DeepMind's Pushmeet Kohli and Biohub's Sal Candido Explain Why AlphaFold Didn't Solve Protein Folding

Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido argue that AlphaFold did not solve protein folding, according to the title of a new Latent Space podcast episode. The episode runs 31 minutes 55 seconds.

What the source confirms

The source page for the episode, "Why AlphaFold Didn't Solve Protein Folding," lists:

• Guests: Pushmeet Kohli (Google DeepMind) and Sal Candido (Biohub)
• Format: Audio podcast, Latent Space: The AI Engineer Podcast
• Runtime: 31:55

The page as available carries no transcript, show notes or summary. We cannot report the specific arguments, data points or benchmark figures discussed. Nothing below should be read as a quote or paraphrase from the guests.

Established background

The episode title responds to a widely held belief that AlphaFold ended the protein folding problem. The public record shows the following:

• AlphaFold 2 was recognized for its performance at the CASP14 structure prediction assessment in 2020.
• DeepMind and EMBL-EBI released the AlphaFold Protein Structure Database, which made predicted structures broadly available to researchers.
• AlphaFold 3 was announced in May 2024, extending prediction to complexes involving proteins, nucleic acids, ligands and other molecules.
• The 2024 Nobel Prize in Chemistry was shared by Demis Hassabis and John Jumper of Google DeepMind and David Baker, for protein structure prediction and computational protein design.

These milestones concern predicting a protein's three-dimensional structure from its amino acid sequence. The episode title suggests the guests are drawing a line between that achievement and the broader scientific question of folding.

What remains unverified

The following are not available from the source:

• The specific limitations Kohli and Candido identify
• Any new models, datasets or tools announced in the episode
• Any collaboration details between Google DeepMind and Biohub
• Any benchmark scores, pricing or release dates

No model is released or updated in the source material.

What this means

The framing matters for AI-for-science builders. Structure prediction is a solved-enough input for many workflows, but it is not the same as modeling how proteins fold, move, interact or function in cells. Researchers have long noted that static predicted structures do not capture conformational dynamics, folding pathways or behavior in cellular context. If the guests make a similar argument, the field's attention is likely shifting from accuracy on static structures toward dynamics, function and experimental validation.

The pairing also stands out. Kohli represents a lab that built the best-known structure prediction system, and Candido comes from a biology-focused organization. A model developer and a biology institution discussing the gap is a useful signal of where wet-lab integration may matter next.

We will update this article if a transcript or detailed show notes become available.