Deep Learning Dynamic Allostery of G-Protein-Coupled Receptors.

G-protein-coupled receptors (GPCRs) make up the largest superfamily of human membrane proteins and represent primary targets of ∼1/3 of currently marketed drugs. Allosteric modulators have emerged as more selective drug candidates compared with orthosteric agonists and antagonists. However, many X-ray and cryo-EM structures of GPCRs resolved so far exhibit negligible differences upon the binding of positive and negative allosteric modulators (PAMs and NAMs). The mechanism of dynamic allosteric modulation in GPCRs remains unclear. In this work, we have systematically mapped dynamic changes in free energy landscapes of GPCRs upon binding of allosteric modulators using the Gaussian accelerated molecular dynamics (GaMD), deep learning (DL), and free energy prOfiling Workflow (GLOW). GaMD simulations were performed for a total of 66 μs on 44 GPCR systems in the presence and absence of the modulator. DL and free energy calculations revealed significantly reduced dynamic fluctuations and conformational space of GPCRs upon modulator binding. While the modulator-free GPCRs often sampled multiple low-energy conformational states, the NAMs and PAMs confined the inactive and active agonist-G-protein-bound GPCRs, respectively, to mostly only one specific conformation for signaling. Such cooperative effects were significantly reduced for binding of the selective modulators to "non-cognate" receptor subtypes. Therefore, GPCR allostery exhibits a dynamic "conformational selection" mechanism. In the absence of available modulator-bound structures as for most current GPCRs, it is critical to use a structural ensemble of representative GPCR conformations rather than a single structure for compound docking ("ensemble docking"), which will potentially improve structure-based design of novel allosteric drugs of GPCRs.