DeepMind today announced that its AI system, AlphaFold, has been recognised as a solution to the 50-year-old grand challenge of protein folding by the organisers of the Critical Assessment of protein Structure Prediction (CASP). At CASP14, the system achieved a median Global Distance Test (GDT) score of 92.4 across all targets, a level of accuracy described as competitive with experimental methods. For the hardest targets, the median score was 87.0.

The protein folding problem, first posed by Christian Anfinsen in 1972, asks whether a protein’s three-dimensional structure can be predicted solely from its linear sequence of amino acids. The problem had resisted solution for decades because the number of possible conformations is astronomically large, a dilemma known as Levinthal’s paradox. DeepMind’s new architecture uses an attention-based neural network that iteratively refines a spatial graph of protein residues, trained on roughly 170,000 known structures.

The result has been met with astonishment in the scientific community. CASP co-founder John Moult called the achievement ‘a very special moment,’ while Nobel laureate Venki Ramakrishnan called it ‘a stunning advance.’ Andrei Lupas of the Max Planck Institute reported that AlphaFold allowed his team to solve a membrane protein structure that had eluded them for years.

The implications extend immediately into drug discovery and understanding disease. DeepMind has already predicted structures of SARS-CoV-2 proteins, demonstrating potential for pandemic response. Still, the company acknowledges limitations: not every prediction is perfect, and the system does not yet handle protein complexes or interactions with DNA and small molecules. The company says the system could complement existing experimental methods and help accelerate drug discovery.