Google DeepMind and Biohub Explain Why AlphaFold Didn't Close the Protein Folding Problem

In 2020, AlphaFold 2 scored a median GDT of about 92 out of 100 at the CASP14 competition, and in 2024 Demis Hassabis and John Jumper received half of the Nobel Prize in Chemistry for it. Now Pushmeet Kohli, Google DeepMind's science lead, and Sal Candido of Biohub say something different on the Latent Space podcast: the protein folding problem is only partially solved. The episode was released on October 10, 2026.

The episode announcement lays out two threads. The first is Richard Sutton's "bitter lesson": general methods backed by large-scale compute eventually beat hand-engineered knowledge. The second is the unsolved puzzles of folding. The source has no full transcript, so below we rely on the announcement's talking points and publicly known context.

What AlphaFold Actually Solved

AlphaFold 2 learned to predict a protein's three-dimensional structure from its amino acid sequence with accuracy comparable to experiment. For many targets, this cut the path from sequence to model from months of crystallography work to minutes. Together with EMBL-EBI, DeepMind released AlphaFold DB, which holds more than 200 million predicted structures. That is practically the entire known protein universe.

In May 2024, AlphaFold 3 widened the scope: proteins, DNA, RNA, and small-molecule ligands in a single model. For ligands, the authors claimed an improvement of at least 50% over existing docking methods.

Why "Solved" Is Too Strong a Word

A structure in the database describes a single conformation, usually the most probable one. A real protein lives in motion: it changes shape, interacts with partners, and stays partly disordered. The classic list of gaps looks like this:

• dynamics and conformational ensembles instead of a single snapshot;
• the folding pathway: how a chain arrives at its shape, not just which shape it ends up with;
• the effect of point mutations on stability and function;
• disordered regions that have no single structure;
• large complexes and interactions inside the cellular environment.

Predicting the end point and understanding the process remain different tasks. A model can guess the outcome brilliantly and still say nothing about the mechanism. For biology and medicine, this is a fundamental gap: a drug works in the dynamics of a living cell, not on a static picture.

Scaling and Its Limits in Biology

Sutton's lesson has held up in language models: more data and compute produce more capable systems. In biology, this recipe has its own limit. The Protein Data Bank holds roughly 200 thousand experimental structures, while the volume of text on the internet is several orders of magnitude larger. Data on dynamics and function is even scarcer, and every data point costs laboratory effort. So any discussion of understanding biology inevitably arrives at a combination: large-scale models plus new experimental data collected specifically to target the models' weak spots.

The participants' backgrounds speak for themselves. DeepMind represents the path from AlphaFold to AI systems for science. Biohub is a research organization where models are tied into biological experiments.

Key Facts

• AlphaFold 2, CASP14, 2020: median GDT of about 92.
• AlphaFold DB: over 200 million predicted structures.
• AlphaFold 3, 2024: proteins, nucleic acids, and ligands in a single model.
• 2024 Nobel Prize in Chemistry: Hassabis and Jumper (half), David Baker.

How to Use This

A team in biotech or pharma should treat an AlphaFold prediction as a hypothesis, not a result. It speeds up target selection and experiment planning, but assessing mobility, binding, and the effect of mutations requires molecular dynamics and laboratory validation. Teams building their own models should decide up front what data on dynamics and function they will be able to obtain, because parameter scale alone is not enough here.