Despite enormous efforts to structurally characterise GPCRs, experimental elucidation of their 3D structures remains a challenge. This is in part because they are dynamic, metastable membrane proteins. In order to obtain high-resolution crystals, GPCR engineering is usually required to minimise conformational heterogeneity and maximise crystal contacts and stability (Kobilka and Deupi, 2007). In this study, we explored the use of computational approaches to guide rational GPCR engineering. Initially, we performed extensive manual curation from the literature. In total, we curated 1808 mutations across 46 different GPCRs with experimental measured stability data. This was used to assess the performance of 8 different computational approaches for screening the effects of mutations on protein stability. The maximum accuracy and Matthew’s correlation coefficients across these methods were just 0.65 and 0.30, respectively, despite some of these mutations present in their training data. This highlighted the complexity of using generic computational approaches to model mutations in GPCRs. We next sought to develop a predictor capable of identifying mutations that improve GPCR stability without compromising function. Over 622 features describing both the protein sequence and 3D structure were generated. Applying supervised machine learning, our initial models have achieved a Matthew's correlation coefficient of 0.30 when evaluated against a non-redundant test set. Interpreting the model, the major contribution to performance were sequence-based quantitative measures (substitution probability tables; Pandurangan et al., 2017). Moving forward, we are working on understanding on what mutations the model works well, in order to further improve our predictive performance, and to develop rules to help guide GPCR engineering