The paper introduces AlphaFold, a deep learning-based system for predicting protein structures from amino acid sequences. AlphaFold uses a convolutional neural network to predict distances between pairs of residues, which provides more detailed information about the protein's structure compared to contact predictions. These distance predictions are used to construct a potential of mean force that accurately describes the protein's shape. The potential is optimized using gradient descent, resulting in high-accuracy protein structures without the need for complex sampling procedures. In the Critical Assessment of Protein Structure Prediction (CASP13) assessment, AlphaFold achieved high accuracy for 24 out of 43 free modeling domains, outperforming other methods. The system's accuracy is attributed to the high precision of distance predictions, which are influenced by factors such as data augmentation, feature representation, and auxiliary losses. AlphaFold's performance in CASP13 demonstrates its potential to advance protein structure prediction and enable insights into protein function and malfunction, especially for sequences without known homologous structures.The paper introduces AlphaFold, a deep learning-based system for predicting protein structures from amino acid sequences. AlphaFold uses a convolutional neural network to predict distances between pairs of residues, which provides more detailed information about the protein's structure compared to contact predictions. These distance predictions are used to construct a potential of mean force that accurately describes the protein's shape. The potential is optimized using gradient descent, resulting in high-accuracy protein structures without the need for complex sampling procedures. In the Critical Assessment of Protein Structure Prediction (CASP13) assessment, AlphaFold achieved high accuracy for 24 out of 43 free modeling domains, outperforming other methods. The system's accuracy is attributed to the high precision of distance predictions, which are influenced by factors such as data augmentation, feature representation, and auxiliary losses. AlphaFold's performance in CASP13 demonstrates its potential to advance protein structure prediction and enable insights into protein function and malfunction, especially for sequences without known homologous structures.