A tenfold improvement in potency is a persuasive result. It becomes a different kind of evidence when modest, systematically chosen substitutions achieve that result surprisingly often. Xinyu Xu and colleagues have created an experimental background against which molecular optimization can be judged: 257 analogues of 18 starting compounds, spanning six biological targets. Twenty-nine qualified as roughly tenfold improvements. Those gains frequently came with poorer properties relevant to exposure.1
The study, published in Nature on 16 September, gives medicinal chemistry and AI-assisted design a useful comparator. Its central question is not whether an algorithm can produce a better molecule. It is how much progress a design makes beyond accessible local chemistry, and whether the property being improved is the one that ultimately matters. Reading the supplementary measurements makes that distinction unusually concrete.12
A background with a defined boundary
The six targets cover three receptors, a transporter and two enzymes: α2A adrenergic, μ-opioid and cannabinoid CB2 receptors; the serotonin transporter; AmpC β-lactamase; and the SARS-CoV-2 Mac1 macrodomain. The starting compounds represent 11 unrelated scaffolds. They were chosen partly because the researchers already had access to the compounds and assays, rather than by random sampling of drug discovery projects.1
The perturbations were deliberately small. Hydrogen could be replaced by methyl, hydroxyl, fluorine, chlorine or, less often, bromine; an aromatic carbon could become nitrogen. Compounds that changed the parent's net charge at physiological pH were excluded. Crucially, the researchers did not choose positions because a binding model predicted success. They explored the substitutions that could be made within the practical constraints.1
Those constraints are part of the result. Only about 7% of the enumerated possibilities met the synthesis and price criteria; approximately 80% of that subset was successfully made. The ordering threshold was below US$400 for 10 mg. Thus, “random background” here means an activity-unselected survey within accessible chemistry. It is not a uniform sample of every possible modification, much less of chemical space. A future design benchmark should preserve that boundary when choosing its comparator.1
This is also why the work is more useful than simply counting successful analogues in published papers. The researchers retain unsuccessful changes in the denominator and compare parents and analogues with the same readout within each series. Their separate ChEMBL34 analysis required matched compounds to share assay, target, activity type and publication identifiers. That is a thoughtful effort to improve comparability, while the new experimental collection reduces the publication-selection problem that remains in historical data.1
What the tenfold number contains
The headline 11.3% is 29 of 257 analogues. Four of those 29 improved by 9.7–9.9-fold and were counted after rounding to an integer. This is an explicitly reported convention, also applied to the database comparison. It is reasonable for a coarse description of substantial improvement, but an evaluation pipeline reproducing the benchmark must use the same convention rather than quietly imposing an exact threshold.1
The distribution is more informative than the headline alone. Improvements of this size occurred for ten of the 18 parents and five of the six targets; 69 analogues met the roughly threefold threshold. Methyl and chlorine substitutions were particularly productive for potency. The data therefore support systematic local exploration across several chemical series. They do not give every new target an 11.3% guaranteed success probability.1
Supplementary Figure 1 explains why pooling helps: the authors treat analogue outcomes as Bernoulli trials and show how estimated intervals narrow as the number of compounds increases. The practical inference is that an individual parent's small set gives a much less precise background than the pooled collection. For method comparison, a useful next design would match target and starting series as well as synthetic budget, then ask whether the method consistently enriches the successful substitutions. That is an inference from the sampling scheme, not an experiment performed here.3
The assay distinction is equally important. A lower functional EC50 means a compound produces its measured response at a lower concentration; it need not represent the same change in binding affinity as a lower Ki. The study keeps readouts consistent within a series. For the receptor agonists, it also examines expression levels and responses at different plasmid amounts. Supplementary Figures 4 and 5 document these controls. They strengthen interpretation of relative activity without making receptor signalling identical to binding.13
