U-Net-based reactive center loop-identifier for serpins.

Motivation: Serine protease inhibitors (serpins) are a large family of proteins conserved in all animals. A prominent enigma about serpins is that, while members of this family share an archetypal folding scheme, individual proteins have diverse functions and are involved in a variety of biological processes, including controlling inflammation, maintaining hormone homeostasis, and regulating osteogenesis. The reactive center loop (RCL) region plays a crucial role in determining the functional specificity of serpins. Yet, of the >48 000 serpins recorded in the UniProt database, only 78 have the RCL annotated. This is because the RCL, due to its high variability, cannot be identified using standard motif-matrix-based annotation strategies. Results: To overcome this deficiency, we tested neural network-based approaches for automatic RCL annotation. Using an expert-annotated training/validation dataset, we tested combinations of encodings and model architectures. U-Net-based models stood out as the best approach for this task. On the independent test dataset, U-Net-based models achieved ∼98% accuracy in identifying the RCL at the per-sequence level. Availability and implementation: Source code, training and testing datasets, and models are freely available at https://github.com/leizhou69/RCL-identifier or https://github.com/JasonLiang19/RCL-identifier under the MIT license.