Every protein design pipeline ranks candidates by how confident a structure model is.
We designed 297 proteins three ways and measured all of them.
The set selected hardest on structural confidence, had the highest confidence, and the fewest binders: 15 of 90.
The lowest-scoring set: 80 of 96.
Useful controlled comparison from Aether. On their lanmodulin metal-binding task, designs ranked by a model trained on measured binding data hit ~83%, while designs filtered mainly by structural confidence (RFdiffusion-style + Boltz-2 metrics) hit only ~17%. Absolute hit rates aren’t unique though. Chai-3 is reportedly showing ~40-50% on harder zero-shot antibody design against novel targets.
The tasks differ, so the numbers aren’t directly comparable. What stands out is the within-task gap: geometric confidence was a weak or misleading filter here. Structure models keep improving and may capture more functional signal implicitly, but for now the result suggests measured function is still the stronger ranking signal on this problem.
What would clarify things is repeating this kind of head-to-head on more diverse functional tasks with the newest structure and co-design models, and seeing how much the gap shrinks.
An important paper to read. We who work in chemistry, biology or drug discovery have always known that structure and function are two distinct things, but that distinction often gets lost in the breathless enthusiasm around new protein structure prediction techniques. Studies like this put the enthusiasm in perspective.
"The designs that looked most confident produced the fewest working proteins. The design set with the lowest confidence scores produced the most working proteins, and the only designs that beat what nature already offers."
https://t.co/YYc1wKFe49
The protein discovery industry has ballooned to billions of dollars for good reason: nano-engineering new more performant, efficient materials is a market larger than we can conceive today.
But the approach of the big AI labs and industry "leaders" has been misguided. Using an LLM to predict protein function is inherently flawed, because the training data of LLMs (the alphabet) is not the foundation for chemical reactions.
See Aether's new research ⤵️
Anthropic lists Thursday. It bought a biotech and runs a wet lab.
We published a result today: the protein designs the structure models were most confident in were the ones that didn’t work. 84% of them didn’t bind anything.
Structure confidence does not predict function. A model trained on measured function does.
I’m on @MTSLive later today to get into it.
Our CEO @SynBioMars will be on @MTSlive today to talk about Aether's mission: machines the size of molecules, factories the size of shipping containers.
We've built the 1st foundation model for chemistry to reshape what's possible in protein discovery & nano-engineering.
New research on our work ⤵️
After 6 years in stealth, Aether's first paper is out.
We designed and measured hundreds of proteins. The ones that leading models were most confident in were least likely to work: 84% didn't bind.
A confident predicted shape is not evidence of function. https://t.co/fJIFn5uqAv
After 6 years in stealth, Aether's first paper is out.
We designed and measured hundreds of proteins. The ones that leading models were most confident in were least likely to work: 84% didn't bind.
A confident predicted shape is not evidence of function. https://t.co/fJIFn5uqAv
And a model on top of it, one that ranks any protein's function, not just the ones it designed. Scaling on two axes: more proteins, and more functions measured per protein.
A foundation model for protein function.
It doesn't care what a protein looks like. Only what it does.
Our full paper: https://t.co/tFvLD90mWT
Every protein design pipeline ranks candidates by how confident a structure model is.
We designed 297 proteins three ways and measured all of them.
The set selected hardest on structural confidence, had the highest confidence, and the fewest binders: 15 of 90.
The lowest-scoring set: 80 of 96.
So why does the field use a plausibility filter as a function filter?
Because there isn't another one. 250,000 protein structures sit in the PDB. There's no comparable record of what any of them actually do.
So we built one. In-house, at scale, under controlled conditions.
ICYMI: December was big for Aether. $15M raised to reindustrialize America with AI-designed proteins, including a nod in @crunchbasenews's December roundup of interesting startup deals.
https://t.co/ZjRAfq6rPV
At Aether, we're building a future of abundance: creating drones, armor, and more, faster and better than ever. Our founder & CEO @SynBioMars explains 👇🎥
With our AI, we’re able to develop polymers that are among the strongest on the planet.
These super materials deliver aerospace-grade aluminum strength at half the weight, while remaining easy to injection mold or 3D print – an unprecedented feat in manufacturing.
Our Founder & CEO @SynBioMars breaks it all down👇🎥
Our mission: To create more abundance, economic freedom, and upward mobility, was inspired by our CEO & Founder @SynBioMars at a young age. Today, we're backed by leading investors and moving toward commercialization. @UniteAi has more 👇: https://t.co/pHptk1wOGa