Overview
On November 30, 2020, DeepMind announced that its AI system AlphaFold 2 had achieved near-experimental accuracy in predicting protein structures — effectively solving one of biology’s grand challenges that had resisted human effort for over 50 years.
At CASP14 (the 14th Critical Assessment of Protein Structure Prediction competition), AlphaFold 2 scored a median GDT (Global Distance Test) of 92.4 out of 100 — a score so high the scientific community initially questioned whether the evaluation was flawed. It wasn’t. The biology field had simply been watching AI cross a threshold they had not expected to witness in their careers.
The Problem
A protein is a chain of amino acids. Its 3D shape — the way that chain folds — determines its function: whether it catalyzes a reaction, builds cell structure, fights infection, or causes disease. For 50 years, biology operated under a fundamental constraint: we could read the genetic sequence of a protein but not reliably predict its shape.
Determining a protein’s structure experimentally (via X-ray crystallography or cryo-electron microscopy) costs months of laboratory work and hundreds of thousands of dollars. The Human Genome Project gave us the sequence of ~20,000 human proteins. Knowing their shapes is what allows us to build drugs, understand diseases, and engineer biology.
As of 2020, only ~170,000 protein structures had been experimentally determined over decades. An estimated 200 million proteins existed in nature, shapes unknown.
How AlphaFold 2 Works
Unlike AlphaFold 1 (2018), which approached the problem modularly, AlphaFold 2 uses a novel architecture combining:
- Multiple Sequence Alignment (MSA): Leverages evolutionary information — if two species both have a functional protein, their amino acids that co-evolve suggest they’re physically close in 3D space
- Evoformer: A specialized Transformer-like module that jointly reasons over sequence and pairwise distance relationships
- Structure module: Directly predicts 3D coordinates using equivariant representations, respecting the physics of rotation and translation
The model is trained on the Protein Data Bank — ~170,000 experimentally-determined structures — and implicitly learns the physical laws governing how amino acids interact.
Impact
Within one year of AlphaFold 2’s release:
- DeepMind published the AlphaFold Protein Structure Database (July 2021) with structures for ~350,000 proteins including the entire human proteome
- By 2022, the database expanded to 200 million protein structures — covering nearly every known protein in biology
- Researchers used AlphaFold predictions to accelerate: vaccine development, cancer research, antibiotic discovery, understanding genetic diseases, and industrial enzyme engineering
In 2024, Demis Hassabis and John Jumper (AlphaFold’s lead researchers) were awarded the Nobel Prize in Chemistry — the first Nobel given for AI-driven scientific discovery.
Why This Matters
AlphaFold 2 marked the first unambiguous demonstration that AI could make a fundamental scientific discovery — not assist human scientists, but solve a problem that human scientists could not solve alone. It collapsed the distinction between AI as a tool and AI as a scientific agent.
If the Transformer was the architecture that unlocked language and reasoning, AlphaFold was proof that the same class of methods could unlock biology, chemistry, and eventually physics. Every subsequent “AI for science” project — AlphaMissense, AlphaGenome, GNoME (materials), weather forecasting — traces its intellectual lineage to November 2020.