From Sequence to Shape: How AlphaFold Cracked a 50-Year Problem

Proteins are long chains made of small building blocks called amino acids. The order in which the amino acids are arranged in the chain is called the amino acid sequence, and each position where a single amino acid is placed—every spot along the chain—is called a residue. Inside the cell, this chain naturally folds together into an intricate three-dimensional shape. This process where a single-file chain bundles up into a specific spatial shape is called folding, and the resulting shape is called the protein's structure.

The folded shape determines what that protein does. Enzymes speeding up chemical reactions, antibodies catching pathogens, and receptors on the cell surface passing along signals all come from that folded shape. That is why figuring out what 3D structure a protein will fold into just by looking at its amino acid sequence—in other words, structure prediction—was one of biology's biggest unsolved problems for over half a century.

In July 2021, John Jumper's team at DeepMind published "Highly accurate protein structure prediction with AlphaFold" in Nature (PMID 34265844). This article reads that single paper from start to finish. We will follow, step by step, why this paper is considered a landmark, what questions it asked, what methods it used to answer them, what it left behind, and what it still cannot do. Where previous attempts hit a wall before reaching this paper is covered in "Why It Took Fifty Years: Protein Structure Prediction and Its Long Trail of Failures".

Sequences Were Abundant and Structures Were Scarce — The Gap This Paper Narrowed

The root of the problem is old. A chain can take an astronomical number of shapes, but an actual protein does not go through all those shapes one by one; it finds its own unique shape within a biologically reasonable time. This puzzle is Levinthal's paradox. The paradox itself was explained by the folding funnel theory (PMID 36659994).

The principle that sequence determines structure was also known early on. However, translating that principle into a prediction algorithm that 「takes a sequence and outputs coordinates」 repeatedly ran into walls (PMID 10550209). Coordinates are numbers recording where each individual atom is placed in space, and the paper calls this task the structure prediction part of the protein folding problem (PMID 34265844).

There are ways to find structures through experiments. They are X-ray crystallography, nuclear magnetic resonance, and cryogenic electron microscopy (cryo-EM). However, figuring out the structure of a single protein takes anywhere from months to years. As a result, proteins whose structures have been revealed by experiments stop at around 100,000 species, while known sequences reach into the billions (PMID 34265844). Sequences were exploding while structures were stuck in the bottleneck of experiments, and filling this massive gap with computation was the goal of structure prediction.

The reason this paper is a landmark is that it actually narrowed that gap. The fairest stage to gauge how accurate predictions are is CASP (Critical Assessment of protein Structure Prediction). CASP is a blind structure prediction exam held every two years. It gives out structures that have just been solved by experiments but not yet made public as problems, and each team submits predictions without knowing the answers (PMID 34265844).

Looking at one scoring method from CASP14 in 2020, it goes like this. The predicted structure is superimposed onto the actual structure, and it measures how much the backbone atoms running along the chain, in other words the backbone, are misaligned on average. This is a value measured by excluding a small number of sites with unusually large misalignments, and the smaller it is, the more accurate. The middle value when lining up the values of multiple problems in order of size, in other words the median, was 0.96 Å (angstroms) for AlphaFold. The gap becomes clear when placed side-by-side with the runner-up method's 2.8 Å. The Å is a unit of length about the size of a single atom, and the width of a single carbon atom is about 1.4 Å. AlphaFold's misalignment, viewed as a median, was smaller than the width of a single carbon atom, an accuracy approaching the atomic level. The authors described this as accuracy rivaling experimentally determined structures in the majority of cases (PMID 34265844).

The independent evaluation committee of the competition also confirmed this leap. A separate paper analyzing CASP14 evaluated that with the application of the latest deep learning approaches, the 「high accuracy」 category expanded to all targets, and the major leap in high accuracy came from AlphaFold2. A target is a protein that came out as a problem in that round. The same evaluation also recorded the fact that the second-place method of CASP14 surpassed the first place of CASP13. The evaluation committee saw this as showing how community benchmarking, where multiple teams compete through the same exam, plays a role in the development of the structure prediction field (PMID 34218458).

