Cryo-electron microscopy has transformed structural biology over the past decade, but the software that sharpens and enhances the raw density maps has a blind spot: it offers no way to know which parts of the enhanced map are reliable. A neural network can turn noise into apparent structure, and the user has no way to tell the difference.
LocScale-2.0, published in Nature Communications on July 17 by researchers at Delft University of Technology, addresses that gap with voxel-wise confidence scores that tell structural biologists which features to trust and which to treat with caution.
What LocScale Does
Raw cryo-EM maps are blurry. Sharpening them, enhancing high-resolution detail while suppressing noise, is an essential post-processing step. LocScale 1.0, released in 2017, performed this sharpening locally using an existing atomic model as a reference. Version 2.0, developed by Alok Bharadwaj, Reinier de Bruin, and Arjen Jakobi, introduces three major advances.
First, it now works without any atomic model at all. A model-free mode uses false-discovery-rate-controlled density thresholding to estimate the molecular envelope, placing pseudoatoms in unmodelled regions. A hybrid mode combines partial atomic models with pseudoatoms for the rest.
Second, a completely new deep-learning workflow called LocScale-FEM (Feature-Enhanced Maps) operates in real space using a Bayesian 3D U-Net with Monte Carlo dropout. Unlike the Fourier-filter approach of version 1.0, this affects both amplitudes and phases, akin to density modification.
Third, and most importantly, the Monte Carlo dropout generates an ensemble of predictions whose variance quantifies uncertainty. This is converted into a voxel-wise confidence score (pVDDT, on a -100 to 100 scale) that correlates with local phase errors. Scores of ±80 or above indicate reliable features; ±95 or above indicate likely over-emphasis or suppression.
What the Benchmarks Show
In automated model building tests using ModelAngelo across 50 map-model pairs below 4 Å resolution, LocScale-2.0 maps achieved 82% sequence coverage before pruning, compared with the deposited maps. The largest single improvement was on glutamate synthase, where 1,306 additional residues (a 12% increase) were modelled correctly.
Volumetric recall, a measure of how well the enhanced map recovers true molecular density, was significantly higher than DeepEMhancer and EMReady (p < 0.0004 and p < 0.0002 respectively). Crucially, LocScale-2.0 preserved contextual structures that competing methods suppressed: lipid belts around membrane proteins, the central plug in bacterial secretin channels, A-site tRNA and nascent peptide chains in ribosomes, and cholesteryl ester plates inside LDL particles.
The pVDDT confidence scores allow users to distinguish, for example, the well-ordered pyranose core of a ligand from its flexible phenylethyl tail, information that is invisible in a conventional sharpened map.
Why This Matters for the Field
Cryo-EM is producing structures at an accelerating rate, and the field increasingly targets complex systems (membrane proteins in native lipids, in situ subtomogram averages, endogenous complexes), where map quality varies spatially. Deep-learning sharpening tools that produce a single “best” map without uncertainty estimates create a risk of misinterpretation, particularly by inexperienced users.
LocScale-2.0’s confidence-guided approach makes the uncertainty visible. The software is available as part of the CCP-EM Doppio suite and as a standalone Python package. The caveat: in 17% of tested cases, deposited maps outperformed LocScale-2.0 maps for automated model building, and the training data for the neural network covers only 89 EMDB entries, meaning features outside this distribution may not be handled optimally.
Sources
1. A. Bharadwaj, R. de Bruin, A.J. Jakobi, “Confidence-guided cryo-EM map optimization with LocScale-2.0,” Nature Communications (2026). DOI: 10.1038/s41467-026-75327-8