Summary
A Nature study systematically made 257 small chemical changes to 18 ligands across six targets. Eleven-point-three percent improved measured affinity or potency at least tenfold, but those gains often came with worse in-vitro pharmacokinetic properties.
Researchers have created a reference point for judging how often small, unplanned chemical changes improve a drug-like molecule’s activity. In an open-access Nature study, systematic single-atom modifications improved measured affinity or potency by at least tenfold in 29 of 257 analogue molecules, or 11.3% of the tested set.
The result offers a background expectation for ligand optimization—the process of changing an initial active molecule so that it binds its biological target more strongly or produces a more useful response. It also exposed a recurring trade-off: stronger target activity often came with poorer pharmacokinetic properties, which influence a molecule’s stability, distribution and exposure in the body.
A controlled background for ligand optimization
Drug discovery usually begins with a “hit” that shows activity against a target. Hundreds of related molecules may then be designed and synthesised before a candidate is selected. The efficiency of that process is difficult to measure because published examples tend to favour molecules that were deliberately improved.
The researchers addressed this by constructing what they call a random background set. The term refers to a reference set in which changes were applied systematically rather than chosen because they were expected to work. They started with 18 parent ligands covering six targets: three G-protein-coupled receptors—the α2A adrenergic receptor, μ-opioid receptor and cannabinoid receptor 2—along with the serotonin transporter, AmpC β-lactamase and the Mac1 domain of SARS-CoV-2.
Each parent was modified one atom at a time where synthesis was practical and the molecule’s net charge at physiological pH would not change. The changes included replacing a hydrogen with methyl, hydroxyl, chlorine, fluorine or, less often, bromine, and replacing selected aromatic carbon atoms with nitrogen. The team synthesised and tested 257 analogues, measuring activity through the appropriate binding, signalling, transport or enzyme assays for each target.
Twenty-nine analogues improved activity tenfold or more. The study treated four changes measuring between 9.7-fold and 9.9-fold as tenfold improvements, using the same convention in its analysis. The resulting estimate was 11.3%, with a 95% confidence interval of 7.5% to 15.2%. These large improvements appeared for five of the six targets and for 10 of the 18 parent ligands.
Methyl substitutions were the most likely to produce tenfold improvements, followed by chlorine substitutions. The strongest gains were not simply a consequence of making molecules more hydrophobic: the analogues with tenfold improvements also increased ligand efficiency and lipophilic efficiency, measures that relate activity gains to the chemical changes used to achieve them.
Potency gains came with pharmacokinetic trade-offs
The team also measured six in-vitro properties relevant to pharmacokinetics: stability in liver microsomes, plasma stability, the fraction unbound to plasma proteins, solubility, membrane permeability and inhibition of the hERG potassium channel. Individual properties could improve after a small chemical change, but improvements were rarely coordinated.
None of the 29 analogues with tenfold-or-greater activity improvements improved all three of the key measures examined together: microsomal stability, permeability and plasma fraction unbound. Across the full set, 41.4% of analogues deteriorated by more than threefold in at least one of these properties, while 34.5% deteriorated in all three. The pattern shows why a molecule that binds more strongly is not automatically a better drug-discovery lead.
Computational methods captured some of the activity changes but could not replace experimental testing in this benchmark. For 219 analogue–parent pairs, free-energy calculations using FEP+ placed 188 predictions within 2 kilocalories per mole of the measured binding free-energy change and 129 within 1 kilocalorie per mole. The mean unsigned error was 1.06 kilocalories per mole. Models for in-vitro pharmacokinetics correlated with measured permeability and plasma fraction unbound, while predictions of microsomal stability, plasma stability and solubility were essentially uncorrelated with the experiments.
One example showed how medicinal chemistry can navigate the trade-off. Adding a single methyl group to the α2A receptor ligand ‘3629 produced analogue ‘4905, which had 52-fold higher potency in the receptor assay. Compared with its parent, however, ‘4905 had a 57% lower fraction unbound and a microsomal half-life reduced by 41%. In mice, its cerebrospinal-fluid concentration was 36% lower and its half-life fell from 196 minutes to 15.2 minutes.
Even with those disadvantages, ‘4905 produced stronger analgesic effects than ‘3629 in tail-flick, neuropathic-pain and hotplate tests, and was at least as potent as PS75 in the reported comparisons. These were randomized and blinded experiments in male C57BL/6J mice aged 7–8 weeks, making the result preclinical evidence for lead optimization rather than a human treatment result.
The benchmark remains limited to 257 analogues and six targets selected partly for experimental access. Most pharmacokinetic measurements were in vitro; liver microsomes can miss metabolic routes such as glucuronidation, and in-vivo behaviour is only approximately predicted by these assays. The estimate of an 11.3% improvement rate may therefore change as the approach is extended to more targets, scaffolds and chemical modifications.