Best for the Eye, Not for the Algorithm: Anisotropy in Fitting Atomic Models in Cryo-EM

Most atomic model refinement methods in cryo-EM fit models to the reconstructed density map and effectively treat Fourier voxels as equally reliable. However, the uncertainty in the estimation of Fourier coefficients is highly anisotropic, primarily due to the common variability in SNR in different frequency shells and the distribution of particle images across viewing directions. First-principles arguments suggest that atomic models should be fitted to particle images rather than volumes; this strategy may be computationally demanding. We show that under certain modeling choices, fitting atomic models to weighted volumes is equivalent to fitting directly to particle images. Furthermore, we argue that various proxies can be used to capture this and other sources of uncertainty and distortions. We propose that the principle can be implemented in most atomic model-fitting software with relative ease, using information readily available in existing pipelines. As a proof of concept, we extracted the necessary information from standard RELION runs and fed it into a modified version of Servalcat in which we implemented a reinterpreted version of the idea.