Can It Get the Right Answer Even Without a Similar Model? — The Question Posed by the Paper

The core question was clear: Can protein structures be predicted at atomic-level accuracy using only the amino acid sequence, even when no similar structures are known at all?

The caveat "when no similar structures are known" sets the weight of this question. Related proteins that diverged from the same ancestor are called homologs, and homologs resemble each other in both sequence and shape. Therefore, if a homolog with an already solved structure exists, it can be used as a model, or template, to draw the shape of a new sequence. Existing methods fell far short of atomic-level accuracy, especially when homologs were absent (PMID 34265844). What AlphaFold targeted was precisely this difficult side.

The paper takes two traditions of structure prediction as its background. One is the approach of simulating physical interactions between atoms. It is theoretically attractive, but calculations become unmanageable even for medium-sized proteins. The other is the bioinformatics approach of utilizing evolutionary history. This approach has risen to prominence, driven by the growth of the Protein Data Bank (PDB), a public repository of experimental structures; the explosion of genomic sequences; and deep learning—namely, neural networks, which are multi-layered computational networks that learn patterns on their own from many examples (PMID 34265844).

The design philosophy of AlphaFold is to combine these two. It embeds the physical, geometric, and evolutionary constraints on protein structures directly into the architecture of the neural network itself. Instead, handcrafted features chosen and predefined by humans are minimized, allowing the model to learn directly from PDB data.

From Alignment Table to Three-Dimensional Shape — The Two Steps Inside AlphaFold

The AlphaFold neural network receives two inputs. One is the primary amino acid sequence of the protein to be predicted, which is the exact order in which the amino acids are laid out. The other is a multiple sequence alignment (MSA), which is created by lining up the homologous sequences of that protein together. A multiple sequence alignment is a table where a batch of related sequences are gathered and matched residue by residue so they stand side-by-side vertically and horizontally. By reading pairs of positions in this table that have changed together through evolution, one can obtain constraints regarding the structure. The output is the three-dimensional coordinates of every atom excluding hydrogen, meaning every heavy atom, and AlphaFold outputs these coordinates directly (PMID 34265844). The network is broadly divided into two steps.

The first step is the trunk of the network, which is built by stacking many computational blocks called Evoformer. Evoformer is a block newly proposed in this paper, viewing structure prediction as a graph reasoning problem in three-dimensional space. A graph is a network of relationships consisting of points connected by lines, where the points are residues and the lines are relationships between nearby residues.

Evoformer carries around two types of representations. A representation is a numerical memo that the neural network writes down inside after reading the inputs. The MSA representation is a memo transcribing the alignment table, containing one residue of the input sequence per column and one sequence where those residues appear per row. The pair representation is a table that records the relationship between the two for every possible pair of residues (PMID 34265844).

The core of Evoformer is constantly passing information back and forth between these two representations. The MSA representation updates the pair representation at every block. The pair representation in turn adds a bias to the attention that reads the MSA representation, closing the loop of information flow. Attention is a method where the neural network gathers information by placing greater weight on parts of the many inputs that are important for the current calculation. This means that the relationships between residues written in the pair representation are reflected in where the alignment table should be focused and read.

Inside the pair representation, 「triangle」 updates appear. For three distances connecting three residues to actually be placed as a single three-dimensional structure, they must satisfy constraints like the triangle inequality. The triangle inequality is the rule that the sum of the lengths of two sides in a triangle cannot be shorter than the remaining side. If residues A and B are close and B and C are also close, A and C cannot be located very far apart. Utilizing this intuition, 「triangle multiplication updates」 that update a third side using two sides and attention arranged in a triangular shape were introduced (PMID 34265844).