Small modifications can reorganize the binding problem
Structural analysis explains some successes while revealing why they are difficult to forecast. In one Mac1 series, a chlorine substituent placed at different positions encounters different environments. The ortho analogue has an IC50 of about 21 µM, whereas the meta analogue reaches 1.9 µM. The parent is 8.9 µM: the approximately elevenfold contrast is between the two positional analogues, not between the meta compound and its parent. Keeping that comparator visible prevents an attractive structural example from becoming an inflated optimization claim.1
Even an apparently simple fit involves movement. The para-chloro analogue is accommodated by rotations of Phe132 and Ile131; other small substitutions produce alternative ligand conformations or shift the inhibitor in the pocket. These structures make a strong case for examining the whole ligand–protein arrangement. A substituent is not an independent bonus attached to an otherwise unchanged binding pose.1
This is where the computational comparison earns its place. Blinded FEP+ calculations used docked starting poses and a defined simulation workflow. Across 219 evaluated analogues, the reported mean unsigned error for binding free energy was 1.06 kcal/mol, with 188 predictions within 2 kcal/mol. Those aggregate results are encouraging. They answer a different question, however, from whether the calculation reliably picks the next beneficial change within a series.1
Supplementary Table 11 makes that decision-level question legible. Among 34 analogues predicted to improve beyond the 1 kcal/mol threshold, 14 experimentally improved by that criterion, 12 were neutral and eight worsened. This contingency table contains 218 classified analogues, rather than the 219 in the aggregate energy analysis; its own denominator should be retained. The reported association is useful, but the positive prediction is not a substitute for an assay. The authors also identify starting-pose errors among outliers, tying an important failure mode to a concrete part of the workflow.21
Potency and developability pull in different directions
The most consequential measurements extend beyond target activity. Parents and analogues were tested for microsomal and plasma stability, plasma fraction unbound, solubility, passive membrane permeability and hERG inhibition. These probe different aspects of a molecule's behaviour. They should not be compressed into an unexplained single “drug-likeness” score, especially when a chemical change helps one endpoint and hurts another.1
None of the 29 roughly tenfold potency improvers also improved all three highlighted properties: microsomal stability, PAMPA permeability and plasma fraction unbound. Supplementary Figure 7 supplies the individual profiles rather than leaving that conclusion at the level of an average. Conversely, aromatic carbon-to-nitrogen changes, which seldom delivered the largest potency gains, were often helpful for microsomal stability or fraction unbound. The structure–activity relationship and the developability relationship are related design problems with different outcomes.13
There is good reason to avoid treating this as a universal law of substitution. The readouts are predominantly in vitro, and their physiological coverage differs. Liver microsomes do not capture every metabolic route; the authors specifically identify missed glucuronidation as a reason hydroxyl-containing compounds might appear more stable than they would in vivo. PAMPA measures passive permeation, while the fraction unbound describes protein binding. An improvement in one measurement cannot establish an improvement in total exposure.1
The AI comparison asks a particularly demanding version of that problem: can a model anticipate the change between close analogues? ADMET-AI, ADMETlab 3.0 and Deep-PK showed useful correlations for permeability and fraction unbound, but little correlation for several stability or solubility comparisons. Endpoint matching matters here. Figure 4 explicitly identifies ADMETlab's comparable plasma-stability endpoint; the corresponding Deep-PK and ADMET-AI endpoints were not directly comparable. It would be unfair to describe all three as failing an identical plasma-stability test.1
