Poster L12 - CoDIAC (Comprehensive domain interface analysis of contacts) : A bioinformatics toolkit for feature extraction and analysis

Introduction: Protein domains are conserved structural units that play a critical role in cell signaling and in mediating protein-protein interactions through the regulation of cellular processes such as cell growth and differentiation. Non-catalytic domains involve in the formation of protein complexes by recognizing specific peptide motifs, which are the key to signal transduction that are important to development and homeostasis of human tissue. Identifying interactions between protein domains and its ligand partners is essential in understanding the structural and functional relationships of proteins. Mutations and PTMs present on the protein domain regions can modulate the regulatory role of the proteins and interfere with cellular signaling cascades. Interactions between various protein-domain interfaces are essential in governing the interaction networks that are regulated by cell signaling. To identify these molecular interactions from protein structures, we developed a bioinformatics pipeline CoDIAC for feature extraction and analysis. We wished to develop a generalizable approach for extracting the many possible domain interaction interfaces of modular domains in order to assess the possible determinants of canonical binding to domain-motif interactions, as well as understand the larger regulatory structure of a protein. Along with this, we wished to understand if subunit architectures retain conserved binding interactions as they are reused in larger, more diverse protein domain architectures (PDA).

Materials and

Methods: In this work, we develop a computational pipeline that uses controlled ontology of domains from InterPro and primarily pulls information about domain-associated proteins, its structure and sequence information from annotation databases and 3D structural details from experimental archive for proteins of interest. We demonstrate the functionality of CoDIAC using the SH2-pTyr domain-motif interaction module. Though this framework is generalizable to any protein domain, we use it here to deeply explore the binding regions between SH2 domains and their ligands (canonical interface), SH2 domains and other intra-protein domains, and the secondary binding regions (SH2 domains and lipids) to analyze how these overlap with PTMs. Since there are 120 SH2 domains and 46,000 tyrosine phosphorylated residues in human proteome, computational approaches will be helpful to advancing our understanding of how SH2 – pTyr ligand complexes occur in space and time and how they become dysregulated in diseases like cancer. This work paves way for building a machine learning model that can predict the Sh2-pTyr interactions that uses homology modeling, sequence-based methods, and binding affinities of SH2 domain and pTyr ligands as inputs to be trained.

Results, Conclusions, and Discussions: From a global analysis of ligand binding interactions across PDB structures, we find that the contact maps are highly correlated with the specificity of ligand selection. The ligand contacts between the closely evolved proteins are highly identical within the non-tandem SH2 domain proteins but interestingly the tandem SH2 domains (not always evolutionarily closely related) of the same protein share highly identical feature sets. We also find that if an inter-domain interface is part of identical PDA across proteins, then the contacts also remain conserved. Using AlphaFold structures, we found that the SH2-Kinase interfaces that combine with diverse domains forming different gene products, are not conserved across identical PDAs. But within other super architectures like SH2-SH3, we find high conservation across identical PDAs.
Currently available experimental data can give us binding site information for only 14% of the SH2ome and this clearly necessitates the development of global prediction algorithm for SH2 domain interactions. A machine learning framework would allow us to predict and explore the interaction space that cannot be tested by available experimental methodologies. This ML model design will enable us to identify key regulatory interactions in pTyr-signaling system and analyze its effects on the downstream signaling outcomes. Further, this kind of systems-level predictions of SH2-pTyr interactions will allow us to not only classify the key interactions but also to predict the critical nodes in the network.
For example, anti-PD1 therapies are successful to reactivate immune surveillance of tumors in 25% of the cases and there is an equal chance of accelerating the disease using this immunotherapy. This suggests that we don’t well understand the molecular mechanisms that shapes TCR activation and signaling. Such complex systems comprising of hundreds and thousands of pTyr sites changing in response to TCR and cofactor stimulation by PD1 require a systems level understanding of how and what interactions occur, and this knowledge will be the chief component in predicting patient response. Hence, extracting structure-based features and predicting SH2-centric interactions using systems-level approaches will help us gain insights to understand pTyr-mediated signaling of TCR activation and how its cofactors reshape this sufficiently.

Acknowledgements (Optional):