The statistic haunts every pharmaceutical executive: nine out of ten drugs fail. Billions flow into clinical trials, and most yield nothing. But a small cohort of biotech startups now claims they can change that calculus—not by discovering better molecules, but by predicting, before a trial even begins, which patients will actually respond.
Their tool? "Virtual cells"—AI models trained on millions of biological samples that attempt to simulate how human tissue will react to a drug. The promise is audacious: smaller trials, rescued compounds gathering dust in pharma's vaults, and perhaps a dent in that punishing failure rate.
In June 2026, a three-person team from San Francisco called Atlas Discovery emerged from Y Combinator with a retrospective analysis claiming their model could have trimmed 458 patients from a major ulcerative colitis trial without losing statistical power—though this remains the company's own assertion without independent verification. They're hardly alone in making such claims. And increasingly, the pharmaceutical industry appears willing to listen.
Whether any of this actually works in the real world—prospectively, not in hindsight—is the question that will define the next phase of this race.
A Crowded Field, Divergent Approaches
The term "virtual cell" has proliferated across biotech over the past year or so, though it describes a spectrum of technologies rather than one unified thing. At its simplest, the concept involves training AI on vast cellular datasets—gene expression profiles, perturbation experiments, imaging data—to build predictive simulations of biological behavior.
Atlas Discovery positions itself as building "foundation models of patient drug response," integrating preclinical and clinical data. According to their June blog posts, the team conducted a retrospective analysis of ustekinumab in ulcerative colitis, achieving an AUROC of 0.76 in predicting responders from baseline biopsies taken during the Phase 3 UNIFI trial. That level of accuracy, they argue, could enable dramatically smaller, more efficient trials.
But the company is joining a field already thick with competitors. Xaira Therapeutics launched its X-Cell virtual cell model in March 2026, reportedly trained on 25.6 million perturbed single cells across 16 biological contexts. The model uses a diffusion-language architecture with multimodal priors, per their preprint. Recursion Pharmaceuticals has spent years working on what it calls the "virtual cell," leveraging its BioHive-2 supercomputer and massive image-based profiling datasets. Insitro's "Virtual Human" platform, built around human-derived cell data and machine learning, formed the basis of an expanded collaboration with Bristol Myers Squibb announced in March 2026, targeting ALS.
The institutional players are equally busy. Arc Institute launched its Virtual Cell Atlas in February 2025 with over 300 million cells. By June 2025, it had released STATE, a virtual cell model trained on roughly 170 million observational cells and more than 100 million perturbational ones. Arc also hosted community benchmarking through its Virtual Cell Challenge, attempting to set comparative standards.
Then came Illumina's announcement in January 2026 of its Billion Cell Atlas—CRISPR perturbations across more than 200 disease-relevant cell lines, designed explicitly to train internal AI models for target validation and indication prediction. Two months later, the Chan Zuckerberg Initiative and Biohub unveiled a $500 million, five-year Virtual Biology Initiative to build what they described as an "open data foundation for AI-accelerated biology."
It's a land grab, essentially, for the underlying data infrastructure.
Why Now?
Three forces converged to make virtual cells plausible. Data scale, model sophistication, and—perhaps most surprisingly—regulatory acceptance.
The numbers on the data front are staggering. Stanford's AI Index tracked a surge in publications on AI for drug discovery, from 612 papers in 2018 to 3,311 in 2025. More importantly, the datasets themselves have reached giga-scale. The Tahoe-100M perturbation atlas, open-sourced in February 2025, contains over 100 million cells. Illumina's Billion Cell Atlas adds another order of magnitude. The Chan Zuckerberg Initiative's CELLxGENE portal provides standardized access to single-cell data; the Human Cell Atlas consortium continues building reference maps.
These datasets feed increasingly sophisticated models. Atlas Discovery's ExpressionVAE, detailed in June 2026, introduced what the company calls discrete-latent perturbation modeling with FSQ tokenization, reportedly delivering three-to-twenty-fold improvements on distributional metrics versus continuous-latent baselines. Arc Institute's collaboration with NVIDIA on the BioNeMo Agent Toolkit, announced in June 2026, is enabling agentic workflows on life-sciences models—allowing AI systems to iteratively design experiments and refine predictions.
But perhaps the most critical shift is regulatory. The FDA Modernization Act 2.0, signed in December 2022, updated statutory definitions to accept "nonclinical tests" including in vitro, in silico, in chemico, and non-human in vivo methods—explicitly enabling computational model evidence. In June 2026, the FDA finalized ICH M15 guidance on "General Principles for Model-Informed Drug Development," establishing a framework for using quantitative models to inform regulatory decisions. A draft guidance on Bayesian methods, issued in January 2026, opens the door to adaptive trial designs—statistical techniques that pair naturally with predictive models.
The FDA's own AI hub notes that more than 500 drug submissions with AI components were received between 2016 and 2023. Regulators, it seems, are at least curious.
Show Me the Data
The most compelling evidence comes from companies demonstrating real-world applications, though validation remains—how to put this delicately—uneven, according to experts tracking the field.
Atlas Discovery's ustekinumab case study offers the clearest example. Using baseline biopsy data from the UNIFI Phase 3 trial in ulcerative colitis, their model retrospectively predicted responders with an AUROC of 0.76. The company calculates that this predictive power could have reduced enrollment by 458 patients while maintaining statistical power. It's a backtest, not a prospective validation, which is a meaningful distinction. But if replicated in prospective trials, it would represent significant cost and time savings. The company is backed by Y Combinator, Pear, and Glasswing; funding amounts remain undisclosed.
