AlphaFold 3 Changes Biology Forever

Yesterday, Google DeepMind made a major announcement. In collaboration with its spin-off lab, Isomorphic Labs, the top-tier AI research laboratory unveiled the latest iteration of its protein predictor model – AlphaFold 3. And to cut right to the chase—it’s an absolute game-changer.

Building upon the legacy of its predecessors, AlphaFold 3 (AF3) not only extends our capability to predict protein structures but also “predicts the structure and interactions of all of life’s molecules.” In effect, AF3 is able to comprehensively model not just proteins but also the interactions of these proteins with other molecules like DNA, RNA, and ligands.

It’s a transformative breakthrough. And altogether, it promises to fundamentally reshape our approach to medicine, drug discovery, and bioengineering. So allow me to break it down for you—how AlphaFold has developed up to this point, the now-enhanced capabilities of AlphaFold 3, and what this will mean for the world of biology and medicine—because this is truly an enormous step forward for science, AI, and technological progress as a whole.

Revolutionary Predecessor

Before diving headfirst into AF3, it’s worth quickly covering how DeepMind got here. AlphaFold’s journey begins back in 2016. Shortly after DeepMind's success with AlphaGo, they turned their attention to solving one of biology's grand challenges: the protein-folding problem. AlphaFold was developed with the goal of predicting protein structures solely from amino acid sequences—a question that had long puzzled scientists and was fertile ground for the introduction of some artificial intelligence.

By leveraging massive datasets and advanced AI techniques, DeepMind trained AlphaFold on over 170,000 protein structures​​. AlphaFold then first demonstrated its potential at the CASP13 competition in 2018. There, it significantly outperformed other methods in predicting protein structures, marking a breakthrough in computational biology. But it still had a ways to go.

Then came AlphaFold 2 (AF2). Introduced in 2020, it achieved an unprecedented, revolutionary level of accuracy that shook the world of biology. Ultimately, it would prove instrumental in transforming biological research and pharmaceutical development.

Today, AF2’s predictive capabilities have been utilized by millions of researchers worldwide, leading to significant discoveries in areas like malaria vaccines, cancer treatments, and the design of enzymes critical for industrial and medical applications. The scientific community has embraced it - reflected in over 20,000 citations and multiple accolades, including the prestigious Breakthrough Prize in Life Sciences.

From AF2 to AF3: “Revolutionary” to Total Game-Changer

If AlphaFold 2 was already a groundbreaking, “revolutionary” model that made waves throughout the worlds of biological and medicinal research, and modeled the structures of “nearly every known protein,” how is it possible that the latest iteration, AlphaFold 3 is that much of a step up? By going even further.

What can AlphaFold 3 do? According to the (extensively-detailed) research paper DeepMind released at the time of their announcement, the new model “is capable of joint structure prediction of 31 complexes including proteins, nucleic acids, small molecules, ions, and modified residues.” In other words, AF3 not only predicts protein structures – it can also accurately predict how those proteins interact with other molecules, such as DNA, RNA, and small molecules known as ligands, which is a category that encompasses many drugs.

This is an absolute game-changer. As DeepMind writes in their blog post:

Inside every plant, animal and human cell are billions of molecular machines. They’re made up of proteins, DNA and other molecules, but no single piece works on its own. Only by seeing how they interact together, across millions of types of combinations, can we start to truly understand life’s processes.

And if you take away nothing else from this breakdown it’s exactly that—protein structure prediction was a great start, but the interactions between biomolecules are far, far more complicated and difficult to model. For example, given a molecule like ATP and a target protein, AF3 would not just be able to predict how the protein folds, but it would also predict where the ATP would bind to the protein. In a specific example included in their blog post, DeepMind shows how AF3 is able to predict exactly how 7R6R, a DNA binding protein, would bind to the double helix of DNA in a “near-perfect match to the true molecular structure discovered through painstaking experiment.”

So, how are they able to do so? And what opportunities does this open up?

