Networks for tissue-specific phenotype prediction from genomes

点击翻译按钮获取中文摘要

Network science excels at modelling relationships at different levels of systems organization. An emerging area of research is in predicting phenotype from genomic data in multicellular organisms. This context uses similarity networks, where nodes are samples and edges are weighted by pairwise similarity for a type of genomic data. Similarity networks provide a conceptually intuitive framework to integrate heterogeneous data by converting each to the common space of sample similarities. They outperform clustering and classification by concatenation, as they preserve correlation structures in each data layer. Classification tasks use machine learning to find generalizable patterns in complex data, but making models interpretable, needed for mechanistic insight, remains an open challenge. We previously developed a network-based classifier that integrates multi-'omic data and adapts recommender systems such as those used in Netflix ("find movies like this one"), to phenotype ("find samples like this one"). In addition to excellent performance, this algorithm, netDx, provides the ability to create biologically-interpretable features, such as pathways. netDx currently lacks several advantages of the deep learning framework, a fast-growing area of machine learning with superior non-linear discriminability and the ability to use pre-trained models for new applications. Separately, after a model identifies predictive genes or pathways, gene prioritization is used to select molecules for phenotypic perturbation experiments. Current prioritization methods are agnostic to tissue-specific differences in molecular interaction networks, which increase risk of off-target effects. Interpretable phenotype prediction from multi-modal genomic data. We will adapt the netDx classifier to deep learning, using graph convolutional networks. We will evaluate feature design strategies for transcriptomic, epigenomic and chromatin accessibility data to identify general principles for creating interpretable predictors from the coding and non-coding genome. We will evaluate the ability to transfer pre-trained models from one tissue to a related tissue, or a related species. Tissue-specific perturbation modeling for drug prioritization. We will develop a network-based gene prioritization algorithm that uses tissue- and cell-specific molecular interaction networks to find direct and indirect drug targets, and optimize tissue-specificity of candidate genes. Significance, impact, applications, training opportunities: Our methods will be applicable to any problem involving phenotype prediction from multi-`omic data, or to predicting tissue-specific impact of pharmacological perturbation. This work will advance knowledge of how tissue-specific genome regulation impacts phenotype. This research program will provide training opportunities for Canadian HQP in machine learning, network science, and in interpretable predictive modeling of biological systems.