Non-uniform refinement: Adaptive regularization improves single particle cryo-EM reconstruction

Ali Punjani; Haowei Zhang; David J. Fleet

In [25] , a B-factor is fit to the rotationally-averaged power spectrum of the reconstructed density map. The B-factor is composed of two parts, B image and B computation . B image captures the decay of highfrequency signal intrinsic to the imaging process. B computation captures the additional high-frequency signal lost due to computational errors in alignment or reconstruction from the image data. Figure 7A (left) shows overall B-factors fit to reconstructions for the four datasets discussed above, with uniform and non-uniform refinement, given the same image data. The Guinier plot in Fig. 7 (right) illustrates the difference in high-frequency signal decay between the two types of refinement, here for the Na v 1.7 channel dataset. For all datasets, non-uniform refinement B-factor magnitudes are smaller than for uniform refinement. Since B image is the same in both methods, it follows that non-uniform refinement has lower B computation , losing less high frequency signal to misalignment. This difference also suggests that using non-uniform refinement should in general require less image data to reach the same resolutions as uniform refinement, consistent with what we observed with the STRA6-CaM dataset above.