Hiding some evolutionary clues helped AlphaFold 3 find missing protein shapes

A protein model can be impressively accurate and still hide the shape a researcher needs.

That matters because proteins are moving molecular machines, not rigid ornaments. A transporter may open toward one side of a membrane, bind its cargo, then rearrange to release it on the other. Enzymes close around reactants. A promising drug-binding pocket may appear in one conformation and disappear in another. A single predicted structure is therefore closer to a still photograph than a complete account of the machine.

A study published September 18 reports a counterintuitive way to coax more shapes from AlphaFold 3: obscure some of the evolutionary evidence supplied to it. Across 107 proteins with two experimentally determined conformations, three methods for disrupting that evidence improved the model’s resemblance to the harder-to-find state in roughly one-fifth of cases apiece, while causing comparatively few declines of the same size (Communications Chemistry).

The family record inside a prediction

AlphaFold 3 can receive a multiple-sequence alignment: a table in which the amino-acid sequences of related proteins are lined up. Patterns in that table carry useful clues. If changes at two positions repeatedly occur together over evolution, those positions may interact in the folded protein.

But the strongest pattern can favor one familiar conformation. The researchers tested three ways of weakening or separating that dominant signal:

• Subsampling gave AlphaFold 3 a shallower alignment containing fewer sequences.
• Clustering divided related sequences into groups and supplied those groups separately.
• Column masking hid selected amino-acid positions behind an unknown-residue symbol.

None changes the protein sequence being modeled. Each changes the evolutionary context from which the model draws clues. The working idea is that partially muffling the loudest signal can make a quieter alternative easier to sample.

A thousand attempts, then a demanding comparison

The benchmark contained 107 proteins, each represented by two structures previously determined through experiments. For every protein and method, the researchers generated exactly 1,000 predictions, except that clustering produced at least 1,000.

They compared predictions with each experimental structure using TM-score, a whole-structure similarity measure on which 1 represents a perfect match. Because these methods deliberately generate varied answers, the study evaluated the mean score of the best 1%—in most cases, the best ten structures out of a thousand.

Against ordinary, unperturbed AlphaFold 3, column masking improved that top-1% score by at least 0.05 for the alternative conformation of 24 proteins and worsened it by that amount for one. Clustering produced 18 improvements and two declines; subsampling produced 16 improvements and no comparable declines.

Those counts are more informative than saying simply that the methods “worked.” Most proteins did not cross the chosen improvement threshold. The result is a useful expansion of sampling in a substantial minority of cases, not a universal solution.

The calcium pump that came into view

A calcium-transporting ATPase makes the payoff concrete. This membrane protein uses energy from ATP and changes shape while moving calcium ions against a concentration gradient.

Unperturbed AlphaFold 3 closely reproduced two experimentally observed states of the transporter, with mean top-1% TM-scores of 0.95 for each. Yet it did not sample the known state in which ATP and ions are bound nearly as well; that state scored 0.78.

With columns in the evolutionary alignment masked, the generated set broadened. It included structures resembling the missing ATP-and-ion-bound conformation, raising its mean top-1% score to 0.91. The predicted rearrangement included movement in a domain outside the membrane and shifts in two membrane-spanning helices. The other tested methods did not match that result.

This is an observation about resemblance to a known structure. It does not show AlphaFold 3 simulating the pump’s motion, discovering the route between states or calculating how often the real protein occupies each state.

Why the result is promising—and not yet prediction on demand

The benchmark has three important limits.

First, researchers could recognize a success because they already possessed the experimental answer. Selecting the best 1% from a thousand attempts is not the same as reliably identifying an unseen conformation. The calculations also have a real cost: the paper reports that 1,000 AlphaFold 3 predictions on one specified GPU took from about half an hour for smaller proteins to roughly ten hours for the largest examples, excluding alignment generation.

Second, only nine of the 107 proteins had both reference structures released after AlphaFold 3’s training cutoff. The improvement trend persisted in that small subset, but nine proteins cannot carry the evidential weight of the entire benchmark. For many other cases, the study cannot fully exclude influence from structures represented in training data.

Third, a collection of generated structures is not automatically a physical ensemble. The original AlphaFold 3 paper explicitly cautions that repeated random samples do not approximate how biomolecules are distributed among shapes in solution (Nature). Similarity to deposited structures establishes neither the probability of each state nor the path and timing of transitions between them.

The appropriate near-term use is therefore as a candidate generator. An alternative model might help researchers interpret lower-resolution measurements, plan an experiment or examine a possible binding pocket. Those are prospects to test, not outcomes demonstrated by this benchmark.

The work is unusually inspectable. The authors deposited their alignment-manipulation program, analysis code, raw similarity scores, protein data and selected predictions on Zenodo. That makes the central comparison reproducible: perturb the evolutionary record, generate many structures, and ask whether the resulting set contains better matches to experimentally established states.

The charming twist is methodological. More information usually sounds better. Here, selectively hiding information sometimes let a capable model reveal more of the machine.