AlphaFold 2

AlphaFold 2 is a protein structure prediction program developed by Google-owned AI company DeepMind, first unveiled in November 2020 at the CASP14 competition . It was the first system to solve the decades-old problem of predicting a protein's 3D structure from its amino acid sequence with near-experimental accuracy, earning it recognition as one of the most important breakthroughs in life sciences of the preceding decade .

The system is entirely data-driven, relying on multiple sequence alignment (MSA) — a technique tracing to 1993 work by German scientist Chris Sander — rather than physical models . By comparing evolutionarily related protein sequences across species, AlphaFold 2 learns which amino acids vary together, inferring spatial proximity in the folded structure. Trained on roughly 170,000 known protein structures, it achieved what FreeS Fund later described as "ten-thousand-fold efficiency gains," predicting hundreds of millions of structures in under three years compared to the 200,000 experimentally determined over six prior decades .

In the view of academician Weinan E, AlphaFold 2's greatest significance was proving that machine learning could "almost thoroughly resolve" protein folding — a demonstration effect that shifted the entire field's confidence . The tool's release catalyzed the AI-for-science wave and, per FreeS Fund's 2024 retrospective, helped validate AI's potential in drug discovery broadly, with DeepMind open-sourcing the code in 2021 and expanding capabilities through AlphaFold 3 by 2024 .