Why AlphaFold Did Not End Protein Folding: Pushmeet Kohli and Sal Candido on Data, Models, and Trustworthy Biological AI

1. Guest Background

This episode of Latent Space: The AI Engineer Podcast is hosted by Alessio Fanelli and Shawn swyx Wang. Its title frames the central question directly: why AlphaFold did not solve protein folding. The indexed evidence identifies two guests, Pushmeet Kohli and Sal Candido. Pushmeet is supported as a speaker from Google DeepMind, and Sal is supported as a speaker from Biohub. The episode was uploaded by Latent Space on 2026-10-10 and runs 1,961 seconds, a little over thirty-two minutes.

The available guest context supports a narrow but useful description of their roles. Google DeepMind is discussed through AlphaFold and AI modeling. Biohub is described as working openly with the community on biological data and modeling. The episode uses that pairing to make the conversation concrete: Pushmeet repeatedly grounds his answers in the AlphaFold and scientific modeling experience, while Sal emphasizes protein language models, metagenomic sequence data, data-generation choices, and community work.

The subject is not a generic introduction to AI in biology. It is an episode analysis of a specific debate: what AlphaFold’s success does and does not establish, how biological AI teams should think about scaling data, when scientific priors help or hurt, why interpretability may mean different things to different users, and what kind of biological understanding would be needed before AI produces large acceleration in drug discovery or clinical translation.

2. What the Episode Covers

The episode opens by cutting into a popular compression of the AlphaFold story. Pushmeet Kohli accepts that, at a conceptual level, protein folding has seen major advances. But he rejects the stronger implication that scientists now know the real ground state of proteins or the full distribution of structures proteins can take. His description of AlphaFold’s task is deliberately narrower: someone obtained a structure experimentally, deposited it in the PDB, and the model learned to reproduce that structure. That task is useful, but it is not the same as understanding all protein dynamics.

The hosts then shift to a “bitter lesson” question for data. In AI, the bitter lesson is often summarized as the eventual victory of scalable methods. Sal Candido’s response is more conditional than a simple “more data wins.” He says the right data is necessary, but scaling laws do not automatically exist everywhere. A large part of the work is finding situations where more compute, more data, and the right architecture reliably produce better results. If the data does not contain the information statistics required by the target problem, a model cannot extract the understanding the team wants.

Sal makes that point less abstract with protein language models. He says these models can be trained on metagenomic sequences that are not pristine data; much of that data may not even be a complete real protein. Yet it can still improve model performance for designing real proteins and understanding known proteins. The lesson is not that any easy-to-generate data should be scaled indefinitely. Sal warns that this is the trap: teams may expand the data they can readily produce rather than ask what data is needed to solve the problem. He connects Biohub’s open, community-oriented work to this need for model builders and data generators to make those choices together.

Pushmeet reframes the bitter lesson as a research posture. The danger, in his view, is a religious identification with one tool category: “I am a modeler” or “I am a data-generation person.” The problem has to come first. If the scientific goal requires modeling, do modeling; if it requires data collection, collect data; if it requires domain expertise, include it. AlphaFold illustrates one side of that principle: DeepMind could not simply expand the PDB by an order of magnitude, so the high-leverage path was to get more value from existing structure data through modeling. In cell genomics and virtual-cell work, however, analysis of cell-by-gene data suggested the opposite bottleneck: the data was not yet sufficient for the grand ambition.

3. Core Views: Reasoning, Examples, and Limits

The first load-bearing view is that AlphaFold’s success should be treated as a bounded achievement, not a closing argument. Pushmeet does not minimize the advance; he says there has been progress. His reasoning is that the task AlphaFold solved is shaped by the available experimental structures in the PDB. Replicating a deposited structure is enormously valuable, but it does not identify the true ground state of every protein, map the full distribution of structures, or explain protein dynamics. The episode’s strongest corrective is therefore conceptual: the public phrase “protein folding is solved” is an acceptable shorthand only if one remembers what was actually solved.

That boundary matters because proteins are not inert building blocks. Pushmeet explicitly resists the static-block metaphor, even while acknowledging that he uses it. Proteins can be disordered, complex, and context dependent. Their shapes may change according to the biological environment. This is why the episode keeps returning to function, dynamics, and design as still-open problems. A model that is useful for one structural prediction task may be an input into those problems without becoming a complete model of them.

The second view is that data scaling has to be tied to information, not volume. Sal’s reasoning is precise: scaling becomes powerful when a team has found a setting where adding data and compute improves results under an appropriate architecture. Without the right information statistics in the data, the model cannot learn the desired capability. The metagenomic-sequence example complicates any simplistic idea of “clean data.” Imperfect data can still contain useful evolutionary or protein signals. But Sal’s warning is just as important: once teams learn that available imperfect data can help, they may over-scale the data that is easiest to generate rather than seek the data that actually closes the scientific gap.

The third view is that “models versus data” is a false strategic identity. Pushmeet treats the bitter lesson as an argument against tool tribalism. AlphaFold and virtual-cell work point in different directions because the constraints differ. For AlphaFold, the PDB represented a historically accumulated dataset that DeepMind could not cheaply expand by an order of magnitude, so model innovation was the rational investment. For cell genomics, the cell-by-gene data did not yet support the desired virtual-cell ambition, so the bottleneck looked more like missing data. The episode’s reasoning is pragmatic: the right allocation of effort depends on the scientific objective, available resources, constraints, and impact target.

