Protein Structure Prediction After AlphaFold: What Was Solved and What Was Not

For several decades, the primary constraint on molecular biology was the physical labour required to determine the shape of a single protein. The problem is a matter of energetic scales. A protein is a long chain of amino acids, often hundreds or thousands of units long, which must collapse into a specific three-dimensional arrangement to perform its task. While the sequence of these acids is dictated by the genetic code, the final shape is determined by a complex interplay of hydrogen bonds, hydrophobic interactions, and van der Waals forces.

Until recently, determining these shapes relied almost entirely on experimental methods such as X-ray crystallography. This process involves purifying a protein, coaxing it to form a solid crystal lattice, and bombarding it with high-energy radiation to infer its geometry from the resulting diffraction patterns. Many proteins, particularly those embedded in cell membranes or containing disordered regions, resist crystallisation for years. The gap between the number of known protein sequences and the number of solved structures grew wider with every new genome sequenced.

Computational models attempted to bridge this gap by simulating the physics of folding from first principles, but the sheer number of possible configurations—famously described by Levinthal’s paradox—rendered direct simulation impractical for all but the smallest molecules. The arrival of deep learning models, specifically the AlphaFold system developed by Google DeepMind, fundamentally altered this landscape. By treating the protein folding problem as a pattern recognition task rather than a pure physics simulation, these models achieved levels of accuracy that rival experimental methods.

However, the declaration that the protein folding problem is solved warrants careful qualification. While the geometry of most proteins can now be predicted with high confidence, the transition from a static coordinate map to a functional understanding of biology is not linear. A structure is a snapshot of a molecule at rest, yet life is governed by movement, chemical affinity, and the transient interactions between different molecular species. Predictable geometry is the beginning of the inquiry, not the resolution.

The transition from crystallography to computation

The shift from experimental observation to computational prediction represents a change in the evidentiary standards of biology. In the classical model, a structural biologist spent a career perfecting the conditions necessary to freeze a protein in time. This resulted in the Protein Data Bank (PDB), a repository of roughly 200,000 structures that served as the primary training set for modern machine learning. The PDB is a record of physical truth, but it is also biased toward proteins that are easy to stabilise.

Deep learning systems changed the methodology by moving from a bottom-up physical approach to a top-down probabilistic one. Earlier software, such as Rosetta, attempted to find the global free-energy minimum by sampling different angles of the protein backbone. These methods were computationally expensive and often failed for large proteins because the search space was too vast. The current generation of models bypasses this by learning the hidden rules of spatial relationships directly from the archives of previously solved structures.

This transition has effectively commoditised structural information. Where a PhD student might have spent five years determining the structure of a single bacterial enzyme, a researcher can now generate a high-confidence model of that same enzyme in seconds. The limitation is no longer the availability of the structure, but the interpretation of its features. This move toward computation has shifted the bottleneck of the field from structural biology to functional proteomics.

Mechanism of the transformer architecture in spatial folding

The technical foundation of this shift is the transformer, a neural network architecture originally developed for natural language processing. In language, the meaning of a word depends on its context within a sentence; similarly, in a protein, the role of an amino acid depends on its proximity to others in three-dimensional space. The transformer uses an attention mechanism to weigh the importance of these relationships, regardless of how far apart the amino acids are in the linear sequence.

In the context of AlphaFold, the model treats the protein as a graph where each amino acid is a node. The attention mechanism allows the network to iteratively update the position of each node by looking at all other nodes simultaneously. It identifies which residues are likely to be in contact by examining patterns of co-evolution and geometric constraints. This is not a single-step calculation but an iterative process of refinement, where the model constantly adjusts the angles and distances between residues until they converge on a stable configuration.

The transformer architecture allows the model to identify long-range dependencies in a sequence that traditional physics-based simulations often overlook.

The system outputs a set of three-dimensional coordinates for every atom in the protein backbone, along with a confidence score known as pLDDT. This score is a critical feature of the mechanism, as it allows the model to signal when it is guessing. High-confidence regions typically correspond to rigid, well-defined domains, while low-confidence regions often correlate with intrinsically disordered sections of the protein that do not have a fixed shape in nature. The mechanism is a sophisticated form of spatial reasoning that mimics the constraints of physics without explicitly calculating every force.

Evolutionary sequence alignment as a predictive foundation

The success of these models rests heavily on a biological concept known as co-evolution. If two amino acids in a protein are physically touching, they exert selective pressure on one another. If one amino acid mutates to a larger size, the other must often mutate to a smaller size or a different charge to maintain the protein's stability. By aligning thousands of similar protein sequences from different organisms—a process called Multiple Sequence Alignment (MSA)—researchers can identify pairs of amino acids that tend to change in tandem.

