Thread by @fayyazhere

What are the #histopathology image #WSI patterns associated with gene expression in cancer? Can we discover them with #AI and use them for spatial profiling and #precision medicine?
Our preprint explores limitations & possibilities of #cpath for this.
biorxiv.org/content/10.110…

Existing methods aim to predict expression of individual genes from #WSIs but ignore any inter-dependencies in expression across genes. We show that it is typically not possible to predict expression of a single gene due to such codependencies as shown in fig for a group of genes

To formalize this, we first used mutual information to identify groups of genes whose expression in breast cancer patients are significantly inter-dependent and form the basis of important cancer and biological pathways and can be used for precision medicine.

Furthermore, these gene groups also carry prognostic information and significant association with known histological patterns and subtypes of breast cancer.

We then predicted the activation status of these gene groups from histological WSIs with a custom graph neural network pipeline. We found that it is possible to predict the gene expression of a significant number of gene groups with high AUC-ROC.

However, it is not possible to predict overall gene expression state, i.e., status of ALL gene groups of a cancer patient, with high accuracy (cosine similarity between true and predicted states < 0.6). This is a fundamental limitation of image based gene expression prediction.

Akin to spatial transcriptomics, we show that it is possible to use the predictor to spatially profile the expression of certain important gene groups (e.g., G3 which is associated with #immune response).

We then identified histological patterns in terms of cellular composition and mitosis counts for each gene group which agreed with pathologist observations.

We also show that the predicted status of gene groups from WSIs with the proposed machine learning pipeline can be used for multiple downstream tasks such as receptor status and subtype prediction.

In short, we have performed data driven discovery of important gene groups in breast cancer and their histological correlates which can be used for routine WSI based spatial profiling of gene expression and precision medicine.

See our interactive demo: tiademos.dcs.warwick.ac.uk/bokeh_app?demo…

Thanks to all the colloaborators @nmrajpoot and @MostafaJahanif1 @TIAwarwick @PathLAKE_CoE , Kim Branson @GSK, Louise Jones @QMUL, @LawrenceYoung11, Asa Ben-Hur @ColoradoStateU, and the lead by our PhD student @muh_dawood.

@nmrajpoot @MostafaJahanif1 @TIAwarwick @PathLAKE_CoE @GSK @QMUL @LawrenceYoung11 @ColoradoStateU @muh_dawood special thanks to Mark Eastwood (for visualization) and @TIAToolbox teams.