642686 C sert 23.33 1.83 -7.83 -5.96 Z8727393896 -1.87 -6.64 1.00 -0.95 1.19 0.91 Z8735711409 O sert 1.04 40.96 -5.99 -5.96 Z8727393896 -0.02 -5.19 1.42 0.49 0.79 0.52 Z8735711431 Cl sert 6.03 7.08 -7.03 -5.96 Z8727393896 -1.06 -6.20 0.98 -0.52 0.82 0.54 Z3034773283 C sert 20.47 0.09 -9.61 -7.83 Z8731642686 -1.79 -9.35 1.32 -1.51 0.26 0.28 Z3034773290 C sert 1.75 1.05 -8.16 -7.83 Z8731642686 -0.33 -8.86 0.97 -1.02 0.70 0.69 Z3034773359 Cl sert 4.14 0.44 -8.67 -7.83 Z8731642686 -0.84 -8.49 1.22 -0.65 0.17 0.19 Z3034773386 Br sert 27.20 0.07 -9.78 -7.83 Z8731642686 -1.96 -9.45 1.29 -1.61 0.33 0.35 Z8727393902 Br sert 7.28 0.25 -9.00 -7.83 Z8731642686 -1.18 -8.75 1.23 -0.91 0.25 0.27 Z8741436112 F sert 6.98 0.26 -8.98 -7.83 Z8731642686 -1.15 -7.57 1.57 0.27 1.41 1.42 Z8836317766 Cl sert 5.87 0.31 -8.88 -7.83 Z8731642686 -1.05 -9.03 1.18 -1.18 0.15 0.13 Z8843617521 O sert 0.17 10.72 -6.78 -7.83 Z8731642686 1.05 -7.17 1.44 0.68 0.39 0.37 Z8843617522 O sert 0.39 4.71 -7.27 -7.83 Z8731642686 0.56 -8.15 1.60 -0.31 0.89 0.87 Z8843617525 O sert 0.53 3.46 -7.45 -7.83 Z8731642686 0.38 -9.65 1.36 -1.80 2.20 2.18 Z8890819769 F sert 0.35 5.27 -7.20 -7.83 Z8731642686 0.63 -7.55 1.50 0.29 0.35 0.33 Z8891248505 C sert 3.56 0.51 -8.58 -7.83 Z8731642686 -0.75 -7.44 1.14 0.41 1.14 1.16 Z8904328909 N sert 0.20 9.05 -6.88 -7.83 Z8731642686 0.95 -5.86 1.27 1.99 1.02 1.04 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 60 Name modification protein affinity/potency- fold Affinity/potenc y_uM exp_G Parent exp G parent name exp_G FEP_G fep_unc fep_G fep_G err_abs fep_G err_abs Z8904328910 F sert 2.19 0.84 -8.29 -7.83 Z8731642686 -0.46 -8.82 1.12 -0.97 0.53 0.51 Z8904332022 O sert 11.48 0.16 -9.27 -7.83 Z8731642686 -1.45 -8.01 1.60 -0.17 1.26 1.28 Z8915636241 C sert 4.56 0.40 -8.72 -7.83 Z8731642686 -0.90 -8.44 1.09 -0.59 0.29 0.31 Z3034773358 Cl sert 23.52 0.08 -9.70 -7.83 Z8731642686 -1.87 -9.38 1.27 -1.54 0.32 0.33 a, ΔGexp (kcal mol⁻¹): Experimental free energy calculated from assay-derived K (in M) as ΔGexp = 1168 RT ln K, with R = 1.987 × 10⁻³ kcal mol⁻¹ K⁻¹ and T = 298.15 K. 1169 b, ΔGexp,parent (kcal mol⁻¹): ΔGexp of the corresponding parent compound for the same target. 1170 c, ΔΔGexp (kcal mol⁻¹): Relative experimental free energy versus parent, ΔΔGexp = ΔGexp,analog − 1171 ΔGexp,parent. 1172 d, ΔGFEP (kcal mol⁻¹): Absolute binding free energy predicted by FEP+. 1173 e, σFEP (kcal mol⁻¹): Uncertainty reported by the FEP+ workflow for ΔGFEP. 1174 f, ΔΔGFEP (kcal mol⁻¹): Relative FEP+ free energy versus parent, ΔΔGFEP = ΔGFEP,analog − 1175 ΔGFEP,parent. 1176 g, |ΔGFEP − ΔGexp| (kcal mol⁻¹): Absolute error in ΔG. 1177 h, |ΔΔGFEP − ΔΔGexp| (kcal mol⁻¹): Absolute error in ΔΔG.Positive ΔΔG\Delta\Delta GΔΔG 1178 indicates decreased potency/affinity relative to the parent; negative ΔΔG\Delta\Delta GΔΔG indicates 1179 increased potency/affinity. Blank entries indicate values not available for the corresponding metric. 1180 Full chemical structures (including SMILES) for all parent compounds and analogs are provided in 1181 Supplementary Table 1. 1182 1183 1184 1185 1186 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 1187 1188 Extended Data Table 2. The effect of the analogs, relative to parents, on in vitro pharmacokinetics* 1189 parent target Total compound s Microsome stability T1/2, min fold change (≥3x) Plasma stability T1/2, min fold change (≥3x) ppb % of unbound compound fold change (≥3x) Solubility, µM fold change (≥3x) hERG approximate IC50 µM fold change (≥3x) permeabilit y cm/s fold change (≥3x) Z4407498716 MOR 16 0 0 0 0 0 0 Z2750653629 alpha 2 31 7 2 4 5 5 0 Z3034773248 alpha 2 18 1 1 3 16 7 0 Z4376630014 alpha 2 10 5 1 7 2 1 0 Z8727395870 alpha 2 18 3 0 6 16 11 0 Z2610488449 ampc 10 1 0 2 0 2 4 Z52076138 cb2 7 0 1 0 1 0 0 Z6969215903 cb2 6 1 2 2 0 0 0 Z818469891 8 cb2 20 1 4 0 0 5 0 Z1039063794 mac1 12 4 0 1 0 3 0 Z5398393122 mac1 5 4 1 0 0 0 4 Z7534253453 mac1 22 1 10 0 0 4 9 Z2009978218 sert 14 5 4 3 0 2 0 Z2573292480 sert 11 3 0 3 0 0 0 Z2573292509 sert 9 3 2 4 6 3 0 Z6971277399 sert 15 1 3 1 0 1 0 Z8727393896 sert 14 0 3 0 0 0 0 Z8731642686 sert 19 0 3 3 0 1 0 * The MOR parent series and its 16 analogs are not included in this table because they were not profiled 1190 for in vitro PK properties. 1191 1192 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint 1193 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted May 13, 2026. ; https://doi.org/10.64898/2026.05.10.724162doi: bioRxiv preprint