FlowBack: A Generalized Flow-Matching Approach for Biomolecular Backmapping.

Area of Science:

• Biomolecular Modeling
• Computational Biophysics
• Structural Biology

Background:

• Coarse-grained (CG) models accelerate molecular simulations of slow dynamics like protein folding.
• Restoring all-atom (AA) detail from CG trajectories is crucial for mechanistic insights but remains challenging.
• Existing methods for CG to AA reconstruction often lack accuracy and efficiency.

Purpose of the Study:

• Introduce FlowBack, a deep generative model for accurate and efficient CG to AA structure reconstruction.
• Develop a flexible prior distribution for FlowBack, independent of CG mapping and molecular type.
• Evaluate FlowBack's performance on protein and DNA-protein complex systems.

Main Methods:

• Utilized a flow-matching objective for a deep generative model (FlowBack).
• Constructed a molecular type-agnostic prior distribution for the generative model.
• Trained protein-specific and DNA-protein specific models on Protein Data Bank (PDB) data.

Main Results:

• Achieved state-of-the-art performance on structural metrics for proteins compared to existing methods.
• Demonstrated excellent reconstruction and generative capabilities for DNA-protein complexes.
• Showcased FlowBack's effectiveness on static structures, AA simulations, and CG dynamical trajectories.

Conclusions:

• FlowBack provides an accurate, efficient, and user-friendly tool for recovering all-atom detail from CG simulations.
• The model exhibits superior robustness and fewer steric clashes than previous approaches.
• FlowBack is released as an open-source Python package for community use.