AlphaFold—for predicting protein structures

A daunting problem

The body’s proteins execute a plethora of vital roles inside cells. Their diverse capabilities are intimately linked to the forms that they take after they fold from linear amino-acid chains into three dimensions. Insights into structure can illuminate function and unlock biological mysteries.

More than 60 years ago, the late Christian Anfinsen (National Institutes of Health) showed that an unfurled protein could regain its shape unaided and concluded that its amino-acid sequence encodes its final organization. As a nascent chain configures itself, it cannot try every possibility: Sampling all arrangements would take longer than the age of the universe, even for a modest-sized protein. Yet inside cells, folding can occur within milliseconds, so nature somehow deciphers the problem. In theory, at least, scientists could discern the guidelines that steer amino-acid chains into correct conformations.

Using multiple approaches, hordes of investigators forged tactics that they hoped would capture this information well enough to mold a protein’s architecture from its sequence. They attempted to express physical interactions in energy equations and looked toward X-ray crystallography and, eventually, other methods to produce templates that could serve as blueprints for related proteins. They also combined knowledge about the chemical proclivities of specific amino acids—whether they carry a charge, for instance—with their location along a chain to gain hints about a protein’s structural features.

In 1994, John Moult (University of Maryland) and colleagues began to track structure-prediction progress by assessing techniques through a community enterprise called Critical Assessment of Structure Prediction (CASP). Every two years, participants receive amino-acid sequences of proteins whose structures have been worked out in the lab, but not yet released. Contestants apply the system that they developed and generate predictions. Those models are then compared with the experimental answer and scored.

Over decades, performance plodded upward in small increments, sometimes stalling or even backtracking. One of the early approaches was developed by David Baker (University of Washington), who used short segments from a shared worldwide database called the Protein Data Bank (PDB) to predict local architecture within proteins. Although helpful in some cases, this fragment-based strategy was time consuming and had limited applicability for the vast majority of proteins.

In the meantime, experimentally solved structures accumulated. In 2014, the number of listings in the PDB surpassed 100,000—but that was still a tiny fraction of the tens of millions of protein sequences available at that time.

Bringing AI into the fold

By 2018 and the 13^th CASP competition, scientists had introduced machine learning into their prediction schemes. In contrast to traditional AI approaches that rely on pre-conceived logic, machine-learning systems discover patterns for themselves from data. By making machine learning the central component of their protein-structure prediction network, Hassabis and Jumper’s team won CASP13 with a hefty lead in accuracy over the runner-up and almost a 50% improvement since the best of CASP12. Despite this success, the DeepMind researchers were unsatisfied: They wanted a tool that experimentalists would find useful, with errors of less than one angstrom, the size of an atom.

Hassabis, Jumper, and the AlphaFold team started over and brainstormed intensively. They added geometric and genetic concepts, and they integrated established wisdom about proteins. Atoms have characteristic radii, for instance, and bonds have characteristic angles. The group aimed to include these factors in such a way that they didn’t interfere with the system’s power to learn for itself.

The researchers devised ways to extract maximal information from limited experimental data, and they deployed strategies that force AlphaFold2 to learn efficiently. They allowed the network to adjust calculations anywhere in its process—all the way back to the beginning—as it toils. This innovation avoided a previous pitfall of locking in early errors. Throughout, the system iteratively hones its developing structural model by re-feeding itself tentative solutions.

Hassabis, Jumper, and colleagues also discarded principles that had guided traditional algorithms. For example, they ignored linear proximity in favor of three-dimensional relationships, as amino acids that lie hundreds of subunits away can reside together in a folded protein. Moreover, the team boosted the importance of physical closeness by inventing a mechanism that pays special attention to amino acids that are in contact.

No single element was decisive on its own; rather, many ingenious new ideas combined to achieve a breakthrough performance.

AlphaFold2 takes shape

AlphaFold2 starts with a sequence and crawls databases to find similar ones. It lays out these evolutionary family members as strings of amino acids, one atop the others. It also creates a matrix of information about every possible pair of amino acids within the protein of interest, beginning with their identities, their linear distance, and which direction they lie relative to one another in the chain. These two data sets—the multisequence alignment (MSA) and the pair representations—are processed in parallel during the first stage of AlphaFold2, called the Evoformer. If structures for related proteins have been determined, the system can use those too. Early on, the Evoformer develops a crude structural hypothesis, which it tests and refines. At every step, it tries to fit its model to what it has figured out thus far.

Evoformer measures, for instance, distances between each pair among a set of three amino acids and then assesses whether they form a triangle whose sides meet. If not, this impossibility will eventually need to be resolved, but the discrepancy can be put aside and revisited. In this way, geometry inspires but does not constrain the Evoformer’s activities.

AlphaFold2 incorporates an especially powerful innovation that allows the MSA track, which reflects evolutionary relationships, to communicate with the pairwise representation track, which reflects spatial relationships. As information flows, each path can leverage knowledge that the other one has gained and thus sharpen its own work.

