Machine Learning-Driven Multiscale Modeling: Bridging the Scales with a Next-Generation Simulation Infrastructure.

Area of Science:

• Computational biology
• Biophysics
• Molecular modeling

Background:

• Interdependence across biological scales is critical, particularly in cancer signaling pathways involving RAS and RAF proteins.
• Simulating RAS/RAF protein-membrane interactions requires methods that capture atomic detail over long timescales and large spatial scales.

Purpose of the Study:

• To develop and validate a multiscale modeling approach for simulating RAS/RAF protein-membrane interactions.
• To identify specific lipid-protein interactions that influence RAS/RAF complex formation and orientation on the plasma membrane.

Main Methods:

• The Multiscale Machine-Learned Modeling Infrastructure (MuMMI) employs an automated, ensemble-based multiscale approach.
• MuMMI integrates three scales: continuum (milliseconds, μm²), coarse-grained Martini (protein-lipid interactions), and all-atom (specific lipid-protein interactions).
• Dynamically coupled scales use machine learning for forward sampling and backward feedback, enhancing simulation fidelity.

Main Results:

• MuMMI successfully resolves RAS/RAF protein-membrane interactions at the plasma membrane.
• Identified specific lipid-protein fingerprints that promote protein orientations favorable for effector binding.
• Demonstrated efficient scalability from few to many compute nodes and generalizability to diverse systems.

Conclusions:

• MuMMI provides a powerful, automated tool for simulating complex biological systems at multiple scales.
• This multiscale modeling infrastructure is poised to advance the study of intricate biological questions, including cancer signaling.
• Advancements in computing and multiscale methods will drive the common use of automated simulations for scientific discovery.