Pushmeet Kohli of Google DeepMind has a line he uses all the time: proteins are the building blocks of life. He also says he doesn't believe it. "I say proteins are the building blocks all the time," Kohli said, "but actually, like, I don't believe in it." Proteins, he said, "are not blocks." They are "extremely complex," often disordered, and their shape "might change depending on the context."
That confession came during a panel on AI for biology, published as an episode of the Latent Space podcast on October 10, 2026. Kohli shared the stage with Sal Candido of Biohub, a research organization whose mission, as the episode describes it, is to cure all disease. Brandon Anderson moderated. The discussion started from a gap: news reports say protein structure prediction has been solved, but working scientists say function, dynamics and design are still wide open, about five years after AlphaFold 2 was announced. Is another big leap coming, and what is in the way?
What AlphaFold actually does
Proteins are chains of building units called amino acids. Each chain folds into a three-dimensional shape, and that shape largely decides what the protein does in the body. AlphaFold, DeepMind's AI system, predicts that shape from the protein's amino-acid sequence.
Its big public moment came in CASP14, a blind competition in which teams predict structures whose lab-measured answers have not yet been released. DeepMind reported a median score of 92.4 out of 100 on GDT, a measure of how many parts of the predicted chain land close to their measured positions, across the competition's targets, and 87.0 in the hardest category. Those are medians for particular sets of competition targets, not a share of all proteins solved. The 2021 Nature paper describing the system, led by John Jumper, reported a median backbone accuracy of 0.96 ångströms, measured as Cα root-mean-square deviation at 95% residue coverage, across 87 protein domains, against 2.8 ångströms on the same measure for the next-best method. It also noted that accuracy fell when a protein's shape depended mainly on contact with other protein chains.
"What are we trying to solve?"
Kohli said he and Jumper used to put that question to each other. Their honest answer was narrower than the headlines. Nobody knows the true resting shape a protein takes, Kohli said, or the full range of shapes it moves between. What AlphaFold set out to do was more modest: "someone got a structure, deposited it in the PDB, and we are trying to replicate and get the same structure. That's what we did." And, he added, "it just so happens that it's useful."
The PDB, or Protein Data Bank, is the public archive of experimentally determined structures that AlphaFold was trained on; the Nature paper says training used entries released by April 30, 2018. A PDB entry is not a photograph of a protein. According to the archive's own documentation, each starts with a physical sample measured by a technique such as X-ray crystallography, nuclear magnetic resonance or electron microscopy. Scientists then compute a three-dimensional model from those measurements, and the record notes the sample conditions, instruments and processing choices behind it. A prediction that matches a PDB entry is matching one model made under one set of conditions.
That is why Kohli separated reproducing structures from understanding how proteins move. Matching deposited structures "does not mean that we have understood all of protein dynamics," he said. EMBL-EBI, Europe's major bioinformatics institute, makes a similar point in its guide to AlphaFold 2's limitations. The model gives static snapshots, while many proteins change shape as they work. The guide's example is hexokinase, an enzyme with one shape when bound to sugar and another without it; AlphaFold's prediction resembles the sugar-free form. Some protein regions are intrinsically disordered, meaning they may have no single fixed shape at all. AlphaFold's low-confidence scores can help flag those regions but cannot describe them as a single structure.
Kohli said the simpler story is easier to tell the public, but the scientists in the room know how much remains. "There's a lot to celebrate," he said. "But let's not stop the funding of protein structure prediction and protein dynamics because we are just getting started."
The magic wand: go back to the raw images
Asked what he would wish into existence if he could wave a magic wand, Kohli went back to his own career. He described a varied path: security research, then computer vision, Bayesian theory, deep learning, AI for coding and finally science. The computer-vision researcher in him, he said, gets excited about cryo-EM micrographs.
Cryo-electron microscopy, or cryo-EM, freezes many copies of a molecule and photographs them with an electron beam. Those images are then processed into a single three-dimensional structure. Kohli's thought was that the PDB holds only that end product. "I should be working at the source," he recalled thinking. In his view, the finished structures must be missing much of what the original images record.
Archive practice suggests there is something to work with. EMBL-EBI's EMPIAR archive stores raw micrographs and stacks of individual particle images, separately from the final reconstructed maps kept elsewhere. Its documentation says raw data allows later reprocessing. It also says particle images and their alignment data support analysis of a molecule's flexibility, and that keeping particles left out of the final model helps researchers study variation between copies.
Kohli said models that could work at that scale, with enough data, and "extract all the dynamics and distributional information that is captured there would be an amazing thing." This is a hope, not a demonstrated method. "I tried it, but it requires more work," he said. He still thinks it is a route worth taking: "maybe, like, people better than me would take a stab at it, and we'll get somewhere."
From spokes to bicycles
Candido agreed that the problem is bigger than a single structure, and gave an analogy. Today's protein models, he said, are "quite useful, but they're not exactly the problem that most people want to solve." They were built for a narrow purpose, and researchers have found other uses for them, such as designing new proteins. But he compared the current approach to someone trying to understand how a bicycle works by modeling one spoke.
The spoke models can keep improving, Candido said, but the field has to move "from models of like spokes to wheels to maybe whole bicycles," because the whole system is what people want to understand. Stretching the analogy, he said people seem to want to use the bicycle model to design a part for a pickup truck. Placing protein models in the biological setting where proteins actually work, he said, should let researchers learn about interactions on a broader scale.
He did not argue for abandoning the narrow problem. "I think we should keep trying to work on folding models," Candido said, "because I think they're going to keep getting better and better."