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

Summary

• AlphaFold 2 reproduced a valuable experimental target; it did not explain protein behavior in full. Pushmeet Kohli says the model learned to recreate structures deposited in the PDB. The discussion leaves the true ground state, structural distribution, dynamics, function, and design as open questions. “That’s what we did, right? And it just so happened that it turned out to be useful.”
• The relevant scaling law depends on having data with the right statistics, not merely more observations. Sal Candido’s counterexample is metagenomic sequence data: much of it may not represent complete proteins, yet it improves models for understanding and designing real proteins. The trap is concluding, “Let’s just scale the data that we can easily generate.”
• Data generation and modeling are complements whose allocation should follow the problem. Sal Candido criticizes method-first identities—“I am a modeling specialist” or “I am a data-generation specialist”—as almost religious. He says AlphaFold worked with the existing PDB because expanding it was beyond the team’s expertise and resources, whereas cellular genomics showed that CELLxGENE data remained insufficient for the more ambitious goal of a virtual cell.
• Scientific priors remain valuable, particularly when data is scarce, but incorrect inductive bias can eventually become a constraint. Pushmeet says AlphaFold’s design used biophysical intuition that residues influence one another, giving it an “unfair advantage” in data efficiency. Candido says priors matter especially with less data; Brandon Anderson adds that scaling itself requires substantial algorithmic and architectural skill.
• The next frontier is richer biological context, not another isolated structural snapshot. Kohli’s “magic wand” would model raw cryo-EM micrographs directly to recover distributions and dynamics that processed structures may discard. Candido’s analogy is sharper: today’s models resemble one bicycle spoke, while researchers need “wheels, and maybe even whole bikes.”
• Model reliability matters more immediately than complete human-readable mechanistic interpretation. AlphaFold 2 achieved about 90 GDT on the cited dataset, but Kohli argues even 95 GDT would be dangerous without calibrated pLDDT: confidently wrong predictions could waste a year. The operational requirement is knowing “what it can and can’t do.”
• AI is already embedded throughout drug development, but neither guest offers a date for an end-to-end AI-created medicine. Kohli says AI is already used across drug development and that 10x or 100x acceleration requires specifying the bottleneck—target discovery, compound optimization, preclinical work, or toxicology—then improving the underlying biological models. Candido expects rapid progress but preserves the uncertainty: “I don’t know. It’s hard, hard to say for sure.”

Deep dive

1. Useful scaling begins with the information a problem requires

• Brandon Anderson reframes AI’s “bitter lesson” for biological data: do methods that scale ultimately win? Candido’s answer is conditional—scaling laws do not exist “everywhere, all the time”; researchers must discover when additional computation and data reliably improve results.

• Candido’s best specimen is protein-language-model training on metagenomic sequences, much of which he says may not even be “real complete proteins.” Despite that impurity, the data improves both protein understanding and the design of functioning proteins.

• Anderson warns that a community made up only of model builders or only of data generators will default to what already exists or what it can generate. He argues for open collaboration to ask, “What data do we need to solve this problem?” and search for the next packet of information, not just the next batch of rows. Candido’s complementary lesson is to choose modeling or data generation according to the problem rather than a fixed method identity.

2. The problem should choose the method, budget, and expertise

• Candido’s reinterpretation of the bitter lesson is methodological humility: treating modeling or data generation “almost like religion” makes failure more likely. Kohli’s practical advice is to regard oneself as a multidisciplinary expert: start with the desired scientific outcome, then decide whether the binding need is modeling, computational scale, new experiments, or domain expertise.

• Candido says AlphaFold focused on the existing PDB because dramatically expanding it was beyond the team’s expertise and resources. Cellular genomics produced the opposite conclusion: after much work with CELLxGENE, the available data still appeared insufficient for the ambition of a virtual cell.

• Kohli’s practical advice is to understand constraints and “try to fail quickly.”

3. Handcrafted science and scalable learning remain complements

• Anderson characterizes AlphaFold 2 as “a work of art,” built from carefully chosen functions, and asks whether scarce funding should now favor universal scaling over artisanal solutions. Kohli rejects the binary: scientific intuition about residue interactions gave the model useful structure instead of forcing it to rediscover everything.

• Data curation is itself an intellectual problem. “If you just duplicate the same amount of data, you won’t get anywhere,” Kohli says; progress requires identifying which observations change what the model can learn.

• Candido says inductive bias is especially valuable with limited data. Anderson adds that a wrong bias can become a constraint as data grows, and rejects the idea that scaling is unskilled: it requires algorithmic and architectural work, including faster inference, training on more data, doing more computation, and modifying or extending architectures for greater efficiency.

4. AlphaFold solved the deposited-structure task, not protein reality

• Kohli disputes the “protein folding is solved” headline by separating benchmark success from biology. Scientists do not know a protein’s true ground state or full distribution of adopted structures; AlphaFold recreated a structure someone obtained and placed in the PDB.

• The familiar “proteins are building blocks” language is itself misleading: “Proteins are not blocks, they don’t act like blocks.” They can be disordered and change form with context, leaving dynamics, function, and design unresolved.

• Given a magic wand, Kohli would work from raw cryo-EM micrographs rather than processed structures, building models capable of extracting the dynamics and distributions still present in the source data. He tried this direction and concedes plainly, “It takes more work.”

• Candido says current folding models are useful but likens them to modeling one spoke while trying to understand a bicycle—and then using that bicycle model to design pickup-truck parts. His proposed progression is from isolated components to contextual interactions, wheels, and eventually “whole bikes.”

5. Calibration converts black boxes into decision tools

• Anderson invokes Feynman’s “What I cannot create, I do not understand,” then notes that AI can now create without understanding. Candido says black-box design predates AI and can remain useful, but insists that understanding matters and asks how to extract knowledge from machines. Kohli adds that current models contain more information than we know: protein language models reveal structure, function, and movement while compressing information from evolution.

• Kohli distinguishes mechanistic interpretability from behavioral understanding. AlphaFold 2’s roughly 90 GDT was not enough by itself: without calibrated pLDDT, a model could express certainty, send a scientist down a year-long path, and prove wrong. Users therefore need mapped strengths, limitations, and uncertainty.

• Interpretability also depends on the observer. AlphaFold 2 may exceed human cognitive limits while remaining interpretable to a future, larger model given access to its activations; meanwhile, its unexpected ability to identify disordered regions after release demonstrated genuine generalization.

• On clinical timing, Kohli corrects the premise: AI already operates across drug development. Transformative 10x or 100x acceleration over the next few years remains contingent on solving deeper biological problems; Candido argues that pursuing both 10% improvements and first-principles 10x paths reveals solutions incremental thinking may miss.