With SFCalculator you can also *train* structure prediction / generative models directly on X-ray data. As
James Holton has shown, the data are much more accurate than the models. The result: experiment-based generative models for ensembles (not yet general!) 3/4

We can refine protein structures by searching the latent space of pre-trained generative models. Using a Boltzmann Generator (@franknoe) as an example, doing so yields models with good physics & data fit! 2/4

Crystal structures are *not* God-given truth. They approximate, w/ flaws & errors, X-ray diffraction data. AlphaFold etc. have been trained on structures, not data. SFCalculator now differentiably connects structures to diffraction data. What does this enable? 🧵 1/4 www.biorxiv.org/content/10.1...

In evolution, members of a protein family are often variations on a theme. The dynamic states resemble stable states of other channels: The sequence of steps of ion permeation is conserved from channel to channel with variations in relative energy. 7/8

Does the channel itself move? Yes! It moves as the ions pass, differently depending on the direction of ion flow. This suggests how rectification (asymmetric currents) can evolve. 6/8

What do we see? K+ ions coming in, hopping, and leaving. We see the K+ ions move, putting on a new “coat” of water molecules as they exit. 5/8

The experiment in simple: a crystal between two electrodes; X-rays coming in from the side: 4/8

Crystals of NaK2K, (H/T Youxing Jiang @ UTSW), form an ideal starting point: all ion channel pores are easily aligned with electric field (E): half of them “up”, half “down”. The red spheres are K+ ions: 3/8