Why a protein's shape is everything

Proteins do almost all the work in your cells, and what a protein does is governed by how its chain folds into a precise 3D shape. Predicting that shape from the sequence of building blocks alone was an open problem for roughly fifty years. [1]

Near-experimental accuracy

In a 2020 blind assessment known as CASP14, DeepMind's AlphaFold predicted protein structures with an accuracy competitive with laboratory experiments for most targets — a median backbone error under one ångström, versus about 2.8 for the next-best method. [1]

It then predicted structures for essentially every catalogued protein and released them in an open database, growing from about 360,000 structures at launch in 2021 to more than 214 million by 2023 — coverage of nearly all known proteins. [2][3]

A new starting point for medicine

Structures that once took months or years of painstaking lab work can now begin as a computational prediction in seconds. That compresses the first step of research across basic biology, disease understanding, and drug discovery. [2]

The achievement was recognized with a share of the 2024 Nobel Prize in Chemistry, awarded for protein structure prediction and computational protein design. [5]

A prediction is not an experiment

Every AlphaFold model comes with confidence scores, and low-confidence regions should not be treated as truth. The tool usually predicts a single static shape, without the water, ions, or drug molecules that often decide how a protein actually behaves. [4]

For drug design specifically, that matters: in one benchmark, docking against raw AlphaFold models performed clearly worse than against experimental structures. The models also struggle with flexible, disordered regions — a large share of human proteins — and a predicted shape does not tell you a protein's function, its partners, or how a mutation will change it. Structure is not the same as function. [4]

Where this is heading

Newer systems aim to predict how proteins interact with drugs and with each other, and to represent more than one shape at a time. Paired with laboratory validation, structure prediction is becoming a standard first step in biology rather than a final answer. [2]

You will not use AlphaFold directly. But it is quietly accelerating the science behind future diagnostics and medicines — while reminding us that even the most impressive AI predictions still have to be proven in the real world.

What the evidence supports

The useful signal.

  • AlphaFold solved a roughly fifty-year problem: predicting a protein's 3D structure from its sequence, with near-experimental accuracy in a blind test.
  • It released open structures for nearly all known proteins — more than 214 million by 2023 — and shared the 2024 Nobel Prize in Chemistry.
  • This gives biology and drug discovery a powerful new starting point, turning months of lab work into a fast hypothesis.
  • But predictions are models, not experiments: single static shapes, no drug molecules, weaker on flexible regions, and structure does not equal function.
  • The honest frame is a transformative accelerator that still depends on experimental validation.

References reviewed · July 2026

Sources

  1. 01Highly accurate protein structure prediction with AlphaFoldNature (Jumper et al.) · 2021
  2. 02AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy modelsNucleic Acids Research (Varadi et al.) · 2022
  3. 03AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequencesNucleic Acids Research (Varadi et al.) · 2024
  4. 04How good are AlphaFold models for docking-based virtual screening?iScience, Cell Press (Scardino et al.) · 2022
  5. 05AlphaFold wins Nobel Prize in Chemistry 2024EMBL · 2024