Speaker
Description
An intriguing question in cancer biology is the complex relationship between molecular profiles, as provided by transcriptomics, and their phenotypic manifestations at the cellular and the tissue scale. Understanding this relationship will enable us to comprehend the functional impact of transcriptomic deregulations and identify potential biomarkers in the context of precision medicine.
One way to address this question is to investigate to what extent the transcriptome at the tissue level can be predicted from purely morphological data, such as Whole Slide Images (WSI), Gigapixel images of stained tissue sections. Deep Learning models trained to predict an overall patient-level transcriptomic profile from WSIs often lack specificity due to the multitude of different tissue types in a sample and intra-tumoral heterogeneity.
Here, we present a study where we predict molecularly defined subtypes of Muscle-Invasive Bladder Cancer (MIBC) exhibiting intra-tumoral molecular subtype heterogeneity (VESPER trial; N=417). For this, we selected homogeneous regions in the WSI and performed RNA-seq for these regions. WSI are very large and need to be subdivided into smaller images (tiles) that can be processed by Deep Neural Networks. We designed a two-step workflow involving proxy labeling to learn tile predictions with region-based RNAseq as ground truth. The predictions of a first model are used as proxy labels for a second model, following a ""clean"" tile filtering step. This refined model was then applied to whole slides to generate subtype heterogeneity maps.
Our model predicted consensus molecular subtypes with a ROC AUC of 0.89, demonstrating that phenotypic manifestations predict underlying transcriptomic deregulation. Subtype maps revealed diverse heterogeneity profiles, quantified as the percentage of tumor tiles assigned to each subtype.
| Authors | Alice Blondel*, Clémentine Krucker, Clarice S. Groeneveld, Yves Allory, Jacqueline Fontugne, Thomas Walter |
|---|---|
| Keywords | Cancer Biology, Subtype Heterogeneity Maps, Digital Pathology, Deep Learning, Proxy Labeling |
