"AlphaFold 3: Revolutionizing Molecular Structure and Interaction Prediction"

Since its launch in 2020, over 2 million researchers have turned to Google DeepMind's AlphaFold 2 to predict protein structures, aiding in breakthroughs like vaccine development and cancer treatments. This model tackled a challenge that had stumped scientists for over half a century. But the team at Google DeepMind didn't stop there; instead of resting on their laurels, they rolled up their sleeves and started working on AlphaFold 3.
Launched in May by Google DeepMind and Isomorphic Labs, AlphaFold 3 takes things up a notch. It doesn't just predict protein folding; it also forecasts the structure and interactions of all sorts of life's molecules, including DNA, RNA, and ligands—those tiny molecules that attach to proteins.
"With AlphaFold 2, we made huge strides in solving the protein folding puzzle, but the scientific community has moved on to more complex stuff," explains Jonas Adler, a research scientist at Google DeepMind. "Researchers are now diving into details like how small molecules bind, or how RNA works, areas where AlphaFold 2 fell short. To keep up with the latest in biology and chemistry, we needed a tool that could handle all kinds of biomolecules."
"Everything" here includes ligands, which are crucial because they make up about half of all drugs. "At Isomorphic Labs, we're seeing the huge potential of AlphaFold 3 for designing drugs rationally, and we're already putting it to use every day," says Adrian Stecula, a research leader at Isomorphic Labs. "We're exploring how new small molecules bind to new drug targets, figuring out how proteins interact with DNA and RNA, and studying how chemical tweaks affect protein structures—the new model opens up all these possibilities."
Incorporating these extra molecule types meant dealing with a ton more combinations. "Proteins are pretty straightforward; there are only 20 standard amino acids," Jonas notes. "But small molecules? They're all over the place, with endless possibilities. They're super diverse."
Realizing that a comprehensive database was out of the question, the team launched AlphaFold Server. This free tool lets scientists input their own sequences, and AlphaFold generates the molecular complexes for them. Since going live in May, it's been used to create over 1 million structures.
"It's like Google Maps but for molecular complexes," says Lindsay Willmore, a research engineer at Google DeepMind. "Even if you're not a coder, you can just copy and paste your protein, DNA, RNA, or small molecule sequences, hit a button, and wait a few minutes. You'll get your structure along with confidence metrics to help you evaluate the prediction."
To make AlphaFold 3 work with such a broad range of biomolecules, the team expanded the training data to include DNA, RNA, small molecules, and more. "We figured, 'Why not train on everything in our dataset that helped us with proteins and see how far we can go?'" Lindsay explains. "Turns out, we went pretty far."
A key change in AlphaFold 3 was in the final part of the model that creates the structure. While AlphaFold 2 used a complex custom geometry-based module, AlphaFold 3 switched to a simpler generative model based on diffusion, similar to other advanced image generation models like Imagen. This change streamlined how the model deals with the new types of molecules.
However, this shift brought a new challenge: the diffusion model would mistakenly create an "ordered" structure for protein regions that are actually "disordered"—think of it as trying to organize a pile of chaotic spaghetti into a neat spiral.
So, the team turned to AlphaFold 2, which excels at predicting these disordered interactions. "We used those predictions from AlphaFold 2 to train AlphaFold 3, teaching it to recognize and predict disorder," Lindsay says.
"We have a saying: 'Trust the fusilli, reject the spaghetti,'" adds Jonas with a chuckle.
An example of a prediction from AlphaFold 3, showing ordered “fusilli” regions in blue and disordered “spaghetti” regions in orange. The colors indicate the model's confidence in the prediction's accuracy.
The team is excited to see how AlphaFold 3 will be used in fields like genomics and drug design. "It's amazing to see how far we've come," Jonas says. "What was once tough is now easy, and what was impossible is now within reach. There are still tough nuts to crack, but we're thrilled about what AlphaFold 3 can help us achieve."
Related article
Billionaires Discuss Automating Jobs Away in This Week's AI Update
Hey everyone, welcome back to TechCrunch's AI newsletter! If you're not already subscribed, you can sign up here to get it delivered straight to your inbox every Wednesday.We took a little break last week, but for good reason—the AI news cycle was on fire, thanks in large part to the sudden surge of
NotebookLM App Launches: AI-Powered Tool for Instant Knowledge Access Anywhere
NotebookLM Goes Mobile: Your AI-Powered Research Assistant Now on Android & iOSWe’ve been blown away by the response to NotebookLM—millions of users have embraced it as their go-to
Google’s AI Futures Fund may have to tread carefully
Google’s New AI Investment Initiative: A Strategic Shift Amid Regulatory ScrutinyGoogle's recent announcement of an AI Futures Fund marks a bold move in the tech giant's ongoing qu
Comments (15)
0/200
GeorgeMartinez
April 10, 2025 at 12:00:00 AM GMT
