Fig. 7. Validation of autoencoder-generated conformations for ChiZ by experimental SAXS and chemical shift data.

a Comparison of experimental and calculated SAXS profiles. MAPE was calculated as $<\mid O_i-E_i\mid /E_i{>}_{\mathrm{all}\mathrm{data}\mathrm{points}}\cdot 100\%$, where $E_i$ and $O_i$ are experimental and predicted scattering intensities, respectively, and $<\cdots {>}_{\mathrm{all}\mathrm{data}\mathrm{points}}$ denotes the average over all the data points. b Comparison of experimental and calculated secondary chemical shifts. The experimental data and the MD simulations are reported previously¹⁴. RMSE was calculated as the square root of $<{\left(O_i-E_i\right)}^2{>}_{\mathrm{all}\mathrm{data}\mathrm{points}}$. Calculations were done on either the test set comprising 12,180 conformations sampled from 12 MD runs, or on an autoencoder-generated set comprising the same number of conformations, after refinement. The autoencoder was trained on a combined training set comprising 52,200 conformations sampled from the 12 MD runs.