Earlier work introducing ADMET-AI helps explain the change in question. Kyle Swanson and colleagues trained Chemprop-RDKit models using 41 Therapeutics Data Commons datasets, combining a graph representation with 200 calculated molecular features. Ten datasets concern regression and 31 classification. That paper evaluates broad property prediction; Xu and colleagues ask whether predicted changes track measured changes inside local chemical neighbourhoods. Success on the former does not logically guarantee the latter. This is complementary evidence about use in lead optimization, rather than a contradiction that erases the earlier benchmark.41
A correlation can still help rank experiments even when individual decisions remain uncertain. The new study's strongest reported fraction-unbound correlation was 0.53, yet numerous analogues were assigned the wrong direction of change. The constructive response is to use predictions to prioritize measurement and retain endpoint-specific uncertainty. The measured background is valuable precisely because it does not replace difficult experimental outcomes with additional model estimates.1
Progress can survive a worse property
The α2A agonist example prevents the trade-off story from becoming fatalistic. Adding a methyl group to parent ‘3629 produced ‘4905, with a reported 52-fold improvement in functional potency. Its microsomal half-life and fraction unbound deteriorated. Nevertheless, the analogue showed improved activity in mouse nociception experiments, illustrating that a sufficiently large potency gain can remain useful despite losses elsewhere.1
The supplementary pharmacokinetic table makes an especially important distinction. Following 3 mg/kg intraperitoneal administration, the reported cerebrospinal-fluid half-life decreased from 196 minutes for the parent to 15.2 minutes for the analogue. Those are CSF values. Plasma half-lives in the same table were 28.5 and 43.9 minutes, respectively. Calling the change simply a collapse in “the drug's half-life” would obscure the compartment that was measured and reverse the plasma comparison.2
The behavioural evidence also deserves its own boundaries. Mouse allocation was randomized and behavioural assessors were blinded; the reporting summary describes two independent cohorts and no formal advance power calculation. The hotplate comparison between analogue and parent was not statistically significant at the tested 5 mg/kg dose, despite a higher numerical response for the analogue. The strongest interpretation is therefore assay-specific improvement, supported by a chemically intelligible potency change, rather than uniform superiority across every pain endpoint or evidence of human benefit.51
What this study changes is the standard of comparison. A successful substitution needs an accessible chemical background; a promising calculation needs a local decision test; a potency gain needs the relevant exposure measurements. The work gives all three questions experimental substance. For AI-assisted medicinal chemistry, that is a valuable foundation: progress becomes easier to recognize when the background, endpoint and comparator are kept in view.1
What this does not establish
- The accessible, price-constrained analogue collection is not a uniform sample of medicinal chemistry; pooled success rates do not establish a rate for every target.
- Binding and functional activity readouts are kept consistent within series but are not interchangeable across series.
- Most developability evidence is in vitro; the mouse example is preclinical and endpoint-specific.
- Primary experiments, simulations and model inference were inspected, not independently rerun.
Claims and evidence
A tenfold improvement in potency is a persuasive result. It becomes a different kind of evidence when modest, systematically chosen substitutions achieve that result surprisingly often. Xinyu Xu and colleagues have created an experimental background against which molecular optimization can be judged: 257 analogues of 18 starting compounds, spanning six biological targets. Twenty-nine qualified as roughly tenfold improvements. Those gains frequently came with poorer properties relevant to exposure. 1
The study, published in Nature on 16 September, gives medicinal chemistry and AI-assisted design a useful comparator. Its central question is not whether an algorithm can produce a better molecule. It is how much progress a design makes beyond accessible local chemistry, and whether the property being improved is the one that ultimately matters. Reading the supplementary measurements makes that distinction unusually concrete. 12
