Table 4:

Impact of encoder network B architecture on tomoDRGN heterogeneous network reconstruction.

Box size (px)	EncA architecture	EncA-EncB intermediate dimensionality	EncB architecture	Dec architecture	# Trainable parameters (encoder)	# Data points per particle (encoder)	# Trainable parameters (decoder)	# Data points per particle (decoder)	Training time per 1k particles (min)	VRAM per particle (GB)
64	128 × 3	32	64 × 3	256 × 3	577,568	131,528	280,066	96,776	0.41	2.10
64	128 × 3	32	128 × 3	256 × 3	715,040	131,528	280,066	96,776	0.41	2.11
64	128 × 3	128	64 × 3	256 × 3	841,856	131,528	280,066	96,776	0.41	2.10
64	128 × 3	128	128 × 3	256 × 3	1,231,232	131,528	280,066	96,776	0.41	2.10
128	128 × 3	32	64 × 3	256 × 3	1,812,000	526,932	329,218	194,064	0.79	3.95
128	128 × 3	32	128 × 3	256 × 3	1,949,472	526,932	329,218	194,064	0.85	3.96
128	128 × 3	128	64 × 3	256 × 3	2,076,288	526,932	329,218	194,064	0.87	3.95
128	128 × 3	128	128 × 3	256 × 3	2,465,664	526,932	329,218	194,064	0.86	3.95
256	128 × 3	32	64 × 3	256 × 3	6,750,240	2,108,712	427,522	378,516	2.23	9.43
256	128 × 3	32	128 × 3	256 × 3	6,887,712	2,108,712	427,522	378,516	2.22	9.43
256	128 × 3	128	64 × 3	256 × 3	7,014,528	2,108,712	427,522	378,516	2.18	9.43
256	128 × 3	128	128 × 3	256 × 3	7,403,904	2,108,712	427,522	378,516	2.16	9.43

Summary statistics for tomoDRGN heterogeneous network training using the simulated ribosome 4-class particles at various box and pixel sizes, sweeping the encoder A output layer size and the encoder B architecture (number of nodes per layer and number of layers).