DeepMind and the European Molecular Biology Laboratory’s European Bioinformatics Institute released a database in July 2021 containing predicted structures for nearly every protein expressed by the human genome. The initial collection covered roughly 20,000 human proteins and included predictions for 20 other organisms important to biological research.
The release transformed AlphaFold from a striking demonstration into accessible scientific infrastructure. Researchers could inspect predictions freely rather than install a complex system or generate every structure themselves. For laboratories studying disease, evolution and drug targets, that dramatically shortened the route from a protein sequence to a useful structural hypothesis.
Shape Is Central to Protein Function
Proteins are chains of amino acids that fold into three-dimensional forms. Their shapes influence how they interact with other molecules and perform work inside cells. Knowing a structure can reveal active sites, explain the effect of mutations and suggest how a potential medicine might bind.
Experimental methods such as X-ray crystallography, nuclear magnetic resonance and cryogenic electron microscopy remain essential, but they can require substantial time, specialist equipment and suitable samples. Many proteins are difficult to characterise experimentally.
The long-standing computational challenge was predicting a reliable structure directly from the amino-acid sequence. The number of possible configurations is enormous, making simple search infeasible.
AlphaFold Changed Expectations
At the 2020 Critical Assessment of Structure Prediction, AlphaFold produced results that organisers described as a major advance. Its predictions approached experimental accuracy for many targets, establishing that deep learning could capture important relationships between sequence and spatial structure.
The model used evolutionary information and an attention-based architecture to reason about which amino acids were likely to sit near one another. It also generated confidence estimates, allowing researchers to distinguish well-supported regions from uncertain ones.
Those estimates were crucial. A predicted structure is not one uniform fact. Some domains may be highly reliable while flexible or disordered regions remain difficult. Displaying confidence helps scientists decide whether a prediction can guide an experiment or needs further verification.
The Database Expanded Access
Partnering with EMBL-EBI placed the predictions inside an established public bioinformatics ecosystem. Researchers could search for a protein, view its structure and download data without paying for access.
The first release was designed to grow. DeepMind said the database would eventually cover hundreds of millions of protein sequences, making structural information available at a scale no experimental programme could quickly match.
Open access increased the chance of unexpected uses. A specialist in a neglected disease, a crop scientist or a small laboratory without large computing resources could work from the same predictions as a major pharmaceutical company.
Predictions Did Not Replace Experiments
AlphaFold predicts an individual protein structure from sequence, but biology is dynamic. Proteins change shape, interact in complexes, respond to chemical modifications and behave differently within cells. A static prediction may omit precisely the state relevant to a disease or treatment.
Experimental evidence therefore remained the standard for many decisions. Predictions could prioritise targets, help interpret data and accelerate the design of experiments, but they did not automatically establish biological function or clinical safety.
Researchers also needed to consider training-data bias. Proteins similar to well-characterised families might be easier to predict than genuinely unusual structures. Confidence scores and independent evidence were necessary for responsible use.
Medicine Was a Promising Application
Structural insights can support drug discovery by identifying pockets where molecules might bind and by clarifying how mutations alter a target. AlphaFold offered a starting point for proteins whose structures had been unknown.
That did not make drug development automatic. A viable medicine must reach the right tissue, behave safely in the body, produce a therapeutic effect and survive rigorous clinical testing. Protein structure addresses only part of that process.
The technology could nevertheless reduce early uncertainty. It also helped research on diseases that attract limited commercial investment, because freely available predictions lowered the cost of asking structural questions.
A Model Became Public Scientific Infrastructure
The most important feature of the July 2021 announcement was not only the number of proteins. It was the decision to distribute the output through a trusted open database. AI research often attracts attention through benchmark performance, while its wider impact depends on whether people can use the result.
AlphaFold showed how a specialised model could augment a global scientific community. Its predictions supplied hypotheses at extraordinary scale, and scientists supplied the judgement, experiments and domain knowledge needed to test them.
That division of labour is a powerful model for AI in research. The system did not replace biology; it changed which questions laboratories could investigate first and how quickly they could begin.