The six targets cover three receptors, a transporter and two enzymes: α2A adrenergic, μ-opioid and cannabinoid CB2 receptors; the serotonin transporter; AmpC β-lactamase; and the SARS-CoV-2 Mac1 macrodomain. The starting compounds represent 11 unrelated scaffolds. They were chosen partly because the researchers already had access to the compounds and assays, rather than by random sampling of drug discovery projects. 1
The perturbations were deliberately small. Hydrogen could be replaced by methyl, hydroxyl, fluorine, chlorine or, less often, bromine; an aromatic carbon could become nitrogen. Compounds that changed the parent's net charge at physiological pH were excluded. Crucially, the researchers did not choose positions because a binding model predicted success. They explored the substitutions that could be made within the practical constraints. 1
Those constraints are part of the result. Only about 7% of the enumerated possibilities met the synthesis and price criteria; approximately 80% of that subset was successfully made. The ordering threshold was below US$400 for 10 mg. Thus, “random background” here means an activity-unselected survey within accessible chemistry. It is not a uniform sample of every possible modification, much less of chemical space. A future design benchmark should preserve that boundary when choosing its comparator. 1
This is also why the work is more useful than simply counting successful analogues in published papers. The researchers retain unsuccessful changes in the denominator and compare parents and analogues with the same readout within each series. Their separate ChEMBL34 analysis required matched compounds to share assay, target, activity type and publication identifiers. That is a thoughtful effort to improve comparability, while the new experimental collection reduces the publication-selection problem that remains in historical data. 1
The headline 11.3% is 29 of 257 analogues. Four of those 29 improved by 9.7–9.9-fold and were counted after rounding to an integer. This is an explicitly reported convention, also applied to the database comparison. It is reasonable for a coarse description of substantial improvement, but an evaluation pipeline reproducing the benchmark must use the same convention rather than quietly imposing an exact threshold. 1
The distribution is more informative than the headline alone. Improvements of this size occurred for ten of the 18 parents and five of the six targets; 69 analogues met the roughly threefold threshold. Methyl and chlorine substitutions were particularly productive for potency. The data therefore support systematic local exploration across several chemical series. They do not give every new target an 11.3% guaranteed success probability. 1
Supplementary Figure 1 explains why pooling helps: the authors treat analogue outcomes as Bernoulli trials and show how estimated intervals narrow as the number of compounds increases. The practical inference is that an individual parent's small set gives a much less precise background than the pooled collection. For method comparison, a useful next design would match target and starting series as well as synthetic budget, then ask whether the method consistently enriches the successful substitutions. That is an inference from the sampling scheme, not an experiment performed here. 3
The assay distinction is equally important. A lower functional EC50 means a compound produces its measured response at a lower concentration; it need not represent the same change in binding affinity as a lower Ki. The study keeps readouts consistent within a series. For the receptor agonists, it also examines expression levels and responses at different plasmid amounts. Supplementary Figures 4 and 5 document these controls. They strengthen interpretation of relative activity without making receptor signalling identical to binding. 13
Structural analysis explains some successes while revealing why they are difficult to forecast. In one Mac1 series, a chlorine substituent placed at different positions encounters different environments. The ortho analogue has an IC50 of about 21 µM, whereas the meta analogue reaches 1.9 µM. The parent is 8.9 µM: the approximately elevenfold contrast is between the two positional analogues, not between the meta compound and its parent. Keeping that comparator visible prevents an attractive structural example from becoming an inflated optimization claim. 1
Even an apparently simple fit involves movement. The para-chloro analogue is accommodated by rotations of Phe132 and Ile131; other small substitutions produce alternative ligand conformations or shift the inhibitor in the pocket. These structures make a strong case for examining the whole ligand–protein arrangement. A substituent is not an independent bonus attached to an otherwise unchanged binding pose. 1
