Deep residual networks for crystallography trained on synthetic data

Mendez, D.; Holton, J.M.; Lyubimov, A.Y.; Hollatz, S.; Mathews, I.I.; Cichosz, A.; Martirosyan, V.; Zeng, T.; Stofer, R.; Liu, R.; Song, J.; McPhillips, S.; Soltis, M.; Cohen, A.E.

doi:10.1107/S2059798323010586

Acta Crystallographica Section D
Acta Crystallographica
Section D STRUCTURAL BIOLOGY

IUCr IT WDC

home archive editors for authors for readers submit subscribe open access

view article

Figure 5
Model architecture. (a) Raw data were downsampled as described in Section 2.2

, forming four 512 × 512 quadrants. Quadrants were then passed through a ResNet architecture, resulting in 1000 features. Next, a series of fully connected layers (FC1, FC2) was used to convert the 1000 features into a scalar value. If predicting resolution (b), this was converted to an inverse resolution using the diffraction wavelength (λ), downsampled pixel size (p) and sample-to-detector distance (d). If predicting overlapping lattice scattering (c), this scalar was passed through a sigmoid function and then rounded, such that 0 and 1 indicated single and overlapping lattice scattering, respectively. The image and line plots in (a) are from a real experimental image as it was passed through the fully trained resolution model. The inferred resolution in this case was 1.67 Å. Table 2

describes the number of parameters in the different model stages (ResNet, FC1, FC2). One quadrant was sufficient to predict the quantities of interest; however, repeated model passes with the second, third and fourth quadrants can provide a measure of uncertainty in the predicted values.

STRUCTURAL
BIOLOGY

ISSN: 2059-7983

Volume 80| Part 1| January 2024| Pages 26-43

https://doi.org/10.1107/S2059798323010586

Open

access

Follow Acta Cryst. D

Search IUCr Journals		doi		Advanced search
Author		volume	page