Countless resources and time are expended globally to research and advance novel therapies that end in clinical failure and no patient benefit. Imagine a world where it is possible to predict any potential drug’s effect on translatable biological models yet unavailable for lab-based testing, while accurately representing patient biology.
Turbine is a biotech company that has built a biology-first, AI-powered “virtual lab” — the world’s first interpretable cell simulation platform.
Turbine’s Simulated Cell platform creates virtual cells that mimic molecular-level behaviour — modelled on real patient biology. It enables in silico experimentation to accelerate drug discovery and development across oncology and beyond.
I spoke to Szabi Nagy, co-founder and CEO of Turbine, to learn more about the company's origin, product offering, and growth to a team of 60+ data scientists, computational and molecular biologists drawn from Budapest’s deep tech talent pool.
Predicting drug performance before the first clinical trial
The Simulated Cell platform models the fundamental protein signalling logic that decides cell fate and facilitates in silico experiments at scales that are impossible in the physical world. In other words, Turbine.ai constructs a digital twin of cellular systems—a biologically representative virtual model that mirrors real cells, enabling simulation and prediction in a controlled, iterative environment.
As a result, billions of simulated experiments can be run in the time it takes to complete even a single test in a wet lab to empower the biopharma industry by identifying and confirming disease-driving mechanisms.
And now, virtual cell models aren’t just for big pharma anymore. In April Turbine launched a virtual lab, which lets scientists use the company’s powerful cellular simulation technology to tackle some of the toughest R&D challenges – including understanding how a potential drug will perform in humans before embarking on a multi-million dollar clinical trial. For the first time, smaller biotech teams are jumping in and using it hands-on to run experiments faster, smarter, and in ways that just weren’t possible before.
From origins in network science to machine learning
Turbine actually started out not as an AI company, but as a network science company. The idea was to represent a cell as a network — nodes (proteins) connected by edges (interactions).
According to Nagy, “there’s a rich mathematical theory about how networks behave. Our thought was: can we represent biology as a dynamic network and simulate how it responds when something changes?”
The advantage of this approach is interpretability. For example, if you administer a drug that inhibits a protein, you can see how the network predicts downstream effects — whether the cell survives, dies, or changes behaviour.
However, Nagy admits that while scientists loved this because they could intuitively understand the mechanism and generate hypotheses, the problem was scalability.
“These networks were built manually by experts reading literature and setting parameters. After a point, predictivity plateaued, because human bias was baked in.”
That’s when the company turned to machine learning — and with its use became more predictive and flexible, able to model more drugs and diseases. But this didn’t always win over the scientists. Nagy admits, “We became more of a black box. Biologists hated that — they felt they couldn’t trust the predictions without understanding the reasoning.”
Simulating biology is harder than training an LLM
Simulating biology is no easy task. The team spent the first four of five years making the technology scale.
Nagy admits that, “biology is incredibly complex and happens at a microscopic scale, so our ability to generate data is limited.”
The company grappled with challenges such as, how to represent biology at a level where you can learn something universal? Is it possible to create a model that can mimic many different types of cells and eventually tissues?
Then, when it comes to deep learning, there’s the challenge of finding an abstraction level where you can actually start training the model and simulating experiments in a way that’s useful, without trying to “boil the ocean” and spending decades and billions on data generation.
Unlike LLMs, which can be trained to a reputable knowledge of words and grammar, “with cells you only get superficial snapshots. You have very little data and need to learn a very complex system.” The second challenge for the company was the sheer number of possible experiments.
Turbine began with a foundational machine learning model trained to learn fundamental rules of biology — such as how proteins interact with each other, and how molecules (like drugs or microenvironmental signals) acting on proteins can alter their function.
The model is trained on data from real biological systems — including lab-based experiments, animal studies, and human samples. This typically consists of genomic information and protein-level measurements, most often transcriptomics, as it is easier to generate.
These snapshots give the model its training targets: how the cell looked before an intervention, and how it looked after. The model then infers the wiring that led from A to B, across millions of experiments.
Where do biotech startups go wrong?
Drug discovery is often held up as one of AI’s most credible applications. When critics dismiss AI hype or “slop,” they still concede: at least it might help cure cancer.
But turning promising science in the lab into a viable, commercial product that patients and healthcare systems will actually adopt is anything but straightforward.
With Turbine’s years in the field, I was curious to hear Nagy’s view on why so many biotech startups struggle to turn scientific breakthroughs into commercial success. Nagy attributes it to four main factors:
Business model. Early AI drug discovery companies were funded by tech investors who thought Pharma would pay millions for better molecules. However, he contends that Pharma didn’t see it that way.
“They saw these tools as point solutions in a long process, not something worth huge deals. So many companies had to pivot into biotech — building their own pipelines, because VCs wanted a path to returns.”
Narrow focus. Most companies concentrated on molecule design. But most clinical failures aren’t because the molecule isn’t “good enough.” “They fail because of target choice, patient selection, biomarkers, or combination strategies. These are harder problems that many ignore.”
Data dogma. According to Nagy, there’s a strong belief that “data equals value.” Companies raised hundreds of millions to generate proprietary datasets. Nagy suggests that “AlphaFold showed the opposite: it didn’t create new data, it just applied better machine learning to existing public data. Data is useful, but not a moat.”
Speed is not everything. Machine learning and AI are also correlating as bringing speed to drug discovery, but Nagy contends that this approach is misleading. “Faster to failure isn’t a business model.
“Many failures don’t happen because the chemistry was too slow, but because drugs don’t benefit patients. Most failures occur in Phase II, when you need to show efficacy.
Molecule design — AI or not — is rarely the bottleneck. The real issue is whether you picked the right target, patient population, and dose.”
Rethinking pharma economics: fewer lab tests, less animal use, cheaper drug discovery
Most biomedical digital twins today focus on patients — for trial design or treatment optimisation. Turbine is virtualising the steps leading to the patient: preclinical experiments and drug discovery workflows. Nagy asserts that over time these approaches will connect:
“Imagine a patient digital twin generating hypotheses, which feed into our simulated cell or tissue models. You can then run experiments virtually to identify the best treatment strategy for that patient. That feeds back into the patient twin, creating a feedback loop.”
Turbine’s platform is already used to guide the pipelines of pharma companies such as AstraZeneca, Ono Pharmaceutical, Cancer Research Horizons, and Bayer.
Turbine has applied simulations to almost 30 programs — from target identification to indication expansion.
Nagy explained:
“We’ve shown that if you run simulations first and then test what the simulations suggest, you’re two to three times more likely to succeed in real experiments.
That alone may not revolutionise a single step — but if you apply it across dozens of decision points, the cumulative effect is transformative.”
Nagy contends that if platforms like Turbine succeed, instead of running endless lab experiments with limited predictive value, you’d have more computational scientists modelling experiments — and only confirming the most promising ones in the lab.
That could reduce reliance on animal experiments, accelerate discovery, and change the economics of pharma.
“Today, only a few dozen large companies dominate because they’re best at managing clinical trials, approvals, and global commercialisation.
If drug discovery itself becomes cheaper and more predictable, smaller players could innovate and still reach patients. That could expand innovation and potentially lower drug prices.”
In August, Turbine entered into a one-year research partnership with MSD (Merck & Co.) to simulate otherwise hard-to-study cancer patient populations. The collaboration aims to uncover new therapeutic dependencies — insights that can help MSD prioritise drug targets, biomarkers, and combination strategies for validation in wet-lab experiments.
Would you like to write the first comment?
Login to post comments