SandboxAQ Launches AI-Powered Virtual Screening Platform That Predicts GPCR Drug Mechanism of Action
核心洞察
SandboxAQ (搜索) launched a new virtual screening solution for GPCR (搜索) drug discovery that predicts not only binding but also whether a molecule acts as an agonist, antagonist, or inverse agonist.
The platform uses large quantitative models and physics-based modeling to generate receptor structures, screen compound libraries, and infer functional activity before wet-lab testing.
In a retrospective benchmark, the solution correctly identified every antagonist in the test set, and its initial ML binder-screening step achieved 79% accuracy and 83% specificity.
SandboxAQ (搜索) today announced the launch of a new virtual screening solution designed to accelerate drug discovery against G protein-coupled receptors (GPCRs), a family of targets that accounts for roughly one-third of all approved drugs yet remains notoriously difficult to drug. The GPCR (搜索) Virtual Screening Solution, accelerated by NVIDIA (搜索)'s BioNeMo Agent Toolkit, goes beyond traditional computational methods by predicting not only whether a molecule binds a receptor, but also what the molecule does once bound—specifically, whether it activates or blocks receptor signaling.
"For decades, the field has largely been able to ask one question computationally: will this molecule bind?" said Andrea Bortolato, VP of Drug Discovery at SandboxAQ (搜索). "Another important question for a drug discovery program is what the molecule does once it binds. By modeling the physics of how a compound shifts a receptor between its active and inactive states, we can predict whether it activates the target or blocks its activation, and we can do it before molecules are made in the lab."
A Three-Step Workflow Bridging Structure and Function
The platform operates through a sequential three-step process that progressively refines candidate selection. The first step generates biologically relevant GPCR (搜索) structures using SandboxAQ (搜索)'s large quantitative models (LQMs), producing high-quality structural models of both active and inactive receptor states. This structural foundation incorporates multiple state-of-the-art protein structure prediction approaches and can integrate additional biological context, such as G protein interactions, when necessary.
"An important part of GPCR (搜索) discovery is getting the receptor structure right at the start," Bortolato explained. "Our workflow uses cofolding to model the active and inactive receptor states that determine downstream signaling behavior. Because we build on OpenFold3, we also benefit from advances such as cuEquivariance, and we are now testing TensorRT acceleration for Pairformer inference through our agent skills to make this step faster and more scalable."
The second step employs machine learning to rapidly screen full compound libraries and flag likely binders. In an initial ML binder-screening evaluation for a complex GPCR (搜索) target, the method reached 79% accuracy and 83% specificity, demonstrating effectiveness at filtering out unlikely candidates before costly follow-up work.
The third and most distinctive step applies rigorous physics-based modeling to the narrowed set of candidates, determining whether each leading compound stabilizes the active or inactive conformations of the target. This enables the platform to predict whether a candidate acts as an agonist, antagonist, or inverse agonist—connecting static structural data to actionable pharmacology.
Benchmark Performance and Clinical Relevance
In a retrospective mechanism-of-action benchmark, SandboxAQ (搜索)'s Virtual Screening Solution correctly identified every antagonist in the test set, successfully separating receptor-blocking molecules from activators by modeling how each compound shifts the receptor's energy between its active and inactive states.
The practical implications for drug development are substantial. A single drug program can require synthesizing and testing thousands of compounds, each consuming time and financial resources. By predicting molecular behavior before synthesis, the solution allows research teams to narrow the field before entering the laboratory, reducing wasted effort, lowering costs, and accelerating the decisions that move programs forward.
Strategic Vision and NVIDIA (搜索) Collaboration
NVIDIA (搜索)'s BioNeMo Agent Toolkit provides the computing foundation for SandboxAQ (搜索)'s large-scale modeling and screening workloads. "At SandboxAQ we see significant potential in the BioNeMo Agent Toolkit. We look forward to exploring the agent toolkit for accelerating our GPCR (搜索) Virtual Screening workflows as well as additional projects," said Bill Fitzgerald, Vice President of Growth and Ecosystems at SandboxAQ.
Looking beyond the immediate launch, SandboxAQ (搜索) plans to extend the framework across more GPCR (搜索) classes and toward some of the field's most challenging problems. "What excites us most is where this leads," Bortolato added. "We're focused on extending the framework across more GPCR classes, and over time toward some of the hardest open problems in the field, including orphan receptors that have no known partner molecule and have so far resisted successful drug discovery."
In the near term, the company aims to demonstrate the broad applicability of its agonist-versus-antagonist prediction framework across GPCR (搜索) families. Building on that foundation, SandboxAQ (搜索) intends to expand toward higher-order challenges including receptor deorphanization, identification of novel therapeutic targets, and systematic discovery of allosteric binding sites—areas where predictive, structure-based approaches could significantly reduce reliance on large-scale empirical screening and unlock new therapeutic opportunities.