Constraining template matching scores improves specificity in particle assignment

To tap into the full potential of in situ cryogenic electron microscopy it is necessary to annotate many macromolecular structures to reveal 3-dimensional organization. Template matching (TM) can be useful to localize single instances of known macromolecules with high precision by relying on exhaustive translational and orientational searches. However, high signal-to-noise ratio artifacts in images can often interfere with the detection of weakly scattering biological material. Firstly, its important to model the template with a tilt-dependent weighting function and consider defocus gradients in tomogram reconstruction. We show here updated implementations in the pytom template matching module to model these microscope functions. Secondly, we explore a tophat transform to constrain annotations to sharp peaks in TM correlation maps, relying on the observation that correlations with artifacts are often smooth in space, resembling Gaussian blobs. It turns out that a tophat transform is efficient at separating these values. The tophat kernel size was first optimized in a simulated dataset. We then show how it can aid annotations in automated batch processing, leading to clean sets of ribosomes in a ER-microsome dataset. We then show the limits in publicly available lamella of Chlamydomonas Rheinhardtii cells, assessing detection on proteasomes.