I did this at home with a 5090 on a different target a few weeks ago. Cannot generate 318,000 designs but Claude chose the nearly same tooling, except I used BindCraft2 which is better for multi target designs in the beginning. This isn't that special, its just doing tool calls. I got 100 or so trajectories running 24hrs just for fun.
"Throughout the process, Claude generated 318,484 protein designs and selected 180 of them for testing. Our team added 31 more, with Claude’s help."
"lab-tested all 211 designs to see if they would bind to the toxins they were designed against, and 17 designs met our criteria for a binder. "
Alright how many of those 17 were 100% Claude, I know mine were. These mini-proteins the tooling I'm told is pretty good for these. I'm worried that Claude might just choose the same tooling for everyone if this is the path we go down and designs stagnate?
I hope this produces some useful anti-venom that is cheaper and widely available. My kid was stung by a small bark scorpion on his thumb here in Arizona and I would not wish that experience on anyone. For hours afterwards he was in excruciating screaming pain, and he even lost his vision for while. That lasted about 24 hours and then it took weeks for him to recover feeling in his hand. Nothing seemed to help with the pain and there was very little we could do for him. I would love to have had a more available anti-venom, but apparently now it is so expensive they only keep it around for the very worst cases.
A single shot of rabies costs about ~$3 in Asian countries where street dogs are common. It’s crazy how expensive post exposure rabies treatment is in the US.
Unfortunately, available does not mean cheap these days. Parents (good ones anyway) would go into a decade of debt to prevent their child from having an experience like you just described. Drug giants see that as a business opportunity.
Can someone explain whether design of protein binders is an important rate-limiting step in drug creation (of which antivenom is a very small market)? Or AI-driven drug discovery in general? It feels like something that will definitely help the drug market but is nowhere near the rate-limiting step, but that's from my own researcher as someone without a background in it.
My understanding is that FDA trials (specifically phase 2 + 3) are the death valley for most drugs and finding a biomarker is something mostly done through a mix of gene bank scraping + looking at old biomarker databanks. The majority of the cost and time is in failed drugs that do not pass FDA trials (which average 6-7 years end-to-end).
Note: I am asking this in earnest. I am not trying to ask a leading question and saying it is not important at all.
My understanding is it is not - this is something we have models for, and there was a flurry of machine learning based drug discovery companies about 8-10 years ago. But they ran into the testing bottleneck real fast because they would spit out thousands of candidates for binders and so on that all needed to be tested in vitro before they could even start the real drug dev circuit. Furthermore the drugs need to be manufacturable and that's not a trivial problem.
It's nice to (as the blog post suggests) give the opportunity to non medical experts, say in this case a snake/venom research center, to pursue a problem and produce some realistic candidates that they could pay to test. I do think it's good to democratize what parts of the process can be, but it's more of a resource prioritization problem than a molecular design one.
Have a comment below on this, but generally things don't work out in the clinic because we didn't understand the biology adequately, and not because we didn't have a molecule.
A better binder only helps if you know what to bind to and that if you bind to it, you'll have the desired effect.
This is why the industry has crowded around a very limited set of "validated" targets. It's because we know if we do something with them, they will work. Obviously we're likely missing a lot of great ideas though.
AI might help find some of those ideas, but we won't know until someone tests them in a human. Essentially, without more extremely expensive and difficult data to acquire, it's not likely to be that much better at picking winners.
I'm in the sector and I believe strongly in AI helping (I've bet my entire career on it). It will get there eventually, but it will take much longer than people seem to think at the big labs.
While Clinical Trials are the most expensive step and have a 90% failure rate, it is not the *source* of the failures, it's just where failures are revealed.
Much of whether or not a candidate drug will succeed was determined all the way back in the design/discovery stage. If we can design better drugs, then we would enjoy a higher probability of success in the later clinical trials.
The single biggest improvement to pharma drug development would be from increasing the probability of success in clinical trials.
Is it fair to say the source of the failures is we simply do not understand the complex relationship between the drug design (chemistry, cell therapy, etc.) and the biology (the complete system of the living thing). If we could understand that relationship, we could design better drugs? I know this is very hand-wavy. Probably so much so as to not be a useful thing to say, but I'm just trying to understand.
Correct. The on-target effects of antivenin can be well-modeled along with a known subset of off-target effects, but the latter is far too complex and unknown to model completely.
