A Princeton team's deep learning classifier turns the shape of tiny membrane less droplets inside cells (biomolecular condensates) into a drug response signal, surfacing new morphology categories such as a previously unseen 'flower' form tied to topotecan, an FDA approved…
A yellow droplet inside a human cell, lit up by fluorescent tags, sits in the shape of a sphere. Researchers dose the cell with an antiviral. Within hours, the droplet warps into something irregular, and the virus inside stops replicating.
That visual is roughly the door into a new Cell paper from Cliff Brangwynne's lab at Princeton: a deep-learning system trained to read those shape changes as a readout of what a drug is doing to a cell, one cell at a time, without first knowing which gene the drug targets.
The shapes in question come from biomolecular condensates, which are tiny, membrane-less droplets inside cells that concentrate the molecules driving transcription, RNA processing, and other gene-regulation machinery. Their unusual physics is implicated in Alzheimer's, ALS, and cancer, which is why their behavior under drug treatment matters.
In the paper, first author Anita Donlic and colleagues imaged nucleoli, the largest condensate in the nucleus, across hundreds of human cells exposed to a panel of small molecules. A neural network sorted the resulting microscopy images into four morphological categories. Two known anti-cancer drugs reliably produced a "cap" shape. Other stressors produced a "necklace." And topotecan, an FDA-approved topoisomerase-1 inhibitor used in chemotherapy, triggered a shape the team had never seen before, which they named "flower." That morphology change pointed to TOP1's role in organizing nucleolar architecture and RNA processing, a relationship the field had not previously named from images alone (GEN coverage of the work; primary paper in Cell00569-6)).
The approach generalized. When the same classifier was retrained on nuclear speckles, condensates tied to mRNA processing, its morphology predictions tracked dose and response to perturbations. When the team turned it to respiratory syncytial virus condensates treated with an antiviral, the predicted shape changes correlated with disruption of viral replication. In each case, the AI was not discovering a drug. It was reading what a drug had already done to cellular architecture.
That distinction matters. The paper is a method, not a screening platform, and the trade-press framing of "AI predicts gene regulation" overstates what shape alone can show. Morphology is a hypothesis generator: it points researchers toward the functional state a cell has entered, and that hypothesis still has to be tested with biochemistry, genetics, or both. The tool sits alongside those assays, not above them.
Condensate biology itself is also a still-contested framework. If condensates exist and matter, the classifier can be trained against them. If the framework is wrong, the predictions lose their anchor. Brangwynne's broader research program treats condensate physics as a real organizing principle of the cell, and the new paper is best read as a way of putting questions to that framework in image form rather than as a verdict on it.
What to watch next: whether the network and its training data are released openly, whether the "flower" morphology reproduces in other labs and in vivo, and whether the same approach scales to the much messier condensate systems implicated in neurodegeneration. If it does, researchers studying Alzheimer's, ALS, and cancer gain a fast, image-based way to ask which drugs are doing what to the cell's hidden architecture. If it does not, the paper is still a useful demonstration that a microscope, a fluorescent tag, and a well-trained neural network can surface biological categories that human eyes had missed.