Extended Data Fig. 7. CNN detection and validation of GEM2 particles.

a, Full view of the tomographic slice from Fig. 3a. b, Maximum projection of raw CNN probability scores. c, Maximum projection of post-processed CNN probability scores. Scores were thresholded at 0.5, filtered by size (connected-component size cluster of 5000–50000 pixels at 13.7 Å/pixel), and masked with a lamella mask to exclude false positives. This tomogram was not used for CNN training. Numbered boxes correspond to curated particles, shown in d. d, Left columns, tomographic slices at the indicated positions in b and c, 6.74 nm in thickness. Right columns, slices through the thresholded subtomogram average pasted in the tomogram based on the refined position and orientation. Of the 15 particles, 9 corresponded to peaks in the post-processed scores, and 6 more (2, 6, 7, 8, 11, 15) were annotated based on visual inspection of lower scoring peaks. e, Left, cross-sections of simulated densities at 30 Å resolution based on the in vitro structure of the encapsulin scaffold (PDB 6X8M). Right, cross-sections of the GEM2 subtomogram average presented in Fig. 3b.