An Epistemic Engine for Biology

Every therapy begins as a conjecture about a living system: if we intervene here, the patient will change in this way, because of that reason. Knowledge advances when we imagine what might be true, reason about it, then test whether it survives contact with biology. For most of history, a well-reasoned, testable idea was rare and hard-won. That era is ending.

Today, AI agents can scan the literature, propose disease mechanisms, rank targets, generate drug candidates, and analyze data. A single scientist working with this technology will be able to cover literature, data, and design space in a fraction of the time that it currently takes for a whole department. They will be able to follow a question past the edge of their expertise, and produce confident conjectures faster than anyone can test them.

That is a real change. Agentic and generative AI are beginning to collapse the distance between question and action: search the databases, run the code, call the model, and preserve the trail. For a field fragmented across papers, notebooks, file formats, assays, and software, that integration is not a convenience. It is new scientific infrastructure.

Few fields punish confidence like biology

When it is easier to act on a hypothesis, the cost of weak reasoning goes up. Biology is unforgiving. Correlations pose as causes. Confounders hide. Heterogeneity is the only rule. Every dataset is a thin, noisy slice of a system far larger than itself, so a result can look strong in the data and still be wrong. When the goal is to develop a therapy, the risk hides in the seams. The molecule that looks perfect in the dish dies in toxicology. The trial that looks clean asks the wrong question. The finding that held in the lab does not survive regulatory-grade evidence.

A wrong but fluent idea is the hardest kind to catch, and we usually catch it late, after carrying it into the clinic and watching it fail, years and hundreds of millions of dollars later. A flood of confident ideas that fail only once they reach patients is not progress. It is a slightly faster way to fail at much greater cost.

A better way to reason, and a cheaper way to fail

In a field where the ultimate exam is a clinical trial, testing will never keep pace with hypothesis generation. Even as we use AI to streamline clinical trials, the gap between what we and our agents can imagine and what we can test will widen.

We believe the larger promise of AI is to help scientists decide which hypotheses are worth the next experiment, by building, comparing, criticizing, and refining causal explanations of disease and treatment. The standard is not whether a candidate seems plausible. It is whether the explanation says what causes what, reaches beyond the dataset that suggested it, is hard to vary without losing its force, and survives attempts to break it. A system that asks: What must be true for this therapy to work? What would falsify it? Where does it generalize, and where does it break? Which experiment would expose the error fastest? It should make the cheapest and fastest place to be wrong a screen, not a patient.

That means connecting models to mechanisms, mechanisms to interventions, interventions to experiments, experiments to criticism, and criticism back into better explanations. It means moving beyond AI as a productivity layer, toward AI as an epistemic engine for biology.

Zatoona Labs

Zatoona Labs is our integrated environment for therapeutic development: one place where scientists and agents can search, reason, analyze, design, and keep track of the decisions behind a program. It is built for the full path from first mechanistic hypothesis to regulatory submission, so the biology, evidence, experiments, patient context, and development plan stay connected instead of scattering across a dozen tools.

At its center is Olea, an AI co-scientist built on conjecture and refutation. Olea brings the new power of agentic biology into one workspace: reading the literature, ranking targets, proposing mechanisms, designing molecules, running analyses, drafting protocols, and carrying context across the work. That alone changes what a scientist can attempt.

But therapy development needs more than execution. Olea treats every output as a conjecture to be criticized. It asks what the claim depends on, what would falsify it, and which test would expose the error fastest. It then runs the claim through a refutation engine equipped with data, causal methods, and in-silico simulation, before anyone spends the next dollar in the lab or clinic. What it cannot break, you can build on.

Where this goes

As the testing sharpens, more of the loop runs on its own, improving as it goes. The scientist stops running each step by hand and starts directing a relentless, skeptical collaborator. Their job is no longer to generate ideas, but to decide which survivors earn the next, costlier test. The destination is a single line from first conjecture to clinic, run inside a workspace that compounds knowledge instead of scattering it. It will be faster and cheaper, but speed was never the point. The point is better therapies: fewer that fail late, and more that work.

Zatoona is the word for an olive. The tree that bears it is slow to grow and can live for a thousand years, because it was made to endure. We believe knowledge should be built the same way.