MorphoITH: a framework for deconvolving intra-tumor heterogeneity using tissue morphology

Tumor evolution, driven by the emergence of genetically and epigenetically distinct subclones, enables cancers to adapt to selective pressures and become more aggressive, posing a major challenge in oncology. Multi-regional sequencing has been the primary means of studying tumor evolution and the resultant intra-tumor heterogeneity (ITH), but its high cost, resource-intensiveness, and limited scalability have hindered clinical utility.

Here, we present MorphoITH, a novel framework that aims to infer molecular ITH from routinely collected histopathology slides by quantifying phenotypic diversity. MorphoITH integrates a task-agnostic, self-supervised deep learning similarity measure to capture phenotypic variation across multiple dimensions (cytology, architecture, and microenvironment) along with rigorous methods to eliminate spurious sources of variation.

Applying MorphoITH to clear cell renal cell carcinoma (ccRCC), a disease notably shaped by ITH, we show that it captures clinically significant biological features such as vascular architecture and nuclear grade. MorphoITH also recognizes morphological changes associated with subclonal alterations in key driver genes (BAP1, PBRM1, SETD2). Finally, in a multi-regional sequencing dataset, we find that the morphological trajectories revealed by MorphoITH largely mirror underlying patterns of genetic evolution.

MorphoITH provides a scalable and rigorous approach to quantify morphological ITH, serving as a potential proxy for underlying genetic ITH and tumor evolution. By linking histopathology with genomic insights, it lays the foundation for more refined phenotypic profiling in support of precision oncology.