Evaluating Prediction of Stabilising Mutations in GPCRs
Andrew Boardman1, Daniel Jones2, Jens Kleinjung3, Chris de Graaf 3 and Andreas Bender1
1 Center for Molecular Informatics, University of Cambridge, Cambridge, UK
2 Harvard Medical School, Harvard University, Cambridge, MA
3 Sosei Heptares, Granta Park, Cambridge, UK
Mutations that stabilise G protein-coupled receptors (GPCRs) in selected conformations have enabled structural and biophysical studies of key GPCR drug targets [1]. However, the experimental screening of candidate stabilising mutations is expensive, and computational models developed for globular proteins often transfer poorly to membrane proteins [2]. While recent work has proposed specialised machine learning models for membrane protein stability prediction [3–5], the sparsity of available data means these models are trained to identify neutral rather than strongly stabilising mutations, which we argue does not reflect practical utility. Instead, using published datasets, we evaluate the predictive performance of pretrained models for mutations that lead to significant increases in thermal stability, using classification and regression metrics. We show that features underlying published methods decline in performance significantly when stability thresholds are raised; for example, the AUROC of exposed surface area goes from 0.68 for neutral mutations to 0.55 for the strongest stabilisers. However, RaSP, a pretrained model based on structural features trained for sequence recovery, maintains an AUROC of 0.63 when the stability threshold is raised [6]. We present a combined pipeline for stability prediction on membrane proteins and test its performance when applied to models of GPCRs in different conformational states and from different sources (PDB, AlphaFold-2 [7]), and to common stabilising mutations in GPCRs.
References
[1] M. Congreve, C. de Graaf, N. A. Swain, and C. G. Tate, ‘Impact of GPCR Structures on Drug Discovery’, Cell, vol. 181, no. 1, pp. 81–91, Apr. 2020, doi: 10.1016/j.cell.2020.03.003.[2] B. M. Kroncke, A. M. Duran, J. L. Mendenhall, J. Meiler, J. D. Blume, and C. R. Sanders, ‘Documentation of an Imperative To Improve Methods for Predicting Membrane Protein Stability’, Biochemistry, vol. 55, no. 36, pp. 5002–5009, Sep. 2016, doi: 10.1021/acs.biochem.6b00537.
[3] P. Popov et al., ‘Computational design of thermostabilizing point mutations for G protein-coupled receptors’, eLife, vol. 7, 2018, doi: 10.7554/eLife.34729.
[4] S. Muk et al., ‘Machine Learning for Prioritization of Thermostabilizing Mutations for G-Protein Coupled Receptors’, Biophys. J., vol. 117, no. 11, pp. 2228–2239, 2019, doi: 10.1016/j.bpj.2019.10.023.
[5] S. P. D. Harborne, ‘IMPROvER: the Integral Membrane Protein Stability Selector’, Sci. Rep., p. 18, 2020.
[6] L. M. Blaabjerg et al., ‘Rapid protein stability prediction using deep learning representations’, bioRxiv, p. 2022.07.14.500157, Jan. 2022, doi: 10.1101/2022.07.14.500157.
[7] G. Pándy-Szekeres et al., ‘GPCRdb in 2023: state-specific structure models using AlphaFold2 and new ligand resources’, Nucleic Acids Res., vol. 51, no. D1, pp. D395–D402, Jan. 2023, doi: 10.1093/nar/gkac1013.
