>DeepMind started developing AlphaFold in 2018. In 2020, it was recognized as a solution to humanity's 50-year-old "protein folding problem," which sought to answer how amino acids automatically fold into complex 3D shapes.
This is not even close to reality. AlphaFold is not a solution to the protein folding problem, it's (useful) pattern matching to the end state of solved folded protein (in situations that can be pattern matched). If this is considered "solving the protein folding problem" then X-ray crystallography solved it first, 75 years ago.
There is essentially zero "how" information coming to us from AlphaFold. This type of reporting is incorrect, and irresponsible to the folks that are still working on that how problem.
Pissed me off since day one that got by with that "solving" shit. Not to downplay what they did, but it had little to do with the problem as it's been understood to mean.
Yes, they have a probabilistic model, not an analytical solution.
OTOH I don't think they made up the "solved" label - the CASP benchmark had always considered reaching the AlphaFold level of predictive accuracy (within error tolerance of our ability to experimentally verify via X-ray crystallography etc) as "solving" the problem.
For what? I thought your trying to argue only baker deserved solo compared to what happened?
If you're asking about solving the actual folding problem? Probably Dill, Onuchic, Wolynes, maybe Sosnick would be top of my list off the top of my head. Anfinsen if he was alive.
The kind of solution you want may not be possible.
There is no guarantee that a tractable method exists to analytically invert protein folding. Data-driven methods like alphafold may be the only option.
AlphaFold 1 and 2 are open, and freely licensed, weights, and they're still online and downloadable. Alphafold 3 is open weight.
I'm guessing Google can't make money on them.
You can do more with the weights than just the protein structures. You can design new proteins. So no, the database is not enough. But the weights and a university cluster are enough.
Nope. Alphafold 1 & 2 can be used for whatever you want. Also this has resulted in them being used to train better models than Alphafold 3.
> And Isomorphic are their commercial arm, and they partner with the big drug companies.
Yep. And given how Pharma works, being the "descendants" of Nobel-prize winning research I'm sad to say this will probably work and pay pretty well (not by Google standards, but by anyone else's standard, pretty well), despite the fact that there are better models.
It's probably because they hit a dead end with their approach in terms of improvements. You can only get so far with trying to model a physical system with an insane number of degrees of freedom from simulation (and augmented) data.
John Jumper and the rest of the AlphaFold core team left for Anthropic a while ago. And of course Hassabis is handling larger AI efforts at Google. So it was only a matter of time before the effort fell apart.
As per the article they're slowly shuttering the project, but I assume AlphaFold itself is still available and usable?
I assume that AlphaFold isn't perfect, but surely after so many years on it most of the useful juice had been squeezed in terms of making it a useful tool?
Guessing they won't take any of the data or code offline ?
Given we have heard a lot about limited compute available at Google internally to employees I would guess they are freeing up resources wherever they can
The dedicated research team is getting moved to other areas, the Alphafold system itself and its database will likely continue to get maintained, but this means there probably won't be an "Alphafold 4".
The obvious question to any business leader is "why?" Why deploy resources to a project? Is this project central to our current or future revenue streams? If not, toss it out.
It has justifiably generated a lot of good will for Google. Applying AI to help science rather than to replace people's jobs. I think this is why you are now seeing OpenAI and Anthropic try to do more science work, but in their case, coming so late, it seems a cynical PR move as opposed to heartfelt from Hassabis.
It seems that Isomorphic Labs is somewhat a continuation of the AlphaFold work, and the first attempt to commercialize it, so maybe they will get more than just good will out of it.
> DeepMind started developing AlphaFold in 2018. In 2020, it was recognized as a solution to humanity's 50-year-old "protein folding problem," which sought to answer how amino acids automatically fold into complex 3D shapes. Those shapes determine the biological role of a protein. Scientists had identified the structures of roughly 170,000 proteins over the past 50 years, using tools and techniques like X-ray and nuclear magnetic resonance. The AlphaFold team took information from those previous work and then fed it to their AI to train AlphaFold.
