Synaptic vesicle glycoprotein 2C (SV2C) is a vesicular protein highly expressed in dopaminergic neurons of the basal ganglia that modulates dopamine storage and release. Genetic and pathological evidence links SV2C disruption to Parkinson’s disease (PD), yet selective small-molecule probes for SV2C were not available prior to this study. The absence of a full-length high-resolution SV2C structure presented a challenge for structure-based discovery.
The authors implemented an AI-enhanced virtual screening (VS) workflow to discover SV2C ligands from a large general-purpose commercial library. A convolutional neural network-based scoring function (CNN_VS) was central to the pipeline and was applied in a multi-stage funnel to prioritize compounds for experimental testing. The initial library comprised 5.96 million Mcule in-stock compounds; these were sequentially filtered for central nervous system relevance, narrowing the set to 3.19 million molecules for docking and rescoring.
Because no full-length high-resolution SV2C structure was available, the team constructed a homology model of SV2C using available SV2A cryo-EM structures as templates. They characterized the conformational landscape of the SV2C model using classical molecular dynamics (MD) and Gaussian accelerated MD (GaMD) simulations in both the apo state and in complex with known SV2 ligands: plosaracetam, levetiracetam, brivaracetam, and padsevonil. These simulations informed the virtual screening by sampling relevant receptor conformations.
Prior to prospective application, the CNN_VS scoring function was retrospectively validated on a manually curated SV2A benchmark of 39 ligands, where it achieved a correlation of r = 0.72 versus experimental pIC50 values. This retrospective performance supported the use of CNN_VS within the multi-stage funnel applied to the filtered compound set.
From the VS-prioritized list, 94 candidates were selected for follow-up and 71 compounds were experimentally profiled. The orthogonal primary assay cascade combined a thermal shift assay (TSA) with a [3H]-padsevonil scintillation proximity assay (SPA). Key hits were further characterized by Ki determination and isoform selectivity profiling against SV2A and SV2B. This medium-throughput biophysical strategy allowed efficient triage and confirmation of binding activity.
Experimental testing yielded 22 active molecules, corresponding to a 31% hit rate among the 71 screened compounds. The active molecules naturally segregated into two categories: compounds that competed with padsevonil for the primary site (padsevonil-site competitors), and compounds that did not show competition at that site. A subset of the competitor compounds also produced thermostabilization in the TSA, consistent with ligand engagement of the modeled binding pocket.
Among the confirmed hits, two compounds were highlighted as attractive leads. Compound 36 exhibited a Ki of 24.6 µM at SV2C, while compound 56 had a Ki of 3.25 µM at SV2C. Both leads showed greater than 10-fold selectivity versus SV2A, and compound 56 additionally maintained approximately 12-fold selectivity relative to SV2B. These properties position the compounds as starting points for hit-to-lead optimization and for generating tools to probe SV2C function.
Docking analyses of hits suggested a common binding mode anchored by conserved tryptophan residues within the SV2 binding pocket, described as a tryptophan cage. This predicted binding arrangement was independently supported by an unpublished SV2A–plosaracetam cryo-EM structure, which showed a 0.76 Å binding-site Cα root-mean-square deviation (RMSD) relative to the SV2C model and complete conservation of the tryptophan cage motif. The authors note that subtle differences in luminal domain and transmembrane regions among SV2 isoforms likely account for observed isoform selectivity.
The study demonstrates that an integrated AI-driven virtual screening pipeline, tightly coupled with medium-throughput biophysical assays, can deliver selective ligands for a membrane protein lacking a full high-resolution structure. The identified chemotypes include both padsevonil-site competitors and non-competitors, and provide multiple starting points for medicinal chemistry optimization, functional characterization, and tool development to probe SV2C biology and its role in dopaminergic signaling and PD. The authors position these hits as templates for further hit-to-lead work and functional studies.
Several authors are employees of SandboxAQ, and the AQFEP method used in the work is the subject of a SandboxAQ patent. Remaining authors declared no competing interests. Funding sources declared include the National Institutes of Health and the SPARK NS Translational Research Program. Details beyond these declarations were reported in the source.