SandboxAQ has released a new AI model designed to solve one of drug discovery’s oldest bottlenecks: figuring out which molecules are worth testing in the lab when scientists don’t have a detailed map of the target they’re trying to hit.
On August 19, 2026, the Palo Alto-based company announced the general availability of AQPotency, a “Large Quantitative Model” (LQM) built to predict how strongly a candidate drug molecule will bind to and act on a disease-related protein. Unlike many computational screening tools, AQPotency doesn’t require a solved 3D structure of that protein to make its predictions — a distinction that, if it holds up, could open the door to targets that structure-based methods simply can’t reach.
The launch arrived alongside the general availability of a second SandboxAQ model, AQCat Adsorption Spin, aimed at materials and catalyst discovery. Both are being distributed primarily through Anthropic’s Claude, using the Model Context Protocol (MCP), with availability on Google Cloud’s Marketplace described as forthcoming.
Table of Contents
Why It Matters — The Structure Problem
Early-stage drug discovery is largely a process of elimination. Researchers have to decide which molecules, out of enormous chemical libraries, are worth the time and expense of lab testing. Guess wrong too often, and a program stalls before it produces anything useful.
The computational tools traditionally used to narrow that field — docking simulations, free-energy calculations, and similar structure-based methods — share a common requirement: they need a detailed 3D structure of the target protein to work. According to SandboxAQ, many of the most promising disease targets don’t have one, which means those methods never get a chance to help. That’s the gap AQPotency is designed to fill.
It’s worth being clear that this framing of the problem comes from SandboxAQ itself. It’s a reasonable and widely recognized description of a real limitation in structure-based drug design, but it’s the company’s own rationale for why its product matters — not an independently adjudicated industry consensus.
How AQPotency Differs From Structure-Based Tools
AQPotency ranks pairs of molecules and biological targets by predicted binding activity, without needing a crystal structure or predicted 3D model of the target. That sets it apart from structure-based approaches associated with tools like AlphaFold and its commercial successors, which start from a physical or predicted structure and work outward.
The model also runs in the opposite direction. Given a single molecule that shows a promising effect — even one whose mechanism isn’t understood — AQPotency can scan a broad panel of proteins across the body and return a ranked list of likely targets. That kind of target deconvolution is useful when a compound works but researchers don’t yet know why.
Each prediction comes with a confidence or reliability score, which SandboxAQ positions as a meaningful improvement over tools that return a single score with no indication of how much to trust it.
None of this should be read as AQPotency outperforming structure-based methods. No independent, head-to-head comparison between AQPotency and tools from companies like Isomorphic Labs or Schrödinger has been published. The two approaches are better understood as complementary — one for targets with known structures, one for targets without.
| Requires a Solved 3D Target Structure? | Primary Method | |
|---|---|---|
| AQPotency (SandboxAQ) | No | Physics-based quantitative model (LQM) |
| Isomorphic Labs | Yes | AlphaFold-lineage structure prediction |
| Schrödinger | Yes | Physics-based simulation (docking, free-energy calculation) plus machine learning |
The “Quantitative,” Not “Quantum,” Model
The name Large Quantitative Model invites an obvious question: does this run on quantum computers?
The answer, based on how SandboxAQ describes its broader technology platform, is no — not yet, and not as a requirement. The company’s LQMs run today on classical GPU and high-performance computing infrastructure, including NVIDIA’s DGX Cloud. SandboxAQ describes the approach as “quantum-inspired” and “quantum-ready,” meaning it borrows techniques and physics-based training data associated with quantum chemistry, and is built to take advantage of quantum hardware if and when it becomes practical for this kind of molecular modeling — but it doesn’t depend on that hardware existing today.
That distinction is worth holding onto, because “quantum” in the marketing sense and “quantum computing” in the literal sense are easy to conflate. AQPotency’s underlying architecture hasn’t been detailed in a dedicated technical paper, so this description is drawn from how SandboxAQ characterizes its LQM platform as a whole rather than from AQPotency-specific documentation — but it’s consistent across the company’s own statements and independent reporting on the platform.
Claimed Performance and Evidence
SandboxAQ says AQPotency returns predictions in seconds and can cost as little as $1 per 1,000 molecule-target comparisons. Andrea Bortolato, the company’s VP of Drug Discovery, said in the announcement that the model “has already been successfully used in eight customer programs with experimentally validated impact.”
Those are notable claims, and they come with an important caveat: they’re company-reported, attributed to a named SandboxAQ executive, without independent benchmarking or a breakdown of what “experimentally validated” means in each case. No figure for how many compounds AQPotency can screen per hour or per day has been made public.
The strongest evidence in the announcement comes from two named academic collaborations rather than the unspecified customer count. Professor Dario Alessi of the University of Dundee’s MRC Protein Phosphorylation Unit credited SandboxAQ’s models with helping his team “explore a much larger biochemical space” in Parkinson’s disease research. Separately, Dr. Gary W. Miller of Columbia University’s Mailman School of Public Health described a collaboration that used AQPotency to identify selective binders for SV2C, a difficult membrane target linked to dopaminergic signaling, as a foundation for potential Parkinson’s therapeutics.
Those are real, on-the-record academic partnerships — stronger footing than an anonymous customer count. But no peer-reviewed publication validating AQPotency’s accuracy has been identified. Until one exists, the model’s performance claims should be read as promising but unverified by outside researchers.
Distribution Strategy — Why Claude and MCP
The choice to launch through Claude, rather than a specialized scientific software platform, is itself part of the story. Using the Model Context Protocol, researchers can query AQPotency in natural language, without writing custom code or maintaining separate infrastructure.