AlphaFold 3 is mind-blowing! It's incredible how it's helping with vaccine development and cancer treatments. I'm not a scientist, but even I can see the potential here. The only thing is, it's a bit complex for non-experts. Still, a huge leap forward!
0
RoySmith
April 10, 2025 at 12:00:00 AM GMT
AlphaFold 3は驚くべきツールです!ワクチン開発やがん治療に役立っているなんて信じられません。私は科学者ではありませんが、その可能性が見て取れます。ただ、専門家以外には少し複雑かもしれません。それでも、大きな進歩ですね!
0
WilliamCarter
April 10, 2025 at 12:00:00 AM GMT
AlphaFold 3은 정말 놀랍습니다! 백신 개발과 암 치료에 도움이 된다는 것이 믿기지 않아요. 저는 과학자가 아니지만, 그 가능성을 느낄 수 있어요. 다만, 전문가가 아닌 사람에게는 조금 복잡할 수 있어요. 그래도 큰 도약입니다!
0
JustinWilliams
April 10, 2025 at 12:00:00 AM GMT
AlphaFold 3 é impressionante! É incrível como está ajudando no desenvolvimento de vacinas e tratamentos contra o câncer. Não sou cientista, mas até eu posso ver o potencial aqui. A única coisa é que é um pouco complexo para não especialistas. Ainda assim, um grande avanço!
0
WalterThomas
April 10, 2025 at 12:00:00 AM GMT
¡AlphaFold 3 es alucinante! Es increíble cómo está ayudando en el desarrollo de vacunas y tratamientos contra el cáncer. No soy científico, pero incluso yo puedo ver el potencial aquí. Lo único es que es un poco complejo para los no expertos. Aún así, un gran avance!
0
WillLopez
April 11, 2025 at 12:00:00 AM GMT
AlphaFold 3 is a lifesaver for us in the lab! It's mind-blowing how it predicts molecular structures so accurately. Helped us speed up our research big time. Only wish it was a bit faster, but hey, can't complain too much about a tool that's changing the game!
0
Since its launch in 2020, over 2 million researchers have turned to Google DeepMind's AlphaFold 2 to predict protein structures, aiding in breakthroughs like vaccine development and cancer treatments. This model tackled a challenge that had stumped scientists for over half a century. But the team at Google DeepMind didn't stop there; instead of resting on their laurels, they rolled up their sleeves and started working on AlphaFold 3.
Launched in May by Google DeepMind and Isomorphic Labs, AlphaFold 3 takes things up a notch. It doesn't just predict protein folding; it also forecasts the structure and interactions of all sorts of life's molecules, including DNA, RNA, and ligands—those tiny molecules that attach to proteins.
"With AlphaFold 2, we made huge strides in solving the protein folding puzzle, but the scientific community has moved on to more complex stuff," explains Jonas Adler, a research scientist at Google DeepMind. "Researchers are now diving into details like how small molecules bind, or how RNA works, areas where AlphaFold 2 fell short. To keep up with the latest in biology and chemistry, we needed a tool that could handle all kinds of biomolecules."
"Everything" here includes ligands, which are crucial because they make up about half of all drugs. "At Isomorphic Labs, we're seeing the huge potential of AlphaFold 3 for designing drugs rationally, and we're already putting it to use every day," says Adrian Stecula, a research leader at Isomorphic Labs. "We're exploring how new small molecules bind to new drug targets, figuring out how proteins interact with DNA and RNA, and studying how chemical tweaks affect protein structures—the new model opens up all these possibilities."
Incorporating these extra molecule types meant dealing with a ton more combinations. "Proteins are pretty straightforward; there are only 20 standard amino acids," Jonas notes. "But small molecules? They're all over the place, with endless possibilities. They're super diverse."
"It's like Google Maps but for molecular complexes," says Lindsay Willmore, a research engineer at Google DeepMind. "Even if you're not a coder, you can just copy and paste your protein, DNA, RNA, or small molecule sequences, hit a button, and wait a few minutes. You'll get your structure along with confidence metrics to help you evaluate the prediction."
To make AlphaFold 3 work with such a broad range of biomolecules, the team expanded the training data to include DNA, RNA, small molecules, and more. "We figured, 'Why not train on everything in our dataset that helped us with proteins and see how far we can go?'" Lindsay explains. "Turns out, we went pretty far."
A key change in AlphaFold 3 was in the final part of the model that creates the structure. While AlphaFold 2 used a complex custom geometry-based module, AlphaFold 3 switched to a simpler generative model based on diffusion, similar to other advanced image generation models like Imagen. This change streamlined how the model deals with the new types of molecules.
However, this shift brought a new challenge: the diffusion model would mistakenly create an "ordered" structure for protein regions that are actually "disordered"—think of it as trying to organize a pile of chaotic spaghetti into a neat spiral.
So, the team turned to AlphaFold 2, which excels at predicting these disordered interactions. "We used those predictions from AlphaFold 2 to train AlphaFold 3, teaching it to recognize and predict disorder," Lindsay says.
"We have a saying: 'Trust the fusilli, reject the spaghetti,'" adds Jonas with a chuckle.
The team is excited to see how AlphaFold 3 will be used in fields like genomics and drug design. "It's amazing to see how far we've come," Jonas says. "What was once tough is now easy, and what was impossible is now within reach. There are still tough nuts to crack, but we're thrilled about what AlphaFold 3 can help us achieve."



