At Princeton University, Cliff Brangwynne and his team have trained artificial intelligence to read the shapes of invisible cellular structures with a precision that rivals—and surpasses—the human eye, opening a new window into how drugs reshape the microscopic world inside our cells. The breakthrough centers on biomolecular condensates, those tiny droplets floating within living cells that orchestrate everything from gene regulation to protein assembly, and have been implicated in Alzheimer's, ALS, and cancer.
The challenge that Brangwynne, the June K. Wu '92 Professor of Chemical and Biological Engineering, and his team set out to solve is fundamental to biology itself: how do ordered structures emerge from billions of molecular interactions? Their answer was to build a machine-learning tool capable of sorting what human researchers find nearly impossible to classify—microscope images of condensate shapes in hundreds of human cells under various drug treatments.
The focal point of their study, published in Cell in June 2026, was the nucleolus, the cellular factory responsible for assembling the tiny machines that manufacture proteins. Using an advanced microscope, the team captured images of nucleolar shape changes across hundreds of cells exposed to different drug-controlled conditions. When fed through their neural network, the images revealed something remarkable: instead of the three shape categories the researchers expected, the AI identified four distinct patterns—normal, cap, necklace, and an entirely unexpected fourth that was completely new to science.
That unexpected discovery proved to be the most illuminating. Postdoctoral researcher Anita Donlic, the paper's first author, found that two known anticancer drugs triggered cap-shaped nucleoli, a phenomenon never before reported for those particular drugs. This suggests the medications are disrupting cellular machinery in previously unrecognized ways. For another drug, topotecan, the network flagged something it couldn't quite fit into the existing categories: a shape the team christened the "flower" morphology.
What makes this discovery particularly significant is what it revealed about the drug's mechanism. Topotecan was already known to inhibit an enzyme called TOP1, used in DNA replication. But by identifying the flower shape as a direct result of TOP1 loss, Donlic uncovered an entirely new role for the enzyme—maintaining nucleolar organization by regulating RNA processing. The neural network had flagged something invisible to human analysis, revealing biology nobody had seen before.
The team didn't stop there. They tested the same neural network on other cellular condensates linked to RNA processes, including nuclear speckles—hubs for messenger RNA activity—and condensates from respiratory syncytial virus. In each case, the network detected dose-response patterns, showing how different drug concentrations triggered proportional changes in cellular shape and function. This consistency across different condensate types suggests a robust, generalizable system for monitoring how cells respond to therapeutic interventions at the single-cell level.
The implications are far-reaching. As Brangwynne notes, the central problem in biology is understanding how emergent structure arises from molecular interactions. This tool provides a way to learn from images and classify the patterns that emerge. "You could be missing other important features," Donlic reflected, capturing the essence of why this work matters. "Things that could tell you there's new biology." In mapping molecular shape to cellular function, the Princeton team has opened a door to a level of biological understanding that was previously locked behind the limits of human perception.