This is where the computational comparison earns its place. Blinded FEP+ calculations used docked starting poses and a defined simulation workflow. Across 219 evaluated analogues, the reported mean unsigned error for binding free energy was 1.06 kcal/mol, with 188 predictions within 2 kcal/mol. Those aggregate results are encouraging. They answer a different question, however, from whether the calculation reliably picks the next beneficial change within a series. 1
Supplementary Table 11 makes that decision-level question legible. Among 34 analogues predicted to improve beyond the 1 kcal/mol threshold, 14 experimentally improved by that criterion, 12 were neutral and eight worsened. This contingency table contains 218 classified analogues, rather than the 219 in the aggregate energy analysis; its own denominator should be retained. The reported association is useful, but the positive prediction is not a substitute for an assay. The authors also identify starting-pose errors among outliers, tying an important failure mode to a concrete part of the workflow. 21
The most consequential measurements extend beyond target activity. Parents and analogues were tested for microsomal and plasma stability, plasma fraction unbound, solubility, passive membrane permeability and hERG inhibition. These probe different aspects of a molecule's behaviour. They should not be compressed into an unexplained single “drug-likeness” score, especially when a chemical change helps one endpoint and hurts another. 1
None of the 29 roughly tenfold potency improvers also improved all three highlighted properties: microsomal stability, PAMPA permeability and plasma fraction unbound. Supplementary Figure 7 supplies the individual profiles rather than leaving that conclusion at the level of an average. Conversely, aromatic carbon-to-nitrogen changes, which seldom delivered the largest potency gains, were often helpful for microsomal stability or fraction unbound. The structure–activity relationship and the developability relationship are related design problems with different outcomes. 13
There is good reason to avoid treating this as a universal law of substitution. The readouts are predominantly in vitro, and their physiological coverage differs. Liver microsomes do not capture every metabolic route; the authors specifically identify missed glucuronidation as a reason hydroxyl-containing compounds might appear more stable than they would in vivo. PAMPA measures passive permeation, while the fraction unbound describes protein binding. An improvement in one measurement cannot establish an improvement in total exposure. 1
The AI comparison asks a particularly demanding version of that problem: can a model anticipate the change between close analogues? ADMET-AI, ADMETlab 3.0 and Deep-PK showed useful correlations for permeability and fraction unbound, but little correlation for several stability or solubility comparisons. Endpoint matching matters here. Figure 4 explicitly identifies ADMETlab's comparable plasma-stability endpoint; the corresponding Deep-PK and ADMET-AI endpoints were not directly comparable. It would be unfair to describe all three as failing an identical plasma-stability test. 1
Earlier work introducing ADMET-AI helps explain the change in question. Kyle Swanson and colleagues trained Chemprop-RDKit models using 41 Therapeutics Data Commons datasets, combining a graph representation with 200 calculated molecular features. Ten datasets concern regression and 31 classification. That paper evaluates broad property prediction; Xu and colleagues ask whether predicted changes track measured changes inside local chemical neighbourhoods. Success on the former does not logically guarantee the latter. This is complementary evidence about use in lead optimization, rather than a contradiction that erases the earlier benchmark. 41
A correlation can still help rank experiments even when individual decisions remain uncertain. The new study's strongest reported fraction-unbound correlation was 0.53, yet numerous analogues were assigned the wrong direction of change. The constructive response is to use predictions to prioritize measurement and retain endpoint-specific uncertainty. The measured background is valuable precisely because it does not replace difficult experimental outcomes with additional model estimates. 1