The second step is the structure module. The structure module is the part that takes the representations processed by the trunk and molds the actual three-dimensional structure. It views the backbone of each residue as a small frame, and represents it by which way that frame points in space (rotation) and where it is placed (translation). The frames of the residues are not bound in a chain and move separately like gas molecules, which is why this representation is called a residue gas. Starting from a meaningless state where all residues are overlapped at the origin, it rapidly develops into an elaborate structure (PMID 34265844).

Invariant point attention (IPA), newly introduced in the structure module, calculates attention by placing three-dimensional points in the local coordinate system of each residue. The local coordinate system is a scale based on each residue's own frame. Therefore, the result does not change even if the entire structure is rotated or translated as a whole. Since a protein is the same protein no matter which side you look at it from, this means the answer does not waver depending on the viewing direction.

The neural network is modified during training to reduce the loss function. The loss function is a score that grades how much the prediction missed the correct answer. AlphaFold's final loss function, FAPE (frame-aligned point error), compares the predicted atom positions with the actual positions while overlapping them according to various criteria. As a result, a strong bias is applied so that atoms are placed correctly based on each residue's local coordinate system (PMID 34265844).

There is one more principle repeated throughout the network. Recycling is a method where the output of the network is fed back into the same module to be refined repeatedly. It noticeably boosts accuracy without significantly increasing training time (PMID 34265844).

There is also a device in the training method. AlphaFold reaches high accuracy solely through supervised learning, which learns from examples with correct answers attached, meaning solely from experimental structures in the PDB. It goes a step further with self-distillation, which uses its own predictions as study material again. It predicts the structures of roughly 350,000 sequences from a sequence collection called Uniclust30 using the trained network, and uses a high-confidence subset among them as new data to train from scratch again. Vast sequence data without structural answers attached are utilized this way (PMID 34265844).

It also trains together on a BERT-style objective that masks some residues of the MSA and has the network guess the original residues. BERT is a language model that learns by guessing masked words in a sentence, and that training method has been brought over to the alignment table. Doing this induces the network to interpret on its own the phylogenetic relationships of sequences and the relationships between positions that change together (co-evolution).

Above all, AlphaFold assesses for itself residue by residue how much its own predictions can be trusted. pLDDT is a confidence score pre-assigned by AlphaFold indicating how accurate the prediction of each residue will actually be. The fact that this self-evaluation matches actual accuracy well holds great significance later on (PMID 34265844).

A Misalignment Smaller Than the Width of a Single Atom—And the Confidence It Assigned Itself

Summarized in numbers, it looks like this. In CASP14, the median backbone misalignment was 0.96 Å for AlphaFold and 2.8 Å for the runner-up method. Even when measured based on all atoms rather than just the backbone, it led with 1.5 Å versus 3.5 Å (PMID 34265844). It even predicted a gigantic protein of 2,180 residues with no structural homologues with accurate domains and domain arrangement. Domains are lumps that fold separately inside a large protein.

Even outside the special stage called CASP14, this accuracy was maintained in a sample of PDB structures newly released after the training data cutoff (PMID 34265844). This means the performance carried over even to structures that could not be seen during training.

The confidence score pLDDT reliably predicted the actual accuracy (lDDT-Cα). lDDT-Cα is a score that measures how well the distances around the central carbon (Cα) of each residue backbone match between the predicted and actual structures. The TM-score, a global metric that checks how well the entire structure matches when overlaid all at once, was also well estimated. The property of outputting "how much this prediction can be trusted" alongside the prediction is the decisive feature that made AlphaFold a tool that is actually used. This is because by marking low-accuracy parts in advance, researchers can judge how far to trust them and from where to verify them with experiments.

