AI Tool Deep-Phase Links Cellular Droplet Shape to Biochemical Function

Inside our cells, tiny droplet-like structures called biomolecular condensates compartmentalize and organize essential biochemical processes. Now, researchers have developed an artificial intelligence framework called Deep-Phase that uses microscopy images of these droplets to determine how specific drugs alter their inner workings. The study, reported today in Cell, was led by senior author Clifford P. Brangwynne at Princeton University, with Anita Donlic as the lead researcher.

Deep-Phase analyzes the complex physical shapes of condensates, specifically focusing on the nucleolus (a multi-layered droplet where ribosomes are assembled), nuclear speckles, and viral condensates in human cell models. By tracking how a condensate's structure changes over time and at varying drug concentrations, the algorithm automatically generates dose-response curves. The team demonstrated that these image-derived structural shifts closely mirror traditional biochemical measurements of drug potency for inhibitors of RNA transcription and processing.

This matters because linking the physical appearance of a condensate to the nanoscale molecular interactions happening inside has traditionally been a major bottleneck in cell biology. Deep-Phase provides an unbiased, automated way to discover new disease pathways and evaluate potential therapeutics without relying on predefined visual features that might miss subtle biological shifts.

Applying Deep-Phase in a chemical screen, the researchers discovered a previously uncharacterized "flower"-shaped nucleolar structure. Investigating this unique morphology revealed an unexpected role for the enzyme DNA topoisomerase 1 (TOP1) in processing ribosomal RNA and maintaining the physical boundaries within the nucleolus.

While Deep-Phase efficiently identifies strong correlations between condensate shape and molecular function, the authors note that these structural signatures still require experimental follow-up to fully map the underlying mechanisms. Additionally, like many deep learning applications, the model operates somewhat as a "black box," making it challenging to definitively pinpoint which exact visual features drive its classifications.

Moving forward, Deep-Phase could serve as a highly scalable platform for high-content drug screening, allowing scientists to rapidly infer complex biochemical states simply by observing cellular structures.

Reference:
Donlic, A., et al. "Deep learning of functional perturbations from condensate morphology." Cell, 2026. DOI: 10.1016/j.cell.2026.05.010