UGA Cancer Center Brown Bag Lunch Seminar Series presents "Integrative Statistical Learning Approach for Cancer Genomics and Microbial Science."

Abstract: The modeling of interrelated responses utilizing correlated predictors within high-dimensional contexts emerges as a compelling conundrum in the realm of integrative statistical learning, resonating across diverse scientific inquiries. This dilemma is manifest in scenarios such as genomics, where insights into the modulation of gene expression in yeast cells across multiple temporal points via transcription factors are sought; ocean microbiome investigations, delving into the influence of environmental and geochemical variables on microbial abundance; gut microbiome analyses, unraveling the interplay between host attributes and microbial prevalence; and cancer genomics, probing the interrelations among omics datasets, including microbiome and metabolomics information. Effectively addressing this challenge involves harnessing a multivariate analysis framework, synergized with a low-rank, sparse coefficient matrix, to infer the latent associations underpinning the phenomena. However, the simultaneous imposition of orthogonality and sparsity constraints renders the decipherment of such a decomposition arduous. To surmount this, we introduce a divide-and-conquer strategy for inferring the coefficient matrix from data, entailing decomposition into unit-rank matrices with sparse left and right singular vectors, estimated via a sequential (greedy) approach. Our method accommodates diverse outcome types, assuming conditional independence and adherence to an appropriate exponential dispersion family. Implementation is realized through R packages "secure" (Gaussian outcomes), "GO-FAR" (exponential outcomes), "NB-FAR" (negative binomial outcomes) and “TARO”, with demonstrated efficacy on multi-omics data encompassing colorectal cancer, yeast cell cycle, and Hepatocellular carcinoma datasets.