That distinction is central to CEO and co-founder Andrei Georgescu’s argument about the limitations of current AI approaches to biology. Existing datasets can provide extensive information about individual cells, proteins and biological states, but they do not necessarily capture the sequence of events that produced those states.
“All the training is done on static snapshots of these cells, and the models are not conditioned at all by the how a cell got to that state,” Georgescu said. “In other words, the model learns ‘this is cell state A,’ ‘this is cell state B,’ but never ‘cell state B is the effect of inflaming cell state A.’”
Georgescu points to a Nature Methods study published last month that found no clear data scaling laws when generative AI models were trained using existing cellular datasets. His contention is that adding more of the same kind of biological information may not be enough to produce models capable of understanding how interventions cause changes in human tissue.
That limitation also intersects with a longstanding problem in pharmaceutical development. About 90% of drug candidates that perform well enough in animal testing to reach clinical trials ultimately fail to receive regulatory approval for humans. Vivodyne is attempting to generate evidence from human tissue earlier in that process.
The company says its laboratory-grown tissues have already demonstrated strong predictive performance. Its liver models have reached 94% predictive accuracy against human toxicity trials, while airway tissue has matched real human tissue behavior 96% of the time. Vivodyne also reports 100% concordance from its bone marrow models across tests involving 20 chemotherapy drugs.
The newly opened facility is intended to expand that experimentation substantially. Vivodyne says it is already running experiments at twice the throughput of all animal trials taking place in the U.S., while its HIVE systems are monitoring hundreds of thousands of ongoing experiments involving diseased tissue exposed to different stimuli.
For drugmakers, the immediate application is identifying candidates with a better chance of succeeding before committing to human trials, which typically cost tens of millions of dollars. Vivodyne says it is working with several major pharmaceutical companies, although it has not publicly identified them.
Georgescu compares the objective with safety testing in the automotive industry, where manufacturers can develop substantial confidence in a vehicle before submitting it to formal testing. Drug developers, by contrast, routinely enter expensive clinical trials despite high failure rates.
Vivodyne’s ambitions extend beyond evaluating individual drug candidates. The company wants the experimental output from HIVE to become training material for AI models designed around causal relationships in human biology.
That could become particularly relevant when researchers are evaluating therapies that affect several biological pathways simultaneously. Testing every possible combination experimentally becomes increasingly difficult as the number of potential interventions expands.
“If we want combination therapies, the space that has to be searched explodes — it can’t be an experimental approach,” Georgescu said. “You have to say, ‘I want this effect to happen, so what cause should I invoke?’ Establishing causality in human biology is the basis of all of this.”
Vivodyne emerged from the University of Pennsylvania in 2021 after Georgescu completed his bioengineering PhD there. The startup has raised just under $80 million across two rounds led by Khosla Ventures.
For Georgescu, the immediate challenge for AI-assisted medicine is therefore less about making increasingly ambitious predictions and more about supplying models with experimental evidence that reflects human biology. “Absent human testing, what are these models going to do?” Georgescu said. “They’re going to cure cancer in mice.”
This analysis is based on reporting from Bitcoin World & TechCrunch.
Image courtesy of Business Wire.
This article was generated with AI assistance and reviewed for accuracy and quality.
