Diagnostic Evidence Gauge of Spatial Transcriptomics (DEGAS): Using transfer learning to map clinical data to spatial transcriptomics in prostate cancer

Abstract

ABSTRACTBackgroundSpatial and single-cell transcriptomics have revealed significant heterogeneity in tumor and normal tissues. Each approach has its advantages: The Visium platform for spatial transcriptomics (ST) offers lower resolution than single-cell analysis, but histology enables the examination of cell morphology, tissue architecture, and potential cell-cell interactions. Single-cell transcriptomics (SC) provides high resolution, but manual cell-type annotation depends on incomplete scientific knowledge from heterogeneous experiments. When investigating poorly defined phenomena, such as the transition from normal tissue to cancer and metaplasia, researchers might overlook critical and unexpected findings in downstream analysis if they rely on pre-existing annotations to determine cell types, particularly in the context of phenotypic plasticity.ResultsWe employ our deep-transfer learning framework, DEGAS, to identify benign morphology glands in normal prostate tissue that are associated with poor progression-free survival in cancer patients and exhibit transcriptional signatures of carcinogenesis and de-differentiation. We confirm this finding in an additional ST dataset and use novel published methods to integrate SC data, showing that cells annotated as cancerous in the SC data map to regions of benign glands in another dataset. We pinpoint several genes, primarily MSMB, with expression closely correlated with progression-free survival scores, which are known markers of de-differentiation, and attribute their expression specifically to luminal epithelia, which are the presumed origin of most prostatic cancers.DiscussionOur work shows that morphologically normal epithelia can have transcriptional signatures like that of frank cancer, and that these tissues are associated with poor progression-free survival. We also highlight a critical gap in single-cell workflows: annotating continuous transitional phenomena like carcinogenesis with discrete labels can result in incomplete conclusions. Two approaches can help mitigate this issue: Tools like DEGAS and Scissor can provide a disease-association score for SC and ST data, independent of cell type and histology. Additionally, researchers should adopt a bidirectional approach, transferring histological labels from ST data to SC data using tools like RCTD, rather than only using SC cell-type assignments to annotate ST data. Employed together, these methods can offer valuable histology and disease-related information to better define tissue subtypes, especially epithelial cells in the process of carcinogenesis.ConclusionsDEGAS is a vital tool for generating clinically-oriented hypotheses from SC and ST data, which are heterogeneous, information-rich assays. In this study, we identify potential signatures of carcinogenesis in morphologically benign epithelia, which may be the precursors to cancer and high-grade pre-malignant lesions. Validating these genes as a panel may help identify patients at high risk for future cancer development, recurrence, and assist researchers in studying the biology of early carcinogenesis by detecting metaplastic changes before they are morphologically identifiable.