DeepMM

DeepMM reconstructs de novo atomic-accuracy all-atom protein structures from cryogenic electron microscopy (cryo-EM) maps to enable high-resolution structural interpretation.

Key Features:

• Deep Learning Framework: Employs Densely Connected Convolutional Networks (DenseNets) to predict main-chain positions, Cα atom locations, amino-acid types, and secondary structure elements from cryo-EM maps and to generate all-atom models.
• Semi-automatic model building: Automates model construction while permitting manual intervention and refinement during the model-building process.
• Validation and Performance: Validated on 40 simulated maps at 5 Å resolution, 30 experimental maps at 2.6–4.8 Å resolution, and an EMDB-wide dataset of 2,931 experimental maps at 2.6–4.9 Å resolution, demonstrating superior performance relative to RosettaES, MAINMAST, and Phenix.
• Accuracy and Coverage: Improves both atomic accuracy and sequence coverage when constructing full-length protein structures across tested datasets.

Scientific Applications:

• De novo atomic model building: Converts cryo-EM density maps into full-atom protein structures without relying on prior templates.
• Bridging EM maps and structural models: Addresses the gap between deposited cryo-EM maps and corresponding 3D atomic models for structural databases.
• Structural interpretation: Enables atomic-level analysis of protein function, interactions, and dynamics from high-resolution cryo-EM data.

Methodology:

DeepMM uses DenseNets to predict main-chain positions, Cα atom locations, amino-acid types, and secondary structure elements from cryo-EM maps and integrates these predictions into coherent full-atom models.