The paper presents a method to infer the 3D structure of proteins from evolutionary sequence variation. The authors use a maximum entropy model to infer residue pair couplings from multiple sequence alignments, which are then used to predict the 3D structure of proteins. They find that the strength of these inferred couplings is a strong predictor of residue-residue proximity in folded structures. The method is tested on 15 test proteins of different sizes and fold classes, and it successfully predicts the 3D structure with an accuracy of 2.7–4.8 Å Cα-RMSD error over at least two-thirds of the protein. This approach provides insights into essential interactions that constrain protein evolution and has potential applications in protein and drug design, as well as the identification of functional genetic variants.The paper presents a method to infer the 3D structure of proteins from evolutionary sequence variation. The authors use a maximum entropy model to infer residue pair couplings from multiple sequence alignments, which are then used to predict the 3D structure of proteins. They find that the strength of these inferred couplings is a strong predictor of residue-residue proximity in folded structures. The method is tested on 15 test proteins of different sizes and fold classes, and it successfully predicts the 3D structure with an accuracy of 2.7–4.8 Å Cα-RMSD error over at least two-thirds of the protein. This approach provides insights into essential interactions that constrain protein evolution and has potential applications in protein and drug design, as well as the identification of functional genetic variants.