Live·Open questions in longevity research
All news
AI in medicineScientific Computing

AI will see wider adoption as an assistant to biologists sooner than as an accurate model of a living cell, predicts scientific AI developer Kenny Workman

15 November 2026· 261115003

AI will see wider adoption as an assistant to biologists sooner than as an accurate model of a living cell, predicts scientific AI developer Kenny Workman

On November 13, the co-founder of LatchBio, which develops AI for biology, again proposed training models to reason by learning from scientists. People draw useful conclusions from imperfect experimental data despite an incomplete understanding of cellular mechanisms. Workman argues that decades of this work provide examples for training and evaluating AI, and that reproducing human reasoning promises faster economic returns.

In his proposed approach, an AI agent receives results from a real experiment, runs analysis software, and offers an explanation. The researcher's judgment based on the same data serves as its reference. A cell model must predict the outcome of an intervention, such as what happens when a gene is switched off. As Workman explained in his October 6 discussion, measurements must capture changes in cellular processes so that the model can predict how a cell will behave under new conditions. He therefore sees progress in cell models as also depending on better measurement methods.

The paths to adoption also differ: an agent can be integrated into an existing laboratory. It handles interpretation between experiments, while the experiments themselves take as long as the underlying physical processes require. Workman argues that the owner of an accurate cell simulator has an incentive to develop drugs in-house: advancing a drug candidate further through development could yield a larger return from a partnership deal. In his scenario, the company eventually progresses to human trials.

“This means building companies that do more experimental work, require more capital, and take on more clinical risk. That takes time.”

On November 13–14, Ron Alfa challenged Workman's argument. Alfa is a co-founder of Noetik, which builds AI models of tumors. Before Noetik, Alfa led research at the biotechnology company Recursion. He acknowledged that automating analysis saves scientists time, but disputed the conclusion that success on data analysis tests shows that AI has learned biological principles:

“But I don't think incremental improvements in analysis methods will lead to models that learn biology.”

Workman responded with Noetik itself as an example: researchers collect data on the locations of cells and molecules in new tumor samples to discover previously unknown features of the disease. In Workman's view, an agent should also master this scientific task of explaining new observations. Noetik deliberately masks some measurements from tumor tissue during model training and trains its models to reconstruct them from data on neighboring regions. The company's premise is that solving this task should help the model learn biological principles.

An agent's performance can be tested by giving it data from a completed experiment and evaluating its conclusion and choice of the next experiment. A cell model can be given a new intervention in advance, and its predicted cellular response can then be compared with the experimental result. These capabilities can complement each other within a laboratory: the agent helps interpret data and choose the next experiment, while the cell model helps assess the consequences of an intervention before it is carried out.

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-agents#cell-models#experimental-design#biological-reasoning#tumor-models