For instance, if the MSA track identifies two amino acids that don’t change over the course of evolution or that co-vary, it can alert the pair-representation track that these amino acids might physically interact. Conversely, if the pair-representation track homes in on possible amino-acid neighbors, it can tell the MSA, which can check whether analogous amino acids in related proteins have coevolved in ways that support that scenario. In this way, crosstalk between the two tracks helps each improve its hypotheses.

After the Evoformer decrypts as many portions of the structure as it can, it passes them to the so-called structural module, which assembles them into a coherent three-dimensional protein. As the structural module jiggles the pieces around, they continue to morph. Initially, it gives every amino acid a position and an orientation that builds a nonsensical conglomeration: They all lie in the same place.

Step by step, it rotates and moves the amino acids, still ignoring which ones are linearly adjacent. Eventually, the protein’s backbone emerges and the system places the chemical side chains that characterize each amino acid. AlphaFold2 predicts not only the entire 3D structure, but also reliability scores for each part.

Rigorous training

To train the system, Hassabis and Jumper’s team used the PDB’s experimentally established structures. AlphaFold2 repeatedly compared its proposals to the real answer and gradually nudged its solutions closer to reality. By repeating this process on every member of the training set, the algorithm absorbed principles of protein structure.

The researchers exploited tricks that pushed the network to learn better. For instance, they hid amino acids in the MSA and asked it to fill in the gaps. In this way, they demanded that the system master rules of evolutionary relationships. They also recursively supplied outputs from any given step, which provided many opportunities for AlphaFold2 to reconsider and refine.

AlphaFold2 also computed how much to trust its predictions, and these confidence ratings allowed the researchers to squeeze more information from the available data and thus, to enhance its performance. After they fed it the roughly 140,000 PDB sequences, they ran another set whose structures had not been solved. From the predictions, they plucked the most reliable 350,000 sequence/structure pairs and trained the system on those data as if they had been experimentally verified.

Retooling protein science

In 2020, AlphaFold2 soared past the competition in CASP14. Its predictions were accurate to atomic precision and it generated excellent results in minutes even for proteins that lacked a template. This was the first approach that could construct high-resolution predictions in cases where no similar structure is known.

In July 2021, Hassabis and Jumper published their method as well as structure predictions of almost every human protein. In only two years, their manuscript’s impact has vaulted over almost all of the 100,000 research articles that have been published in Nature since 1900. It ranks 50^th, having been cited in more than 7000 papers from top journals.

In collaboration with the European Molecular Biology Laboratory’s European Bioinformatics Institute, Hassabis and Jumper have shared the program and the database with the scientific community, and more than a million investigators have used these resources. The DeepMind team has since expanded its catalog to almost every known protein in organisms whose genomes have been sequenced. Listings include proteomes of, for instance, viruses that pose epidemic threats and the World Health Organization’s high-priority pathogens.

The technology has already made a dramatic impact in myriad biomedical spheres and beyond. It helped researchers fill holes in their visualization of the nuclear pore complex, an enormous and complicated molecular machine that controls transport into and out of the nucleus. Scientists used the tool to analyze a bacterial syringe that shoots molecules into insect cells. By applying understandings that AlphaFold2 revealed, the investigators reengineered the protein to target human cells, opening a new avenue toward medication delivery and gene therapy. Academic laboratories and companies are harnessing AlphaFold2 to develop vaccines, design drugs, craft enzymes that chew up pollutants, and much more. The prospects are endless.

By letting their imaginations and talents fly, Hassabis, Jumper, and their team completed a quest that had flummoxed scientists for half a century. This triumph has launched a new era in studying and manipulating proteins. It has already catalyzed substantial advances, and its impact and reach promise to explode as workers in a vast range of fields dream up new ways to mine its potential.

by Evelyn Strauss

Selected Publications – Demis Hassabis and John Jumper

Senior, A.W., Evans, R., Jumper, J …..…. Kavukcuoglu, K. and Hassabis, D. (2019). Protein structure prediction using multiple deep neural networks in the 13^th Critical Assessment of Protein Structure Prediction (CASP13). Proteins: Structure Function and Bioinformatics. 87, 1141-1148.

Senior, A.W., Evans, R., Jumper, J …..…. Kavukcuoglu, K. and Hassabis, D. (2020). Improved protein structure prediction using potentials from deep learning. Nature. 577, 706-710.

Tunyasuvunakool, K., Adler, J …..…. Jumper, J. and Hassabis, D. (2021) Highly accurate protein structure prediction for the human proteome. Nature. 596, 590-596.

Jumper, J., Evans, R …..…. Kohli, P. and Hassabis, D. (2021). Applying and improving AlphaFold at CASP14. Proteins: Structure Function and Bioinformatics. 89, 1711-1721.

Jumper, J., Evans, R., Pritzel, A …..…. Kohli, P. and Hassabis, D. (2021). Highly accurate protein structure prediction with AlphaFold. Nature. 596, 583-589.

Jumper, J. and Hassabis, D. (2022). Protein structure predictions to atomic accuracy with AlphaFold. Nature Meth. 19, 11-26.