journal menu
![[Figure 6]](bar5003fig6mag.jpg)
|
|
Figure 6
MMSE pose estimation improves heterogeneity reconstruction over MLE. (a) Ground-truth density maps from a synthetic dataset simulating a one-dimensional conformational transition, a standard benchmark for methods such as cryoDRGN (Zhong et al., 2021  ) and RECOVAR (Gilles & Singer, 2025). 50 equally spaced states approximate the continuous trajectory, with colors indicating positions along the pathway (only five states are shown in the figure for clarity). (b) Example of a noisy projection image from the high-noise regime used in this study, where even ground-truth poses recover only ∼40% of the variance for the top 30 components due to finite-sample and high-noise effects. (c) Comparison of reconstructed structures using different pose priors (ground truth, MMSE, MLE), showing that MMSE yields structures closer to the true conformation state. (d) First five principal components estimated using ground-truth poses, MMSE pose estimates and MLE pose estimates. MMSE results closely match the ground truth in structural detail, whereas MLE reconstructions degrade beyond the first modes, becoming blurrier and less representative of the underlying variability. (e) Subspace accuracy, measured as the percentage of total variance captured (as defined in Section 5.3). Ground truth achieves ∼40%, MMSE ∼30% and MLE ∼25%. (f) Eigenvalue recovery: both methods estimate the largest true eigenvalues well, but MMSE remains accurate for smaller eigenvalues, while MLE substantially overestimates them. These results show that MMSE pose estimation consistently yields more accurate recovery of conformational variability than MLE. |
![[Figure 6]](bar5003fig6.jpg)
 | STRUCTURAL BIOLOGY |
ISSN: 2059-7983
access
