AlphaFold 2 at CASP14: Structure Accuracy and Confidence

Predicting shapes before the answers were public

On November 30, 2020, DeepMind announced AlphaFold 2's CASP14 results. Teams predicted three-dimensional protein structures from amino-acid sequences before the experimental answers were public. This blind assessment tested predictions against independently determined structures. [2]

DeepMind reported a median Global Distance Test score of 92.4 across targets, and 87.0 in the hardest free-modelling category. GDT scores structural agreement from 0 to 100; 92.4 is not a probability of correctness. [2]

The first AlphaFold had led CASP13 in 2018. AlphaFold 2's redesigned system predicted atomic positions more accurately. It did not simulate every step by which a protein folds. [2][3]

Combining evolutionary clues with geometry

The July 2021 paper describes two interacting representations: an alignment of related protein sequences and information about pairs of amino-acid residues. Related sequences provide evolutionary clues; pair information helps constrain which parts of a chain can fit together in space. [3]

The network's Evoformer blocks exchange information between these representations. A structure module turns the result into three-dimensional coordinates, and the network feeds its predictions back through itself to refine them. This iterative process, called recycling, improves the proposed structure. It learns from known structures and sequences rather than exhaustively trying every possible fold. [3]

A convincing picture can contain uncertain regions

The confidence estimates matter as much as the rendered shape. DeepMind's July 2021 guidance distinguishes pLDDT, confidence in local structure along the chain, from predicted aligned error, confidence in how different parts sit relative to one another. Two domains can each look reliable while their relative placement remains uncertain. [4]

Atlas interpretation: For example, a researcher considering an apparent pocket between two domains should check confidence in their relative positions before treating that pocket as a useful experimental target. A well-predicted local fold alone does not settle the question. The confidence measures help decide which parts of a structural hypothesis deserve investigation. [4]

The result came before the public tool

The detailed method paper appeared on July 15, 2021, with code, trained weights and an inference script made available. The November announcement had still described publication and broader access as work ahead. Releasing the system let other groups generate predictions for their own sequences. [3][2]

Atlas interpretation: The 2020 announcement left open questions about protein complexes and interactions with other molecules. Predicting a structure supplied a stronger starting point for experiments, not a finished explanation of biological behavior. [2]