Machine Learning in Complex Hierarchical Domains: Navigating the Genomic Landscape to Unravel the Biology of Blood Cancers

Abstract

Blood cancers comprise over 160 distinct entities that include leukemias, lymphomas, and plasma cell myelomas. While the World Health Organization (WHO) has organized subtypes into a hierarchy reflecting shared biology, studies often analyze subtypes in isolation or treat them as unrelated labels. This thesis explores viewing the blood cancer taxonomy itself as a computational object. Searching its leaves can uncover subtype-specific insights while the whole tree can be leveraged to learn generalizable representations.Chapter 1 explores the leaf level through a multi-institutional cohort of extranodal marginal zone lymphoma (EMZL), integrating mutation, expression, fusion, and copy-number (CN) analyses across tumors from diverse anatomic sites. The study identifies recurrent pathway-level themes while showing how shared alterations coexist with site-specific transcriptomic and microenvironmental profiles.Chapter 2 addresses a recurring bottleneck in extracting biological insights in such analyses. Namely, transcriptomic interpretation often relies on large, redundant Gene Set Enrichment Analysis (GSEA) outputs that require principled aggregation. This chapter thus introduces gtCLIP, a contrastive learning framework that aligns gene sets with pathway text descriptions using a soft-target objective. gtCLIP improves GSEA interpretability by learning reusable embeddings that facilitate clustering the enriched pathways into coherent biological communities, enabling consistent cross-study interpretation.Chapter 3 expands Chapter 1’s leaf-level view to the full hierarchy with BLOOM, a transfer learning pipeline that learns diagnosis-aligned multi-omic embeddings (DAMEs) across a large cohort spaning the whole blood cancer taxonomy. BLOOM uses these representations for clinically robust, hierarchy-aware diagnosis as well as survival analysis to stratify patients into clinically meaningful risk groups.Together, the chapters connect a single-subtype “leaf” investigation (EMZL), an interpretability tool that strengthens pathway-based analyses (gtCLIP), and a hierarchy-aware modeling framework (BLOOM) into a unified multi-scale computational exploration of the family of hematologic malignancies.