Bespoke library docking for 5-HT2A receptor agonists with antidepressant activity

The investigators generated a bespoke virtual library of THPs from defined, commercially available building blocks and a convergent one-pot synthetic sequence (C–H alkenylation, electrocyclization and reduction) that yields three THP subtypes with up to six derivatisation sites. Inputs were filtered for single-isomer, achiral reagents and lead-like properties; initial reactant mass cut-offs and an assembled-product cap of ≤400 amu and cLogP ≤3.5 produced 4.3 billion virtual THPs. To prioritise molecules with high ligand efficiency for the initial screen, the library was further restricted to ≤350 amu, yielding a 75 million compound set for docking. Because no experimental 5-HT2A R structure was available at the time, the team built 1,000 homology models based on the 5-HT2B R–LSD X-ray structure, retaining LSD to preserve a ligand-competent orthosteric site. Models were evaluated for enrichment of 34 known 5-HT2A ligands versus decoys and for reproduction of key interactions (for example, salt bridge to Asp3.32). The best-performing model was used for docking with DOCK3.7: each library molecule was represented by presampled conformers (mean ~92) and sampled in ~23,000 orientations, yielding ~7.46 trillion sampled complexes. The top 300,000 docked molecules were clustered (ECFP4 Tc > 0.5) and manually triaged for strain, unsatisfied donors and appropriate interactions; 205 candidates were filtered for topological novelty (ECFP4 Tc < 0.35 versus ~28,000 aminergic ligands) and 30 compounds were prioritised for synthesis, of which 17 were successfully made. Syntheses used the convergent one-pot sequence from imines and alkynes to obtain THP subtypes, with options for late-stage N-substituent modifications; final products were purified and characterised by NMR, HRMS and chiral HPLC when relevant. Pharmacology employed radioligand displacement assays (3H-LSD or 3H-mesulergine) and functional assays including calcium-flux and BRET-based G-protein and β-arrestin recruitment. MD and free-energy perturbation (FEP) calculations were applied during optimisation to refine poses and predict relative affinities. For structural validation, the 5-HT2A R–(R)-69 complex was formed with a miniGαqi heterotrimer and scFv16, purified and subjected to single-particle cryo-EM; the final map was refined to an overall nominal resolution of ~3.4 Å. In vivo pharmacokinetics and a battery of behavioural assays in mice (head-twitch response, prepulse inhibition, open field locomotion, conditioned place preference, behavioural sensitization, tail suspension and learned helplessness, plus sucrose preference and elevated zero maze) were used to assess psychedelic-like, locomotor and antidepressant-like effects. Behavioural statistics used ANOVA and related methods with P < 0.05 as the threshold for significance.

