Deep Learning for Secondary Structure Elements Prediction From Cryo-EM Data
Abstract
Understanding protein structure is vital for drug design and advancing medical knowledge, playing a crucial role in modern medicine. Cryo-Electron Microscopy (Cryo-EM) is a leading method for imaging proteins, yet challenges persist, particularly at mid-level resolutions (4 to 10 angstroms), compromising clarity and hindering the identification of secondary structure elements like Helix, Sheet, and Coils.Artificial Neural Networks (ANNs) emulate human neurons and have been early tools for protein structure prediction. Deep learning, a form of ANN, excels in pattern detection.The Inception architecture, introduced by Google in 2014, addresses computational limitations in deep neural networks by utilizing convolutional filters of varying sizes (1x1, 3x3, and 5x5) simultaneously. This approach captures both local and global features efficiently, proving powerful in pattern detection.We present a novel deep learning framework for annotating secondary structures in intermediate-resolution cryo-EM maps, employing a three-dimensional Inception architecture. Evaluation on diverse datasets, including maps with authentic resolutions, demonstrates its accuracy and robustness. Comparative analysis against state-of-the-art frameworks reveals superior performance across secondary structure elements, with notable F1 scores of 0.657 for helix, 0.712 for coil, and 0.596 for sheet predictions. Certain helix and sheet predictions achieved impressive F1 scores of 0.881.
Subject Area
Computer science|Bioinformatics|Artificial intelligence|Molecular biology
Recommended Citation
Mohammad Bataineh,
"Deep Learning for Secondary Structure Elements Prediction From Cryo-EM Data"
(2024).
ETD Collection for Tennessee State University.
Paper AAI30995494.
https://digitalscholarship.tnstate.edu/dissertations/AAI30995494