These correlations act as a set of evolutionary breadcrumbs. If two positions in a sequence consistently mutate together across millions of years of evolution, it is highly probable that they are close to each other in the folded structure, even if they are hundreds of units apart in the sequence. AlphaFold uses these alignments as a primary input, extracting spatial constraints that narrow the search space for the final structure.

This reliance on evolutionary history means that the models are most accurate when there are many related sequences available in public databases. For entirely synthetic proteins or proteins with no known relatives, the accuracy of the prediction tends to drop. The model is essentially a very sophisticated interpolator of biological history; it predicts how a protein folds because it has seen how nature has solved similar puzzles across the tree of life.

The distinction between structure and biochemical function

A common misconception is that knowing the shape of a protein is equivalent to knowing what it does. Structure provides the stage, but function is the play. A three-dimensional map can reveal a pocket where a small molecule might bind, but it does not necessarily reveal the strength of that binding, the speed of the resulting chemical reaction, or the effect of the protein’s activity on the wider cellular network.

Many proteins are pleiotropic, meaning they perform different tasks depending on their environment or their interaction partners. A static structure from a predictive model represents the protein in isolation, usually in its most stable state. In a living cell, proteins are subject to post-translational modifications, such as the addition of phosphate or sugar groups, which can radically alter their shape and activity. These modifications are not currently predicted by standard folding models.

Furthermore, the goal of drug discovery is often to find a molecule that disrupts a protein's function. While a predicted structure helps identify potential binding sites, it does not solve the problem of binding affinity or specificity. A drug must not only fit into a pocket but must also stay there long enough to be effective and avoid binding to similar pockets in other proteins. Predicting the geometry of the target is merely the first step in a much longer and more uncertain engineering process.

• Prediction models offer a static snapshot of a molecule at rest.
• Biological function is defined by dynamics, conformational changes, and chemical affinities.

The distinction is particularly clear in the study of enzymes. An enzyme may have a well-defined structure, but its function is defined by its ability to transition through high-energy intermediate states during a reaction. These states are fleeting and often involve significant structural distortions that are not captured by models trained on stable, ground-state configurations. Therefore, while the folding problem is largely addressed in terms of geometry, the broader problem of predicting biological behavior remains an open and active area of research.

The constraints of static geometry

Biological systems do not operate in a state of equilibrium. The utility of a protein is rarely found in its rigid shape but rather in its capacity to move. A protease must clamp down on its substrate; a motor protein must swing its lever arm; a membrane channel must open and close its pore in response to electrochemical signals. Current computational models typically produce a single, high-confidence coordinate set for the atoms in a sequence. While this provides a template for the protein's resting state, it frequently obscures the ensemble of configurations that the molecule adopts during its work.

Proteins behave more like vibrating machines than frozen sculptures. They possess intrinsic flexibility, where different segments of the chain move at timescales ranging from picoseconds to milliseconds. A static model might show a pocket that appears closed, yet in a living cell, thermal fluctuations occasionally open that pocket to allow a small molecule to enter. This phenomenon, known as transient opening, is difficult to capture when the training data consists primarily of crystals where the protein is locked into a single position by the forces of the crystal lattice itself.

The discrepancy between a predicted structure and a functional structure becomes acute when considering allostery. This is the process by which a molecule binding at one site on a protein triggers a structural change at a distant site, often several nanometres away. Predicting these long-range communications requires an understanding of how energy propagates through the protein backbone. A static model can identify the two ends of the communication line but often fails to explain the mechanism of the signal transmission. Without an account of these dynamics, the ability to predict a fold does not equate to an ability to predict how that fold will react to its environment.

The difficulty of molecular recognition

Predicting how a protein interacts with a drug candidate or a natural ligand remains a significant hurdle. In pharmaceutical development, the goal is to find a small molecule that fits precisely into a protein pocket, forming specific hydrogen bonds and van der Waals interactions. While AlphaFold and its successors can often predict the shape of these pockets, they are less reliable at predicting the binding affinity, which is the strength of the connection between the protein and the ligand.

Binding is a thermodynamic event. It is governed by the change in free energy when a ligand moves from being dissolved in water to being nestled within a protein. This calculation involves more than just the geometry of the two participants. It requires accounting for the displacement of water molecules, the entropy lost when the ligand stops moving freely, and the subtle shifts in the protein's own shape to accommodate the guest. Current geometric models do not inherently calculate these energetic costs.

Errors of even a few tenths of an angstrom in the positioning of side chains can lead to vastly different predictions of binding strength. Furthermore, many proteins undergo a process called induced fit, where the shape of the binding pocket changes significantly only after the ligand begins to enter. Predicting the structure of the empty pocket is insufficient if the functional pocket only exists in the presence of the drug. Consequently, the pharmaceutical industry still relies heavily on experimental screening and expensive, high-level physics simulations to supplement the predictions made by machine learning models.

