AlphaFold is a deep learning system that improves protein structure prediction by using distance predictions from neural networks. The system constructs a potential of mean force that accurately describes protein shape and can be optimized using gradient descent. AlphaFold achieved high accuracy in the Critical Assessment of Protein Structure Prediction (CASP13), predicting structures for 24 out of 43 free modelling domains with TM-scores of 0.7 or higher, outperforming other methods. The system uses evolutionary covariation data to predict residue distances, which provides more detailed structural information than contact predictions. AlphaFold's neural network predicts distances between residues, which are used to construct a protein-specific potential. This potential is optimized using gradient descent, allowing accurate structure prediction without complex sampling procedures. The system also incorporates torsion angle predictions and uses a combination of distance and torsion potentials to model protein structure. AlphaFold's performance in CASP13 demonstrated its ability to predict previously unknown folds and outperformed other methods in accuracy. The system's success is attributed to its ability to model detailed interactions and its use of deep learning to predict residue distances. AlphaFold's results highlight its potential for understanding protein function and malfunction, especially in cases where no homologous proteins have been experimentally determined. The system's accuracy improvements enable better insights into protein function, protein-protein interactions, and binding pockets. Overall, AlphaFold represents a significant advance in protein structure prediction, with the potential to impact various areas of protein science.AlphaFold is a deep learning system that improves protein structure prediction by using distance predictions from neural networks. The system constructs a potential of mean force that accurately describes protein shape and can be optimized using gradient descent. AlphaFold achieved high accuracy in the Critical Assessment of Protein Structure Prediction (CASP13), predicting structures for 24 out of 43 free modelling domains with TM-scores of 0.7 or higher, outperforming other methods. The system uses evolutionary covariation data to predict residue distances, which provides more detailed structural information than contact predictions. AlphaFold's neural network predicts distances between residues, which are used to construct a protein-specific potential. This potential is optimized using gradient descent, allowing accurate structure prediction without complex sampling procedures. The system also incorporates torsion angle predictions and uses a combination of distance and torsion potentials to model protein structure. AlphaFold's performance in CASP13 demonstrated its ability to predict previously unknown folds and outperformed other methods in accuracy. The system's success is attributed to its ability to model detailed interactions and its use of deep learning to predict residue distances. AlphaFold's results highlight its potential for understanding protein function and malfunction, especially in cases where no homologous proteins have been experimentally determined. The system's accuracy improvements enable better insights into protein function, protein-protein interactions, and binding pockets. Overall, AlphaFold represents a significant advance in protein structure prediction, with the potential to impact various areas of protein science.