journal menu
![[Figure 10]](sor5003fig10mag.jpg)
|
|
Figure 10
Entanglement of heterogeneity vector fields with pose. Inaccuracies are entangled with pose due to information loss of the projection of density from 3D → 2D. After training, vector fields for (test-set) images from two thyroglobulin states were inferred from their respective images (1960 in state A; 1960 in state B). (a) Cosine similarity between predicted latent variables (predicted per-image vector field and predicted per-image density) is significantly higher within states (0.75 ± 0.20, 0.66 ± 0.23) than between states (0.30 ± 0.12), as indicated by the block structure. (b) The distributions of the accuracy measure, conditioned on ground-truth heterogeneity, are not completely separated, and we show the 6% overlapping region in red (dark red for state A, light red for state B). (c) Images with more accurately inferred heterogeneity (blue) have a pose that reveals more about the conformational heterogeneity (see Fig. 9 ![[link]](../../../../../../logos/arrows/d_arr.gif){kind=small} a; axis 0), while images in the less accurate region (red) have a view that obscures the conformation (see Fig. 9a; axes 1 and 2). (d) The pose conditioned on inferred heterogeneity shows the entanglement: the pose distribution of images with more accurately inferred heterogeneity is different from the pose distribution of less accurately inferred heterogeneity. Thus, while the ground-truth distribution of pose and heterogeneity are independent, the inferred heterogeneity is entangled with pose. |
![[Figure 10]](sor5003fig10.jpg)
 | STRUCTURAL BIOLOGY |
ISSN: 2059-7983
access
