PNPShift: an NMR chemical shift predictor for purine-nucleotide-bound and unliganded proteins

Accurate prediction of protein chemical shifts is important for NMR-based structural and ensemble modeling. However, most existing predictors are designed for unliganded proteins and cannot quantitatively calculate the chemical shifts of purine nucleotide-bound proteins, which are encoded by approximately 10% of the human genome and play critical roles in signal transduction, energy metabolism, and other cellular processes. Here, we introduce PNPShift, an artificial neural network-based program designed to improve chemical shift prediction accuracy for purine nucleotide-bound proteins. By incorporating the PNPShift-X_PN correction, PNPShift reduces ¹H^N prediction errors in nucleotide-binding regions by up to ~ 30% across 28 tested purine nucleotide-bound proteins. PNPShift also improves ¹H^N predictions for unliganded proteins, achieving a ~ 6% reduction in RMSE across an independent test set of 60 proteins. The utility of PNPShift in structural and dynamics modeling is demonstrated by its superior performance in maximum-entropy (MEM)-based ensemble refinement and its greater ability to discriminate native states from decoy conformations compared with other leading programs. PNPShift not only extends accurate chemical-shift prediction to purine nucleotide-bound proteins critical for diverse cellular functions, but also illustrates a general framework for developing predictors for other liganded protein systems.