Segmentation and Quality Assessment  of Protein Secondary Structure

Segmentation of protein secondary structures

Summary: This project employs a convolutional neural network (CNN)-based architecture to discern the precise organization of protein secondary structure elements in medium-resolution cryo-electron microscopy (cryo-EM) images. The proposed model effectively addresses class imbalance by utilizing selectively cropped and masked density images at the chain level. Trained on approximately 1300 chains derived from around 500 experimental cryo-EM density maps, the model demonstrates robust performance achieving residue-level F1 scores of 0.76 for helix detection and 0.60 for $\beta$-sheet detection on average across a set of testing chains. Notably, the framework exhibits comparable or superior performance to traditional image processing tools, such as SSETracer, which requires significant manual intervention and pre-processing steps. Additionally, the proposed method performs competitively with another deep learning-based architecture, EMap2Sec. Furthermore, this research sheds light on the impact of secondary structure quality—specifically, adherence to expected shapes—on detection performance, underscoring the necessity for meticulous evaluation of data quality. 

Quality assessment of alpha-helix in medium resolution image

Summary: Cryo-EM density maps at medium resolution (5–10 Å) show secondary structural features like α-helices and β-sheets, lacking detailed side chain information crucial for direct structure determination. Over 800 entries in the Electron Microscopy Data Bank (EMDB) link medium-resolution density maps with atomic models, displaying diverse similarities. This study introduces and assesses the F1 score, a local similarity criterion, to validate atomic models and classify structural features. The F1 score, ranging from zero to one, evaluates cylindrical agreement between helical density and atomic models. Analysis of 30,994 helices across 3,247 protein chains reveals F1 scores ranging from 0.171 to 0.848, indicating effective stratification of data. Higher F1 scores correlate with regions exhibiting high and spatially homogeneous local resolution (5–7.5 Å) in helical density. The proposed F1 scores serve as a discriminative classifier for validation studies and a ranking criterion for cryo-EM density features in databases.