Trials are the acceptable middle ground that uses animal and human models to explore those unknowns.
Some trials fail because we didn't have a "better" drug for a given target, many fail because we don't know the right biology. Biology and knowing precisely how to modulate the right thing is more gating than "design a better compound".
Better for what is the question and while sometimes the answer is "solubility, metabolism, off target toxicity", more often it's "you misunderstood something about the biology of this disease".
Maybe you needed more potency than your assays said you needed.
Maybe you missed something else the target does.
Maybe you had the right effect on the target but the body found a way to compensate (in an undesirable way).
Maybe the target looks like it matters in all sorts of model systems, but in real life it doesn't matter at all!
Many such cases. AI might marginally help in these scenarios, but doesn't look like it will eliminate the problem. AI can't help us with what we fundamentally don't know and have zero data for, or for which data is extremely hard to get.
The impression I have, which admittedly is purely as a layman reading articles, is that by far the rate-limiting factor in all drug creation is the cost of doing clinical trials in humans.
my general question with any drug isn't its effectiveness. It is what are the side effects? Unless somehow AI is solving this without clinical trials, there isn't much of note here
It orchestrated available open source tools to de novo generate and assess the designs. It is still really impressive that it has enough theory to use the tools correctly, but it isn't generating the residues from scratch within the model itself.
I got Opus5 to write me a enzyme design pipeline on AWS last month using available tools and databases. It works fine. It's not as handsoff as you might think when you start looking at more complex proteins with multiple binding pockets and sequential catalytic process within multimeric structures. The bottleneck is still actually testing the candidates out in the lab.
This is somewhat funny given that not that long ago they were doing an article instigating fears of what Claude could do, and one of them was a researchers screening venom peptides along other things
Perhaps he could have worked on a similar subject? Who knows.
This sounds awesome but I wish these would be published in a peer-reviewed journal or something of that nature. I can't believe these press releases they put out, even though I'm sure it's an achievement.
They're trying to catch up to OpenAI which has been doing these a lot, but I agree, can we actually get them peer reviewed as a hard requirement? I think OpenAI would have some looked over by experts before posting their blog posts.
How do you determine which one's ahead in what? OpenAI isn't just known for their models being used in mathematical discoveries, they've also been used in other areas as well? Anthropic on the other hand has not had anywhere near the amount of news coverage as OpenAI gets for their discoveries / efforts.
If I were at Anthropic, I'd be making sure the higher ups highlight discoveries made using the models they make, because OpenAI is definitely highlighting those.
"Our prompt specified the targets, the design sets and their sizes, the rules for each design, the protein design and prediction tools, the scores and ranking formulas, and the novelty and diversity rules."
I think most scientists would consider LLM to be an orchestration tool in cases like this. LLMs do work great for automating this type of drug discovery workflow! You definitely don't need the latest and greatest model, either, as long as you're careful in your specification. But attributing the designs to the tool (Claude), not the scientists who came up with the workflow, is distasteful and misleading.
Neat, but doesn't change anything about the fact that even with this protein design headstart nobody is going to pay for the human trials or the cost to produce and distribute the possible antivenom.
In practical terms this is like Claude designing you a new home. Cool, you now have blueprints for a new home. You can print it out and hang it on the wall of your studio apartment.
I don't see scientists reacting to this with the same affront as Terrance Tao does for frontier LLM mathematical proof dumping. Just imagine the Salk Institute declaring "we are on a war footing" after learning of such LLM aided discoveries.
I did this at home with a 5090 on a different target a few weeks ago. Cannot generate 318,000 designs but Claude chose the nearly same tooling, except I used BindCraft2 which is better for multi target designs in the beginning. This isn't that special, its just doing tool calls. I got 100 or so trajectories running 24hrs just for fun.
"Throughout the process, Claude generated 318,484 protein designs and selected 180 of them for testing. Our team added 31 more, with Claude’s help."
"lab-tested all 211 designs to see if they would bind to the toxins they were designed against, and 17 designs met our criteria for a binder. "
Alright how many of those 17 were 100% Claude, I know mine were. These mini-proteins the tooling I'm told is pretty good for these. I'm worried that Claude might just choose the same tooling for everyone if this is the path we go down and designs stagnate?
