AutoLoop is an autoregressive deep learning method for protein loop prediction with high accuracy

Protein structure prediction remains a key challenge in biology, with loop regions being essential for protein function yet difficult to model accurately. Herein, we propose AutoLoop, a computational model designed to automatically generate accurate loop backbone conformations. Uniquely, AutoLoop employs a bidirectional training approach while merging atom- and residue-level embedding, thus bolstering its robustness and precision. To validate its efficacy, we compare AutoLoop with thirteen established methods, including FREAD, NGK, AlphaFold2, and AlphaFold3. The results indicate that AutoLoop consistently outperforms other methods, achieving a median root-mean-square deviation (RMSD) of 1.12 Å on the CASP15 dataset, also maintaining its superior performance on the HOMSTARD dataset. The addition of a post-processing module enhances AutoLoop's performance slightly, reflecting the sound reliability of the predicted backbone structures. Additionally, the case study demonstrates AutoLoop's ability to capture one or several dominant loop conformations. These advancements hold great promise for protein engineering and drug discovery.