LEDGER | 26.3
NEWSLETTER
THE BIOLOGICAL AWAKENING
A dispatch from the century of biology
№ 01 | AI × BIOLOGY
The Move We Cannot Compute.
A New Look at Artificial Intelligence in Biology
This is the short version of our latest issue. The full essay — with every source and the field notes behind it — lives on The Biological Awakening on Substack.
ISSUE 2026.3
BERLIN
“Nature opens its eyes in us.”
When I named this newsletter The Biological Awakening, I was quoting Schelling: nature opens its eyes in us. But an eye that has only just opened is not yet an eye that sees well. Ours is still short-sighted — we see the single tree and not the millions of chemical signals traded beneath it. That is about to change, because a new player has entered the scene. And we have been misreading what it is and what it can do.
Within the old paradigm, artificial intelligence promised to be the ultimate means to create life, at last, according to our wishes — to make biology “programmable,” to build “machines to read, write, and think in the language of nucleotides.” We should acknowledge what this dream has achieved: in 2024 the Nobel Prize in Chemistry went to computational protein design and structure prediction, and by August 2026 the first AI-designed bacteriophages able to infect living bacteria had become a reality. And yet the limits of these models have grown just as visible as their promise.
01
THE LIMITS
The limits are becoming visible — and so is our misunderstanding
The numbers are sobering. Of the AI-written phage genomes, only a hit rate below six percent produced something viable. De novo antibodies land between zero and two percent per target. And the confidence scores the models hand us are no rescue — across hundreds of characterized de novo proteins there is no threshold above which success reliably begins.
A 2024 paper in PNAS named the truth behind these numbers: protein language models do not model the physics of folding; they store the evolutionary statistics of interacting sequence motifs. They remember what has occurred together across four billion years. Which is why the instrument works beautifully where life repeats itself, and badly where life does something new. AlphaFold changed everything and nothing: it predicts structure with an accuracy that reorganized the field, and tells us almost nothing about how a protein actually moves to get there. The achievement is real — but it is not the same thing as understanding.
“
I’m of the opinion that deep learning models cannot go beyond interpolating from the data they’ve seen. If a model does well at a task that seems ‘out of distribution,’ to me that means that we didn’t understand existing relationships between the data and tasks well enough to start with. But that’s not necessarily bad — in much of biology, we’re still trying to figure out those relationships. To me that’s why it’s interesting to be doing modeling in biology at all.
HANNAH WAYMENT-STEELE
Assistant Professor of Biochemistry
University of Wisconsin–Madison
TWO VOICES FROM THE FIELD
The dialogue continues.
Over the coming issues I’ll put these questions to Hannah Wayment-Steele, who co-authored the PNAS study at the heart of this piece, and Jonathan Funk, who builds the open tools designers depend on. Both conversations will run as essays and as podcast episodes. If you have a question you’d like put to either of them, bring it to the comments on Substack.
Join the conversation on Substack →
02
THE METAPHOR
Caught in a metaphor trap
In 1944 Erwin Schrödinger gave the gene the name that stuck, calling the chromosome a “code-script.” What almost nobody quotes is the sentence that follows: the chromosomes, he wrote, are “law-code and executive power — or, to use another simile, architect’s plan and builder’s craft — in one.” Eighty years later we have built machines that hold the code-script and nothing else.
So let me correct myself. The genetic code is a code. But life is more than code — because a code has to be read by someone, and enacted, and the reading is not in the script. A genome without transcription is just a molecule; it becomes life in the reading. The model has the ink, but it is the cell who reads and gives the final answer.
03
THE DIALOGUE
Only life can answer. Only we can change the question.
Once we see the limits of the code metaphor, it becomes clear that we understand AI in biology far better if we think dialogically — and divide the labor in this new communication triangle exactly.
01
Life answers, and never explains.
It is the only party that can say yes or no to this design, here, now. What it cannot do is give reasons. The cell says yes or no and never why — which is why we are still needed after the experiment, not only before it.
02
The model speaks without answering.
It holds the whole prior conversation — four billion years of it — at once, which no human and no cell can do. But it has no access to the case in front of it, and it cannot let a result change what it is looking for. It turns the archive of past answers into a supply of next questions.
03
We change the question.
Nature answers the question it was asked. The model paraphrases questions already asked. Only we can ask a different one — and only we can be held to it.
And this reframes what once looked like failure. Within the dream of programming life at will, a cell that refuses our design is a defeat. Within a dialogue, that same refusal becomes a widening of the view. In a real conversation I cannot expect to convince my partner every time; sometimes my argument fails and I am taught better. That is the moment I learn.
№ 04 | THE WET LAB
Which is why the wet lab just became the most valuable place in the world.
Sequences are abundant; measurements are scarce. To predict function — turnover rate, stability, solubility — we need numbers, and those are absent from the gene databases. With text models the training material was also the supervision: the next word is a free label. In biology there is no free label for function. Somebody has to go out and measure.
This kind of data — functional measurement, not sequence — is the real bottleneck now. It explains why the OpenAI Foundation recently granted forty million dollars to a single university for biological data, and is buying up the datasets of insolvent biotechs — precisely the negative results no one ever published. It explains the SynBio bust, too: Ginkgo and Zymergen fell hard not because biology does not work, but because the old paradigm assumed the design was the hard part. Reading and writing DNA got 357 times cheaper since 1999. Understanding did not.
“
I often get feedback from designers before they take samples to the lab. Experienced designers can often tell by intuition if designs are likely to work or not. However, nature is the ultimate judge, jury, and executioner.
JONATHAN FUNK
Digital Biotechnology Lab
Technical University of Denmark
05
THE AGENDA
What this opens up
This is where the realistic view stops being sobering and turns into an agenda. If the design is no longer the scarce thing, then value moves to where the measurements are made — to the lab, the assay, the negative result finally worth keeping. Whole new fields open around this bottleneck: functional data as an asset class, foundries built for learning rather than only for making, ventures whose moat is a proprietary library of what life actually answered. For research, for industry, for founders and the VCs backing them, the question is no longer how do we program the cell but how do we build the space where the dialogue happens fastest.
That space is exactly what we built the SynBioReactor to be: a transdisciplinary place in Berlin where academia, venture capital and industry sit at the same table, so that the questions worth asking — and the data worth generating — find each other early. The most valuable move in all of this is choosing which question to ask next. That is the one thing no machine can compute, and the one thing this room exists to do together.
Read the full essay
With all sources and further reading, on Substack
→
A TASTE OF WHAT’S NEXT
Who Is Prompting Whom?
If we have to persuade a cell rather than command it, then it can also persuade us. What could life — the longest-running pilot programme nobody is running — teach us about our own hardest problems: about supply chains, markets, a circular Interspecies Economy? The catch: a survivor is not an adviser. What lasted is not the same as what works best. That is where I take this next — the full essay is on Substack.
READ ON SUBSTACK →
See you where biology awakens,
in
Felix Ernst
AUTHOR & EDITOR
IN COLLABORATION WITH THE SYNBIOREACTOR TEAM
WITH THANKS TO OUR ORGANISERS
www
in
www
in
The SynBioReactor is made possible by GASB | the German Association for Synthetic Biology and biobeyond. Our sincere thanks to both teams for the work behind this summit, and for keeping the dialogue going all year round.
THE BIOLOGICAL AWAKENING · 2026.3
You are receiving this because you asked to walk this path with us. The essays arrive as they are finished, and nothing else does.
Sponsoring
|
Contact | sbr@ga-sb.de
|
Legal Notice
synbioreactor.de
© 2026 GASB | German Association for Synthetic Biology. All rights reserved.