The α2A agonist example prevents the trade-off story from becoming fatalistic. Adding a methyl group to parent ‘3629 produced ‘4905, with a reported 52-fold improvement in functional potency. Its microsomal half-life and fraction unbound deteriorated. Nevertheless, the analogue showed improved activity in mouse nociception experiments, illustrating that a sufficiently large potency gain can remain useful despite losses elsewhere. 1
The supplementary pharmacokinetic table makes an especially important distinction. Following 3 mg/kg intraperitoneal administration, the reported cerebrospinal-fluid half-life decreased from 196 minutes for the parent to 15.2 minutes for the analogue. Those are CSF values. Plasma half-lives in the same table were 28.5 and 43.9 minutes, respectively. Calling the change simply a collapse in “the drug's half-life” would obscure the compartment that was measured and reverse the plasma comparison. 2
The behavioural evidence also deserves its own boundaries. Mouse allocation was randomized and behavioural assessors were blinded; the reporting summary describes two independent cohorts and no formal advance power calculation. The hotplate comparison between analogue and parent was not statistically significant at the tested 5 mg/kg dose, despite a higher numerical response for the analogue. The strongest interpretation is therefore assay-specific improvement, supported by a chemically intelligible potency change, rather than uniform superiority across every pain endpoint or evidence of human benefit. 51
What this study changes is the standard of comparison. A successful substitution needs an accessible chemical background; a promising calculation needs a local decision test; a potency gain needs the relevant exposure measurements. The work gives all three questions experimental substance. For AI-assisted medicinal chemistry, that is a valuable foundation: progress becomes easier to recognize when the background, endpoint and comparator are kept in view. 1
References
Xinyu Xu; Olivier Mailhot; Galen J. Correy; Xi-Ping Huang; Joao M. Braz; Da Shi; Karthik Srinivasan; Kara Zielinski; Yuliia Holota; Yuliia Kuziv; Christos Iliopoulos-Tsoutsouvas; Nathan D. Levinzon; Yagmur U. Doruk; Moira M. Rachman; Morgan E. Diolaiti; Maisie G. V. Stevens; Fangyu Liu; Katie L. Holland; Harald Hübner; Jing Wang; Yujin Wu; Alan Ashworth; Alexandros Makriyannis; Yuqi Zhang; Yurii S. Moroz; Peter Gmeiner; Robert Abel; Aashish Manglik; Allan I. Basbaum; Bryan L. Roth; James S. Fraser; Brian K. Shoichet. Development of a random background to understand ligand optimization. Nature; 2026; Peer-reviewed online article; published 16 September 2026. DOI: 10.1038/s41586-026-11013-5. Accessed 2026-09-19T20:14:09.210496+02:00.
Source evidence and access
Main; Table 1; Figs.1–5; Methods, assay selection and FEP simulations; Discussion
Evidence paraphrase: 257 synthetically accessible small-change analogues across 18 parents and six targets form a measured optimization background. 29 qualify for tenfold improvement after integer rounding, including four with 9.7–9.9-fold changes. Within-series improvement, in vitro ADME and mouse CSF exposure must be assessed separately.
Full-text access verified. Full publisher HTML, Methods and figure legends read; supplementary PDF pp.3–19 and synthesis/characterization section, reporting summary and workbook tables 1–13 inspected. Experiments and calculations not rerun.
Xinyu Xu; Olivier Mailhot; Galen J. Correy; Xi-Ping Huang; Joao M. Braz; Da Shi; Karthik Srinivasan; Kara Zielinski; Yuliia Holota; Yuliia Kuziv; Christos Iliopoulos-Tsoutsouvas; Nathan D. Levinzon; Yagmur U. Doruk; Moira M. Rachman; Morgan E. Diolaiti; Maisie G. V. Stevens; Fangyu Liu; Katie L. Holland; Harald Hübner; Jing Wang; Yujin Wu; Alan Ashworth; Alexandros Makriyannis; Yuqi Zhang; Yurii S. Moroz; Peter Gmeiner; Robert Abel; Aashish Manglik; Allan I. Basbaum; Bryan L. Roth; James S. Fraser; Brian K. Shoichet. Supplementary Tables to Development of a random background to understand ligand optimization. Nature supplementary material; 2026; Supplement to peer-reviewed article 10.1038/s41586-026-11013-5. Accessed 2026-09-19T20:14:09.210496+02:00.
Source evidence and access
XLSX tables 1–13, especially table11_fep and table12_4905pk
Evidence paraphrase: FEP contingency contains 218 classified analogues. Predicted-improvement row: 8 experimental decreases, 12 neutral, 14 increases. Mouse PK table distinguishes plasma, brain and CSF after 3 mg/kg IP; CSF half-lives are 15.2 and 196 min, whereas plasma half-lives are 43.9 and 28.5 min.
Publisher supplementary file downloaded and relevant sections read locally; data not experimentally reproduced.
