Lab Newsletter — July 23, 2026: AI Designs the Parts

For years AI mostly read biology — predicting structures, classifying images, scoring perturbations. Today’s three items are about the flip side: AI writing it — designing the parts, and finally being able to build them.

🔬 AI-designed proteins become new eyes inside the cell

The standout is NovoTags (in Science): fluorescent protein tags designed from scratch to bind Janelia Fluor dyes and light up specific proteins in living human cells. Out of the Baker Lab’s design pipeline (RFdiffusion → LigandMPNN → AlphaFold/RoseTTAFold filtering) came orthogonal far-red, orange and green binders that enable multicolor imaging, live STED and fluorescence-lifetime microscopy — enough, in principle, to track up to 30 proteins at once. A split version, NovoSplit, acts as a dye-triggered “molecular switch,” and the tags point toward inducible cryo-CLEM for in-cell structural biology. Why it matters for the lab: this is protein design in service of seeing — AI designing the very tools our microscopes use. It’s the most on-brand possible fusion of generative design and bioimaging, and the sequences are open to everyone.

🧬 AI can now write genomes

The same generative turn is reaching DNA itself. A Nature feature asks how close AI-written genomes are to synthetic life, and the proof points are piling up: DNA-Diffusion (Nature Genetics) designs compact, cell-type-specific regulatory elements and used one to reactivate the leukemia-protective gene AXIN2 in its native context, while an earlier Cell study showed AI-designed DNA controlling gene expression in healthy mammalian cells for the first time. Why it matters for the lab: designing regulatory DNA is designing the control logic of the cell — a direct handle on the cellular programs a virtual cell tries to predict.

🏗️ And now we can actually build them

Design has been outrunning construction — Evo 2 can write sequences far faster than labs can make them. Sidewinder (Caltech) closes that gap: it assembles many sequences at once with one error per 10 million junctions (versus one per 10–30 conventionally), using cheap oligos tagged with molecular “barcodes” so fragments snap together in the right order. In a demo, the team redesigned a 12,500-letter E. coli stretch with Evo 2 and built it error-free — turning a month of work into a few days. Why it matters for the lab: the bottleneck is shifting from designing biology to building and testing it — which is exactly the design–build–test loop REEF is built to run, now with far more ambitious things to make.

Design the probe, write the DNA, build it for real — AI is moving from describing biology to manufacturing it. The lab’s job is to keep that loop honest: design, build, and test on real cells.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.