The statistic is stark enough that pharmaceutical executives recite it like a mantra: nine out of ten clinical trials fail. It's baked into the spreadsheets, the investor pitches, the carefully hedged language of earnings calls. But what if the problem isn't inevitability—what if it's prediction?
A small team in San Francisco thinks they've found an answer, and they're hardly alone. Atlas Discovery, a three-person startup that participated in Y Combinator's Summer 2026 batch, is building what it calls "virtual cells"—computational models trained on millions of cellular responses to predict, before the first patient ever enrolls, which drugs will work and which won't. It sounds almost too convenient, the sort of pitch that venture capitalists hear a dozen times a week. Except this time, the infrastructure might actually be ready.
From Xaira Therapeutics to Recursion to a growing constellation of well-funded ventures, a race is underway to build AI models sophisticated enough to simulate human biology at the cellular level. The prize isn't just efficiency. It's a fundamental rewiring of how medicines get made.
When Data Doesn't Translate
The economics remain punishing. Bringing a new drug to market can cost north of a billion dollars, and clinical failure drives most of that expense. A compound might look flawless in a petri dish or a mouse model, then collapse when it meets the messy reality of human metabolism, immune systems, genetic variation.
The industry's response has been to generate data at almost incomprehensible scale. The European Molecular Biology Laboratory now maintains more than 120 petabytes of biological information. North of 440,000 clinical trials have been registered worldwide. And yet translating all that information into better predictions? That part has stayed stubbornly out of reach.
Until maybe now. A survey from Deloitte published this year found that 14% of life sciences executives report they've fully implemented AI tools into daily workflows. Another 40% are working toward it. But the shift isn't just about technology getting smarter—it's about regulators getting comfortable. In January, the FDA and Europe's medicines agency jointly published guiding principles for good AI practice in drug development. By midyear, the FDA had finalized guidance on model-informed drug development, effectively opening pathways for companies to submit computational predictions alongside traditional evidence.
Suddenly, the door isn't just cracked. It's open.
What a Virtual Cell Actually Means
The term itself is slippery—less a technology than a goal. At their simplest, virtual cells are models that simulate how real cells respond to disruptions: a drug molecule binding to a receptor, a gene getting switched off, a disease state kicking in. The most advanced versions pull in multiple data streams—single-cell RNA sequencing, spatial maps of gene expression, tissue samples from actual patients—and try to predict what happens next.
Xaira Therapeutics made waves in March 2026 when it unveiled X-Cell, a model trained on more than 25.6 million perturbed single-cell transcriptomes spanning seven biological contexts. The company, which only launched two years prior, is positioning itself as the leader in this particular subfield. That dataset—called X-Atlas/Pisces—represents a threefold leap over Xaira's earlier efforts, and the company is betting the scale advantage will matter.
But throwing more data at the problem doesn't automatically solve it. Atlas Discovery's technical approach illustrates why nuance matters. The team built something called ExpressionVAE, a discrete-latent model using finite scalar quantization to "tokenize" cellular states before feeding them into a prediction engine. In benchmark tests against continuous-latent alternatives, the architecture showed three-to-twentyfold improvements on distributional accuracy using standard datasets. More telling: when tested on a held-out inflammation benchmark, their frozen encoder matched the performance of a well-known model called scGPT while needing less training data.
Translation: it's not just about size. It's about architecture, about which biological signals get prioritized, about how the model learns causality rather than just correlation.
The Retrospective That Matters
Of course, the real test is clinical. Atlas turned to a completed phase 3 trial called UNIFI, which evaluated a drug called ustekinumab for ulcerative colitis. Using baseline colon biopsies from 358 patients who'd received the treatment, their model achieved an AUROC—a standard measure of predictive accuracy—of 0.760 in identifying who would respond.
That number doesn't leap off the page. But here's what it means in practice: Atlas's analysis suggests a biomarker performing at that level could have cut UNIFI's enrollment by 458 patients while preserving statistical power. Smaller trials mean lower costs, faster timelines, fewer people exposed to treatments unlikely to help them.
It's a retrospective claim, not a prospective validation—an important distinction. But it points to something concrete. If virtual cell models can sort responders from non-responders with even moderate reliability, they don't just optimize trials. They change the fundamental math of drug development.
The Field Gets Crowded
The opportunity has attracted a particular breed of company: part computational biology lab, part platform play, part infrastructure bet. Recursion, already public, has been vocal about constructing what it calls "the foundation for the first virtual cell," pulling together its phenomics data with external sources and running it through BioHive-2, its custom computing cluster. In December 2025, the company announced its first clinical validation through a program targeting familial adenomatous polyposis—a genetic condition that leads to colon cancer.
