AlphaFold Five Years On: Real-World Impact, Limits, and What Comes Next

A milestone in protein science

In 2017 John Jumper joined Google DeepMind, drawn by rumors of a secret project to predict protein structures. Three years later he co-led the team that delivered AlphaFold 2, an AI system that predicted protein structures with near-atomic accuracy and vastly reduced timescales for getting results. That leap solved a decades-old challenge in biology and helped earn Jumper and Demis Hassabis a Nobel Prize in chemistry in 2024.

How AlphaFold works and why it succeeded

AlphaFold 2 was built on transformer neural networks, the same architecture powering large language models. Transformers excel at 'paying attention' to specific parts of a larger puzzle, which proved useful for inferring how strings of amino acids fold into complex three-dimensional shapes. The team also emphasized rapid prototyping: producing models that failed quickly made it possible to try adventurous ideas and iterate fast. They trained models on abundant structural and evolutionary information, which dramatically improved predictions.

Rapid rollout and massive scale

After the original breakthrough, DeepMind released AlphaFold Multimer for multi-protein complexes and later AlphaFold 3, the fastest iteration yet. DeepMind also deployed AlphaFold across UniProt, resulting in predicted structures for roughly 200 million proteins — almost the entire catalog known to science. Jumper stresses these are predictions with caveats, not unquestionable facts.

Unexpected and creative uses

Researchers incorporated AlphaFold into areas the team had not foreseen. One lab used it to study disease resistance in honeybees, an application far removed from mainstream biomedical research. Other groups used AlphaFold as a design verifier: when AlphaFold confidently agreed with an intended synthetic protein structure, researchers would proceed to build it; when it signaled uncertainty, they would avoid costly lab work. That practice sped design cycles as much as tenfold.

AlphaFold also became a search tool. In fertilization studies, scientists compared a known egg protein against thousands of candidate sperm-surface proteins and identified a likely interacting partner that was later validated in the lab. Doing thousands of structural comparisons manually would have been impractical without AlphaFold.

Limits and how scientists work with them

Despite its usefulness, AlphaFold is far from perfect. It is less reliable when predicting interactions among multiple proteins or modeling dynamic interactions and binding with small molecules. Users report occasional borderline or misleading predictions — 'it will bullshit you with the same confidence as it would give a true answer,' as one researcher put it — so scientists interpret results cautiously.

Many labs use AlphaFold to run virtual experiments and narrow down hypotheses before committing to wet-lab work. The tool has augmented experiments and saved time, although it has not replaced the need for laboratory validation.

The next wave: tailoring structure prediction to drug discovery

Startups and academic groups are extending AlphaFold's legacy with models targeted at drug discovery. Examples include Boltz-2, which predicts both protein structure and how well small molecules will bind, and Pearl, a model claiming improved accuracy for drug-relevant queries and offering interactive guidance for additional data.

Some companies aim to reduce prediction error from under two angstroms to below one angstrom. That matters because tiny differences at the angstrom scale can change whether chemical forces enable or prevent binding, and small errors can therefore have large practical consequences for drug design.

Where AlphaFold fits into the bigger picture

Jumper is pragmatic about expectations: protein structure prediction is one important piece of biology, not a cure-all. Predicting a structure can make experiments feasible and cheaper, but it does not by itself produce therapies. Still, researchers are trying to make structure prediction a bigger part of broader discovery workflows, using it as a 'big hammer' to accelerate many tasks.

Looking ahead: combining strengths of different AI systems

Jumper is interested in merging AlphaFold's deep, narrow capabilities with the broader reasoning abilities of large language models. Systems that can read scientific literature, propose hypotheses, and then verify them with structure prediction could change workflows. DeepMind teams are already experimenting with approaches that use one model to generate ideas and another to check them, and Jumper expects more LLM impact on science going forward. He also emphasizes focusing on smaller, tractable ideas over chasing another headline-grabbing milestone.