Yonatan Stelzer and Amos Tanay propose building biological AI around processes in living systems
Yonatan Stelzer and Amos Tanay propose building biological AI around processes in living systems
On 28 July, Stelzer and Tanay published an article in Cell about “biological language models.” The authors propose training AI on recurring processes in the body and using a single model to connect molecules, cells, space, and time.
Many existing models already identify patterns in DNA sequences, protein structures, and cell images. They answer specific questions, such as which type a cell belongs to, what a protein looks like, or what risk is associated with a set of features. Experimental researchers often need something different: to trace the sequence of events that brought a cell to its current state and choose an intervention within that sequence.
A cell changes over time and functions alongside other cells in a tissue. Stelzer and Tanay therefore propose organizing the model around a canonical biological process, meaning a recurring sequence of events in a living system. In such a model, molecular changes, cell states, and tissue context should form a single sequence, rather than a set of independent measurements that produce a final label.
The authors describe such a system as a world model for biology. It should connect data to mechanism by showing how molecular and cellular states change across space and time. This connection can be tested experimentally by changing an experimental condition and comparing the result with the model’s prediction. A measurement then becomes part of a causal hypothesis that can be tested.
This idea extends the “virtual cell” research program, in which researchers have already tried to represent and simulate molecules, cells, and tissues in different states. The new article proposes organizing these data around a process with its own dynamics. A developer must define which process the system explains, which measurements correspond to its different stages, and which intervention would test the proposed connection.
Age-related changes also unfold through molecules, cells, and tissues over many years. An estimate of biological age reduces the state of a sample to a single measure. A process model asks a different question: which change in a cell leads to a change in tissue, and at which step can an experiment alter the subsequent course of the process? The authors describe this path from observation to intervention.