Why AlphaFold Didn’t Solve Protein Biology

The shape is a starting point

Start with a protein’s amino-acid sequence: the ordered chain of building blocks from which the protein is made. AlphaFold predicts a three-dimensional structure for that sequence. A researcher can inspect the resulting shape for clues about how the protein might work, then use experiments to investigate those ideas.

Now ask a different question about that same protein: what happens when it meets a partner or changes shape inside a cell? A single predicted shape does not describe every state it might take, and knowing a shape does not by itself establish its function. Those distinctions matter when turning a useful structural prediction into a biological explanation.

In a Latent Space panel moderated by Brandon Anderson, Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido unpack what comes after a breakthrough in structure prediction. Their discussion moves from the data behind a model to its architecture, what its confidence means, and the harder problem of understanding biology in context. Watch from the opening.

Kohli puts the achievement in specific terms: AlphaFold’s team tried to reproduce structures that experimentalists had obtained and deposited in the Protein Data Bank. That result is useful, but a protein can adopt different shapes in different contexts. A predicted structure is therefore one piece of evidence about a molecule, not a complete account of its possible states.

From one structure to biological behavior

The panel separates several questions that can easily be collapsed into “protein folding”:

- Structure: What arrangement of the protein is a model predicting? Kohli describes the target as a structure comparable to one deposited in the PDB. - Dynamics: What other conformations can the protein take, and how does context affect them? Kohli says its distribution of structures remains a separate challenge. - Function and interactions: What does the protein do in a larger biological system? Candido argues for placing models in the context in which molecules operate. - Design: Can we create a protein with a useful property? Candido notes that structure models can help with design, even though design was not the starting task.

Candido’s analogy is a bicycle: a model of one spoke may get better and better while the scientist still needs to understand the wheel and the bicycle. His point is not to abandon folding models—he expects them to keep improving—but to connect molecular predictions to the systems people ultimately want to understand.

Find data that makes the problem learnable

Candido starts with a practical test for a scaling law: find a setting where more relevant data and compute actually produce a better result. That depends on the information in the data, the statistics it captures, and the model architecture. An abundant dataset can be useful without answering the question a scientist most wants to solve.

His example is metagenomic sequences: the data can be messy, and he says some sequences are not complete, real proteins, yet training on them improves his team’s protein language models for designing and understanding proteins. The lesson is not that data quality is irrelevant. It is that useful signal can live in material a conventional search for pristine examples might discard.

The opposite trap is scaling whatever happens to be cheap to generate. Candido wants model builders and data generators to work together on the missing information, rather than letting available data dictate the research agenda.

Kohli describes two different resource decisions. For AlphaFold, the team built on the substantial investment by scientists and organizations in PDB structures and focused on modeling; expanding that dataset by an order of magnitude was beyond its expertise and resources. For a virtual cell, he says the cell-by-gene data they examined was not yet sufficient for the larger ambition. The right investment depends on the problem and its constraints.

Scientific intuition belongs inside the model

AlphaFold 2 was not simply a generic model made bigger. Kohli describes using biophysics and biochemistry to give it a head start: amino-acid residues influence one another, so the model should represent those interactions. In his account, that knowledge makes learning more data-efficient because the model need not rediscover everything from examples.

Candido adds the tradeoff. With less data, a useful inductive bias—an assumption built into a model about the problem—can help. With more data, a wrong assumption can hold a model back or prevent it from finding patterns its designers did not anticipate. He also pushes back on the idea that scaling means less craft: choosing architectures and making larger training runs work are technical work too.

Look closer at the experiments

Asked what additional resource might accelerate work on function, dynamics or design, Kohli turns to cryo-electron microscopy micrographs. He wondered whether models could learn directly from those images, rather than only from derived structures, to recover information about dynamics and the distribution of molecular states. He tried the idea and says it needs more work.

That suggestion changes where the model looks: from a deposited structural result back toward the experimental images used to study molecules. The opportunity is richer information; the open question is whether a model can extract it reliably at the necessary scale.

A useful prediction needs trustworthy confidence

Candido wants to extract biological knowledge from models, not only use them as design tools. He says his team has found information about function and motion in protein-language-model representations alongside structural information.

Kohli emphasizes a different kind of understanding: knowing when to trust a prediction. He gives a researcher’s failure case: a model looks confident, the scientist spends a year pursuing the result, and the answer turns out to be wrong. A confidence score is valuable only when its stated confidence corresponds to observed reliability.

That behavioral understanding is distinct from explaining every internal step by which the model produced an answer. Kohli argues that users need the former—its strengths, weaknesses and calibration—even if its inner workings are not fully intelligible to a human. He speculates that a future, more capable model might help interpret another model; he does not present that as an achieved result.

What would acceleration in medicine mean?

When Anderson asks when AI will reach the clinic, Kohli resists a single date. AI is already being used in drug discovery, he says; a claim of 10× or 100× acceleration needs a named stage. Faster target discovery, lead optimization, preclinical work and toxicology are different outcomes, and he expects larger gains to require progress on hard biological questions.

Candido declines to predict when an entirely AI-made drug will arrive. Instead, he suggests asking what a 10× improvement would require as well as pursuing 10% gains: the larger question can expose routes that an incremental plan misses. Both approaches matter, he says, and both ultimately aim to improve outcomes.