Turbine, a Hungarian startup, raised $25 million in a Series B round in February 2026 and is expanding into immunology with an unnamed top-ten pharma partner. In October 2025, Turbine announced a collaboration with AstraZeneca using its "virtual disease models" to rationalize antibody-drug conjugate discovery—predicting response mechanisms and prioritizing cell lines in a "lab-in-the-loop" strategy. The approach attempts to bridge the gap between computational prediction and wet-lab validation, which is where many of these efforts will live or die.
Insitro has taken a slightly different tack, publishing validation of its POSH (Pooled Optical Screening in Human cells) platform in Nature Communications in December 2025. The work demonstrates high-throughput phenotypic screening that preserves cellular complexity, feeding data into cell-state models. Insitro's March 2026 Bristol Myers Squibb collaboration specifically cites the "Virtual Human" platform and human-derived cell data as the foundation.
Not all efforts center on startups. The Chan Zuckerberg Initiative's $500 million Virtual Biology Initiative, announced in April 2026, represents one of the largest single philanthropic bets on the field. Biohub leadership framed it explicitly as building "the path to accurate predictive models of the cell," partnering with the Human Cell Atlas and Human Protein Atlas consortia. The initiative emphasizes open data and multimodal integration—combining transcriptomics, proteomics, imaging, and perturbation data.
Illumina's strategic positioning is worth noting. The Billion Cell Atlas isn't just a research resource; it's a platform play. By generating the training data for virtual cell models internally, Illumina positions itself to potentially capture value across the AI drug discovery stack—from sequencing to data generation to model training to target validation. Whether that vertical integration strategy pays off remains to be seen.
What Happens Next
The next eighteen months will test whether virtual cells can move from retrospective analyses and pharma partnerships to prospective clinical validation and regulatory acceptance. Several trends seem likely, or at least plausible.
First, expect a wave of prospective trial enrichment studies. Atlas Discovery's UNIFI backtest is provocative, but the field needs forward-looking demonstrations where models select patients before randomization and where predicted outcomes match actual results. The FDA's evolving framework on external controls and model-informed drug development creates a pathway, but sponsors will need to demonstrate rigorous validation and address concerns about selection bias and generalizability. That's harder than it sounds.
Second, watch for consolidation around data infrastructure. The proliferation of billion-cell atlases from Illumina, CZI/Biohub, and Arc Institute suggests intense competition to own the "data gravity" that attracts model builders. Partnerships between sequencing platforms and research institutions hint at vertical integration strategies. Companies that can generate high-quality, disease-relevant perturbation data at scale may capture disproportionate value—assuming the models built on that data actually work.
Third, the benchmarking community is maturing. Arc Institute's Virtual Cell Challenge set a precedent for rigorous, biologically meaningful evaluation beyond simple reconstruction error. Nature Reviews Genetics highlighted in December 2025 that the field is moving toward metrics assessing perturbation generalization, context-dependent gene function, and out-of-distribution performance. As standards converge, it will become clearer which architectural choices actually matter and which are mostly hype.
The regulatory landscape will continue evolving, though perhaps not as rapidly as the technology. ICH M15's finalization in June 2026 provides a framework for model-informed development, but companies will need to educate reviewers on virtual cell predictions as a distinct category—neither traditional PBPK modeling nor pure data mining. The FDA's draft guidance on Bayesian methods and external controls provides tools, but virtual-cell-informed trial designs will likely require extensive pre-submission discussions and careful operating characteristic studies. Regulators move deliberately, for good reason.
Perhaps the most interesting question is whether virtual cells can rescue shelved drugs. Pharmaceutical companies have graveyards full of compounds that failed in clinical trials, often for patient selection reasons rather than fundamental biology. If foundation models can identify subpopulations likely to respond, those assets become potentially valuable again. Atlas Discovery explicitly positions this as part of its thesis. Turbine's work with AstraZeneca on ADC discovery suggests similar thinking—using models to find contexts where previously failed mechanisms might succeed.
There are real technical hurdles remaining, and they're not trivial. Generalization beyond training contexts, batch effects in single-cell data, and the interpretability of billion-parameter models all present ongoing challenges. A Nature Reviews article from December 2025 noted that virtual cell evaluation requires careful attention to biological meaning, not just statistical benchmarks. The National Academies emphasized in a 2024 report on digital twins that verification, validation, and uncertainty quantification remain critical gaps—points that apply directly to virtual cells used in decision-making.
For founders, the opportunity lies in finding specific, validated use cases where virtual cells demonstrably reduce risk or cost. For investors, the question is whether these companies can generate data flywheels that improve models faster than competitors. For pharma executives, it's about managing the integration of these tools into existing R&D processes without overfitting to computational predictions—trusting the model, but not too much.
And for computational biologists, there's fundamental science to be done in understanding what these models actually learn and where they fail.
The Long Game
The nine-out-of-ten failure rate isn't disappearing overnight. That's probably worth stating plainly. But for the first time, there's a plausible—not guaranteed, but plausible—path to predicting which one will succeed. And perhaps more importantly, who it will succeed for.
Whether Atlas Discovery, Xaira, Recursion, Insitro, or some as-yet-unfunded team in a garage ultimately cracks this problem is almost beside the point. The convergence of massive datasets, sophisticated models, and regulatory frameworks willing to at least entertain computational evidence has created conditions that didn't exist five years ago.
That alone makes virtual cells worth watching closely. The promises are bold. The backtests look good. Now comes the hard part: proving it works when the predictions are made in advance, when real patients are enrolled based on model outputs, and when clinical endpoints—not AUROC scores—determine success or failure.
That's the experiment everyone in this space is running, whether they admit it or not.
