SandboxAQ has launched AQPotency, a Large Quantitative Model (LQM) that ranks protein-ligand pairs by predicted binding strength (pIC50) so research teams can prioritize the most promising drug candidates before committing to wet lab testing. Because it needs only a target’s UniProt ID and candidate molecule SMILES — with no experimentally resolved 3D structure — AQPotency runs on standard CPU infrastructure and can score roughly 2,000 protein-ligand pairs in about 15 seconds, at roughly one dollar per thousand on Claude, removing the cost, speed, and structural constraints that limit traditional docking and free energy perturbation workflows. This structure-free approach also opens programs that structure-based methods cannot reach, including membrane targets and proteins that have never been crystallized. On leakage-controlled virtual-screening benchmarks, AQPotency’s early-enrichment ranking is statistically equivalent to docking, with a significant speed advantage.
In this session, Ben Shields and Nihit Pokhrel will walk through how AQPotency supports several core workflows. In single-target virtual screening, researchers evaluate a candidate library against one protein and receive a ranked shortlist with a predicted potency, a per-prediction uncertainty estimate, and an applicability-domain score for each compound. In reverse screening, researchers start with a single active compound and scan it across a proteome-wide panel to generate a ranked list of candidate protein targets for follow-up validation. AQPotency also supports selectivity and off-target analysis across target families and panels such as the kinome and the Bowes safety panel. Unlike opaque scoring tools that return a single number, it ships with a built-in applicability-domain tool that lets teams check for data leakage and benchmark the model against their own data, so scientists know not just what the model predicts but which rankings to rely on. Attendees will also see how natural language prompts can trigger these screening and scanning jobs directly through MCP, without writing code.
Learn how AQPotency ranks protein-ligand pairs by predicted binding strength (pIC50), returning a predicted potency score, a per-prediction uncertainty estimate, and an applicability-domain score for each candidate.
See how AQPotency removes the traditional barriers to virtual screening: it runs on CPU, requires no crystal structure, and can score roughly 2,000 protein-ligand pairs in about 15 seconds, at roughly a dollar per thousand.
Explore AQPotency’s core workflows: single-target virtual screening to rank a candidate library, reverse screening to surface likely protein targets from a single active compound, and selectivity and off-target profiling across target families and safety panels.
Understand how per-prediction uncertainty and applicability-domain scoring — together with a built-in tool for checking data leakage and self-benchmarking — help teams judge which predictions to trust before committing wet lab resources.
See how natural language prompts can launch AQPotency screening and scanning jobs directly through MCP-compatible tooling, without code, GPUs, or existing crystal structures.