AlphaFold 3 achieves its significant leap in capabilities from its predecessor through a major evolution of the AlphaFold 2 architecture and training methodologies in order to better handle a wider array of chemical structures and enhance the efficiency of the learning process. Central to this model is the evolved version of the “Evoformer” module, a deep learning framework crucial for AlphaFold 2’s high-performance protein structure predictions.

In AF3, this module has seen a major upgrade. The shift involves replacing the previous model's reliance on multiple sequence alignments (MSA) with what they term the  “Pairformer” module. This new component simplifies the processing of molecular interactions by focusing directly on pairs of elements within the molecule, rather than the full sequence. This has the additional effect of significantly reducing computational complexity and improving data processing efficiency.

Additionally, AF3 introduces a Diffusion Module – yes, a diffusion model! This new module simulates the diffusion process similar to techniques used in AI image generation, starting with a "cloud" of atomic coordinates and iteratively refining this input into a precise, detailed molecular structure. This allows the model to predict the final atomic arrangement with greater accuracy by progressively reducing the noise introduced at the initial stages, effectively learning the molecular structure at multiple scales—from the fine details of local atomic configurations to the broader arrangement of the molecule as a whole.

The researchers at DeepMind and Isomorphic Labs go into far more detail in their paper, which you can read for yourself if you’re interested. But suffice to say, the implementation of these new techniques enables AF3 to surpass the capabilities of all other existing systems out there in predicting molecular interactions. It computes molecular complexes holistically, integrating diverse types of biochemical data without the need for specific adjustments for different molecule types.

Such a unified approach not only enhances the precision of the model but also broadens its applicability across different fields of biomedical research. And this is the real kicker. Because at this point, many of you may be wondering what opportunities such a model opens up. What sort of research can be done with AlphaFold 3?

Unlocking the Future of AI-Powered Cell Biology

AF3 offers potential breakthroughs in understanding complex diseases, developing new drugs, designing novel enzymes, and much more. In drug design, AF3 can predict the behavior of molecules commonly used in pharmaceuticals, such as ligands and antibodies, revealing how these molecules interact with proteins to influence human health. Such precise understanding of molecular interactions helps to facilitate targeted drug design and can lead to more effective therapies with fewer side effects.

AlphaFold 3 also stands out for its unprecedented accuracy in modeling drug-like interactions. It achieves a 50% improvement over the best traditional methods on the PoseBusters benchmark, a level of precision that’s critical for designing antibodies that can precisely target and bind to specific proteins, or to explore novel therapeutic avenues that were previously too complex to tackle.

As I write this, Isomorphic Labs is already leveraging AlphaFold 3 alongside a suite of complementary AI models to expedite and refine drug design processes. This will enable a more nuanced approach to understanding things like disease mechanisms and tailoring treatments to specific pathological profiles. As a result, they aim to streamline the drug development pipeline and bring effective solutions to market more quickly.

Unfortunately, unlike its predecessors, AlphaFold 3 comes with restricted access, limiting the broader scientific community's ability to run personal versions of the model or to access the underlying code. Now, researchers can use the new AlphaFold Server to input protein sequences and receive structure predictions. But for now, they’re capped at 10 predictions per day and can’t model proteins in conjunction with potential drugs. This limitation is in part to protect the commercial interests of Isomorphic Labs.

Despite these restrictions, the AlphaFold Server should be a massive utility for many researchers. There’s also some optimism about the future possibility of open-source alternatives. Based on the detailed information provided in AlphaFold 3's publications, some experts anticipate the development of independent versions of the model by other teams before the year's end. An open source AF3 could unlock a new golden age of biological research – fingers crossed.

Accelerate the Biological Sciences

AlphaFold 3 is nothing short of a breakthrough, a herald of a new dawn in the fusion of artificial intelligence and the biological sciences. It’s not a stretch at all to say it will undoubtedly change the world. Novel therapeutics. Sustainable biotechnologies. Targeted drug design. AlphaFold 3 will not only accelerate the pace of medical breakthroughs but also fundamentally broaden the horizons of scientific discovery. And I, for one, can’t help but feel the techno-optimism in that, right down to my core. And you should, too.