MOBS – Multi-Omics Brush for Subgraph visualisation

Abstract

One of the big opportunities in multi-omics analysis is the identification of interactions between molecular entities and their association with diseases. In analyzing and expressing these interactions in the search for new hypotheses, multi-omics data is often either translated into matrices containing pairwise correlations and distances, or visualized as node-link diagrams. A major problem when visualizing large networks however is the occurrence of hairball-like graphs, from which little to none information can be extracted. It is of interest to investigate subgroups of markers that are closely associated with each other, rather than just looking at the overload of all interactions. Hence, we propose MOBS (Multi-Omics Brush for Subgraph visualisation), a web-based visualisation interface that can provide both an overview and detailed views on the data. By means of a two dimensional brush on a heatmap that includes hierarchical clustering, relationships of interest can be extracted from a fully connected graph, to enable detailed analysis of the subgraph of interest. D.H. and J.P. are funded through Hasselt University BOF grants (BOF20OWB29 & BOF20OWB33). D.H. also receives funding from VITO (R-11362). Figure: Full overview of MOBS' functionality; in detail exploration of subgraphs from a larger biological correlation network after hierarchical clustering is applied. Relations between proteome (blue), clinome (yellow-orange) and metabolome (red) nodes, that are considered important (absolute value > 0.37), are shown as a subgraph after applying a rectangular brush on the heatmap. Implementation MOBS is a web-application developed in Svelte (https://svelte.dev/), using the D3.js library, and can be either installed and run locally or be used from the online hosted version on vercel (https://mobs.vercel.app/). The source code is publicly available, and detailed instructions on how to use and install the tool are provided in the README.md file in the GitHub repository, as well as the complementary data and scripts generated within the use case to demonstrate and evaluate the functionality of the tool. MOBS contains various functionalities that enable efficient analyses. Different styling options are available for both the adjacency matrix and the node-link diagram (e.g. adjusting color and size of nodes and edges based on node values, edge weights, clusters, or any other variable included in the data). Tooltips are available in the node-link diagrams displaying information on all variables provided in the node list, and edge weights. In addition, node highlighting and dragging is included for easy reference of interactions. The tool comes with several interactive features to optimize the user experience. The key interactive feature of the tool resolves around brushing on the adjacency matrix to extract sub parts of the graph. One can also zoom on both the adjacency matrix and node-link diagram (triggered by holding shift key in combination with mouse events) to obtain more detail. In addition, a threshold can be set to filter links in the node-link diagram based on edge weight. Since the identification of interesting (multi-omics) clusters is crucial to identify parameters with similar interaction patterns, hierarchical clustering is included in the tool using the DRUIDjs library to rearrange the order of the nodes across all data types in the adjacency matrix. Whenever a clustering method is chosen, the corresponding dendrogram will be added on top of the heatmap. A clustering threshold can be set to define the maximum depth of clusters, nodes can be assigned to. Conclusion With MOBS we propose a tool that enables the expert user to include and explore different types of omics data. Scalability is the main focus for future work. This toolset provides a clear visual overview of up to 500 parameters, but above that, the computational load to render the visualisations becomes too high. Hence, additional actions need to be taken to either improve the rendering of larger data files (e.g. use WebGL or Canvas based rendering libraries), or include algorithms that rely on clustering in the preliminary network generation process (e.g. Mapper).