GMFCC-UMA: A Fragment-Based Machine Learning Framework for Scalable Ab Initio-Quality Protein Energies.

Area of Science:

• Computational Chemistry
• Biophysics
• Structural Biology

Background:

• Fragmentation-based quantum chemistry provides accurate protein energetics but is computationally expensive.
• High computational cost of fragment-level quantum mechanical (QM) calculations limits applicability.
• Need for efficient methods to evaluate protein energies for large-scale studies.

Purpose of the Study:

• Introduce GMFCC-UMA, a method combining generalized molecular fractionation with conjugate caps (GMFCC) and a neural network potential.
• Eliminate the need for computationally intensive fragment QM calculations.
• Enable scalable and accurate protein energy evaluation.

Main Methods:

• Integrate GMFCC framework with a fine-tuned foundation neural network potential (UMA).
• Decompose protein energy into adjacent and nonadjacent interaction components.
• Utilize ACE-NME-capped fragments for adjacent terms and hierarchical treatment for nonadjacent interactions (UMA model and molecular mechanics).

Main Results:

• GMFCC-UMA closely matches quantum-based fragmentation references for relative energies.
• Outperforms conventional force fields in error reduction and correlation across benchmark proteins.
• Achieves an order-of-magnitude acceleration compared to QM methods while retaining ab initio fidelity.

Conclusions:

• GMFCC-UMA offers a computationally efficient alternative to traditional QM methods for protein energy calculations.
• The method enables accurate and scalable high-throughput conformational analysis.
• Represents a significant advancement in computational biophysics and structural biology.