Computational models benefit from high quality experimental data. In previous work, we curated the Predicted and Experimental Peptide Binding Information (PEPBI) database, which has structural and isothermal titration calorimetry experimental data for 329 protein-peptide complexes. In this work, we describe how we used partial least squares regression to train a ΔΔG model for how mutations affect peptide-protein binding energies. The resulting model has a Pearson correlation of 0.89 and an R² of 0.79, demonstrating strong predictive performance. Although the predictive power diminishes significantly when the model is applied to entirely novel complexes, performance is rapidly recovered through including just a few experimental data points for a complex. By identifying key binding determinants, this work offers a practical approach to protein-peptide interaction modeling and contributes to data-driven peptide engineering strategies.