Denoising and iterative phase recovery reveal low-occupancy populations in protein crystals.

Advances in structural biology increasingly focus on uncovering protein dynamics and transient macromolecular complexes. Such studies require modeling of low-occupancy species like time-evolving intermediates and bound ligands. In protein crystallography, difference maps that compare paired perturbed and reference datasets are a powerful way to identify and aid modeling of low-occupancy species. Current methods to generate difference maps, however, rely on manually tuned parameters and, when signals are weak due to low occupancy, can fail to extract clear, chemically interpretable signals. We address these issues, first by showing that negentropy - a measure of how different a signal looks from anticipated Gaussian noise - is an effective metric to assess difference map quality and can therefore be used to automatically determine difference map calculation parameters. Leveraging this, we apply total variation denoising, an image restoration technique that requires a choice of regularization parameter, to crystallographic difference maps. We show that total variation denoising improves map signal-to-noise and enables us to estimate the latent phase contribution of low-occupancy states. We implement this technology in an open-source Python package, METEOR. METEOR opens new possibilities, for time-resolved and ligand-screening crystallography especially, allowing detection of low-occupancy states that could not previously be resolved.