AlphaFold3: Revolutionizing drug discovery and development

Newsletter Signup - Under Article / In Page

"*" indicates required fields

Building on the foundations of its hugely successful AlphaFold2, which made a fundamental breakthrough in protein structure prediction in 2020, Google DeepMind and its sister company Isomorphic Labs launched AlphaFold3 in May. After controversially withholding the code for six months, DeepMind finally – and somewhat unexpectedly – open-sourced its platform early last week, meaning academic researchers all over the world can access both the code and training weights for the first time for non-commercial applications. This decision marks a significant advance that could massively accelerate scientific discovery and drug development. So, with the already unavoidable acclaim for AlphaFold3, can the platform really usher in a new era for drug discovery and molecular biology?

What is AlphaFold3?

First, let us start with the basics by looking at what AlphaFold actually is.

DeepMind’s AlphaFold technology was built to help alleviate one of the key challenges that researchers encounter in the scientific process – the ability to determine the structure and sequence of proteins. AlphaFold is essentially designed to unravel the mystery of protein folding, in turn unlocking potential new breakthroughs in science.

AlphaFold2 was seen as revolutionary because it was able to accurately predict the structure of the majority of proteins from their DNA sequence. The platform ended up predicting structures for all 200 million proteins with known DNA sequences, which DeepMind made freely accessible to scientists in a massive database. Before this advancement, only around 100,000 proteins had known structural information.

As the most recent model, however, AlphaFold3 takes protein prediction a step further than its predecessor.

“Building on the remarkable achievements of its predecessors, the new AF3 algorithm expands its capabilities beyond individual protein structure prediction to deliver detailed and accurate insights into complex molecular interactions,” Dr Yvonne Tan, associate director of Product, at Nuclera, told Labiotech. “One of its most exciting advancements is its ability to predict protein-protein interactions with high precision, offering researchers a deeper understanding of how proteins form complexes and carry out their biological functions. AlphaFold3 extends its prediction capabilities to protein-DNA, protein-RNA interactions, protein-ligand, and protein-small molecule interactions.”

As DeepMind and Isomorphic Labs said in their joint release about AlphaFold3, it’s only by seeing how proteins, DNA, and other molecules interact together, across millions of types of combinations, that we can “start to truly understand life’s processes.”

To learn how to puzzle out the precise atomic structures of molecules, AlphaFold3 uses a diffusion model, similar to ones used for popular text-to-image generation models such as OpenAI’s DALL-E 3. Ironically, despite covering far more substances than its predecessor, AlphaFold3 is actually a simpler design than AlphaFold2, with fewer separate components.

It’s also worth noting that, unlike some attempts to create large language models (LLMs) for biology that can be prompted in natural language to produce a formula for a compound with particular properties, AlphaFold3 still needs someone to have a relatively good understanding of biology in order to use it effectively. Furthermore, any suggested molecular structure would still need to be produced or isolated in a lab, which also requires some specialized knowledge.

How AlphaFold3 could revolutionize drug discovery

AlphaFold3’s capabilities are so unique that they essentially mark a shift in computational biology. Not only does it outperform traditional physics-based methods in predicting molecular interactions, but it also aligns with atomic-level physics, making it more efficient and reliable. By working directly with atomic coordinates (the data points that describe molecular structures), it can achieve higher accuracy in predicting interactions.

According to data from DeepMind and Isomorphic Labs, AlphaFold3 is 50% more accurate than the best traditional methods on the PoseBusters benchmark without needing the input of any structural information, meaning that AlphaFold3 is the first artificial intelligence (AI) system to surpass physics-based tools for biomolecular structure prediction.

This makes AlphaFold3 the perfect tool for drug discovery, ultimately enabling scientists to simulate vital molecular interactions with unprecedented speed and precision, and offering a faster, more cost-effective alternative to traditional lab methods. As scientists use AlphaFold3 to better understand protein structures and interactions, they can also create better targets for medications, better understand side effects, and delve into new areas of protein-drug interactions that may not have previously been fathomable.

“The drug discovery industry is constantly evolving to explore innovative approaches that go beyond the traditional ‘one drug per target’ model,” said Tan. “Strategies such as disrupting preformed protein complexes or promoting protein-protein interactions are gaining traction. A critical first step in this process is the ability to produce target proteins, often in pre-formed complexes, to study such interactions effectively.”

As an example, Tan said that cell and gene therapy is an area where AlphaFold3 could make a key difference. Here, early insights into protein-DNA and protein-RNA interactions are vital, and understanding these interactions enables researchers to design nucleases with enhanced binding and recognition capabilities, meaning that AlphaFold3 can pave the way for more effective therapeutic tools such as transcription factors and other DNA/RNA-binding targets.

AlphaFold3 is also particularly accurate at predicting interactions between proteins and small molecules. “By analyzing protein surface charges and predicting druggable pockets, AF3 delivers deep insights into the druggability of potential targets,” explained Tan. “This capability is instrumental in identifying and optimizing therapeutic candidates, significantly accelerating the discovery and development of next-generation treatments.”

Ultimately, AlphaFold3’s impact on drug discovery and development is expected to be substantial. And, while commercial restrictions currently limit pharmaceutical applications, DeepMind’s decision to make the platform open-source means that academic research will inevitably be able to advance scientists’ understanding of disease mechanisms and drug interactions.

Google DeepMind’s decision to open-source AlphaFold3

When AlphaFold3 was initially released in May, DeepMind’s decision to withhold the code while offering limited access through a web interface sparked an outcry from the research community, leading to a protest letter signed by more than 1,000 scientists.

The withholding of the code was largely down to commercial reasons, as co-developer Isomorphic Labs, a DeepMind spinoff company, is using AlphaFold3 internally to speed up its drug discovery efforts. The company currently has partnerships potentially worth $3 billion with Eli Lilly and Novartis aimed at developing multiple drugs.

The eventual open-source release, however, offered a middle path, attempting to satisfy both scientific and commercial needs. While the code is freely available under a Creative Commons license, access to the crucial model weights requires Google’s explicit permission for academic use.

AlphaFold3: The next frontier in AI-driven molecular biology and drug discovery

The sheer potential of DeepMind’s AlphaFold technology, reinforced by the awarding of the 2024 Nobel Prize in Chemistry to its developers Demis Hassabis and John Jumper, shows just how much of an impact AI can really have on molecular biology and overall drug development. In many cases, AI-based predictions like those provided by AlphaFold3 now surpass even the best physics-based models.

In a Forbes article, Jumper described the “AlphaFold story” as a significant milestone in science, especially with regard to how protein structure and folding are understood – what previously took years can now be completed in mere minutes with this AI-based technology.

Other companies are already working on developing AlphaFold3-inspired platforms. According to an article in Nature, Chinese technology giant Baidu and TikTok developer ByteDance have now rolled out their own models, as have San Francisco-based firms Chai Discovery and Ligo Biosciences. While these models come with certain limitations, including not being licensed for commercial use, some teams are also working on versions of AlphaFold3 that don’t come with such limits; for example, a group at Columbia University hopes to have a fully open-source model called OpenFold3 available by the end of the year that would enable drug companies to retrain their versions of the model using proprietary data.

All of this work on AI-based biological models creates an extremely promising future for the area of drug discovery, with AlphaFold3 very much at the forefront of this forward push.

If AlphaFold2 has already opened up new areas of biological research that we thought were near impossible, just imagine what AlphaFold3 could do in the capable hands of academic researchers all over the world; it’s fair to say that these certainly are exciting times for the world of molecular biology and drug discovery.