In September, @AnthropicAI announced that its AI agents had found a reverse transcriptase alongside a novel CRISPR-like genetic array in phage DNA.
What it actually does, the paper said, would “await experimental characterization.”
So we tested it on Ginkgo’s Cloud Lab. Scientists at Ginkgo wrapped Claude Opus 5 in a harness, provided it with the preprint and a catalog of validated workflows in the Ginkgo Cloud Lab, and asked it to generate executable work orders. We ran the first set of these experiments and obtained the first evidence of biochemical activity 13 days after experimental design using about $14,000 of cloud lab services.
Read the full write-up: https://t.co/l38C3b4GtD
You can run your own experiments on our Cloud Lab from anywhere today: https://t.co/VKjWr4OFrv
wtf now my agents found the first bacteria to use new stop codons!!
After this TAG result, a chat w/ friend inspired the question: do any bacteria use other stop codons?
Overnight, my guys found myxobacteria that stop at TTA, normally read as an amino acid.
Evidence:
1) These bacteria lost the tRNA that normally reads TTA as leucine.
2) Their RF2, the protein that recognizes stop codons, has two changes: one in its codon-reading tip (SPF → SPY) and one nearby (L → F). No other species in GTDB has both.
3) We compared the same genes in these bacteria and their closest relatives. Where a relative's gene ends at a normal stop but the gene here keeps going, the very next codon is TTA 11% of the time (252 of 2,339 genes). In the relatives, it's 0.05% (1 of 2,025).
4) TTA swaps with the normal stop codons. Where one of these genomes ends a gene at TTA, another genome from the same group ends that gene at a normal stop 78% of the time.
Method:
Claude agents on my Mac mini scanned all 199,923 bacterial and archaeal genomes in GTDB for amino-acid codons unusually common near the ends of genes, where a hidden stop would sit. In 4,599 genomes, they then matched each gene to corresponding proteins and checked which codon sits exactly where the known protein ends.
Closest prior work:
TTA works as a stop in some mitochondria and plastids (Barcytė et al. 2025), and CGA was proposed as a stop in Asgard archaea (Melnykov 2026, PNAS), but none in bacteria until now.
One of these bacteria, 1-FT3.2, is kept alive in Finland (Demina et al. 2025). Possibly its proteins can be measured to confirm this.
If you think bird flu spreading unchecked in mink is dangerous because of adaptation to respiratory transmission in mammals, wait until you see what happens when it gets to do that in humans
@ProfTomEllis@WillPVGreen NCTC was very easy to deal with. ATCC is a nightmare. They sequence their collection and then refuse to release sequence unless you buy the strain.
@WillPVGreen The solution is to make recreating that strain trivial, with automated design tools and automated laboratories. It's time to move biology from who-you-know to what-you-know.
Let’s compare the state of protein structure prediction circa 2017 to the virtual cell today:
Structure prediction c. 2017:
- deep learning working on sequences (eg seq-to-seq, transformer) and other modalities like images and point clouds (PointNet). Mapping a sequence to a heavily constrained point cloud was entirely feasible.
- Protein Data Bank (PDB) was 46 years old with data mapping amino acid sequences to structures and sequence databases existed with close evolutionary neighbors to these sequences so we knew what positions covary (more info about the seq->struc map).
- Anfinsen’s 1972 Chemistry Nobel for showing a protein’s native structure is the free-energy minimum of its sequence, ie the sequence->structure map is largely deterministic / sequence is a sufficient statistic for predicting structure.
So the problem was well-defined: learn the map seq->struc, which we knew must exist (Anfinsen), and use deep learning which we knew works, on data we had in hand (PDB + homologous sequence data which contain all the necessary signal about the inputs & outputs of that map).
Edit: AF is still impressive and required advanced engineering.
The virtual cell today:
- no agreed definition
- predicting future “cell states” from past states + perturbation is underspecified; what is a “cell state”? What is a perturbation?
- all measurable choices for cell state (eg scRNAseq, multi-omics, imaging, phenomics) are very lossy and almost all methods require destroying the cell to take the measurement (can’t measure multiple time points)
- we have no evidence that any measurable past “states” (which we can’t even define) form a sufficient statistic for future states (can’t rule out the influence of unobservables; no Anfinsen result).
- the only map we know must exist is physics, which is intractable to simulate at cell scale.
- we have the wrong data to learn this undefined state+perturb->state mapping. Even if we take a narrow stance on the inputs/outputs like state = RNA levels & perturbations = single gene knockdowns, the mapping is heavily environment dependent and we have no way to reduce “environment” into a feature vector without losing (most) information.
AlphaFold had to learn a map we knew existed, from data that captured the inputs and outputs of it. The virtual cell is a map we can't define, that may not exist over observable states, and our data captures neither inputs nor outputs fully. These are categorically different initiatives and project risk is worlds apart.
Ginkgo computational biologist Keith Robison clears up where you need to be in order to use Nebula, the world's largest autonomous lab: anywhere.
You can run your own experiments right now at https://t.co/L1igkvy7F9
@leecronin As if those people could come close to inventing a dog. Dogs are WAY beyond our ability to understand or engineer, and will be for a very long time. We don't know in detail how the even the simplest bacteria work, and can't make one without an existing one.
I've been prototyping an automated peer review project. With current models you get human level peer review for about $0.10 of tokens. For $5 you can rederive every result, check every statistical inference, and audit every paper cited in both directions.
There are about 100m papers in the scientific corpus. Most are not important. Many are sloppy. For structural reasons it's essentially impossible to change published work, even to correct a typo.
For $500m today you can thoroughly audit every paper ever written. Every p-hacked result, every plagiarist, every fabrication, every citation club, everything.
I don't have $500m so instead I build a tree, focusing on the 1% of papers that carry 99% of the signal, doing a cheap first pass looking for obvious problems. I can find 90% of the nonsense for a few hundred dollars of tokens.
Any guesses where the biggest surprises will be found?
🤖 My latest post proposes AI safety measures that may sound less insane now than earlier this year. No one else is saying this, but many are thinking it:
@NikoMcCarty If only we had small, viable, low complexity cells, with, say, 500 or so genes, that could be carefully studied and completely understood so that we could solve these problems ;-).