> In 2021, Nature published the papers with AlphaFold's methodology and the structure predictions of the entire human proteome, or the complete set of proteins expressed by our species. DeepMind then launched the AlphaFold Protein Structure Database, giving researchers free access to over 200 million protein structure predictions.
Why is there no information on corroborations of the predictions? Anyone can make predictions. Surely there must have been teams picking predicted structures out of the database and comparing them to actual molecules?
I know you're saying this tongue in cheek. But the reason they're shuttering Alphafold is (likely) to assign those engineers and their expertise to Generative AI initiatives like Gemini.
>DeepMind started developing AlphaFold in 2018. In 2020, it was recognized as a solution to humanity's 50-year-old "protein folding problem," which sought to answer how amino acids automatically fold into complex 3D shapes.
This is not even close to reality. AlphaFold is not a solution to the protein folding problem, it's (useful) pattern matching to the end state of solved folded protein (in situations that can be pattern matched). If this is considered "solving the protein folding problem" then X-ray crystallography solved it first, 75 years ago.
There is essentially zero "how" information coming to us from AlphaFold. This type of reporting is incorrect, and irresponsible to the folks that are still working on that how problem.
Pissed me off since day one that got by with that "solving" shit. Not to downplay what they did, but it had little to do with the problem as it's been understood to mean.
And to be clear I do think it was nobel worthy.
Yes, they have a probabilistic model, not an analytical solution.
OTOH I don't think they made up the "solved" label - the CASP benchmark had always considered reaching the AlphaFold level of predictive accuracy (within error tolerance of our ability to experimentally verify via X-ray crystallography etc) as "solving" the problem.
CASP is explicitly structure prediction. That's not the folding problem.
I do not think it was Nobel worthy for Hassam et. al.
David Baker is the GOAT of that field, it should have been awarded to him only.
I know baker's work quite well, no it shouldn't have
Can you name others that deserved it more?
For what? I thought your trying to argue only baker deserved solo compared to what happened?
If you're asking about solving the actual folding problem? Probably Dill, Onuchic, Wolynes, maybe Sosnick would be top of my list off the top of my head. Anfinsen if he was alive.
The kind of solution you want may not be possible.
There is no guarantee that a tractable method exists to analytically invert protein folding. Data-driven methods like alphafold may be the only option.
This has nothing to do with what I am talking about.
That has nothing to do with the fact that the article is inaccurate.
Tell that to the nobel prize committee.
The article is fine. AlphaFold is widely regarded as "solving" protein folding.
Gemini says it did not solve the physics of how a protein actually folds.
Yes and correct me if I'm wrong, but they never did deeper work on dynamics. I think that's very telling.
they're probably referring to CASP13/14
It's an NP hard problem.
To SOLVE it would easily win the Nobel, possibly the Turing and more.
Designer proteins would change the world and medicine also in unimaginable ways.
> DeepMind then launched the AlphaFold Protein Structure Database, giving researchers free access to over 200 million protein structure predictions.
Is this existing database enough for researchers?
AlphaFold 1 and 2 are open, and freely licensed, weights, and they're still online and downloadable. Alphafold 3 is open weight.
I'm guessing Google can't make money on them.
You can do more with the weights than just the protein structures. You can design new proteins. So no, the database is not enough. But the weights and a university cluster are enough.
> I'm guessing Google can't make money on them.
The weights are free for non-commercial use. And Isomorphic are their commercial arm, and they partner with the big drug companies.
Edit: after reading the article, I realise it's about the disbanding of the team, so probably your original comment is correct.
> The weights are free for non-commercial use.
Nope. Alphafold 1 & 2 can be used for whatever you want. Also this has resulted in them being used to train better models than Alphafold 3.
> And Isomorphic are their commercial arm, and they partner with the big drug companies.
