The bottleneck for AI in biology is the lack of instruments that can read and alter a living cell's state in real time, argues Harvard biologist Alex Plesa
The bottleneck for AI in biology is the lack of instruments that can read and alter a living cell's state in real time, argues Harvard biologist Alex Plesa
On September 15, Plesa published a Substack essay "AI for Bio Doesn't Need More Data", examining why models that predict a cell's response to perturbation have for years failed to outperform simple averaging of already collected data.
The standard way to study a living cell is to kill it and read the activity of 20,000 genes once, yet the cell senses, decides, and acts in real time, like a neutrophil (an immune cell) chasing a bacterium. Plesa spent ten years building models of cellular aging. He never had measurements of a single cell over time, only averages across hundreds, although the goal is to control an individual cell, not the mean of a million dead ones.
Every method in molecular biology sacrifices one of three properties: speed, nondestructiveness, or breadth of coverage. Microscopy labels no more than 3–4 proteins; RNA sequencing yields thousands of measurements but kills the cell and takes days. No existing method delivers all three at once, and without all three, closed-loop control of the cell is impossible.
This has been achieved before: in 1952, Hodgkin and Huxley explained the nerve impulse using a voltage clamp, an instrument that read the voltage across a neuron's membrane and immediately injected current to hold it at a set level, faster than the cell could change. Their equations still predict the impulse accurately. The instrument closed the loop before the computational power to model it even existed.
With cell models the opposite happened: within a single year, independent studies reversed the conclusion several times on whether these models outperform simple sample averaging. The Arc Virtual Cell Challenge tested this separately with more than 1,200 teams: models consistently failed to beat naive baselines, and the winners combined neural networks with statistics, not scale.
"AI in biology has hit an instrumentation wall, not an intelligence wall"
Plesa writes. Even a perfect model would predict only the endpoint, missing the cell's path to it, because Perturb-seq (which labels cells and reads genes activated after perturbation) captures only the start and the finish.
Language, despite its volume, has surprisingly few independent variables: roughly 42 by current estimates, and halving the error requires only a 1,450-fold increase in data. For cells, no one has pinned this number down. Estimates diverge and shift with new data; whether ten times more data is needed or ten million times more remains unknown. Meanwhile, cell atlases yield only percentage-point improvements, not a step change.
Plesa calculated the requirements: reading at least a thousand signals once per minute from a living cell throughout its entire cell cycle, and reversibly altering at least a hundred parameters with latency under one minute. This rate is ten times higher than the rate at which the cell's key signals change.
Both halves of this loop already exist separately. On the reading side, Biohub uses light scattered through a cell to read its state without killing it, matching the accuracy of RNA sequencing; WHOLISTIC scaled this to live fish. On the writing side, JURA Bio, whose cofounder is Plesa's research advisor, tests roughly 10 quadrillion antibodies per experiment. There is also a model that incorporates time: trained on 1.67 million cancer patients, it identifies the longer-surviving patient more accurately than clinical indicators (77.4% versus 56.3%).
"We have half the loop on the reading side and half on the writing side. Both were built separately by different groups, and no one has connected them yet. Yet"
Plesa acknowledges.