The paper also looks into how the network reaches its answer. By attaching a separate structure module to each of the 48 Evoformer blocks in the trunk, it extracted the structure that the network considers most plausible up to that block. Including the recycling steps, it obtained a trajectory where 192 intermediate structures followed one another in sequence. After the first few blocks, the structure was smoothed out remarkably, refined little by little until it could no longer be improved (PMID 34265844). On difficult targets like SARS-CoV-2's ORF8, the network used almost the entire depth to rearrange the secondary structure before settling on a good structure. Secondary structures are small shapes like helices or flat strands that the chain forms here and there.

What Took Months Now Takes Minutes: How Speed Became Scale

The shock of AlphaFold came from its speed. Because it outputs coordinates directly, depending on protein length, a graphics processing unit (GPU) can produce a structure in minutes to hours (PMID 34265844). Work that took months by experiment became a calculation that finishes practically instantly.

This speed soon led to scale. In a companion paper, the same team applied AlphaFold2 to the human proteome, meaning the entire set of proteins that humans possess. As a result, they predicted 98.5% of human proteins, covered 58% of all residues at a reliable level, and among those, a portion corresponding to 36% of all residues had very high confidence. The residues whose structures were determined by experiment were 17% (PMID 34293799). Reliable prediction covered an area vastly broader than the share experiment had covered over decades, all at once.

Another study quantified how much this database increased the structural coverage of human proteins, meaning the share whose structure is known. This share rose from 48% using experimental structures and template-based models alone to 76%. The "dark proteome," whose structures could not be guessed, shrank from 26% to 10%. The authors evaluated that the sequence-structure gap for human proteins has nearly vanished. The contribution was actually even greater in non-human species, such as pathogenic bacteria, than in humans (PMID 35073311).

The ripples spread beyond prediction into new tools. AlphaMissense was created by fine-tuning AlphaFold, meaning training an already learned neural network a bit further to suit a new task. This tool predicts the pathogenicity, meaning the likelihood of causing disease, of every possible single amino acid substitution in humans. Genetic mutations that swap one amino acid for another are called missense variants, and AlphaMissense classified 89% of missense variants as "likely benign" or "likely pathogenic" (PMID 37733863). Clues for judgment were attached to variants whose clinical meanings were mostly unknown.

Structure prediction networks also became the foundation for protein design. RFdiffusion is another structure prediction network, RoseTTAFold, fine-tuned into a diffusion model. A diffusion model is a generative method that produces new shapes by gradually peeling away noise. RFdiffusion actually designed new binding proteins and symmetric assemblies and verified them with experiments (PMID 37433327). This current of drawing and building desired proteins is covered in "From Prediction to Design — After AlphaFold, the Science of Drawing Proteins and Building Them".

In 2024, AlphaFold 3 was released, newly designed on a diffusion basis. This model predicts the structures of complexes where multiple molecules gather as a single mass, not only proteins but also nucleic acids, small molecules, ions, and modified residues (PMID 38718835).

Why Saying It Was "Solved" Is an Exaggeration — Limitations the Paper Itself Wrote Down

The contest organizers' summary evaluated it as a solution to the classic folding problem, at least regarding single proteins. However, calling the entire folding problem "solved" is an exaggeration. The paper itself clearly sets down several limitations.

First is MSA dependence. If the alignment depth, meaning the median number of sequences gathered in the multiple sequence alignment, becomes shallow to roughly less than 30 sequences, accuracy drops significantly (PMID 34265844). This means that proteins with sparse evolutionary neighbors are difficult to predict.

Second is cross-chain contact. In complexes where multiple chains gather, AlphaFold is much weaker for proteins with few contacts within their own chains or among chains of the same kind, but many contacts with chains of a different kind (heterotypic contacts) (PMID 34265844). Proteins whose shapes are determined by interactions with other chains inside a large complex fall into this category. This limitation led afterward to complex-dedicated systems such as AlphaFold-Multimer and AlphaFold 3 (PMID 38718835).