Cellular Intelligence, which rebranded from Somite earlier this year, now describes itself as building "the first universal virtual cell signaling model." Galen, another early-stage entrant, simply positions itself as "the virtual cell company" on its homepage. The branding wars have begun.
Then there are the collaborative efforts, which may prove more durable than any single company's model. Ginkgo Datapoints launched something called the Virtual Cell Pharmacology Initiative late last year, aiming to test more than 100,000 compounds and generate over 12 billion datapoints through free RNA profiling. The Arc Institute, through its Virtual Cell Challenge and PerturbSpace project, is working to establish standardized benchmarks and spatial context for CRISPR screens performed in living organisms.
The dealmaking reflects a sector heating up fast. Isomorphic Labs, the DeepMind spinout, closed a $2.1 billion Series B on May 12, 2026 to scale its AI drug design engine. Insilico Medicine signed a deal with Eli Lilly valued at up to $2.75 billion back in March. Not all of these partnerships center explicitly on virtual cells, but they signal something broader: Big Pharma is willing to write very large checks for computational biology, and to do it now rather than wait.
The Regulatory Tightrope
Still, enthusiasm has limits. A Nature Biotechnology editorial published this year argued that hybrid approaches—marrying AI pattern-finding with mechanistic models grounded in biology—will be necessary to preserve causality and satisfy regulatory gatekeepers. Pure pattern recognition, no matter how accurate it looks in benchmarks, may not suffice when the stakes involve human lives.
The regulatory landscape is evolving, but carefully, almost cautiously. The FDA now runs a Model-Informed Drug Development Paired Meeting Program through fiscal year 2027, offering companies a structured way to discuss computational approaches with regulators before they file submissions. The EU's AI Act, which entered force in mid-2024, began imposing core obligations this past August, with high-risk applications facing additional layers of scrutiny into next year.
Companies deploying virtual cells for evidence generation are increasingly aligning their work to FDA and EMA expectations around credibility, documentation, bias control, explainability—standards that were vague just a few years ago but are now codified in guidance documents that run to dozens of pages.
Technical challenges persist too. Most virtual cell models today train on cell lines—immortalized, homogenized, often decades removed from the patients they're meant to represent. Recursion, Relation Therapeutics, and others are prioritizing patient-derived data, but acquiring and standardizing those samples at scale is both expensive and slow. Relation, which raised a $26 million seed extension in December, is working with Novartis on atopic diseases using patient-tissue multi-omics. The work is painstaking.
And then there's the generalization problem, which looms largest of all. Can a model trained on cancer cell lines predict drug response in fibrosis? Can perturbations observed in liver cells extrapolate to kidney tissue? The Virtual Cell Challenge and emerging benchmarks are starting to provide answers, but they remain incomplete.
What Comes Next
The virtual cell revolution is real. It's also early—perhaps more so than the funding rounds suggest.
The tools work well enough to attract serious capital and serious pharmaceutical partnerships. They don't yet work well enough to replace animal studies or traditional phase 1 trials outright. What seems more likely is a hybrid future: virtual cells informing which molecules advance, which patient populations to target, how to design trials with tighter, more testable hypotheses. A McKinsey analysis from midyear highlights the rise of agentic AI across R&D and clinical operations, predicting near-term maturity in operational applications and growing use of simulation for trial design.
For the founders building in this space, the opportunity may be as much about infrastructure as pure innovation. The companies that succeed won't just build accurate models—they'll build models that pharma partners trust, that regulators find credible, that slot cleanly into existing workflows without requiring a wholesale rethinking of development pipelines. Atlas Discovery's focus on foundation models of patient drug response, backed by Y Combinator, Pear, and Glasswing, reflects this pragmatic calculus.
The broader AI drug discovery market is projected to hit $13.8 billion by the early 2030s, growing at nearly 25% annually, according to Grand View Research. North America held more than half the market share last year, driven by regulatory clarity and computational infrastructure that much of the world still lacks.
Perhaps the most revealing signal isn't the funding rounds or the academic publications—it's the data generation itself. 10x Genomics, a leading provider of single-cell and spatial omics tools, maintained its revenue guidance this year at $600 million to $625 million and reported double-digit growth in single-cell reaction volumes. When the picks and shovels are selling, the gold rush is real.
Virtual cells won't solve that 90% failure rate overnight. Nobody serious claims otherwise. But if they can shave even a fraction off that number—if they can rescue shelved assets, identify responders earlier, reduce trial sizes by hundreds of patients—they'll reshape an industry desperately in need of reshaping.
And the clock, as it always does in drug development, is ticking.
