Confidence-guided cryo-EM map optimisation with LocScale-2.0

Abstract

Cryogenic sample electron microscopy (cryo-EM) maps often display uneven quality, with high-resolution features coexisting alongside weak or poorly ordered regions. Such variation complicates structural interpretation, especially for heterogeneous macromolecular assemblies. Here, we present LocScale-2.0, a context-aware map optimisation framework that operates without prior knowledge of molecular structure or composition. By leveraging general expectations of electron scattering by biological macromolecules, it enhances local detail and connectivity while preserving weak but biologically relevant structural context. We further introduce LocScale-FEM, a Bayesian-approximate deep-learning approach that emulates this optimisation to generate feature-enhanced maps. LocScale-FEM provides voxel-wise confidence scores that give a statistically grounded measure of reliability, are sensitive to local phase error in the feature-enhanced map, and highlight regions where interpretation warrants caution. Selected examples illustrate how such maps can aid biological interpretation and how confidence-guided map optimisation can increase objectivity in cryo-EM density analysis.