A structural prediction provides the map of the lock but does not necessarily reveal how the key must turn.

The challenge of intrinsic disorder

A significant portion of the human proteome does not fold into a stable, three-dimensional structure at all. These are known as intrinsically disordered proteins, or IDPs. Rather than adopting a single shape, they exist as a constantly shifting cloud of configurations. These proteins are not broken or non-functional; their lack of a fixed structure is precisely what allows them to perform their roles. They often act as hubs in signalling networks, capable of binding to many different partners by adopting different shapes for each one.

Traditional structure prediction tools are designed to find the one "correct" fold that minimises energy. When applied to IDPs, these tools often produce low-confidence scores, which researchers initially interpreted as a failure of the model. In reality, the model is correctly identifying that the sequence does not have a single stable state. However, identifying that a region is disordered is not the same as predicting its function.

The behaviour of these disordered regions is governed by the principles of polymer physics rather than the rules of architectural assembly. To understand them, one must predict the statistics of their movement: how expanded or compact the protein cloud is, and how often specific residues come into contact. Because these proteins do not form crystals, there is very little high-resolution experimental data to train models on. This creates a blind spot in our structural understanding of the cell, as many of the proteins involved in diseases like Alzheimer’s and Parkinson’s involve disordered regions that eventually aggregate into toxic clumps.

Data scarcity and multi-protein assemblies

Biological processes are rarely carried out by lone proteins. They occur within large, multi-component machines involving dozens of proteins, nucleic acids, and lipids. While the prediction of individual protein folds has reached a high level of accuracy, predicting how these units assemble into larger complexes is considerably more difficult. The surface area involved in a protein-protein interaction is much larger than that of a drug-binding pocket, and the forces that hold them together are often more subtle.

The primary bottleneck is the lack of diverse training data for complexes. The Protein Data Bank contains many examples of individual folds, but the number of solved structures for large, heteromeric complexes is much smaller. Furthermore, many proteins change their shape when they join a complex, meaning the structure of the isolated protein is a poor guide to its structure when it is functional.

There is also the problem of stoichiometry and specificity. A model might predict that protein A can bind to protein B, but it may not be able to determine if they form a pair, a trimer, or a long chain. In a dense cellular environment, a protein is surrounded by thousands of potential partners. Predicting which ones it will actually bind to requires an understanding of relative concentrations and competing affinities that go beyond the scope of current structural models. The structural library of the cell is being filled, but the manual for how those parts are assembled remains largely unwritten.

Towards integrated molecular dynamics

The next phase of computational biology involves moving from static snapshots to integrated simulations that combine machine learning with classical physics. These hybrid approaches aim to use the speed of models like AlphaFold to generate starting points, which are then evolved over time using molecular dynamics. By simulating the movement of every atom in a protein over microseconds or milliseconds, researchers can observe the transitions between different states and the binding of ligands in real time.

These simulations are computationally expensive, often requiring massive supercomputing clusters or specialised hardware. However, they offer a way to bridge the gap between structure and function. They allow for the inclusion of the solvent—the water and ions that surround a protein—which is essential for realistic behaviour but is usually ignored in simple fold predictions.

Machine learning is also being used to speed up these simulations. Instead of calculating every force from scratch, new models are being trained to predict the "force field" or the energy landscape of a protein, allowing for longer simulations with less computational effort. This could eventually allow for the routine simulation of large complexes and disordered proteins, providing a movie of cellular life rather than a series of still photographs.

The state of the field

The problem of determining the ground-state geometry of a single protein from its amino acid sequence is widely considered to be solved for the majority of folded proteins. This is a monumental achievement that has provided researchers with a vast catalogue of templates for further study. It is now established that deep learning can capture the complex, non-linear relationships between a sequence and its final fold with an accuracy that rivals experimental methods in many cases.

However, several critical areas remain contested or unresolved. The ability of these models to predict the effects of single-point mutations—where one amino acid is swapped for another—is still a matter of debate. While some models can flag these changes as destabilising, they often fail to predict the specific structural consequences, which is a major limitation for clinical genetics. Furthermore, the prediction of how a protein will fold in the complex environment of the cell, aided by chaperone proteins and modified by chemical markers, is not yet possible.

A shift in the field is occurring, moving away from the "static structure" paradigm toward a "dynamic ensemble" paradigm. If researchers can reliably predict the entire range of motion of a protein and its interactions within a crowded environment, it would transform drug discovery from a process of trial and error into a process of rational design. Until then, the structural models provide a necessary foundation, but the transition from knowing a shape to understanding a function remains the primary challenge of modern biophysics.