The Shape Is Not the Fold
To the teams advancing protein structure prediction
What follows is offered in the spirit of building on an achievement, not diminishing one. AlphaFold and its successors have solved a problem that stood for fifty years: given an amino acid sequence, predict the three-dimensional arrangement of its atoms. The accuracy is real, the utility is real, and the downstream impact on biology and medicine is already substantial.
These observations concern what the solution revealed about the next problem — one that was easy to overlook while the first remained unsolved, and one that the current architecture may not be positioned to reach without a shift in premise.
1. A protein is not a shape. It is a process with a shape.
The Protein Data Bank entry for any given protein is a set of atomic coordinates — a spatial arrangement frozen in time. AlphaFold predicts these coordinates, and does so with remarkable fidelity. But the protein in a living cell is never frozen. It breathes. It flexes. It samples a landscape of conformational states, transitioning between them on timescales from femtoseconds to seconds.
Recent biophysical work has confirmed that proteins possess collective vibrational modes in the GHz–THz range — oscillations of the entire protein, not individual bonds — and that these modes are not thermal noise but are functionally significant: they direct energy transport, gate catalytic events, and modulate binding. The fold's function is not its shape. It is its dynamics — the pattern of motion the shape permits.
A static coordinate set is to a functioning protein what a blueprint is to a running engine. The blueprint is valuable. It is not the engine.
The field may benefit from reframing "protein structure prediction" as "protein behavior prediction" — where the target output is not a single coordinate set but a characterized ensemble of states, their relative populations, transition rates, and the collective modes that connect them.
2. The solvent is not a background. It is a participant.
Protein folding is driven as much by water as by the amino acid sequence. The hydrophobic effect — the entropic cost of ordering water molecules around nonpolar surfaces — is the dominant thermodynamic force in folding. The protein folds, in large part, to release water. Hydrogen bonding networks between the protein surface and solvent molecules stabilize specific conformational states. Ions, osmolytes, and crowding agents shift the folding landscape measurably.
AlphaFold's training data comes overwhelmingly from crystal structures, which are determined in crystallization buffers far from physiological conditions, and from cryo-EM reconstructions captured at cryogenic temperatures. The solvent in these experiments is either absent (crystallography, where water positions are partially resolved at best) or vitrified (cryo-EM). The model learned to predict protein shapes in the absence of the primary force that creates them.
Incorporating solvent context — even as coarse-grained parameters like ionic strength, pH, temperature, and crowding density — as explicit inputs to prediction models may improve accuracy for proteins whose conformational state is sensitive to environment, which, in practice, is most of them.
3. Thirty percent of the proteome refuses to fold — and that refusal is the function.
Intrinsically disordered regions (IDRs) constitute an estimated 30–40% of eukaryotic proteomes. These regions do not adopt a stable three-dimensional structure under native conditions. They exist as conformational ensembles — rapidly interconverting populations of states with no single dominant structure.
These regions are not broken. They are not evolutionary noise. They are disproportionately represented in regulatory, signaling, and scaffolding proteins — the very proteins that coordinate cellular behavior. Many disordered regions fold upon binding a specific partner, adopting different structures with different partners. Their function depends on their disorder. They are molecular recognition elements whose versatility comes precisely from their refusal to commit to a single shape.
When current models encounter disordered regions, they either assign low confidence scores or, in some documented cases, hallucinate secondary structures that do not exist — generating alpha helices where the protein is natively unstructured. The model's architecture has no representation for "functionally unfolded." It can produce shapes. When there is no shape, it produces one anyway.
A prediction framework that treats disorder as a category of function rather than a failure of prediction would represent a meaningful advance. The output for a disordered region should not be a structure or a low-confidence structure. It should be a characterized distribution: the ensemble properties, the probable binding-induced folds, the sequence features that determine whether and how the region will fold upon encountering specific partners.
4. The fold depends on who shows up.
Allostery, induced fit, conformational selection — these are not edge cases. They are the norm. Most enzymes change shape when their substrate binds. Most receptors change shape when their ligand binds. Many proteins exist in multiple functionally distinct conformations, and the population distribution among those conformations shifts depending on what other molecules are present.