I hope this produces some useful anti-venom that is cheaper and widely available. My kid was stung by a small bark scorpion on his thumb here in Arizona and I would not wish that experience on anyone. For hours afterwards he was in excruciating screaming pain, and he even lost his vision for while. That lasted about 24 hours and then it took weeks for him to recover feeling in his hand. Nothing seemed to help with the pain and there was very little we could do for him. I would love to have had a more available anti-venom, but apparently now it is so expensive they only keep it around for the very worst cases.
> but apparently now it is so expensive
They saw Shkreli hike prices 100x and decided that was the business model for them. Same thing has happened with rabies.
A single shot of rabies costs about ~$3 in Asian countries where street dogs are common. It’s crazy how expensive post exposure rabies treatment is in the US.
Unfortunately, available does not mean cheap these days. Parents (good ones anyway) would go into a decade of debt to prevent their child from having an experience like you just described. Drug giants see that as a business opportunity.
This might be a job for fentanyl. Carefully monitored in a hospital setting of course.
Leave it to the internet to come up with a new application for a recreational drug.
https://en.wikipedia.org/wiki/Fentanyl
Approved for medical use in 1968
Fentanyl is widely used in surgery
Had to get fentanyl during a gall bladder issue, only thing that stopped the screaming
Can someone explain whether design of protein binders is an important rate-limiting step in drug creation (of which antivenom is a very small market)? Or AI-driven drug discovery in general? It feels like something that will definitely help the drug market but is nowhere near the rate-limiting step, but that's from my own researcher as someone without a background in it.
My understanding is that FDA trials (specifically phase 2 + 3) are the death valley for most drugs and finding a biomarker is something mostly done through a mix of gene bank scraping + looking at old biomarker databanks. The majority of the cost and time is in failed drugs that do not pass FDA trials (which average 6-7 years end-to-end).
Note: I am asking this in earnest. I am not trying to ask a leading question and saying it is not important at all.
My understanding is it is not - this is something we have models for, and there was a flurry of machine learning based drug discovery companies about 8-10 years ago. But they ran into the testing bottleneck real fast because they would spit out thousands of candidates for binders and so on that all needed to be tested in vitro before they could even start the real drug dev circuit. Furthermore the drugs need to be manufacturable and that's not a trivial problem.
It's nice to (as the blog post suggests) give the opportunity to non medical experts, say in this case a snake/venom research center, to pursue a problem and produce some realistic candidates that they could pay to test. I do think it's good to democratize what parts of the process can be, but it's more of a resource prioritization problem than a molecular design one.
Have a comment below on this, but generally things don't work out in the clinic because we didn't understand the biology adequately, and not because we didn't have a molecule.
A better binder only helps if you know what to bind to and that if you bind to it, you'll have the desired effect.
This is why the industry has crowded around a very limited set of "validated" targets. It's because we know if we do something with them, they will work. Obviously we're likely missing a lot of great ideas though.
AI might help find some of those ideas, but we won't know until someone tests them in a human. Essentially, without more extremely expensive and difficult data to acquire, it's not likely to be that much better at picking winners.
I'm in the sector and I believe strongly in AI helping (I've bet my entire career on it). It will get there eventually, but it will take much longer than people seem to think at the big labs.
On crowding in big pharma: https://www.lek.com/insights/life-sciences-pharma/biopharma-...
While Clinical Trials are the most expensive step and have a 90% failure rate, it is not the *source* of the failures, it's just where failures are revealed.
Much of whether or not a candidate drug will succeed was determined all the way back in the design/discovery stage. If we can design better drugs, then we would enjoy a higher probability of success in the later clinical trials.
The single biggest improvement to pharma drug development would be from increasing the probability of success in clinical trials.
Is it fair to say the source of the failures is we simply do not understand the complex relationship between the drug design (chemistry, cell therapy, etc.) and the biology (the complete system of the living thing). If we could understand that relationship, we could design better drugs? I know this is very hand-wavy. Probably so much so as to not be a useful thing to say, but I'm just trying to understand.
Not necessarily. It is often due to a lack of choices (in lead candidates).
Correct. The on-target effects of antivenin can be well-modeled along with a known subset of off-target effects, but the latter is far too complex and unknown to model completely.
Trials are the acceptable middle ground that uses animal and human models to explore those unknowns.
Some trials fail because we didn't have a "better" drug for a given target, many fail because we don't know the right biology. Biology and knowing precisely how to modulate the right thing is more gating than "design a better compound".
Better for what is the question and while sometimes the answer is "solubility, metabolism, off target toxicity", more often it's "you misunderstood something about the biology of this disease".