The fourth view is that scientific priors and scale are complements under conditions, not permanent rivals. Pushmeet argues that AlphaFold2’s crafted design encoded scientific intuitions from biophysics and biochemistry, especially that amino-acid residues influence one another. That gave the model an “unfair advantage” in data efficiency. Sal adds the limitation: with smaller data, inductive bias can be necessary; with more data, models may discover patterns humans did not anticipate, and an inaccurate bias can hold them back. The episode also rejects the idea that scaling is effortless. Larger models, more data, faster inference, transformer modifications, and bespoke architectures all require algorithmic and systems craft.

The fifth view is that interpretability should be separated from trustworthiness. Sal says black-box design was useful before AI, but scientists still want to extract knowledge from models. He points to protein language models that contain not only structural information but also information about function and motion. Pushmeet draws a sharper operational line: users need calibrated uncertainty and behavioral characterization. He says AlphaFold2 was not perfect and cites about 90 GDT on a particular set; even a hypothetical 95 GDT model would be hard to trust if its pLDDT confidence score were uncalibrated. A confidently wrong model could send a researcher down the wrong path for a year.

The limitation in the interpretability discussion is explicit. Pushmeet does not claim humans can fully understand AlphaFold2’s internal mechanism. He says interpretability depends on the interpreter. A human rational system with human cognitive limits may not be able to interpret it, while a future larger model with access to activations might develop a theory of its behavior. That is a possibility, not an established result. What is established for use is more modest and more practical: characterize what the model can do, what it cannot do, where it is strong, and where it fails.

The final view concerns clinical impact. Pushmeet says asking when AI results will appear in the clinic is ill-posed because AI is already used throughout drug discovery. The sharper question is when it will produce 10x or 100x acceleration in specific stages such as target discovery, lead optimization, preclinical work, or toxicology. He ties that acceleration to harder biology and better biological models. Sal is similarly cautious about predicting a fully AI-made drug, while expecting rapid progress because tools are being used. Both guests avoid a countdown narrative and make the uncertainty part of the analysis rather than a footnote.

4. Learning and Application

The most immediate application is to label the task boundary before using a biological AI system. For AlphaFold-like systems, the safe working assumption is not “protein folding is finished.” It is: the model can generate highly useful structure-related predictions grounded in the kind of structures represented in the training and evaluation regime. That makes it valuable for hypothesis generation, prioritization, and experimental planning. The boundary appears when teams move into disorder, context-dependent conformations, dynamics, function, or design. In those settings, model output should be treated as one piece of evidence in a broader workflow, not as a complete biological explanation.

A second application is project design. Teams should begin with the target scientific question and then decide whether the bottleneck is data, modeling, compute, architecture, experimental feedback, or domain expertise. Pushmeet’s AlphaFold example and virtual-cell contrast are a useful diagnostic pair. If the dataset is historically valuable but hard to expand, model work may be the best route. If the ambition requires biological coverage that existing data does not contain, better modeling alone may be a poor substitute for new data. A practical planning review should ask whether the available data contains the required information statistics, whether new data adds coverage rather than duplicates the same distribution, and how quickly a failing path can be detected.

A third application is to avoid a naive clean-versus-dirty data rule. Sal’s metagenomic example shows that imperfect data can be useful if it carries signal relevant to the model’s objective. But the tradeoff is that available data can become seductive. Teams can drift into scaling what is cheap to collect rather than what answers the question. The practical boundary is information gain: each new data source should be justified by the blind spot it reduces, the biological modality it adds, or the decision it improves. Sample count alone is not a strategy.

A fourth application is to treat scientific priors as design choices that need audit trails. In low-data or expensive-data settings, priors from biophysics, biochemistry, structural constraints, or residue interaction patterns can improve data efficiency. But as data and tasks broaden, the same priors should be retested. The engineering version is straightforward: write down the assumed prior, create evaluations where it should help, create stress tests where it might fail, and run ablations when feasible. This preserves the value of domain knowledge without letting inherited assumptions silently cap the model.

A fifth application is to make trust a product requirement. Pushmeet’s pLDDT point matters beyond AlphaFold: for scientific users, calibrated confidence, failure-mode documentation, and behavioral characterization can be as important as average performance. If a model is used to choose experiments, prioritize molecules, or guide drug-discovery work, a confident wrong answer can consume months of effort. The operational standard should include uncertainty calibration, known-strength and known-limit documentation, and decision rules for when model output requires experimental confirmation.

Finally, drug-discovery and translational teams can use the 10x question without turning it into hype. Sal’s point is not that incremental improvements are useless; he says both 10% and 10x approaches matter. The value of the 10x lens is that it forces a broader first-principles view of the biological model, data-generation loop, and experimental system. The boundary is equally important: the episode does not provide a date when an AI-made drug will reach the clinic, nor does it prove that a specific stage will be accelerated by a fixed amount. It gives a better question to ask: which stage, under what biological understanding, with what data and validation, could plausibly move by an order of magnitude?

Source

• Original episode: Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub