Deep Learning for Protein Secondary-Structure Segmentation in Intermediate-Resolution Cryo-EM Maps: A Resolution-Stratified Benchmark Dataset and Controlled Evaluation

Accurate segmentation of protein secondary structures in cryo-electron microscopy (cryo-EM) maps is crucial for protein tracing and atomic model building, yet remains challenging at intermediate resolutions. Although deep learning architectures have recently been applied to this task, there remains a lack of standardized evaluation protocols and systematic resolution-stratified analyses. Consequently, it remains unclear how segmentation performance varies across finer resolution ranges within the intermediate-resolution regime and how different volumetric segmentation architectures compare under controlled experimental conditions. To address these gaps, we present a nonredundant, resolution-stratified benchmark dataset and a controlled evaluation framework for protein secondary-structure segmentation in intermediate-resolution cryo-EM maps. The dataset comprises 972 nonredundant protein-chain samples from 495 cryo-EM protein structures resolved at 4.5-8.0 Å, with voxel-level annotations for helices, sheets, and non-periodic secondary-structure (NPSS) regions. The dataset is stratified into three resolution ranges4.5-5.5, 5.5-6.5, and 6.5-8.0 Åto enable systematic analysis of resolution-dependent segmentation performance. Using this benchmark, we evaluate five representative convolutional and transformer-based volumetric segmentation architecturesU-Net, Residual U-Net, Attention U-Net, UNETR, and Swin-UNETRunder standardized preprocessing, training, and evaluation protocols. Residual U-Net achieves the highest macro-averaged foreground (Macro-FG) F1 score in all three resolution bins, while helices are segmented more accurately than sheets and NPSS regions. All five architectures achieve their highest Macro-FG F1 scores in the highest-resolution bin, whereas performance varies between the two lower-resolution bins across architectures and secondary-structure classes. Overall, this benchmark enables fine-grained analysis of resolution-dependent segmentation performance, supports reproducible method development, and facilitates controlled comparison of future cryo-EM secondary-structure segmentation approaches.