This study introduces a method for improved protein structure prediction by integrating deep learning with Rosetta-based energy minimization. The method predicts both interresidue distances and orientations, and uses these predictions to generate accurate protein structure models. The approach outperforms existing methods in benchmark tests on the 13th Community-Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction (CASP13) and Continuous Automated Model Evaluation (CAMEO) datasets. The method is trained on native proteins but consistently assigns higher probability to de novo-designed proteins, identifying key fold-determining residues and providing an independent measure of protein structure ideality. The method is applicable to a wide range of protein structure prediction and design problems. The method, named transform-restrained Rosetta (trRosetta), combines a deep residual-convolutional network for predicting interresidue geometries from multiple sequence alignments (MSAs) with a Rosetta-based optimization protocol that incorporates predicted restraints. The network is trained on a nonredundant dataset of 15,051 proteins from the Protein Data Bank (PDB), and the trained model is available for download. The method uses covariance matrix inversion to derive residue-residue couplings from MSAs, and the derived couplings are integrated into the network for improved performance. The method's performance is evaluated on CASP13 and CAMEO datasets, where it outperforms existing methods in terms of accuracy. The method is also tested on de novo-designed proteins, where it shows high accuracy and is able to predict the structure of these proteins with high precision. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling allows for the generation of accurate protein structures. The method's success is attributed to its integration of deep learning with Rosetta-based energy minimization, which allows for the generation of accurate protein structures. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling is a key factor in its success. The method is also able to handle a wide range of protein structures, including those with complex topologies. The method's performance is further enhanced by the use of MSA subsampling and alignment selection, which improve the accuracy of the predicted structures. The method's results suggest that the integration of deep learning with Rosetta-based energy minimization is a promising approach for protein structure prediction. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling is a key factor in its success. The method's performance is further enhanced by the use of MSA subsampling and alignment selection, which improve the accuracy of the predicted structures. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling is a key factor in its success. The method's results suggest that the integration of deep learning with Rosetta-based energy minimization is a promising approach for protein structure prediction.This study introduces a method for improved protein structure prediction by integrating deep learning with Rosetta-based energy minimization. The method predicts both interresidue distances and orientations, and uses these predictions to generate accurate protein structure models. The approach outperforms existing methods in benchmark tests on the 13th Community-Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction (CASP13) and Continuous Automated Model Evaluation (CAMEO) datasets. The method is trained on native proteins but consistently assigns higher probability to de novo-designed proteins, identifying key fold-determining residues and providing an independent measure of protein structure ideality. The method is applicable to a wide range of protein structure prediction and design problems. The method, named transform-restrained Rosetta (trRosetta), combines a deep residual-convolutional network for predicting interresidue geometries from multiple sequence alignments (MSAs) with a Rosetta-based optimization protocol that incorporates predicted restraints. The network is trained on a nonredundant dataset of 15,051 proteins from the Protein Data Bank (PDB), and the trained model is available for download. The method uses covariance matrix inversion to derive residue-residue couplings from MSAs, and the derived couplings are integrated into the network for improved performance. The method's performance is evaluated on CASP13 and CAMEO datasets, where it outperforms existing methods in terms of accuracy. The method is also tested on de novo-designed proteins, where it shows high accuracy and is able to predict the structure of these proteins with high precision. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling allows for the generation of accurate protein structures. The method's success is attributed to its integration of deep learning with Rosetta-based energy minimization, which allows for the generation of accurate protein structures. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling is a key factor in its success. The method is also able to handle a wide range of protein structures, including those with complex topologies. The method's performance is further enhanced by the use of MSA subsampling and alignment selection, which improve the accuracy of the predicted structures. The method's results suggest that the integration of deep learning with Rosetta-based energy minimization is a promising approach for protein structure prediction. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling is a key factor in its success. The method's performance is further enhanced by the use of MSA subsampling and alignment selection, which improve the accuracy of the predicted structures. The method's ability to predict interresidue geometries and incorporate these predictions into structure modeling is a key factor in its success. The method's results suggest that the integration of deep learning with Rosetta-based energy minimization is a promising approach for protein structure prediction.