The paper presents a new automated technique for loop modeling in protein structures, which significantly improves the accuracy of loop predictions. The method optimizes the positions of all nonhydrogen atoms of a loop within a fixed environment using a pseudo-energy function that includes spatial restraints from the CHARMM-22 force field, statistical preferences for main-chain and side-chain dihedral angles, and nonbonded atomic contacts. The energy function is optimized using conjugate gradients, molecular dynamics, and simulated annealing. The accuracy of the method is evaluated by comparing it with known structures of 40 loops of various lengths (1 to 14 residues) and by assessing its performance in blind predictions of protein structures. The results show that the method achieves high accuracy, with 100%, 90%, and 30% of 4-, 8-, and 12-residue loop predictions, respectively, having an RMSD error of <2 Å for the mainchain N, Cα, C, and O atoms. The method's accuracy is also assessed in approximately correct environments typical of comparative modeling without misalignment, showing that the average loop prediction error increases by 180%, 25%, and 3% for 4-, 8-, and 12-residue loops, respectively, when the RMSD distortion of the main-chain stem atoms is 2.5 Å. The paper concludes that the method's accuracy is primarily limited by the accuracy of the energy function rather than the extent of conformational sampling.The paper presents a new automated technique for loop modeling in protein structures, which significantly improves the accuracy of loop predictions. The method optimizes the positions of all nonhydrogen atoms of a loop within a fixed environment using a pseudo-energy function that includes spatial restraints from the CHARMM-22 force field, statistical preferences for main-chain and side-chain dihedral angles, and nonbonded atomic contacts. The energy function is optimized using conjugate gradients, molecular dynamics, and simulated annealing. The accuracy of the method is evaluated by comparing it with known structures of 40 loops of various lengths (1 to 14 residues) and by assessing its performance in blind predictions of protein structures. The results show that the method achieves high accuracy, with 100%, 90%, and 30% of 4-, 8-, and 12-residue loop predictions, respectively, having an RMSD error of <2 Å for the mainchain N, Cα, C, and O atoms. The method's accuracy is also assessed in approximately correct environments typical of comparative modeling without misalignment, showing that the average loop prediction error increases by 180%, 25%, and 3% for 4-, 8-, and 12-residue loops, respectively, when the RMSD distortion of the main-chain stem atoms is 2.5 Å. The paper concludes that the method's accuracy is primarily limited by the accuracy of the energy function rather than the extent of conformational sampling.