Lighting Up Individual Proteins in Living Cells: AI-Designed NovoTags Expand Multicolor Microscopy

Identifying specific proteins inside cells involves more than simply illuminating them: multiple labels must remain spectrally distinct, avoid interfering with native functions, and withstand prolonged imaging. Teams from the University of Washington, the HHMI Janelia Research Campus, and the European Molecular Biology Laboratory have now used artificial intelligence to design from scratch a set of small fluorescent tags called NovoTags, seeking to combine the precise localization of genetic labeling with the high brightness and photostability of small-molecule dyes.

The researchers designed binding proteins for the cell-permeable dyes JF657, JF596, and JF494, covering the far-red, orange-red, and green spectra. The three NovoTags selectively bind their respective dyes with low-nanomolar affinity. The research team then attached the tags to target proteins, allowing different cellular structures to become visible after the dyes were added, without relying on the intrinsic fluorescence of natural fluorescent proteins.

The design process was not completed solely on computers. The team used RFdiffusion to build protein backbones, LigandMPNN to specify amino acid sequences, and AlphaFold and RoseTTAFold to screen candidates, followed by stepwise validation through yeast surface display, fluorescence-activated cell sorting, sequencing, and biochemical analysis. The crystal structure of NovoTag657 bound to its dye was also determined at 2.38-angstrom resolution. The monomeric structure, comprising 135 amino acids, closely matched the computational model, providing experimental support that the design was realized at the atomic scale.

In mammalian cells, the three tags can simultaneously label different subcellular compartments. The researchers also demonstrated live-cell STED super-resolution microscopy and fluorescence lifetime imaging in human cell lines. In addition to distinguishing signals by color, they could use differences in emission duration caused by the dyes’ surrounding protein environments to increase the combinations of labels distinguishable within the same field of view. The research team proposed that spectral and lifetime combinations might eventually be used to track more proteins, but identifying as many as 30 labels at once remains a development goal and was not experimentally demonstrated in this study.

Another extension of the system is NovoSplit. The researchers divided a tag into two parts and attached each part to a different protein. After the designated dye was added, it acted like a molecular glue, prompting the two halves to assemble while simultaneously generating a fluorescent signal. The system can therefore both report whether two proteins are in proximity and chemically activate their interaction. The team also created a version capable of forming a covalent bond with the dye to improve labeling durability.

However, NovoTags remain research tools rather than a clinical imaging technology. The published results primarily come from cultured human or other mammalian cells, and their performance in tissues, whole animals, or long-term tracking has not yet been demonstrated. Attaching artificial proteins to the subjects under study and then adding dyes could also alter protein localization, interactions, or cellular state, which must be ruled out for each experiment. Differences in cell permeability, background signal, and toxicity among dyes will likewise limit the conditions under which they can be used.

Published in *Science*, the study shows that the role of generative protein design is moving beyond structure prediction toward the creation of functional cellular tools. If subsequent work can establish more mutually orthogonal tags that are smaller and less toxic, they may further connect with cryogenic correlative light and electron microscopy, allowing researchers to first locate targets in living cells and then investigate their in situ structures. At this stage, their most practical value is to add a programmable option for multicolor imaging and protein-interaction experiments.