Pinned Tweet
Do we even need to fold proteins at all to find drugs?
Or do we just do that to satisfy our 3D brains?
Meet Ptarmigan-1 🧵
- Structure-free drug discovery
- 5000x less compute than co-folding
- Finds known and cryptic pockets
- Actually works on poorly structured targets
Or do we just do that to satisfy our 3D brains?
Meet Ptarmigan-1 🧵
- Structure-free drug discovery
- 5000x less compute than co-folding
- Finds known and cryptic pockets
- Actually works on poorly structured targets
12
32
239
17,675
Just got back from the AI Drug Discovery conference in Boston, and my big takeaway is the rise in the number of AI drug discovery models, how few have lab data behind them, and how much everyone wants to see that data.
Most are benchmarked on the same leaky public datasets, so ones actually work.
We have our own lab, so the experiment is part of our loop, but third party lab validation also should play a big part, @adaptyvbio is a great example for protein design
What should look like in the small molecule space? What targets?
I go back and forth on this. An easy target (a kinase or something) is cheap to test, but is probably just memorization at this point. A hard target (a TF, something without an obvious pocket) would be convincing if it worked. It also usually requires assay dev, and a negative result tells you very little.
And what's the bar? Is it biochemical potency? Selectivity across the family? The proteome? Engagement in cells? Inhibiting protein function?
More questions than answers at this point, but we'll keep sharing how we're thinking about this as we get more data
Most are benchmarked on the same leaky public datasets, so ones actually work.
We have our own lab, so the experiment is part of our loop, but third party lab validation also should play a big part, @adaptyvbio is a great example for protein design
What should look like in the small molecule space? What targets?
I go back and forth on this. An easy target (a kinase or something) is cheap to test, but is probably just memorization at this point. A hard target (a TF, something without an obvious pocket) would be convincing if it worked. It also usually requires assay dev, and a negative result tells you very little.
And what's the bar? Is it biochemical potency? Selectivity across the family? The proteome? Engagement in cells? Inhibiting protein function?
More questions than answers at this point, but we'll keep sharing how we're thinking about this as we get more data
5
1
11
885
Alex Federation retweeted
Jon Clardy was a brilliant scholar and a wonderful man, beloved uniformly in chemistry and chemical biology from Boston to Brazil. My sincerest condolences to those who were lucky to know him well, as mentees and close friends. RIP, Jon, just cannot believe you are gone.
Jon Clardy: 1943-2026
One of the most impactful, creative, and (quietly) irreverent practitioners of chemical biology. Transformed how we think about nature's chemicals: their structures, their role in biology. Loved him -- he had tremendous influence on my thinking.
1
5
48
3,284
We opened the API and MCP for our structure-free drug discovery model (Ptarmigan-1) today.
With that, we're also increasing usage limits by 1,000x
In my first week of grad school I sat through a lecture on ‘undruggable’ targets, which ended up being a frustrating (but nd I've worked the last 15 trying to help remove this term from the trade. Ptarmigan-1 is a first step to let AI target these proteins, powered by native-cell data where these targets are at home.
With the API/MCP, you (or your agent) can now rank up to 10M compounds against a protein using only its sequence, with no 3D structure predict where it will bind.
That's ~10 GPU-years of co-folding
With that, we're also increasing usage limits by 1,000x
In my first week of grad school I sat through a lecture on ‘undruggable’ targets, which ended up being a frustrating (but nd I've worked the last 15 trying to help remove this term from the trade. Ptarmigan-1 is a first step to let AI target these proteins, powered by native-cell data where these targets are at home.
With the API/MCP, you (or your agent) can now rank up to 10M compounds against a protein using only its sequence, with no 3D structure predict where it will bind.
That's ~10 GPU-years of co-folding
4
4
16
761
One cool example on a former "undruggable" target (KRAS) --
Ptarmigan-1 ranked known KRAS G12C compounds in the order of their development generations without ever being trained on affinity, this ranking required the mutation to be present.
Across 337 drug-like complexes, the top-scoring residue sat within 5 Å of the ligand 92% of the time, from sequence and chemistry alone
Ptarmigan-1 ranked known KRAS G12C compounds in the order of their development generations without ever being trained on affinity, this ranking required the mutation to be present.
Across 337 drug-like complexes, the top-scoring residue sat within 5 Å of the ligand 92% of the time, from sequence and chemistry alone
1
0
2
125
Try it and let us know how it goes!
