How Vivodyne’s HIVE Labs Aim to Close AI Drug Discovery’s Human-Data Gap

Vivodyne’s central thesis is that AI drug discovery has a data problem before it has a model problem. AI systems can generate targets, molecules, and experimental ideas quickly, but their predictions remain limited when the underlying data mostly describe animals, isolated cells, proteins, or observations rather than how complex human tissue responds to an intervention. Vivodyne CEO Andrei Georgescu has argued that current AI-designed drugs are not yet substantially better than human-designed drugs because there is not enough relevant laboratory data for models to learn from. 25

The company’s answer is HIVE, a modular autonomous-laboratory platform designed to grow, dose, monitor, and analyze human tissue at scale.

The bottleneck: correlation is not enough

A drug-discovery model may identify an association between a biological feature and a disease outcome. That does not necessarily show what will happen when a drug changes the system, at a particular dose, in a particular tissue context.

Vivodyne is therefore focused on perturbation: deliberately changing a condition and measuring the tissue’s response. In the company’s framing, an experiment can produce a more useful relationship for drug development: if treatment X is applied at dose Y, tissue Z changes in a measurable way. Its platform describes this approach as generating cellular-causality data through controlled experiments on human tissues. 2

That distinction matters because a model trained on incomplete or poorly matched biology can make highly confident predictions that do not translate to patients. Vivodyne’s argument is not that larger models are useless, but that scaling models without scaling the quality and human relevance of experimental data may produce limited gains. 1725

How HIVE works

HIVE combines robotics, microfluidics, tissue engineering, automated dosing, and biological readouts. Vivodyne says its system can cultivate 20 types of human tissue, then administer drugs or other perturbations and track the resulting changes over time. 117

The platform is built around TissueDisks that allow hundreds of independent human tissues to be cultivated and dosed in parallel, according to Vivodyne. The company says each tissue model contains approximately 200,000 to 500,000 cells and can include perfused, tissue-specific three-dimensional vascular networks for more realistic delivery of therapeutics. 24

The practical workflow is intended to be iterative:

1. Grow standardized human-tissue models.
2. Apply a drug, dose, genetic change, or combination of interventions.
3. Measure molecular, cellular, and functional responses.
4. Feed the results into analysis and subsequent experiment design.

This turns the laboratory into a data-production system rather than a collection of isolated assays. The output is not simply whether a tissue survived. It can include how a tissue changed, when it changed, which cells were affected, and how different doses or treatment combinations altered the result.

Why animal and cell-based research can miss the mark

Animal models remain useful for studying whole-body effects, but biology that works in an animal may not behave the same way in humans. Species differences can affect drug metabolism, toxicity, immune responses, disease mechanisms, and the interaction among multiple tissues. Single-cell and protein assays provide valuable detail, but they remove much of the multicellular structure and physiological context that shape a drug’s effects.

That mismatch helps explain why promising preclinical candidates can fail during human development. Vivodyne’s proposed role is to add human-tissue evidence earlier, so that some safety and efficacy problems are identified before a candidate reaches an expensive patient trial. 113

The limitation is that lab-grown tissue is not a complete human body. Tissue models may not fully reproduce whole-body pharmacokinetics, organ-to-organ signaling, long-term immune effects, patient-to-patient variation, or every feature of a disease. They should therefore be viewed as a potential bridge between simplified laboratory tests and clinical research—not as a replacement for clinical trials.

What evidence has Vivodyne reported?

Vivodyne has reported results from testing 20 chemotherapy drugs against several tissue models. The company says its liver model predicted human drug toxicity with 94% accuracy, its airway model matched human tissue behavior at 96%, and its bone-marrow model reached 100% concordance in that test set. 17

Those figures are potentially significant, but they need to be interpreted carefully. The available material identifies them as company-reported performance claims. It does not establish a large, independently replicated, peer-reviewed validation across many drug classes, nor does it show that HIVE predictions have improved clinical approval rates in a prospective comparison with conventional preclinical programs.

Vivodyne has also announced a facility with 12 HIVE laboratories and an annual capacity of 3.1 million controlled experiments on living human tissues. That capacity is a scale claim about the company’s platform, not evidence by itself that the resulting experiments are clinically predictive. 37

Funding and commercial support

The company announced $38 million in total seed financing in 2023, led by Khosla Ventures, with participation from Kairos Ventures, CS Ventures, MBX Capital, and Bison Ventures. 9 Vivodyne later publicized a $40 million funding round to expand its robotics- and AI-driven human-tissue testing infrastructure. 7

The funding supports a strategy that treats experimental biology as infrastructure for AI: generate large, standardized datasets first, then use them to improve therapeutic predictions and prioritize experiments.

What could change for clinical trials?

If the platform proves predictive in independent prospective studies, it could influence drug development in three main ways.

Earlier failure detection

Candidates with weak efficacy or unacceptable toxicity could be deprioritized before patient exposure. That would not eliminate clinical failures, but it could reduce the number of poorly supported candidates entering trials.

Better candidate and dose selection

Testing multiple doses and tissue contexts may help researchers identify which candidates are worth advancing and which dosing ranges deserve clinical investigation. Vivodyne presents its human-data platform as useful across discovery, ADME/toxicity work, and trial design. 13

More informative AI models

Repeated intervention-and-response experiments could give AI systems training data that connect a treatment to a tissue-level biological outcome. Over time, Vivodyne hopes this will support a broader computational model—or “world model”—of human biology. 8

That outcome depends on the quality, reproducibility, diversity, and external validation of the data. More experiments are not automatically better if the tissue models do not represent relevant patient populations or if the measurements do not predict outcomes in people.

Why combination therapies are an important test case

Combination therapies require researchers to explore more than one variable at a time: drug pairings, doses, treatment order, timing, and tissue context. A human-tissue platform could make it more practical to test these combinations systematically, looking for synergistic effects while screening for toxicity before patients are exposed.

This could be particularly useful for diseases in which a single drug affects only part of a complex biological pathway. But the same caveat applies: tissue experiments can reveal interactions within the modeled biology, while clinical outcomes also depend on absorption, distribution, metabolism, immune effects, organ interactions, and patient heterogeneity.

The bottom line

Vivodyne is betting that the next advance in AI drug discovery will come from better experimental grounding, not simply larger algorithms. HIVE is designed to provide that grounding by running controlled, high-volume experiments on 20 types of lab-grown human tissue and converting the results into structured biological data. 117

The approach addresses a real weakness in preclinical translation: models trained or tested on simplified systems may not reliably predict what happens in people. Vivodyne’s early performance figures and claimed operating scale are notable, but they remain company-reported evidence. The decisive test will be whether independently validated HIVE predictions can improve candidate selection, reduce clinical failures, and produce AI models that generalize beyond the tissues and drugs used to train them.