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

Fragmentation-based quantum chemistry offers a practical route to achieving ab initio quality energetics for proteins, but its applicability is often constrained by the high computational cost of fragment-level quantum mechanical (QM) calculations. Here, we introduce GMFCC-UMA, a method that integrates the generalized molecular fractionation with conjugate caps (GMFCC) framework with a fine-tuned foundation neural network potential, thereby eliminating the need for fragment QM calculations. In this approach, the total protein energy is decomposed into adjacent and nonadjacent interaction components. Adjacent terms are assembled in an MFCC II-consistent manner from ACE-NME-capped fragments, while nonadjacent interactions are treated hierarchically: near–contact pairs are evaluated using the fine-tuned UMA model, and longer-range contributions are described via a complementary molecular mechanics (MM) treatment. The underlying neural network model is fine-tuned on a large, protein-focused fragment data set that includes capped monomers, adjacent dimers, and nonadjacent residue pairs and corresponding blocks, ensuring consistent accuracy across all fragment types required by the GMFCC fragmentation scheme. Across benchmark proteins and conformational ensembles, GMFCC-UMA yields relative energies that closely match quantum-based fragmentation references and systematically outperforms conventional force fields in both error reduction and correlation. By replacing fragment-level QM with neural network inference, GMFCC-UMA achieves an order-of-magnitude acceleration while retaining ab initio fidelity, thereby enabling scalable and accurate protein energy evaluation for high-throughput conformational analysis and related applications.