Abstract
Side-chain methyl group NMR spectroscopy provides invaluable insights into macromolecular structure, dynamics, and function, particularly for large biomolecular complexes. Accurate assignment of methyl group resonances in two-dimensional spectra is essential for structural and dynamics studies. Traditional methyl group assignment strategies rely on either transferring assignments from backbone resonance data or NOESY data and high-resolution experimental structures; however, these methods are often limited by molecular size or availability of structural information, respectively. Here, we describe the use of AlphaFold2 structural models as a basis for the manual, distance-based assignment of side-chain methyl group resonances in the folded domains of S. cerevisiae Xrs2. While AlphaFold2 models facilitated initial assignments for the methyl resonances, inaccuracies in the side-chain coordinates highlighted the need for improved structural models. By generating >500 ColabFold-derived models and filtering with methyl residual dipolar couplings (RDCs), we identified structural models with superior agreement to experimental data. These refined models enabled additional methyl group assignments while suggesting an iterative approach to simultaneously improve structure prediction and resonance assignment. Our findings outline a workflow that integrates machine learning-based structural predictions with experimental NMR data, offering a pathway for advancing methyl group assignment in systems lacking high-resolution experimental structures.