NovoTags: Designing Proteins That Recognize Specific Fluorescent Dyes
A paper from the Baker Lab (Institute for Protein Design), the Lavis Lab (HHMI Janelia), and the Mahamid Group (EMBL) introduces NovoTags, a set of small, de novo designed proteins that each selectively bind one of three closely related Janelia Fluor (JF) dyes spanning green, orange, and far-red wavelengths.
The motivation is pretty straightforward once you lay out the tradeoffs fluorescence imaging has lived with for years: fluorescent proteins fuse genetically to your protein of interest but are limited in brightness and photostability; small-molecule dyes are brighter and more photostable but need a mechanism to find their way to the right protein inside the cell; and existing solutions to that problem (HaloTag, SNAP-tag) work well but are relatively large (20–35 kDa) and recognize a shared labeling chemistry rather than the dye itself, making it hard to build several mutually orthogonal tags for multiplexed imaging.
NovoTags combine the genetic programmability of a fusion tag with the brightness and color range of synthetic dyes, plus direct, dye-specific recognition, meaning several tags can coexist in the same cell without cross-talk.
The design workflow
Starting from the structure of a JF dye, the team generated pseudocyclic protein backbones around it using Cα RFdiffusion (the same backbone-generation approach previously used in de novo serine hydrolase design), then used LigandMPNN for sequence design and Rosetta FastRelax to refine geometry. Designs were filtered with Rosetta (ligand interaction) and AlphaFold2 self-consistency, then screened via yeast surface display, FACS, and next-gen sequencing, and finally validated biochemically, structurally, and in cells.
How well did it work?
Out of 16,675 designs screened across the three dyes, 626 came back as initial yeast-display hits, an overall hit rate of about 3.8%, ranging from 1.2% (JF494) to 5.5% (JF657). The selected leads are compact (13–16 kDa), bind their target dye with 1.5–19 nM affinity, and show more than 1,000-fold selectivity over the other two dyes. The NovoTag657 crystal structure matched the computational design almost exactly: 0.58 Å backbone RMSD for the dye-bound structure, 0.70 Å for the apo form.
From binder to platform
What makes this paper more than "three new fluorescent tags" is what the team did once they had working binders. The same scaffold was extended into three-color live/fixed-cell imaging and multiplexed STED super-resolution microscopy; binding-pocket redesign to program fluorescence lifetime (short-, medium-, and long-lifetime variants of the same JF494 dye, distinguishable even though they share a spectrum); a covalent, cysteine-mediated JF657-binding variant; NovoSplit, a split system where the dye itself acts as a molecular glue inducing dimerization of two protein fragments; and a protein-proximity sensor built on NovoSplit with a direct fluorescent readout.
That last set of results is the part I find most interesting: the binding pocket isn't just a targeting module, it's a programmable chemical environment. Once you can shape how a dye behaves once it's bound, you're not just tagging proteins anymore. You're designing the optical behavior of the label itself.
Worth knowing before you get too excited
The often-cited "30 distinguishable labels" figure (roughly 10 colors × 3 lifetime states) is a projection, not something demonstrated in the paper: the actual result is 3 colors and 3 lifetime states for one of those colors. Cellular testing was mostly in HeLa cells (with some E. coli), the covalent labeling reaction is still slow (~3 hour half-time), and the proximity-sensing demo used an engineered heterodimer rather than a native interaction. All reasonable first steps, not a finished toolkit yet.
The collaboration behind this is also worth noting: it's not a purely computational result. It took the Baker Lab's design capability, the Lavis Lab's dye chemistry, and the Mahamid group's imaging expertise (STED, FLIM, and, per HHMI's coverage, an eye toward eventual cryo-CLEM) to turn a binder into a working set of imaging tools. HHMI's writeup also mentions an active effort to expand the dye panel toward about a dozen colors and to explore binders for physiological indicators like calcium and metabolites. That's worth watching, but it's explicitly a roadmap rather than a result in this paper.
For similar tools, see the BioMoDes Protein Engineering Applications page.
References
- Tran, L., Klein, S., Juergens, D., et al. De novo design of orthogonal far-red, orange, and green fluorophore-binding proteins for multiplexed imaging. Science (2026).
- NovoTag design pipeline (GitHub).
- Cα RFdiffusion: paper and code.
- EMBL. NovoTags: AI-designed proteins help scientists see inside living cells.
- HHMI. AI-Designed Proteins Could Transform Fluorescent Cell Imaging.
- Phys.org. AI-designed proteins help scientists see inside living cells.
This post is an AI-reworded, expanded version, in my own voice, of a summary I originally posted on LinkedIn.
I try my best to make the information on this website as accurate as possible.
If you find any errors in the contents of this page or any other page on this website,
I would greatly appreciate that you kindly get in touch with me at
contact[at]abeebyekeen[dot]com.
If you are interested in joining my free weekly “BioModes and Top Reads” newsletter, please subscribe below.