In just a few years, our small team was able to determine the structures of 200 million proteins, compared to the 150,000 structures revealed by tens of thousands of scientists over 50 years.

We released the protein structures to the world knowing that the scientific community would find countless applications for them, far beyond what we could imagine.

Seeing the map of the globe light up as researchers worldwide accessed the protein structures we released was a humbling and amazing experience.

We started with DNA because it is the fundamental information layer of life and where we have the most data available from the scientific community.

AlphaFold 3 represents a significant breakthrough in understanding biomolecular structure, allowing for experimental-level accuracy in predicting protein structures and their interactions with small molecules, DNA, and RNA.

AlphaFold 3 is a significant step towards solving the structure prediction problem, but we still aim for 100% accuracy in understanding molecular interactions.

To achieve transformative drug design, we need several more breakthroughs similar to AlphaFold, focusing on different core concepts of biology and chemistry.

The training of AlphaFold 3 involved overcoming challenges in tokenizing various molecular types, including proteins, DNA, and small molecules, to accurately predict their structures and interactions.

There's enormous potential in specializing general techniques for specific applications, such as protein structure prediction, and we are actively exploring this in our AI for science work.