SFCalculator: connecting deep generative models and crystallography.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Proteins function through dynamic conformational changes, crucial for biological processes.
• X-ray crystallography and machine learning have advanced protein structure prediction, but a gap exists in connecting atomic models to crystallographic data.
• This disconnect limits improvements in prediction accuracy and machine learning applications in experimental structure determination.

Purpose of the Study:

• To introduce SFCalculator, a differentiable pipeline connecting atomistic molecular structures to crystallographic observables.
• To bridge the gap between crystallographic data and neural network-based molecular modeling.
• To enable new analytical paradigms integrating experimental structure determination and machine learning.

Main Methods:

• Developed a differentiable pipeline (SFCalculator) to generate crystallographic observables from molecular structures.
• Incorporated bulk solvent correction for accurate data generation.
• Validated SFCalculator against conventional methods.

Main Results:

• SFCalculator enables accurate molecular model placement within crystal lattices (phasing).
• It facilitates searching generative model latent spaces for conformations consistent with crystallographic data.
• Allows direct use of crystallographic data in training generative models for conformation ensembles.

Conclusions:

• SFCalculator provides a crucial link between crystallographic data and machine learning in structural biology.
• It opens new avenues for enhancing protein structure prediction accuracy and experimental structure determination.
• Enables a new generation of analytical tools integrating experimental and computational approaches.