Highly accurate protein structure prediction with AlphaFold

Predicting a protein's 3D structure from its amino acid sequence has been a central challenge in biology for over 50 years. While experimental methods like X-ray crystallography and cryo-electron microscopy provide high-resolution data, they are time-consuming and expensive. Computational approaches have historically struggled to reach atomic accuracy, especially for proteins lacking known structural templates. AlphaFold addresses this by integrating evolutionary, physical, and geometric constraints into a novel deep learning architecture.

AlphaFold processes protein sequences through a two-stage neural network. The first stage, the 'Evoformer,' uses a novel attention-based mechanism to jointly embed multiple sequence alignments (MSAs) and pairwise residue relationships. This allows the model to reason about spatial and evolutionary constraints simultaneously. The second stage, the 'structure module,' directly predicts the 3D coordinates of all heavy atoms. This module uses an iterative refinement process, allowing the network to build and adjust the protein structure in a 3D space, effectively learning to 'fold' the protein.

AlphaFold was validated in the 14th Critical Assessment of protein Structure Prediction (CASP14), where it significantly outperformed all other methods. It achieved a median backbone accuracy of 0.96 Å, a level of precision competitive with experimental structures. The model also provides per-residue confidence scores (pLDDT), which allow researchers to assess the reliability of specific regions of the predicted structure. By enabling rapid, large-scale structure prediction, AlphaFold has the potential to accelerate structural bioinformatics and facilitate a deeper understanding of protein function at the proteome scale.