The paper presents an improved method for protein structure prediction by integrating deep learning and Rosetta energy minimization. The authors developed a deep residual network to predict interresidue orientations and distances from coevolutionary data, and a Rosetta-constrained energy-minimization protocol to generate accurate protein structures. The method outperforms previous approaches in benchmark tests on CASP13 and CAMEO datasets, achieving higher TM-scores for structure prediction. The network consistently assigns higher probabilities to de novo-designed proteins, identifying key fold-determining residues. The method is useful for a broad range of protein structure prediction and design problems, and the codes are available for further development.The paper presents an improved method for protein structure prediction by integrating deep learning and Rosetta energy minimization. The authors developed a deep residual network to predict interresidue orientations and distances from coevolutionary data, and a Rosetta-constrained energy-minimization protocol to generate accurate protein structures. The method outperforms previous approaches in benchmark tests on CASP13 and CAMEO datasets, achieving higher TM-scores for structure prediction. The network consistently assigns higher probabilities to de novo-designed proteins, identifying key fold-determining residues. The method is useful for a broad range of protein structure prediction and design problems, and the codes are available for further development.