Third is the intrinsically disordered region. A disordered region is a portion of a sequence that does not fold into a fixed shape under normal physiological conditions. Interestingly, this weakness turned into a signal instead. When replacing conventional modeling with the AlphaFold model in Rosetta's predictor (ResidueDisorder) that finds disordered regions in sequences and updating the scoring function together, the performance for distinguishing disordered and non-disordered residues and ranking them (AUC, higher is better) rose from 0.69 to 0.78. The accuracy determined as either of the two also rose from 62% to 74%, reaching a level equal to or better than existing methods (PMID 36251522). The weakness of being an unfolded region became a clue to finding that region.

Fourth is the fundamental limitation of a static structure. AlphaFold produces a single most plausible structure that might appear in the PDB. It does not output the various shapes that proteins actually transition between (conformational ensembles) or the folding stability ($\Delta$G), which indicates how much more stable the folded state is than the unfolded state. In fact, a method that modeled the folded and unfolded states together performed better than AlphaFold-based metrics in selecting de novo designed protein candidates (PMID 41565673).

A recent review synthesizing these limitations acknowledges that AlphaFold has transformed structural biology over the past three years. At the same time, it notes that it remains vulnerable in disordered regions, interactions that proteins form with ligands or cofactors, and very large or transient complexes. A ligand is a small molecule that binds to a protein, and a cofactor is a helper molecule needed alongside a protein for it to work. The review also points out that on community benchmarks, roughly one-third of residues may lack atomic-level precision (PMID 41326937).

And AlphaFold is a tool that assists rather than replaces experiments. The paper itself also emphasized its usefulness in aiding molecular replacement and cryo-EM map interpretation (PMID 34265844). Molecular replacement is a method to solve structures from X-ray crystallography data using a similar model as a clue, and a cryo-EM map is the three-dimensional density map produced by cryo-electron microscopy. The review likewise nails down that orthogonal experimental validation, confirming things separately with experiments based on different principles, is still necessary (PMID 41326937). How these limitations and impacts are read within the lineage of deep learning is continued in the final sections of "From Backpropagation to AlphaFold: What Deep Learning Stood On".

Even so, what this paper left behind is clear. It practically narrowed the half-century bottleneck of the gap between sequence and structure, transforming structural biology into a field where computation takes the lead. New questions after AlphaFold start from here.

---

This article is intended to convey scientific information easily and does not serve as a basis for specific research, diagnostic, or therapeutic decisions. We encourage you to check the original paper and follow-up research directly.

Source Papers

• Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature 596:583–589 (2021). PMID 34265844 — The target paper closely read by this article. The source for all methods, results, and discussions in the text, including the Evoformer, structure module, IPA, FAPE, recycling, CASP14 results, and limitations.
• Pereira J, et al. High-accuracy protein structure prediction in CASP14. PMID 34218458 — Independent evaluation at CASP14.
• Tunyasuvunakool K, et al. Highly accurate protein structure prediction for the human proteome. PMID 34293799 — The companion proteome paper.
• The structural coverage of the human proteome before and after AlphaFold. PMID 35073311 — Quantification of coverage and the dark proteome.
• Review of the Levinthal paradox and the protein folding problem. PMID 36659994; PMID 10550209 — Historical background of the fifty-year challenge.
• Review of AlphaFold applications and limitations (2022–2025). PMID 41326937 — The one-third residue precision limit, IDRs, ligands, and experimental assistance.
• Rosetta ResidueDisorder + AlphaFold2 disorder prediction. PMID 36251522 — Replacing ab initio models with AlphaFold models and updating the scoring function to improve the disorder prediction AUC from 0.69 to 0.78.
• Limits of folding stability ($\Delta G$) prediction. PMID 41565673 — Static structure limitations of AlphaFold-based metrics.
• AlphaMissense: Pathogenicity of missense variants. PMID 37733863 — An AlphaFold-derived tool.
• RFdiffusion: Protein design. PMID 37433327 — Design based on structure-prediction networks.
• AlphaFold 3: Biomolecular interaction prediction. PMID 38718835 — Diffusion-based successor.