Portal: ptarmigan.app
Paper: biorxiv.org/content/10.648…
Email us: ptarmigan at talus dot bio
Portal: ptarmigan.app
Paper: biorxiv.org/content/10.648…
Email us: ptarmigan at talus dot bio
0
0
0
39
Alex Federation retweeted
Another thing: AlphaFold is like a tiny, tiny, slice of all of biology. It's a great technical accomplishment, but it applies to static, well-formed structures (not intrinsically disordered proteins - roughly 30% of the entire proteome), and still cannot handle dynamic multimers well. It also depended on having enough data: 200,000 structures in the PDB and counting.
Saying that AlphaFold works, therefore we could "solve" biology with enough compute, is like saying that because we can predict the shape of every component in a car, we can predict everything the car will do on the road. And saying that compute or reasoning can somehow magically compensate for data is a religious belief.
Saying that AlphaFold works, therefore we could "solve" biology with enough compute, is like saying that because we can predict the shape of every component in a car, we can predict everything the car will do on the road. And saying that compute or reasoning can somehow magically compensate for data is a religious belief.
If you had an exponential amount of compute you could solve simulate biology exactly and wouldn't need to run any experiments.
Some of you will think "of course, but that's an unrealistic assumption" and others will think "no, you need experiments even with that much compute. That's just How It's Done".
You would of course need evaluations once you have a solution, but you don't in principle need experiments to find it: Experiments are a way to save compute, dating from a time when it was much more expensive.
And I think the presentation is exponential but real systems aren't irreducibly exponential. AlphaFold worked, after all.
16
27
280
9,863
Unlocking hard drug targets means searching massive chemical spaces
Since opening a structure-free discovery model 2 weeks ago (Ptarmigan-1), users have already tested 2M+ compounds for free
This would have taken years of compute with co-folding models
Since opening a structure-free discovery model 2 weeks ago (Ptarmigan-1), users have already tested 2M+ compounds for free
This would have taken years of compute with co-folding models
1
1
2
104
Alex Federation retweeted
Build at DNAForge’s AI × Biology Hackathon.
November 14th · Berkeley · $1,500+ in prizes
With @modal @boltz_bio @tamarindbio @talusbio @gxl_ai @benchflow_ai @SynphonyDev @GabriCorso @xdotli @Atelfo
Priority: October 25th
Apply in under 5 minutes:
dnaforge.com/hackathon
November 14th · Berkeley · $1,500+ in prizes
With @modal @boltz_bio @tamarindbio @talusbio @gxl_ai @benchflow_ai @SynphonyDev @GabriCorso @xdotli @Atelfo
Priority: October 25th
Apply in under 5 minutes:
dnaforge.com/hackathon
1
15
22
1,920
Today we’re opening our structure-free AI drug discovery model for all to use!
Ptarmigan-1 predicts which small molecules bind a protein, and where, without building a 3D protein structure. It learns from multiple kinds of data, including measurements of proteins at work inside is is the culmination of a long and winding path to get here
In my first week of grad school, I sat in a lecture hall where generations of chemists had trained and Nobel Prizes hung on display. One of the most accomplished chemists in the world introduced us to the word “undruggable,” describing targets that so many had given up on.
I remember sitting there feeling so indignant. Of course all targets are impossible to drug until the right technology lets you drug it.
Thankfully I found my people in @jaybradner's lab, where the T-shirts said “nothing is undruggable”, and got to work
We focused on gene regulators there, proteins that control which genes turn on and off and can drive cancer. Most are poorly structured, making them difficult targets for traditional tech that focus on a well-defined binding pocket.
Gene regulators are not the only disordered proteins though. Roughly 40% of the human proteome has significant disorder, and these targets remain out of reach while discovery for structured proteins gets faster every year.
With Ptarmigan-1, we’re giving researchers a way to search for molecules against these targets without first having to solve their structure.
I’m grateful to the team and supporters that brought us here. Now we get to see what the community does with it.
Try it and tell us what you think. Link below.
Ptarmigan-1 predicts which small molecules bind a protein, and where, without building a 3D protein structure. It learns from multiple kinds of data, including measurements of proteins at work inside is is the culmination of a long and winding path to get here
In my first week of grad school, I sat in a lecture hall where generations of chemists had trained and Nobel Prizes hung on display. One of the most accomplished chemists in the world introduced us to the word “undruggable,” describing targets that so many had given up on.
