Building a computational model of a cell requires identifying which measurements predict its response to a perturbation, says David Bellamy

Building a computational model of a cell requires identifying which measurements predict its response to a perturbation, says David Bellamy

On October 25–26, AI researcher David Bellamy compared the success of AlphaFold, a model that predicts protein structures, with efforts to develop a “virtual cell” capable of predicting responses to perturbations. The discussion followed a claim by a writer using the pseudonym Mira: AlphaFold’s success shows that biology can be modeled accurately without experiments given sufficient computing resources.

AlphaFold, Bellamy explains, had a defined goal: to predict a protein’s three-dimensional structure from its amino acid sequence. A database of experimentally determined structures provided examples for training and testing. The neural network AlphaFold2 predicted a protein’s structure in a single state, also using sequences of related proteins. When different parts of these sequences changed together during evolution, those changes provided clues about the relative positions of regions within the protein.

For a “virtual cell,” Bellamy asks:

“What is a ‘cell state’? What is a perturbation?”

In his example, researchers measure RNA levels, which reflect gene activity, and try to predict how those levels will change after suppressing the activity of one gene. RNA levels describe only part of a cell’s state, and its response also depends on its environment. Testing whether this description is sufficient for prediction requires comparing the initial measurements and perturbation conditions with the cell’s subsequent response.

Conventional single-cell RNA analysis destroys the cell. In a 2022 study, researchers used a method called Live-seq to extract some of the contents of living mouse immune cells grown in the laboratory and measure their RNA levels. They then observed how those same cells responded to an inflammatory signal. This allowed them to relate initial differences between cells to the strength of their subsequent responses.

A writer using the pseudonym SecondAccount objects: future methods may uncover patterns that allow cell behavior to be predicted from a simplified description. He points to earlier arguments that predicting protein structures would require an impractical amount of computation.

On October 28, Norn Group, which funds aging research, proposed starting to collect data for AI now. Through its Impetus Grants research funding program, the organization is offering to fund observations of disease and treatment effects over time, linking molecular and cellular changes to organ function and health. Norn proposes collecting these data alongside model development.