This wasn’t SandboxAQ’s first move in that direction. In May 2026, the company integrated an earlier LQM, AQCat, with Claude on a waitlist basis. The following month, it announced plans to bring its models to Google Cloud’s Marketplace, with AQCat arriving first and AQPotency expected to follow. August’s announcement completes that sequence for AQPotency, while the Google Cloud listing remains pending.
SandboxAQ’s 2026 rollout, in brief:
- March 2022 — SandboxAQ spins off from Alphabet as an independent company.
- May 2026 — AQCat becomes available on Claude via MCP, on a waitlist basis.
- June 2026 — SandboxAQ announces plans to bring its models to Google Cloud’s Marketplace, with AQCat listed first.
- August 19, 2026 — AQPotency and AQCat Adsorption Spin both reach general availability via Claude/MCP.
Trade coverage has framed this as a deliberate strategic choice: rather than compete purely on model accuracy against well-funded rivals, SandboxAQ is betting that ease of access — letting biologists without machine-learning backgrounds use these tools directly — is the bigger unlock. That framing comes from industry analysis rather than from SandboxAQ itself, and it’s worth noting as context rather than as the company’s own stated strategy.
SandboxAQ’s Corporate Background
SandboxAQ began in 2016 as an internal research group inside Alphabet, known as Sandbox@Alphabet, under Jack Hidary. It spun off as an independent company in March 2022, raising a funding round reported to be in the “nine figures” from investors including former Google CEO Eric Schmidt, who became chairman. Hidary remains CEO.
Alphabet does not own SandboxAQ. Google has continued as an investor and cloud infrastructure partner — a relationship reflected in the planned Google Cloud Marketplace listing — but the company operates independently, with its own leadership, investors, and product roadmap.
Competitive Landscape
AQPotency enters a crowded field of AI-driven drug discovery efforts, each with a different technical approach. Isomorphic Labs, a Google DeepMind spinout led by Demis Hassabis, builds on AlphaFold’s structure-prediction lineage. Schrödinger combines physics-based simulation with machine learning and already has clinical-stage assets developed through pharma partnerships. Insilico Medicine, Recursion Pharmaceuticals, and Relay Therapeutics represent other variations on AI-assisted discovery, ranging from generative molecule design to phenomics-driven screening.
None of these companies has issued a public statement responding to AQPotency’s launch. The competitive picture here is one of parallel approaches rather than direct rivalry playing out in public — SandboxAQ’s structure-free method occupies a different niche than the structure-based tools most closely associated with Isomorphic Labs and Schrödinger, rather than competing head-to-head on the same task.
Limitations and Open Questions
Several details that would help outside observers evaluate AQPotency’s real-world impact simply haven’t been made public. There’s no disclosed throughput figure — how many compounds it can realistically screen in a given period. There’s no named commercial pharmaceutical customer, only the academic collaborators described above and Apheris, a federated-data infrastructure partner whose CEO was quoted in the announcement. And there’s no peer-reviewed study validating the model’s predictions.
None of this means the claims are false. It means they haven’t yet been tested outside the company making them, which is a normal and expected gap at the moment of a product launch rather than evidence of a problem.
What to Watch Next
The most concrete near-term development to watch is AQPotency’s arrival on Google Cloud’s Marketplace, which SandboxAQ has said is coming but hasn’t dated. Given the company’s pattern of rolling out models on Claude before expanding to other platforms — as it did with AQCat — further Large Quantitative Models focused on other domains may follow a similar path.
Independent validation is the bigger question mark. If AQPotency’s predictions hold up in published, peer-reviewed studies or in results from named pharmaceutical partners, it would meaningfully strengthen the case that structure-free screening is a viable complement to existing drug discovery methods. Until then, what’s confirmed is more modest but still real: a working product, generally available, built around a genuinely useful technical premise — with its performance claims still resting on the company’s own word.
Frequently Asked Questions
What is SandboxAQ’s AQPotency? AQPotency is a Large Quantitative Model from SandboxAQ that predicts how strongly a candidate drug molecule will bind to and act on a biological target, without needing a solved 3D structure of that target. It became generally available on August 19, 2026.
How is AQPotency different from AlphaFold-based drug discovery tools? AlphaFold-lineage tools, including those used by Isomorphic Labs, typically start from a known or predicted 3D protein structure. AQPotency is designed to work without one, which SandboxAQ says lets it address targets that structure-based methods can’t currently reach.
Does AQPotency use quantum computers? No. Based on how SandboxAQ describes its Large Quantitative Model platform, AQPotency runs on classical GPU and high-performance computing infrastructure today. The company describes its methods as “quantum-inspired” and its platform as “quantum-ready” for future hardware, but it does not require or run on physical quantum computers.
How much does AQPotency cost to use? SandboxAQ states that AQPotency can cost as little as $1 per 1,000 molecule-target comparisons. This is a company-reported figure that hasn’t been independently benchmarked.
Is SandboxAQ owned by Google or Alphabet? No. SandboxAQ began as an internal Alphabet research group but spun off as an independent company in March 2022. Google remains an investor and cloud partner, but does not own SandboxAQ.
Which companies or researchers are currently using AQPotency? SandboxAQ says the model has been used in eight customer programs, without naming them. The only publicly named collaborators are academic: the University of Dundee’s MRC Protein Phosphorylation Unit and Columbia University’s Mailman School of Public Health, both in Parkinson’s disease-related research.
How can researchers access AQPotency? AQPotency is available now through Claude via the Model Context Protocol and through SandboxAQ’s website. Availability on Google Cloud’s Marketplace has been announced but is not yet live.