I remember sitting there feeling so indignant. Of course all targets are impossible to drug until the right technology lets you drug it.
Thankfully I found my people in @jaybradner's lab, where the T-shirts said “nothing is undruggable”, and got to work
We focused on gene regulators there, proteins that control which genes turn on and off and can drive cancer. Most are poorly structured, making them difficult targets for traditional tech that focus on a well-defined binding pocket.
Gene regulators are not the only disordered proteins though. Roughly 40% of the human proteome has significant disorder, and these targets remain out of reach while discovery for structured proteins gets faster every year.
With Ptarmigan-1, we’re giving researchers a way to search for molecules against these targets without first having to solve their structure.
I’m grateful to the team and supporters that brought us here. Now we get to see what the community does with it.
Try it and tell us what you think. Link below.
19
31
220
22,385
The "p" is silent. Ptarmigan, like pterodactyl
The portal is here: ptarmigan.app
Blog: blog.talus.bio/p/c0a4b0c3-ca1…
The portal is here: ptarmigan.app
Blog: blog.talus.bio/p/c0a4b0c3-ca1…
0
1
5
729
Well in my hands the bio filters have a 100% false positive rate, but I'm no statistician 🤷♂️
Today, we'll resume charging for requests our safeguards block before Claude responds. This only applies in categories with low false positive rates: biology, distillation attacks, and frontier LLM development. We've seen some coordinated attacks on our systems in recent weeks, and this is one layer of defense.
In recent testing, 99.7% of accounts using Claude Code, Claude.ai, or Cowork did not hit any of these newly "billable blocks." The classifiers behind the blocks we’re resuming charging for today are tuned to have a <0.1% false positive rate. We know that's not 0%, and we're going to keep improving them so they interrupt your work less often. If you think a request has been blocked incorrectly, please report it with /feedback in Claude Code. platform.claude.com/docs/en/build-…
1
1
11
805
A molecule for every target that needs one
We’re hearing two kinds of voices on the potential of AI to cure all disease:
One says AI will cure everything, but without a concrete path to get there.
The other says biology is too complex to ever tame, but underestimates the power of the tools that will unlock the full potential of frontier intelligence.
At Manifold Bio, we see a path and we’re already building it: a high-throughput interface into living system biology. Millions of protein designs, measured directly in vivo, to learn how molecules behave in the body and ultimately predict what they’ll do before they reach a patient.
Today we share mBER-2, the latest iteration of our AI protein design model that is a core piece of this mission.
0
0
3
405
Alex Federation retweeted
We’re hearing two kinds of voices on the potential of AI to cure all disease:
One says AI will cure everything, but without a concrete path to get there.
The other says biology is too complex to ever tame, but underestimates the power of the tools that will unlock the full potential of frontier intelligence.
At Manifold Bio, we see a path and we’re already building it: a high-throughput interface into living system biology. Millions of protein designs, measured directly in vivo, to learn how molecules behave in the body and ultimately predict what they’ll do before they reach a patient.
Today we share mBER-2, the latest iteration of our AI protein design model that is a core piece of this mission.
One says AI will cure everything, but without a concrete path to get there.
The other says biology is too complex to ever tame, but underestimates the power of the tools that will unlock the full potential of frontier intelligence.
At Manifold Bio, we see a path and we’re already building it: a high-throughput interface into living system biology. Millions of protein designs, measured directly in vivo, to learn how molecules behave in the body and ultimately predict what they’ll do before they reach a patient.
Today we share mBER-2, the latest iteration of our AI protein design model that is a core piece of this mission.
Today, we’re announcing mBER-2, our latest AI protein design system, and sharing a bit of how we use it to explore biology in vivo at scale.
For a long time, we’ve been working on what I think of as “frontier problems” in medicine and biology. We know a lot about the targets t there is still an enormous amount of biology we haven’t explored, and potential uses for that biology we haven’t discovered.
One of those problems is drug delivery. There are thousands of potential targets, and most receptors remain unexplored as routes for delivering medicines. We want to understand which ones we can use, where they can take a drug, and what kinds of molecules make that possible.
