EMReady2

EMReady2 is a deep-learning tool for improving the quality and interpretability of cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET) density maps. Developed by the Huang laboratory at Huazhong University of Science and Technology and released as a bioRxiv preprint in September 2025 (updated February 2026), it is the successor to the original EMReady map-enhancement method. The model post-processes a reconstructed map to sharpen features, suppress noise, and make density easier to interpret and to build atomic models into.

The core methodological advance is the use of a Mamba-based architecture — a state space model that scales more efficiently than attention for long sequences — embedded in a dual-branch UNet. This design lets EMReady2 jointly capture local detail and global context across a map. A local resolution-guided learning strategy explicitly accounts for the fact that cryo-EM maps are heterogeneous, with resolution varying substantially from region to region within a single reconstruction.

EMReady2 sits in the post-processing and map-improvement niche of the cryo-EM ecosystem, complementing reconstruction and model-building pipelines rather than replacing them. The software is openly available on GitHub under the MIT license, with pre-trained weights distributed via the lab website.

#Key Features

• Mamba dual-branch UNet: A fast Mamba-based architecture captures both local and global features, offering an efficient alternative to attention-heavy enhancers.
• Local resolution-guided learning: Explicitly conditions on local map quality to handle the resolution heterogeneity common in real cryo-EM reconstructions.
• Cryo-EM and cryo-ET support: Enhances both single-particle cryo-EM maps and cryo-ET tomograms within a single framework.
• Nucleic-acid-aware enhancement: Validated on nucleic-acid-containing samples, broadening applicability beyond protein-only maps.
• Open-source release: Python implementation under an MIT license with sliding-window inference, Gaussian patch blending, and automatic voxel-size-based weight selection.

#Technical Details

EMReady2 is built on a Mamba-based dual-branch UNet that processes density maps via overlapping boxes (sliding-window inference) with configurable stride and Gaussian patch blending for aggregation. A local resolution-guided training strategy adapts enhancement to per-region map quality. The method was evaluated on 136 diverse maps spanning roughly 2.0-10.0 Å resolution, with improvements reported in both map quality and downstream interpretability, including for nucleic-acid-containing samples and cryo-ET data. The implementation runs on Linux with CUDA ≥ 11.8 and Python 3.10; two pre-trained checkpoints (for ~0.6 Å and ~1.0 Å voxel sizes) are downloaded separately and selected automatically based on input voxel size.

#Applications

Structural biologists use EMReady2 to improve cryo-EM and cryo-ET maps prior to atomic model building, refinement, and validation, making weak or heterogeneous density easier to interpret. By raising the effective interpretability of intermediate-resolution maps, it can accelerate model building for challenging targets, including assemblies and complexes containing nucleic acids, and extend useful analysis to tomography data.

#Impact

EMReady2 brings state-space (Mamba) modeling to cryo-EM map enhancement, a domain previously dominated by convolutional and attention-based networks, and pairs it with resolution-aware training to address map heterogeneity directly. As an open-source, MIT-licensed tool with released weights, it is readily adoptable within existing cryo-EM workflows. Reported performance reflects benchmark evaluation on 136 maps; broader community validation across diverse experimental datasets will further establish where its enhancements are most reliable.