Maybe you needed more potency than your assays said you needed.
Maybe you missed something else the target does.
Maybe you had the right effect on the target but the body found a way to compensate (in an undesirable way).
Maybe the target looks like it matters in all sorts of model systems, but in real life it doesn't matter at all!
Many such cases. AI might marginally help in these scenarios, but doesn't look like it will eliminate the problem. AI can't help us with what we fundamentally don't know and have zero data for, or for which data is extremely hard to get.
The impression I have, which admittedly is purely as a layman reading articles, is that by far the rate-limiting factor in all drug creation is the cost of doing clinical trials in humans.
my general question with any drug isn't its effectiveness. It is what are the side effects? Unless somehow AI is solving this without clinical trials, there isn't much of note here
Your impression is correct. Clinical trials are the rate limiting step and the primary source of cost and risk.
There would be less failures if problematic therapies were filtered out before the clinical stage.
It orchestrated available open source tools to de novo generate and assess the designs. It is still really impressive that it has enough theory to use the tools correctly, but it isn't generating the residues from scratch within the model itself.
I got Opus5 to write me a enzyme design pipeline on AWS last month using available tools and databases. It works fine. It's not as handsoff as you might think when you start looking at more complex proteins with multiple binding pockets and sequential catalytic process within multimeric structures. The bottleneck is still actually testing the candidates out in the lab.
This is somewhat funny given that not that long ago they were doing an article instigating fears of what Claude could do, and one of them was a researchers screening venom peptides along other things Perhaps he could have worked on a similar subject? Who knows.
Does it cost more than some generic benedril? Otherwise no biggie
This sounds awesome but I wish these would be published in a peer-reviewed journal or something of that nature. I can't believe these press releases they put out, even though I'm sure it's an achievement.
They're trying to catch up to OpenAI which has been doing these a lot, but I agree, can we actually get them peer reviewed as a hard requirement? I think OpenAI would have some looked over by experts before posting their blog posts.
>> They're trying to catch up to OpenAI which has been doing these a lot
Sorry, what? OpenAI is "ahead" in math. Anthropic is "ahead" in life sciences, and even have their own wet lab.
How do you determine which one's ahead in what? OpenAI isn't just known for their models being used in mathematical discoveries, they've also been used in other areas as well? Anthropic on the other hand has not had anywhere near the amount of news coverage as OpenAI gets for their discoveries / efforts.
If I were at Anthropic, I'd be making sure the higher ups highlight discoveries made using the models they make, because OpenAI is definitely highlighting those.
>> How do you determine which one's ahead in what?
We are talking about discoveries the labs have publicly showcased.
From the technical report:
"Our prompt specified the targets, the design sets and their sizes, the rules for each design, the protein design and prediction tools, the scores and ranking formulas, and the novelty and diversity rules."
I think most scientists would consider LLM to be an orchestration tool in cases like this. LLMs do work great for automating this type of drug discovery workflow! You definitely don't need the latest and greatest model, either, as long as you're careful in your specification. But attributing the designs to the tool (Claude), not the scientists who came up with the workflow, is distasteful and misleading.
So it’s ok for Anthropic to make new biologics … but it’s users cannot because it might end humanity.
Let alone the fact this doesn’t even appear to be peer reviewed either from Anthropic.
Yes, they _hope_ they can verify that the outcome of giving someone they trust a gun will not end in the end of the world.
I wouldn't trust you with a gun of mine... I don't know you.
And what makes you trust them, with the “gun”?
It's less likely to have an omnicidal maniac in a group of 100 people than in a group of a few billion people.
That assumes you have no prior knowledge of either group.
I hate how the role of the human scientists involved is so diminished, with credit being given to "Claude," which is at best a family of products.
Neat, but doesn't change anything about the fact that even with this protein design headstart nobody is going to pay for the human trials or the cost to produce and distribute the possible antivenom.
In practical terms this is like Claude designing you a new home. Cool, you now have blueprints for a new home. You can print it out and hang it on the wall of your studio apartment.
I don't see scientists reacting to this with the same affront as Terrance Tao does for frontier LLM mathematical proof dumping. Just imagine the Salk Institute declaring "we are on a war footing" after learning of such LLM aided discoveries.
Is the CSS on the page all messed up for anyone else or just me?
https://i.imgur.com/fRVxjzq.png