DMS/NIGMS 2: Novel Statistical Methods, Algorithms, and Pipelines for Learning Omits Data with Complex Heterogeneity

Confronted with the pressing challenges of cancer and other complex diseases, the pivotal role of heterogeneity learning becomes undeniably clear. This approach offers a fine-tuned lens, enabling us to decipher key disease characteristics, direct precise therapeutic strategies, and mitigate widening treatment disparities among diverse patient subgroups. Addressing this need, our proposal champions the development of cutting-edge statistical methods tailored for high-dimensional heterogeneity learning. Specifically, we intend to innovate in studying subgroups within genomic, transcriptomic, or connectomics data, even amidst potential data shifts and intricate multidimensional-array structures. Traditionally, unsupervised heterogeneity learning has sought to partition unlabeled heterogeneous data into distinct clusters, each expected to conform to a unique distribution. This commonly known “model-based heterogeneity learning” characterizes data as a mixture of distributions. However, the prevailing Gaussian presumption in this approach can be limiting and at times, off-mark. To rectify this, our project introduces a novel class of mixture models, a departure from conventional parametric models. This new class stands out as it refrains from setting strict distributional assumptions, meeting the growing demand for more adaptable heterogeneity learning tools. To validate the effectiveness of our techniques, comprehensive assessments will be carried out using The Gene Expression Omnibus Database, The Cancer Genome Atlas Database, and a study focused on intricate pediatric conditions. The success of this endeavor not only promises breakthroughs in research on life-threatening ailments but also sets the stage for groundbreaking statistical methods for a wider scope in neurosciences and biomedicine.