A full-proteome, interaction-specific characterization of mutational hotspots across human cancers

Abstract

Rapid accumulation of cancer genomic data has led to the identification of an increasing number of mutational hotspots with uncharacterized significance. Here we present a biologically informed computational framework that characterizes the functional relevance of all 1107 published mutational hotspots identified in approximately 25,000 tumor samples across 41 cancer types in the context of a human 3D interactome network, in which the interface of each interaction is mapped at residue resolution. Hotspots reside in network hub proteins and are enriched on protein interaction interfaces, suggesting that alteration of specific protein–protein interactions is critical for the oncogenicity of many hotspot mutations. Our framework enables, for the first time, systematic identification of specific protein interactions affected by hotspot mutations at the full proteome scale. Furthermore, by constructing a hotspot-affected network that connects all hotspot-affected interactions throughout the whole-human interactome, we uncover genome-wide relationships among hotspots and implicate novel cancer proteins that do not harbor hotspot mutations themselves. Moreover, applying our network-based framework to specific cancer types identifies clinically significant hotspots that can be used for prognosis and therapy targets. Overall, we show that our framework bridges the gap between the statistical significance of mutational hotspots and their biological and clinical significance in human cancers.

A multiscale functional map of somatic mutations in cancer integrating protein structure and network topology

“…(ii) Heat delivery ratio α: we used α = 5, which was determined based on experimental evidence. Studies of known disease mutations have shown a 70.6% rate of PPI-perturbing effects for missense mutations at the interfaces, as well as a 13.3% rate for non-interface mutations ( 31, 32 ). These findings suggest that functional missense mutations on the partner-specific PPI interfaces are ~5 times more likely to perturb the corresponding interactions than the interactions with other partners.…”

“…In NetFlow3D, the heat sources are defined as the proteins involved in 3D clusters (or LOF enrichment signals), while in standard PPI network analyses, proteins with any mutations will be considered as heat sources. Second, previous studies have shown a 70.6% PPI-perturbation rate for known disease mutations on the interfaces, as well as a 13.3% rate for non-interface mutations (31,32). Driven by this ~5-fold difference, NetFlow3D uses a 5:1 ratio of heat transfer rate through PPIs when 3D clusters are on vs. not on the interfaces.…”

“…Second, we force the heat to diffuse unevenly from a given protein to its neighbors by designing a 5:1 heat delivery ratio towards the neighbors when 3D clusters are on vs. not on the corresponding PPI interfaces. This design is inspired by recent studies ( 31, 32 ) that report a 5-fold difference in PPI-perturbation rates when known disease mutations are on vs. not on the binding interfaces. Taken together, our network propagation model incorporates the local spatial organization of mutations on the 3D structures of proteins and PPIs, and therefore achieves a higher resolution than the conventional node-and-edge view of the PPI network.…”

“…Importantly, single protein structures alone (even though covering the entire human proteome) are still not enough for the comprehensive identification of all 3D clusters. This is because many driver mutations accumulate at the binding interfaces of cancer-related PPIs ( 22, 32, 33 ). Only looking at individual proteins will split inter-protein 3D clusters into smaller fragments on individual proteins, making them harder to identify.…”

Leveraging Protein Dynamics to Identify Functional Phosphorylation Sites using Deep Learning Models

“…Recent advances in different levels of “omics” data, from proteomics to structural-phosphoprotomics, have exponentially accelerated the identification of PTM sites. This boosted a number of known PTM databases with compiled information, including protein sequence and three-dimensional (3D) structure, − PTM functional annotations, , PTM involved in diseases, PTM crosstalk, and PTM-associated mutations, , etc. Meantime, data-driven and machine learning (ML)-based predictors have been widely recognized as an effective approach to rapidly uncover potential PTM sites with fast speed but low cost.…”