Abstract
G protein-coupled receptors (GPCRs) represent prominent therapeutic targets for neurodegenerative disorders, including Alzheimer's and Parkinson's diseases. However, the high sequence conservation of orthosteric binding pockets across receptor subtypes severely limits traditional drug discovery due to widespread off-target toxicity. Allosteric modulators circumvent this limitation by binding to spatially distinct, less conserved topographies, yet their rational identification remains impeded by the conformational plasticity and transient nature of allosteric sites. Here, we present DeepAllosteric, an integrated computational framework combining equivariant graph neural networks (GNNs) with structure-conditioned latent diffusion models to discover subtype-selective positive allosteric modulators (PAMs). Applied to the human muscarinic acetylcholine receptor M1 (M1 mAChR) and metabotropic glutamate receptor 5 (mGluR5), our framework accurately identified cryptic allosteric pockets within dynamic receptor ensembles derived from microsecond-scale molecular dynamics simulations. De novo generated candidate molecules exhibited optimal drug-like metrics, high synthetic tractability, and sub-micromolar predicted binding affinities. Orthogonal biochemical assays and cellular calcium mobilization experiments validated our top predicted leads, confirming PAM activity with nanomolar potencies (EC50 values of 42 nM for M1 and 78 nM for mGluR5) and over 50-fold selectivity against homologous subtypes (M2–M5 and mGluR1). These findings demonstrate that deep generative modeling conditioned on dynamic allosteric microenvironments significantly accelerates the discovery of highly selective GPCR neurotherapeutics.