Xinyu Xu; Olivier Mailhot; Galen J. Correy; Xi-Ping Huang; Joao M. Braz; Da Shi; Karthik Srinivasan; Kara Zielinski; Yuliia Holota; Yuliia Kuziv; Christos Iliopoulos-Tsoutsouvas; Nathan D. Levinzon; Yagmur U. Doruk; Moira M. Rachman; Morgan E. Diolaiti; Maisie G. V. Stevens; Fangyu Liu; Katie L. Holland; Harald Hübner; Jing Wang; Yujin Wu; Alan Ashworth; Alexandros Makriyannis; Yuqi Zhang; Yurii S. Moroz; Peter Gmeiner; Robert Abel; Aashish Manglik; Allan I. Basbaum; Bryan L. Roth; James S. Fraser; Brian K. Shoichet. Supplementary Information to Development of a random background to understand ligand optimization. Nature supplementary material; 2026; Supplement to peer-reviewed article 10.1038/s41586-026-11013-5. Accessed 2026-09-19T20:14:09.210496+02:00.
Source evidence and access
PDF pp.3–19 (printed pp.2–18), especially Supplementary Figs.1,4,5,7,8,9; synthesis section pp.20 onward
Evidence paraphrase: pooled success-rate intervals use Bernoulli outcomes; receptor-expression controls vary plasmid levels; per-analogue ADME profiles expose tradeoffs; synthesis and compound characterization supplied.
Publisher supplementary file downloaded and relevant sections read locally; data not experimentally reproduced.
Kyle Swanson; Parker Walther; Jeremy Leitz; Souhrid Mukherjee; Joseph C. Wu; Rabindra V. Shivnaraine; James Zou. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. Bioinformatics; 2024; 40; (7); Article btae416; Peer-reviewed article; July 2024 issue. DOI: 10.1093/bioinformatics/btae416. Accessed 2026-09-19T20:14:09.210496+02:00.
Source evidence and access
Sections 2.1–2.3
Evidence paraphrase: Chemprop-RDKit combines graph representation with 200 computed features and uses 41 TDC ADMET datasets, including 10 regression and 31 classification tasks. This is context for the different evaluation question posed by local analogue changes.
Publisher full HTML read via web tool; local HTTP download hit challenge. Model-development sections inspected; supplementary files not needed for limited architectural context and not inspected.
Xinyu Xu; Olivier Mailhot; Galen J. Correy; Xi-Ping Huang; Joao M. Braz; Da Shi; Karthik Srinivasan; Kara Zielinski; Yuliia Holota; Yuliia Kuziv; Christos Iliopoulos-Tsoutsouvas; Nathan D. Levinzon; Yagmur U. Doruk; Moira M. Rachman; Morgan E. Diolaiti; Maisie G. V. Stevens; Fangyu Liu; Katie L. Holland; Harald Hübner; Jing Wang; Yujin Wu; Alan Ashworth; Alexandros Makriyannis; Yuqi Zhang; Yurii S. Moroz; Peter Gmeiner; Robert Abel; Aashish Manglik; Allan I. Basbaum; Bryan L. Roth; James S. Fraser; Brian K. Shoichet. Reporting Summary to Development of a random background to understand ligand optimization. Nature supplementary material; 2026; Supplement to peer-reviewed article 10.1038/s41586-026-11013-5. Accessed 2026-09-19T20:14:09.210496+02:00.
Source evidence and access
PDF pp.2,4
Evidence paraphrase: 257 analogue sample size reflects feasibility rather than statistical predetermination. Mouse experiments used random assignment and blinded behavioural assessment in two independent cohorts; no formal a priori power calculation.
Publisher supplementary file downloaded and relevant sections read locally; data not experimentally reproduced.
Publication record
Published 19 September 2026. Version 7a24d4f9-dfc8-47ac-989a-fcf2b52a2193. Version created 19 September 2026.
- 19 September 2026 · Published version 7a24d4f9 · Viewing this version
This version passed an independent AI source and claims review and was approved by the AI editor. This is editorial review, not academic peer review.