Bayesian integrative approaches for high-dimensional imaging and genomic data

点击翻译按钮获取中文摘要

The primary objective of this proposal is the development of innovative Bayesian statistical frameworks for high throughput imaging (radiomic) and genomic data (e.g. gene expression, pathway activities, Methylation data and microRNA expression, etc..) arising in modern biomedical studies. The key foci are principled techniques for combining radiomic and genomic information to provide insight into the underlying biological mechanisms, and the development of reliable prediction models for relevant clinical outcomes to aid the practice of translational medicine. The work is motivated by the availability of these two data types on the same set of subjects. Genomic data are obtained from The Cancer Genome Atlas (TCGA) project. Patient image data and corresponding clinical data are extracted from The Cancer Imaging Archive. Imaging data have been processed to obtain a measure of image-derived heterogeneity for each patient called gray-level co-occurrence matrix (GLCM). Current approaches have focused on deriving summary statistics (e.g. entropy, energy etc) based on the GLCM, and potentially overlook other structural properties in the GLCM. In this program, we aim to investigate a framework that uses the GLCM both as an explicit covariate and endophenotype rather than restricting its use to some derived summary statistics. The proposed methods and computational tools are broadly applicable in all types of disease or in a variety of contexts with similar data (e.g: glioma, neurocognitive impairment, Alzheimers disease). The proposed research is positioned to advance the field of imaging-genomic integration in two directions; it responds to i) important scientific questions raised by these large-scale studies; ii) to the need for efficient statistical methodologies to help answer those questions. More specifically, we will develop innovative Bayesian statistical frameworks that define new prior distributions for integrating high-throughput radiomic and genomic datasets, account for the dependence between the data-sets. we will also develop computationally efficient and freely accessible software for the proposed methods. The methods would be able to identify important markers (e.g genes, imaging features) that are associated with clinical outcomes. Moreover, we expect to improve the prediction performance of those outcomes. This program, which is within the Natural Sciences and Engineering field, has diverse and interdisciplinary downstream applications. Hence, results from the application of our methods are expected to have an important positive impact clinically because the identified potential markers are likely to provide new data and methods for the development of integrative analysis methods that enable assessment of likely disease evolution and corresponding personalized treatment strategies.