A protein rarely works alone. In a cell, it may bind to another protein, interact with DNA or RNA, or attach to a small molecule. Understanding those encounters can help researchers investigate how biology works—and where it goes wrong. But determining the shapes of molecular complexes can be painstaking work.
AlphaFold 3, presented by Google DeepMind and Isomorphic Labs in 2024, aims to help with that challenge. It uses artificial intelligence to predict the three-dimensional arrangements of many biological molecules, including how they may fit together. The advance is not a machine discovering medicines on its own. It is a new way to generate structural hypotheses that scientists can examine and test.
From protein shapes to molecular relationships
The first AlphaFold made headlines for predicting the structures of proteins from their amino-acid sequences. Proteins are built from chains of amino acids that fold into shapes related to their functions. In 2021, researchers made a large collection of AlphaFold 2 predictions available through the AlphaFold Protein Structure Database, giving scientists a resource for exploring protein structures.
AlphaFold 3 broadens the question. Rather than focusing mainly on the shape of an individual protein, it can predict structures involving proteins alongside DNA, RNA, small molecules, ions and certain chemical modifications. Those components are part of the same intricate molecular environment: a protein’s function can depend on what it binds, and a drug molecule’s potential action depends in part on how it interacts with its target.
The system’s architecture also changed. The research team used a diffusion-based approach to generate molecular structures. In broad terms, diffusion models learn to turn a noisy or incomplete representation into a more organized result. Here, the model produces a proposed arrangement of atoms and molecules, informed by patterns learned from structural data.
What the researchers tested
In their Nature paper, the developers evaluated AlphaFold 3 on prediction tasks involving molecular interactions. They reported improvements over existing methods in several benchmarks, including tests of protein interactions with other molecule types. The performance varied by task; it was not a guarantee that every predicted complex would be correct.
That distinction matters. A structure prediction is a model, not a photograph of a molecule at work inside a living cell. Molecular shapes can shift. The surrounding environment can affect whether a proposed interaction occurs. And a plausible-looking prediction still needs to be assessed against experimental evidence.
Researchers already use several ways to investigate molecular structure, including X-ray crystallography, nuclear magnetic resonance and cryo-electron microscopy. These methods can provide experimental evidence, though preparing samples and collecting or interpreting data can take substantial effort. AI predictions can help researchers decide what to investigate, compare possibilities or make sense of a difficult structural problem. They do not erase the need for laboratory measurements.
A tool for asking better questions
Knowing how molecules may meet can help scientists frame research questions. For instance, a predicted interaction might suggest which parts of a protein to investigate or which candidate molecules merit further study. In drug research, structural predictions can inform a search for compounds that might bind to a biological target.
That is a useful role, but a limited one. A predicted fit does not establish that a compound is safe, effective or suitable as a treatment. Drug development involves many stages of testing, and biological systems are more complicated than a single molecular structure. AlphaFold 3 can contribute information to that work; it cannot turn a prediction into a validated therapy by itself.
The model’s release also raised questions about access. When AlphaFold 3 was announced in May 2024, researchers could use a web server under conditions described by its developers, while the full model was not initially made openly available in the same way as AlphaFold 2. That difference prompted discussion among scientists about how to balance broad access to research tools with the developers’ policies for sharing them. The details of access matter because a tool’s scientific value depends partly on who can use it and for what purposes.
Why the advance matters—and what it cannot tell us
AlphaFold 3 reflects a wider change in biology: AI is increasingly being used to make predictions about complex systems that once required researchers to tackle each question with specialized tools. A model that can consider several kinds of molecules together may help scientists explore more possibilities and prioritize experiments.
But more kinds of predicted structures do not mean that biology has become fully legible to a computer. The model is shaped by its training data and evaluation methods, and its results can be uncertain. Scientists need to judge whether a prediction makes sense for the organism, conditions and biological question they are studying.
The clearest promise of AlphaFold 3, then, is not that it replaces structural biology. It is that it can give researchers another way into difficult problems: a detailed proposal to compare with evidence, challenge and refine. When the model is useful, the next step is still a human one—designing the experiment that can show whether the predicted molecular story holds up.
