DeepEMhancer

DeepEMhancer employs a U-Net style encoder-decoder architecture trained on a synthetic dataset of paired raw and sharpened density maps. The network learns to perform non-linear local filtering that simultaneously applies B-factor sharpening and amplitude scaling in a spatially aware manner.

• Input: Raw, unfiltered cryo-EM half-maps
• Output: Post-processed map with restored high-resolution detail
• Training data: Maps sharpened with phenix.auto_sharpen as ground truth targets
• Key advantage: Learns to avoid amplifying noise in disordered regions while enhancing detail in ordered cores

Unlike global B-factor sharpening which applies a uniform frequency-dependent weighting across the entire map, DeepEMhancer performs voxel-wise amplitude adjustment. The network recognizes structural context and applies stronger sharpening to well-ordered regions while preserving the natural appearance of flexible loops and side chains.

• Global methods (e.g., phenix.auto_sharpen) apply one B-factor to the entire map
• DeepEMhancer learns to modulate sharpening strength based on local density quality
• Prevents the over-sharpening artifacts (ripples, noise amplification) common in globally sharpened maps
• Preserves solvent flattening implicitly through learned amplitude normalization

DeepEMhancer takes two independent half-maps as input rather than a single merged map. This design leverages the gold-standard FSC principle—the network learns to recognize and suppress noise that is uncorrelated between half-sets while enhancing signal that is consistent across both.

• Half-map 1 and Half-map 2: Independently refined reconstructions from split datasets
• The network implicitly performs noise suppression by comparing features across both inputs
• Eliminates the need for explicit solvent masking before post-processing
• Output maintains the FSC-based resolution claims of the input reconstruction

The network is trained using a supervised learning paradigm where raw experimental maps serve as input and their corresponding phenix.auto_sharpen processed versions serve as targets. This creates a massive training corpus without requiring manual annotation.

• Training pipeline: Raw half-maps → phenix.auto_sharpen → sharpened target maps
• Network learns to mimic the phenix sharpening algorithm but with spatial awareness
• Generalizes to novel macromolecular complexes unseen during training
• Can be applied to maps from any reconstruction software (RELION, cryoSPARC, cisTEM)

DeepEMhancer-processed maps show dramatically improved atomic feature definition, enabling more accurate automated model building with tools like ModelAngelo and manual interpretation in Coot. Side chain density, carbonyl bumps, and main chain connectivity become significantly clearer.

• ModelAngelo: Achieves higher sequence assignment accuracy on DeepEMhancer-sharpened maps
• Coot: Improved visualization of side chain rotamers and backbone carbonyls
• ISOLDE/MDFF: Better starting maps for molecular dynamics flexible fitting
• Compatible with RELION and cryoSPARC post-processing workflows

DeepEMhancer's local processing approach is particularly valuable for maps exhibiting conformational heterogeneity. The network can apply different sharpening parameters to rigid domains versus flexible linkers, preserving biological information that global sharpening would either over-enhance or blur.

• Flexible domains: Maintain natural appearance without artificial sharpening
• Rigid cores: Receive stronger enhancement for atomic model building
• Works on maps from 3D variability analysis and multi-body refinement
• Preserves domain motion information that global B-factor application can distort

A post-processing step that applies a B-factor weighting to Fourier amplitudes to restore high-frequency detail attenuated during imaging. Unlike DeepEMhancer, traditional sharpening uses a global B-factor applied uniformly across the entire map. DeepEMhancer improves on this by performing local amplitude scaling that varies spatially, correcting for position-dependent resolution variation and avoiding over-sharpening of flexible regions.

A computational method that calculates a resolution value for each voxel in a cryo-EM density map. Tools like ResMap or MonoRes produce a 3D resolution map identifying rigid core regions (high resolution) versus flexible loops (low resolution). DeepEMhancer leverages this local resolution information implicitly through its convolutional neural network to apply appropriate sharpening at each spatial location, preventing noise amplification in disordered domains.

An atomic model optimization method that directly minimizes the discrepancy between a model's calculated density and the experimental cryo-EM map in real space. Tools like Phenix.real_space_refine use gradient-driven optimization to improve geometry and fit. DeepEMhancer maps are often used as input for refinement because their enhanced interpretability aids in manual model building and improves the convergence of automated refinement algorithms.

An automated atomic model building program that uses a graph neural network to trace the protein backbone and assign amino acid sequences directly into cryo-EM density maps. ModelAngelo performs best on high-quality, well-sharpened maps. DeepEMhancer post-processing is frequently applied immediately before ModelAngelo to enhance side-chain features and improve the accuracy of automated sequence assignment, particularly at resolutions between 2.5–3.5 Å.

A mathematical function describing how the electron microscope's objective lens aberrations modulate image contrast as a function of spatial frequency. CTF correction is performed during 3D reconstruction to restore phase-flipped information. DeepEMhancer operates on CTF-corrected maps, meaning it assumes the input has already undergone phase correction and focuses solely on amplitude scaling and denoising to enhance interpretability.

A resolution estimation method that splits the particle dataset into two independent half-sets for separate 3D reconstruction, comparing their Fourier shells to avoid overfitting. The FSC=0.143 criterion defines the nominal resolution. DeepEMhancer's post-processing does not alter the resolution limit but improves visual interpretability within the resolved frequency range, making it easier to identify secondary structure elements and bound ligands without introducing false features.