A structural machine learning approach for rapid prediction of thermodynamically destabilizing tyrosine phosphorylations

Abstract

Tyrosine phosphorylations are a prominent characteristic of numerous cancers, necessitating the use of computational tools to comprehensively analyze phosphoproteomes and identify potentially (dys)functional phosphorylations. Here we propose a machine learning-based method to predict the thermodynamic stability change resulting from tyrosine phosphorylation. Our approach, based on prediction of phosphomimetic delta-delta-G from structural features, strongly correlates with experimental mutational scanning cDNA proteolysis data (R = 0.71). We predicted the destabilizing effects of all 384,857 tyrosine residues from the Alphafold2 database. We then applied our approach to a pan-cancer phosphoproteomics dataset, comprising over 600 unique tyrosine phosphorylations across 11 cancer subtypes. We predict destabilizing phosphorylations in both oncogenes and tumor suppressors, where the former likely reflects a generalized relief of auto-inhibition or activating conformational change. We find that the number of circuit topological parallel relations with respect to residues contacting the phosphorylated site is greater for autoinhibited oncogenes than for other proteins (Wilcoxon p = 0.03). Utilizing an extreme gradient-boosting machine learning approach, we obtain an AUC of 0.85 for the prediction of autoinhibited phosphorylation states from circuit topological features. The top destabilized proteins from the pan-cancer data are enriched for chemical and oxidative stress pathways. Among metabolic proteins, highly destabilizing phosphorylations tend to occur in more peripheral proteins with lower network centrality measures (Wilcoxon p = 0.005). We predict 58% of recurrent tyrosine cancer phosphorylations to be destabilizing at the 1 kcal/mol threshold. Our approach can enable rapid screening of destabilizing phosphorylations and phosphomimetic mutations.