Integrating high-throughput histology with systems genetics through causal graphical models

点击翻译按钮获取中文摘要

PROJECT SUMMARY The overall objective of this proposal is to develop and validate a deep learning analytical framework to integrate histological traits into systems genetics analysis of complex diseases. Mapping the risk genes for poor health outcomes is a major focus of biomedical research, and new approaches to improve genetic mapping power can have a transformative impact on public health. Genetic disease risk manifests through complex interactions between gene regulation and tissue structure, ultimately influencing organ function. However, quantifying tissue structure for quantitative genetic mapping has not been widely adopted. This is partly because histological scoring has traditionally been labor intensive and error prone, and limited to coarse measures (e.g., discrete categories) that are suboptimal for association testing. In contrast, deep neural networks (DNNs) now routinely automate laborious image quantification tasks for histopathology, making them an ideal platform for integrating histology into genetic analysis. Furthermore, unlike human-defined histological scores, DNN readouts enable objective histological trait discovery as a function of genetic, molecular, and physiological variation. In this project, histological features will be rigorously and robustly quantified using DNNs and these data will be integrated into novel multiscale statistical models that will connect genetic, molecular, and histological variation to physiological outcomes. In particular, novel methods will be developed to integrate histology into three major classes of systems genetic analysis, i.e., heritable trait inference, causal mediation analysis, and molecular quantitative trait locus (mQTL) mapping. These methods will be developed and validated using a data set of genetic, histological, transcriptomic, proteomic, and physiological data from a cohort of Diversity Outbred mice used for the study of age-related kidney disease. By using a model system, complex genetic effects and causal mediation hypotheses can be directly tested to validate and refine the analytical framework. The specific aims of this proposal include: Aim 1: Identify maximally heritable histological traits through deep learning on paired genotypes and histological images. Aim 2: Genetically map histological mediators of complex physiological traits using deep learning on histological images. Aim 3: Identify causal paths connecting molecular QTLs (mQTLs) to outcomes through histological mediators. The outcome of this study will be a validated methodological framework for histological systems genetics that is modular, enabling a wide range of users to incorporate appropriate computer vision tools into state-of-the-art systems genetics workflows for any complex disease.