Yep. And given how Pharma works, being the "descendants" of Nobel-prize winning research I'm sad to say this will probably work and pay pretty well (not by Google standards, but by anyone else's standard, pretty well), despite the fact that there are better models.
To learn more about AlphaFold and its importance, I would recommend the Veritasium video[0]: AlphaFold - The Most Useful Thing AI Has Ever Done
[0]: https://www.youtube.com/watch?v=P_fHJIYENdI
It's probably because they hit a dead end with their approach in terms of improvements. You can only get so far with trying to model a physical system with an insane number of degrees of freedom from simulation (and augmented) data.
More likely because the entire Google organization is refocusing on LLMs.
Also they already got their nobel prize, so the marketing benefit is already maxed out.
They are cracking on with related biology related simulation. Some of it is now proprietary though in Isomorphic Labs.
John Jumper and the rest of the AlphaFold core team left for Anthropic a while ago. And of course Hassabis is handling larger AI efforts at Google. So it was only a matter of time before the effort fell apart.
As per the article they're slowly shuttering the project, but I assume AlphaFold itself is still available and usable?
I assume that AlphaFold isn't perfect, but surely after so many years on it most of the useful juice had been squeezed in terms of making it a useful tool?
google-deepmind/alphafold3: AlphaFold 3 inference pipeline https://github.com/google-deepmind/alphafold3
And then host then database.
AlphaFold did not yet solve the Folding@home protein folding problem.
Folding@home: https://en.wikipedia.org/wiki/Folding@home
Guessing they won't take any of the data or code offline ?
Given we have heard a lot about limited compute available at Google internally to employees I would guess they are freeing up resources wherever they can
The dedicated research team is getting moved to other areas, the Alphafold system itself and its database will likely continue to get maintained, but this means there probably won't be an "Alphafold 4".
Seems like https://alphafoldserver.com/ still works. This is about the team working on the product.
The obvious question to any business leader is "why?" Why deploy resources to a project? Is this project central to our current or future revenue streams? If not, toss it out.
It has justifiably generated a lot of good will for Google. Applying AI to help science rather than to replace people's jobs. I think this is why you are now seeing OpenAI and Anthropic try to do more science work, but in their case, coming so late, it seems a cynical PR move as opposed to heartfelt from Hassabis.
It seems that Isomorphic Labs is somewhat a continuation of the AlphaFold work, and the first attempt to commercialize it, so maybe they will get more than just good will out of it.
Is there a lot more work to be done?
There's a fair amount
Earlier: https://news.ycombinator.com/item?id=49096841
guess we need all that compute for AI?
False title. The shutdown is of the AlphaFold team.
Google REALLY wants Gemini 4 to be the leading LLM.
They need to justify the capex
> DeepMind started developing AlphaFold in 2018. In 2020, it was recognized as a solution to humanity's 50-year-old "protein folding problem," which sought to answer how amino acids automatically fold into complex 3D shapes. Those shapes determine the biological role of a protein. Scientists had identified the structures of roughly 170,000 proteins over the past 50 years, using tools and techniques like X-ray and nuclear magnetic resonance. The AlphaFold team took information from those previous work and then fed it to their AI to train AlphaFold.
> In 2021, Nature published the papers with AlphaFold's methodology and the structure predictions of the entire human proteome, or the complete set of proteins expressed by our species. DeepMind then launched the AlphaFold Protein Structure Database, giving researchers free access to over 200 million protein structure predictions.
Why is there no information on corroborations of the predictions? Anyone can make predictions. Surely there must have been teams picking predicted structures out of the database and comparing them to actual molecules?
There's plenty of information
https://en.wikipedia.org/wiki/CASP
They realised it's time... Gemini to follow?
I know you're saying this tongue in cheek. But the reason they're shuttering Alphafold is (likely) to assign those engineers and their expertise to Generative AI initiatives like Gemini.
AlphaFold found to not be actually prize winning.