In 2021, the journal Nature published a paper on AlphaFold, an AI system from DeepMind that solved a 50-year-old challenge in biology. The system predicts the three-dimensional shapes of proteins with unprecedented accuracy. This achievement signaled a new era of AI-powered discovery and transformed how scientists approach complex biological questions.
The Protein Folding Problem
Every function in the human body relies on proteins, which are large molecules built from chains of amino acids. While scientists can easily determine the linear sequence of these amino acids, predicting the intricate 3D structure that chain folds into has been a major computational hurdle. This challenge is known as the protein folding problem.
A protein’s shape is directly linked to its function. For example, an enzyme’s ability to accelerate a reaction or an antibody’s capacity to bind to a virus depends on its specific 3D architecture. Misfolded proteins can lose their function and lead to diseases like Alzheimer’s or Parkinson’s, making the study of their final shape important for medicine and biology.
For decades, determining a protein’s structure required slow, expensive lab techniques like X-ray crystallography or cryo-electron microscopy. These methods could take years for a single protein and did not always succeed. The number of ways a protein could theoretically fold is too large to solve through brute-force calculation, so a reliable computational method was needed to bridge this gap.
The AlphaFold Method Explained
AlphaFold uses a deep learning approach to learn the physical and biological rules governing how a protein folds. The process begins with a protein’s one-dimensional amino acid sequence. The system then consults public databases of known protein sequences to build a Multiple Sequence Alignment (MSA).
The MSA is a collection of thousands of related sequences from different organisms. By analyzing these alignments, AlphaFold identifies evolutionary patterns. For example, if two amino acids in a sequence consistently change together across species, it suggests they are close to each other in the final folded structure, even if they are far apart in the linear chain.
This information is fed into a neural network called the “Evoformer.” The Evoformer processes both the evolutionary information from the MSA and the spatial relationships between amino acid pairs. It refines its understanding of which parts of the protein are near each other, allowing it to interpret the interplay between genetic sequence and physical structure.
The data is then passed to the Structure Module, which translates this information into a three-dimensional model of the protein. This module generates atomic coordinates for each amino acid, producing a detailed structural prediction. The system is trained end-to-end, improving its predictions by comparing them against a database of experimentally determined protein structures.
Key Findings and Accuracy
AlphaFold’s capabilities were tested at the 14th Critical Assessment of protein Structure Prediction (CASP), an event often called the Olympics of protein folding. In this competition, research groups receive amino acid sequences for proteins whose structures are solved but not yet public. The groups submit their computational predictions, which are then judged against the actual laboratory results.
Performance at CASP is measured using the Global Distance Test (GDT), a score from 0 to 100 that assesses how closely a predicted structure matches the experimental one. A higher score indicates a more accurate prediction. A GDT score above 90 is considered on par with results from experimental methods.
The 2021 Nature paper detailed AlphaFold’s performance at CASP14, where it achieved a median GDT score of 92.4. For the most difficult proteins, its median score was 87.0, significantly higher than any other computational method. This level of accuracy showed that the AI could produce structures with reliability competitive with physical experiments.
Scientific Impact and Accessibility
Following the paper’s publication, DeepMind partnered with the European Molecular Biology Laboratory’s European Bioinformatics Institute (EMBL-EBI) to create the AlphaFold Protein Structure Database. This database made hundreds of millions of high-quality protein structure predictions freely accessible to the scientific community.
This open-access database democratized the field, as obtaining a protein structure was previously limited to well-funded labs with specialized equipment. Now, any researcher can instantly look up a predicted structure for a protein of interest. This has accelerated research in areas like drug discovery and vaccine development by allowing scientists to visualize the proteins central to their work.
The database launched with predictions for nearly the entire human proteome—the full set of proteins expressed by humans—and has since expanded to include millions of proteins from other organisms. This resource allows scientists to form new hypotheses and design experiments more efficiently. It has shifted the bottleneck in many fields from obtaining a structure to interpreting its biological function.
Limitations and Future Directions
The version of AlphaFold from the 2021 paper has limitations. It predicts a single, static 3D structure, but proteins are dynamic molecules that move and change shape to function. AlphaFold does not capture this motion or reliably predict how a single gene mutation, a common cause of genetic disorders, might alter a protein’s structure.
The original system was designed for single protein chains and was not optimized to model how multiple proteins interact to form larger complexes. Its accuracy can also be less reliable for proteins that lack a sufficient number of related sequences in the MSA. This can be an issue for newly evolved or artificially designed proteins.
Subsequent research has begun to address these shortcomings. DeepMind released AlphaFold-Multimer, a version trained to predict the structure of protein complexes. Other groups are developing methods to simulate protein dynamics and understand the effects of mutations. These efforts point toward a future where computational tools predict not only what proteins look like but also how they behave inside a cell.