Introduction

Knowing a protein’s amino-acid sequence is only part of the story. Its three-dimensional structure helps determine what it can bind to and how it behaves. For decades, finding that structure usually meant relying on experiments such as X-ray crystallography, nuclear magnetic resonance, or cryogenic electron microscopy. Those methods remain essential, but they can be slow and difficult for some targets.

AlphaFold, developed by Google DeepMind, changed the practical limits of computational structure prediction. It can often produce a useful structural hypothesis in far less time than an experiment. That does not make the experiment unnecessary; it gives researchers a strong place to start.

The Protein Structure Prediction Problem

A protein folds through interactions among its amino acids and the surrounding environment. The number of possible conformations is enormous, and a sequence alone does not reveal the full folding pathway.

AlphaFold tackles a narrower and more useful task: predicting a likely structure from sequence and related biological information. It predicts an end structure, not a frame-by-frame account of how the molecule folded or how it moves in a cell.

How AlphaFold 2 Works

AlphaFold 2 combines several ideas:

  1. Multiple sequence alignments (MSAs): Related sequences reveal patterns of co-evolution. If two positions tend to change together across species, they may be close in the folded protein.
  2. Pair representations: The model tracks possible relationships between residue pairs as well as information about individual residues.
  3. The Evoformer: Attention-based blocks repeatedly exchange information between the MSA and pair representations.
  4. Invariant Point Attention: The structure module reasons about residues in three-dimensional space while respecting rotations and translations.
  5. Recycling: A prediction is fed back through the network so that the model can refine it.

AlphaFold 2 also reports confidence estimates. The per-residue pLDDT score helps show which local regions are more or less reliable, while predicted aligned error (PAE) helps assess the relative placement of domains or chains. A high local score does not guarantee that every interaction or domain orientation is correct.

From AlphaFold 2 to AlphaFold 3

AlphaFold 3 is not simply AlphaFold 2 with a longer list of supported molecules. It introduces a different architecture, including a Pairformer and a diffusion-based structure module. It is designed to model complexes containing proteins, nucleic acids, small molecules, ions, and modified residues.

This wider scope is useful for studying molecular interactions, but the same caution applies: a prediction is a model-generated hypothesis. Binding affinity, reaction mechanism, cellular context, and molecular dynamics require additional methods and evidence.

Scientific Impact

AlphaFold 2’s performance at CASP14 marked a major step in protein structure prediction. The public AlphaFold Protein Structure Database later made predicted structures available at an unprecedented scale. Researchers use these predictions to generate hypotheses, interpret experiments, compare protein families, and plan protein-engineering work.

In 2024, the Nobel Prize in Chemistry recognized two related achievements. One half went to David Baker for computational protein design. The other half went jointly to Demis Hassabis and John Jumper for protein structure prediction. Baker was not an AlphaFold creator; the award connected complementary advances in predicting structures and designing new proteins.

What AlphaFold Cannot Settle on Its Own

AlphaFold predictions need careful interpretation, especially when the question involves:

  • flexible or disordered regions;
  • alternative conformations and molecular motion;
  • the effect of pH, temperature, membranes, or other cellular conditions;
  • binding affinity and reaction kinetics;
  • uncommon ligands, modifications, or complexes that differ from the training data;
  • subtle stereochemical details.

Confidence scores help, but they are not experimental validation. Researchers should compare predictions with known biology and use suitable structural or biochemical experiments when the conclusion matters.

Conclusion

AlphaFold did not “solve protein folding” in every sense of the phrase. It made high-quality structure prediction practical for a huge range of research questions. Its greatest value is as a scientific tool: it narrows the search space, suggests experiments, and helps researchers ask better questions about molecules that were previously hard to study.

References