Ultimately, we need to test these molecules in vivo to understand what they actually do. We’ve spent the last six years building measurement technologies that let us generate millions of measurements in living systems.
mBER-2 is an AI protein design system built for that scale of in vivo measurement. Our goal is to generate binders across the full range of targets we care about, while systematically exploring the possible binding sites on each one. It’s not just what you bind, but where and how you bind that determines whether a molecule performs it's function. mBER-2 lets us probe those differences at massive scale.
We’re also sharing a look at some of our in vivo data. We’ve designed more than 50 million molecules across over 4,000 targets, and screened millions of molecules in living systems. By probing hundreds of receptors, we’ve found new pathways for delivering genetic medicines into fat that outperform industry benchmarks by a large margin.
This is just the beginning. We call the space of all bindable sites across proteomes the EpiTome. As we explore it, we’re building the data to connect AI design, receptor binding, and what a molecule actually does in a living organism.
Much like the idea of a virtual cell, we envision a Virtual Organism that helps us design medicines with specific properties in mind. The foundation is our own in vivo data, connecting molecular design to outcomes measured in living systems.
The endgame is to use AI to explore more biology, measure what happens, and use what we learn to design better medicines. Every round should deepen our understanding of biology and improve our ability to build molecules that do what we need them to do.
8
27
255
22,903
Alex Federation retweeted
This is fun.
1. Mango Banana - banana on the outside, mango on the inside. Enough said.
2. Plant machine interfaces - we should be able to trade plants N,P,H2O for reduced carbon compounds. Do to them what they did to chloroplasts
3. Solution phase self assembly of 3D memory chips using designed self-assembly of macromolecules - do we really get protein design
4. A brain the size of a building. - what would it be thinking?
5. A five assed monkey - bc South Park
6. Self growing housing - the rent is too damn high
7. Obviously brain machine interfaces - I will judge this by when I can browse the web by thinking about it.
8. Dinosaurs.
9. Cats with dog personalities.
10. A nice smelling microbiome
1. Mango Banana - banana on the outside, mango on the inside. Enough said.
2. Plant machine interfaces - we should be able to trade plants N,P,H2O for reduced carbon compounds. Do to them what they did to chloroplasts
3. Solution phase self assembly of 3D memory chips using designed self-assembly of macromolecules - do we really get protein design
4. A brain the size of a building. - what would it be thinking?
5. A five assed monkey - bc South Park
6. Self growing housing - the rent is too damn high
7. Obviously brain machine interfaces - I will judge this by when I can browse the web by thinking about it.
8. Dinosaurs.
9. Cats with dog personalities.
10. A nice smelling microbiome
In light of the progress in mathematics, we at Edison Scientific and FutureHouse have assembled a set of Millennium Problems for Biology. They are chosen to be very hard to solve but very easy to validate in a simple laboratory environment. Any of these, if solved, would mark a hnology, and most of them would contribute materially towards curing disease. These are, in some sense, the “last reasonable eval” for AI in biology. This was work primarily by @MichaelaThinks and myself, with contributions from many others.
Short descriptions below. The full descriptions of the problems with acceptance criteria are at the Bio Millennium Problems website, linked in the next post. Share more if you have ideas. If they meet our criteria, we’ll add them to our list (with attribution and permission).
20
18
230
27,958
Alex Federation retweeted
1/
What could we discover if we had a binder for every protein in the human proteome?
We are excited to introduce the Bindome, an open and free resource of over 300,000 designed protein binder candidates against over 8,000 human target proteins.
🧵
What could we discover if we had a binder for every protein in the human proteome?
We are excited to introduce the Bindome, an open and free resource of over 300,000 designed protein binder candidates against over 8,000 human target proteins.
🧵
20
113
497
48,684
Alex Federation retweeted
When you actually look at global biotech companies, the number of interesting molecules is very thin. Discovery and early stage has been disregarded for so long because of the biotech bear market. In the end, the funnel has to be better broadly distributed and early stage assets/novel targets are where the big wins are at. Both for patients and profits. $XBI
Discussion w the senior BD leaders at big pharma and one surprising point made is that more than one said that they are focused on earlier stage and are comfortable with clinical risk. They don’t seem to be shaken up by avidity/$NVS recent post deal setback. “In BD, risk is your friend. If you don’t take risk. You’ll pay what others think it’s worth” talking about transacting phase 2 or earlier, pre readout $XBI $IBB
5
5
38
7,303