Docking of the 75 million THP library against the best homology model sampled ~7.46 trillion receptor–ligand complexes; after clustering and manual selection, 17 THPs were synthesised from docking-prioritised clusters. Radioligand displacement assays across 5-HT2A, 5-HT2B and 5-HT2C receptors identified four compounds with Ki values between 0.67 μM and 3.9 μM, corresponding to a 24% hit rate among the 17 synthesised molecules. Functional assays showed these initial hits could act as agonists or antagonists at 5-HT2 subtypes, with agonist potencies at 5-HT2B R in the 1.9–3.0 μM range for some molecules. Docked poses of confirmed ligands consistently suggested a salt bridge between the THP tertiary amine and Asp155 (Asp3.32), hydrogen bonds to Asn3436.55 and contacts with residues on TM3, TM5 and TM6. Structure-based optimisation focused on varying the THP N-substituent and C5 heteroaryl groups. Racemic and then enantiomerically resolved analogues showed strong chiral discrimination; docking correctly predicted instances of stereoselective binding. Exploration of the larger 4.3-billion virtual set led to azaindole substituents at C5, and separation of 14 azaindole enantiomers produced potent partial agonists (R)-69 and (R)-70. In a calcium-flux assay, (R)-69 and (R)-70 had EC50 values of 41 nM and 110 nM, respectively. A selective antagonist (S)-65 with 59 nM activity was also found by substitution. Selectivity profiling showed (R)-69 was 4.6-fold selective for 5-HT2A R over 5-HT2B R (EC50 190 nM at 5-HT2B), and (R)-70 was 6.4-fold selective versus 5-HT2B R; both compounds were 29–51-fold selective over 5-HT2C R. Screening across 318 GPCRs using a β-arrestin-recruitment assay found no measurable agonism for either (R)-69 or (R)-70, and neither compound inhibited hERG up to 10 μM. Ligand displacement assays revealed no substantial binding at 45 other off-targets; (R)-69 had modest SERT activity (0.8 μM) whereas (R)-70 had none measurable. The optimised THPs showed unusually high ligand efficiencies (up to 0.6 kcal per heavy atom) and were G-protein-biased agonists: they robustly activated Gq signalling at mid-nanomolar concentrations, while β-arrestin-2 recruitment was detectable only at higher concentrations. Time-resolved transduction coefficients confirmed stable differences in Gq versus β-arrestin kinetics compared with classical 5-HT2A R ligands. To refine binding hypotheses, the team generated multiple MD-stable poses for (S)-69 and used retrospective FEP against analogue affinities to select the predicted pose, which featured the Asp155 salt bridge, a Ser2425.46 hydrogen bond and packing with Phe2226.52. Cryo-EM of the 5-HT2A R–(R)-69–miniGq/i complex produced a high-quality map at ~3.4 Å that permitted unambiguous modelling of (R)-69. The structure corroborated the salt bridge to Asp155 and revealed a π-stacking interaction between the azaindole of (R)-69 and Phe340, plus hydrophobic contacts with Val1563.33, Val2355.39, Trp3366.48 and Phe339. The experimentally determined pose superposed well with the predicted pose, although the THP ring adopted a slightly different conformation with the C3 methyl pointing more extracellularly. Pharmacokinetic studies in mice after intraperitoneal dosing (10 mg kg-1) showed substantial brain penetration: Cmax ≈ 12 μM for (R)-69 and ≈ 35 μM for (R)-70, plasma:brain ratios of 1.09 and 0.11, and brain half-lives of 111 min and 28.2 min, respectively; exposures after 1 mg kg-1 remained substantial. Behaviourally, both (R)-69 and (R)-70 induced very low head-twitch responses (HTRs) compared with LSD and, at tested doses, significantly blocked LSD-induced HTRs, consistent with low psychedelic-like activity. Neither compound affected prepulse inhibition, and neither produced locomotor stimulation, behavioural sensitisation or conditioned place preference versus cocaine. In VMAT2 heterozygous (HET) mice, which display depressive-like phenotypes, acute administration of fluoxetine (1 mg kg-1), (R)-69 (1 mg kg-1) or (R)-70 (0.5–1 mg kg-1) normalised elevated immobility in the tail suspension test, with effects persisting at 24 h for both THPs. In a learned helplessness paradigm using C57BL/6J mice, a single 1 mg kg-1 dose of (R)-70 produced immediate and sustained increases in sucrose preference in foot-shocked mice for up to 3 days, similar to psilocin and with a faster onset than ketamine in this assay; (R)-70 also reduced immobility in the tail suspension test over multiple days and attenuated anxiety-like behaviour in the elevated zero maze. Escape performance deficits induced by foot shock were not reversed by the treatments, and antagonist studies to map receptor mediation were confounded because the antagonists alone affected immobility.

Kaplan and colleagues present a workflow in which a bespoke, ultralarge virtual library built around an under-sampled but synthetically tractable scaffold (THPs) can yield novel, biologically active aminergic ligands. The authors emphasise three main points: first, a focused 75 million compound THP library explored scaffold space largely absent from general make-on-demand libraries and produced docking-prioritised molecules active at 5-HT2 receptors; second, the cryo-EM structure of the 5-HT2A R–(R)-69 complex confirmed the predicted binding mode and can guide further optimisation; third, optimisation produced partial 5-HT2A R agonists ((R)-69 and (R)-70) that are G-protein-biased, lack overt psychedelic-like and locomotor-stimulating properties in mice, yet exert anxiolytic- and antidepressant-like effects in genetic and learned helplessness models at low doses. The investigators acknowledge several caveats. Rapid synthesis of richly functionalised scaffolds depends on compatible convergent chemistries and so not all scaffolds are amenable to such bespoke libraries. Although the initial hit rate across 5-HT2 receptor subtypes was high (24% for the initial set), the specific hit rate for 5-HT2A R was lower, requiring extensive structure–activity exploration. Docking alone was insufficient to predict the correct binding pose reliably; success required a combination of stable MD geometries and retrospective FEP calculations. The unusual phenotype — antidepressant-like activity without psychedelic-like actions — may relate to the observed G-protein versus β-arrestin functional selectivity, but the authors admit that roles for off-targets, other 5-HT subtypes or active metabolites cannot be excluded. Finally, while the behavioural data suggest therapeutic potential, further optimisation and investigation are needed before these molecules could be considered drug candidates. More broadly, the authors propose that multiple synthetic strategies could be used to create specialised virtual libraries around under-explored scaffolds, potentially illuminating receptor pharmacologies inaccessible to general-purpose libraries. They frame their findings as proof-of-concept that focused ultralarge virtual libraries, combined with iterative synthesis, structural biology and in vivo pharmacology, can deliver novel, selective ligands with distinct signalling and therapeutic-relevant profiles.