Using Tissue Morphology to Infer Intra-Tumor Variation in Molecular State and Response

Intra-tumor heterogeneity (ITH) presents a significant challenge in the effective treatment of cancer. The traditional approaches to characterize ITH, such as multi-regional sequencing, are cost prohibitive for broad adoption. In contrast, histopathology slides are ubiquitous and, as evidenced by their extensive use in clinical diagnosis, capture key aspects of tumor biology. However, the scale and complexity of the morphological phenotypes in tissue slides render manual quantification impractical. This work presents deep learning approaches to understand the basis and implications of ITH through the lens of tissue morphology. These approaches are applied to two cancer types clear-cell renal cell carcinoma (ccRCC) and colorectal cancer (CRC). In ccRCC, we aim to understand the extent to which intra-tumor variation in key driver genes, BAP1, PBRM1, and SETD2, can be inferred purely from the analysis of histopathology slides. Our work demonstrated for the first time that the morphology-genetics relationship was powerful enough for inference of intra-tumor heterogeneity in driver mutation status purely from histopathology images. Our predictive power was especially high in the case of BAP1 -- highest AUC for any gene across all cancers at the time-- and we were able to validate this across a range of cohorts including in PDX models. In CRC, we sought to use early radiation-induced changes in tissue morphology to better understand response to radiation therapy, for which we currently lack any biomarkers. To this end, we developed a novel deep learning framework to analyze changes to the composition, gland structure, and overall tissue phenotype in clinical trial patients two weeks after they were treated with radiation. As most studies profile tissue months after radiation, this work represents the first quantification of radiation response at such an early time point in clinical tissue samples. Together, these projects highlight the ability of deep learning analysis of tissue morphology to provide insights into genetic and response heterogeneity in tumors. Future work will focus on refining these predictive models for greater interpretability and accuracy, alongside cellular and molecular studies aimed at uncovering the underlying mechanisms driving these morphological changes. By establishing a more nuanced comprehension of ITH and its impact on treatment outcomes, this thesis contributes to the evolving landscape of cancer biology, offering new insights into the interplay between form and function.

Author dc:creator

• Acosta, Paul Humberto

Contributors dc:contributor

• Conacci-Sorrell, Maralice
• Montillo, Albert A.
• Lin, Milo
• Malladi, Srinivas
• Rajaram, Satwik

• Carcinoma, Renal Cell
• Colorectal Neoplasms
• Deep Learning
• Genetic Heterogeneity

Language dc:language

en

Identifier

1596185065

OAI identifier oai:identifier

oai:utswmed-ir.tdl.org:2152.5/10819