AlphaFold 3 is mind-blowing! It's incredible how it's helping with vaccine development and cancer treatments. I'm not a scientist, but even I can see the potential here. The only thing is, it's a bit complex for non-experts. Still, a huge leap forward!




AlphaFold 3は驚くべきツールです!ワクチン開発やがん治療に役立っているなんて信じられません。私は科学者ではありませんが、その可能性が見て取れます。ただ、専門家以外には少し複雑かもしれません。それでも、大きな進歩ですね!




AlphaFold 3은 정말 놀랍습니다! 백신 개발과 암 치료에 도움이 된다는 것이 믿기지 않아요. 저는 과학자가 아니지만, 그 가능성을 느낄 수 있어요. 다만, 전문가가 아닌 사람에게는 조금 복잡할 수 있어요. 그래도 큰 도약입니다!




AlphaFold 3 é impressionante! É incrível como está ajudando no desenvolvimento de vacinas e tratamentos contra o câncer. Não sou cientista, mas até eu posso ver o potencial aqui. A única coisa é que é um pouco complexo para não especialistas. Ainda assim, um grande avanço!




¡AlphaFold 3 es alucinante! Es increíble cómo está ayudando en el desarrollo de vacunas y tratamientos contra el cáncer. No soy científico, pero incluso yo puedo ver el potencial aquí. Lo único es que es un poco complejo para los no expertos. Aún así, un gran avance!




AlphaFold 3 is a lifesaver for us in the lab! It's mind-blowing how it predicts molecular structures so accurately. Helped us speed up our research big time. Only wish it was a bit faster, but hey, can't complain too much about a tool that's changing the game!












