[Reading Science] Nobel Prize Technology Released for Free... Could It Be a Game Changer in New Drug Development?

Nobel Chemistry Prize AlphaFold3 Released as Open Source

Google DeepMind's AI-based protein structure prediction model, 'Alphafold 3,' has been released for free, drawing attention to whether it can open a new era in drug development. Although limited to non-commercial academic use, the technology that won the Nobel Prize in Chemistry being made freely available has raised expectations for a new path toward conquering diseases.

Google DeepMind recently announced the full release of 'Alphafold 3.' Now, scientists can download the Alphafold 3 source code, install it on their own computers, and use it for research. This means anyone can use Alphafold AI for protein research for academic purposes.

Unlike the previous 'Alphafold 2,' whose source code could be downloaded and used, Alphafold 3, announced in May, was only accessible through Google's 'Alphafold server.' The types and number of proteins were limited, restricting new protein combination experiments or data input. Google also did not release the core model weights of Alphafold, which contain information about each part of the Alphafold model. Some speculated that Google imposed these restrictions to monopolize Alphafold 3 for research related to themselves. This led to calls for a full release.

This free release was made for non-profit academic research. However, interest from academia and the pharmaceutical industry is also focused on whether commercial use will be possible.

Demis Hassabis, CEO of Google DeepMind, and John Jumper, a researcher who shared the Nobel Prize in Chemistry, said in the scientific journal Nature, "We are very excited to see what people will do with this."

Protein structures, which form the basis of life, are crucial foundational data for drug development, but experimentally determining the 3D structure of proteins has traditionally required long time and high costs. Google DeepMind succeeded in enabling protein structure prediction through AI, and Alphafold has evolved through versions 1, 2, and 3. Notably, Alphafold 3 improved prediction accuracy by more than 50% compared to previous versions and can predict interactions of various biomolecules such as DNA and RNA.

◇ Free to use... Restrictions for pharmaceutical companies = Alphafold 3 is also regarded as presenting new possibilities in computational biology. It can be applied in various fields such as drug development, biotechnology, and agriculture. The source code of Alphafold 3 is freely available for non-commercial use, allowing researchers to access it freely. This means anyone can use it for non-profit academic research. However, the weight model is only available upon request. Korean researchers can also utilize this AI model to study protein structures and various biomolecular interactions.

While academic use is permitted, pharmaceutical companies aiming for commercial use must obtain permission through a separate license agreement with Google DeepMind. In other words, using Alphafold 3 for commercial purposes during drug development is restricted. Google DeepMind is reportedly reviewing requests for commercial licenses on a case-by-case basis.

Life science researchers can greatly improve the efficiency of protein research by leveraging the high prediction accuracy provided by Alphafold 3. Results that previously required long-term experiments can now be obtained accurately in a short time through AI, accelerating research to develop new treatments for various diseases such as genetic disorders, cancer, and infectious diseases. Expectations are especially high for drug development. Understanding the structure and interactions of target proteins precisely is essential in drug development. Alphafold 3 can revolutionize and accelerate this process by highly automating the prediction of complex biomolecular structures and binding sites, complementing traditional experimental research methods.

Other AI tools supporting drug development also exist. Besides Google DeepMind, several companies have released open-source protein structure prediction tools. Chinese search company Baidu introduced Helixfold, and ByteDance, the developer of TikTok, has also joined. The US-based Chai Discovery has entered the competition as well. Alphafold has sparked a wave of competition.

◇ Alphafold also requires high-performance GPUs = Efficient use of Alphafold 3 requires high-performance computing equipment. Alphafold 3 is a deep learning-based model. Predicting protein structures requires large-scale data and complex calculations. High-performance equipment is necessary to quickly and accurately predict the coordinates of each atom in a protein. In research environments where multiple proteins and biomolecular interactions must be predicted, high-performance GPUs are essential. Like most AI, Alphafold 3 also relies on GPU support.

Alphafold 2 could run on general research GPUs or NVIDIA 'V100' or higher GPUs, but Alphafold 3 is known to require 'A100' or higher GPUs, which are successors to the V100, for smooth operation. The A100 is currently the GPU model mainly used in domestic supercomputers.

Given the limited availability of expensive GPU resources in Korean university labs, even though Alphafold 3 is free to use, research may face constraints. Recently, Seoul National University's Bio-AI Research Group accelerated their research by acquiring 56 NVIDIA 'H100' GPUs, which are newer than the A100, a very rare case.