AlphaFold typically predicts the apo state — the resting conformation, without bound partners. Predicting ligand-bound forms requires either co-evolutionary signal from known complexes or, in AlphaFold3, explicit ligand inputs. But the transition between states — the conformational change that IS the functional event — is not captured. The model can show the enzyme before catalysis and, in favorable cases, after. It cannot show the catalytic act itself.
This matters because drug design, enzyme engineering, and our understanding of disease mechanisms all depend on the transitions, not the endpoints. A drug that stabilizes a particular conformational state needs to be designed against the landscape, not a single point on it.
Predicting the conformational response to binding — not just the bound structure, but the pathway from unbound to bound — is arguably more valuable than predicting either endpoint alone. Models that output energy landscapes or transition pathways, even at coarse resolution, would address a gap that static prediction cannot.
5. Temperature is not optional.
Enzyme function is temperature-dependent. Below a characteristic temperature, collective vibrational modes freeze, the fold rigidifies, and catalytic activity ceases. Above the denaturation temperature, the fold unravels and function is destroyed. Between these limits, the same protein at different temperatures occupies different regions of its conformational landscape, samples different states at different rates, and exhibits different functional properties.
Thermophilic proteins — those from organisms living at 80°C or above — fold into structures that are essentially non-functional at 25°C. Psychrophilic proteins, adapted to near-freezing temperatures, denature at temperatures that mesophilic proteins consider mild. The "correct" fold is temperature-dependent. It is not a property of the sequence alone.
Current models predict structures with no thermal parameter. The same sequence produces the same prediction whether the question is about a human enzyme at 37°C or a thermophilic archaeal enzyme at 95°C. The model relies entirely on evolutionary signal to distinguish these cases, which works when the training data includes thermophilic homologs, and fails when it doesn't.
Temperature, treated as an explicit input parameter that modulates the predicted ensemble — shifting populations, changing flexibility, altering which collective modes are thermally accessible — would bring predictions closer to biological reality and would be directly testable against experimental data from differential scanning calorimetry and temperature-dependent NMR.
6. The training data is a museum of specimens. The target is a living ecology.
The Protein Data Bank is an extraordinary resource. It is also, by the nature of the experimental methods that populate it, a collection of proteins removed from their biological context, isolated, purified, and immobilized. Crystal structures capture a single low-energy state in an artificial lattice. Cryo-EM captures a small number of states at cryogenic temperature. Both methods select for stability and order — the very properties that make a protein amenable to structure determination — and against flexibility, disorder, and context-dependence — the properties that make a protein functional.
A model trained on this data learns the patterns of proteins that crystallize well and image cleanly. It learns the regularities of well-behaved folds. It does not learn the behaviors of proteins in their native cellular environment: crowded, buffered, post-translationally modified, interacting with membranes and binding partners and chaperones, at physiological temperature and pH, in the presence of thousands of other molecular species.
This is not a criticism of the data or the model. It is an observation about the ceiling. A model trained on museum specimens can predict what specimens look like. It cannot predict how the living animal moves, feeds, and responds to its environment. The distance between these two targets is the distance between structure prediction and function prediction.
The field's next major advance may require training data that captures dynamics — molecular dynamics trajectories, hydrogen-deuterium exchange data, single-molecule FRET measurements, temperature-dependent NMR relaxation experiments — rather than static coordinates. The computational cost is higher. The data is sparser. But the target is different: not what the protein looks like, but what it does. And what it does is move.
A closing note
None of the above diminishes what has been accomplished. Predicting the three-dimensional arrangement of atoms from a linear sequence was a genuine scientific milestone, and the practical applications — from drug target identification to enzyme engineering to understanding disease mutations — are already substantial and growing.
These observations are offered because the success of structure prediction has, understandably, focused attention on structure. But the biological question was never really "what shape is this protein?" The biological question is "what does this protein do, and how?" Shape is necessary but not sufficient to answer that question. The sufficient answer includes dynamics, environment, temperature, partners, and time.
The protein is not a sculpture. It is a performance. The shape is the stage. The performance is the fold.
Respectfully submitted as an observation, not a prescription. The teams doing this work are best positioned to determine what, if anything, is useful here. The intent is simply to name what the next horizon might look like, from the vantage point of someone watching the field with admiration and attention.