Identification of clinically predictive metagenes that encode components of a network coupling cell shape to transcription by image-omics

Abstract

The associations between clinical phenotypes (tumor grade, survival) and cell phenotypes, such as shape, signaling activity, and gene expression, are the basis for cancer pathology, but the mechanisms explaining these relationships are not always clear. The generation of large data sets containing information regarding cell phenotypes and clinical data provides an opportunity to describe these mechanisms. Here, we develop an image-omics approach to integrate quantitative cell imaging data, gene expression, and protein-protein interaction data to systematically describe a "shape-gene network" that couples specific aspects of breast cancer cell shape to signaling and transcriptional events. The actions of this network converge on NF-κB, and support the idea that NF-κB is responsive to mechanical stimuli. By integrating RNAi screening data, we identify components of the shape-gene network that regulate NF-κB in response to cell shape changes. This network was also used to generate metagene models that predict NF-κB activity and aspects of morphology such as cell area, elongation, and protrusiveness. Critically, these metagenes also have predictive value regarding tumor grade and patient outcomes. Taken together, these data strongly suggest that changes in cell shape, driven by gene expression and/or mechanical forces, can promote breast cancer progression by modulating NF-κB activation. Our findings highlight the importance of integrating phenotypic data at the molecular level (signaling and gene expression) with those at the cellular and tissue levels to better understand breast cancer oncogenesis.[Supplemental material is available for this article.]

Joint analysis of gene expression levels and histological images identifies genes associated with tissue morphology

“…Second, it is not clear how to identify biologically relevant features automati-cally from histological images. Previous work on this subject involved extracting hand-engineered features from images and computing pairwise correlations with gene expression data 12 . Methods exist to analyze histological images automatically, but often these methods extract image features that are not associated with genomic features 13 .…”

KCML : a machine‐learning framework for inference of multi‐scale gene functions from genetic perturbation screens

“…This suggests an important link between cell microenvironment (adhesion) and shape (cytoskeleton) and determination of cell fate (i.e., stemness and differentiation via TGFβ and WNT signalling). Consistent with this, SMAD3, which plays an essential role in TGFβ signalling, has been shown to link shape information to transcription in breast cancer cells (Sailem & Bakal, ), while WNT signalling can be linked to cell microenvironment via the differential localisation of its downstream effector β‐catenin. These results exemplify how KCML can be used to automatically interrogate quantitative phenotypic profiles to identify combinatorial use of modular gene programmes in different contexts.…”

KDML: a machine-learning framework for inference of multi-scale gene functions from genetic perturbation screens

“…stemness and differentiation via TGFβ and WNT signalling). Consistent with this, SMAD3, which plays an essential role in TGFβ signalling, has been shown to link shape information to transcription in breast cancer cells (Sailem & Bakal, 2017), while WNT signalling can be linked to cell microenvironment via the differential localisation of its downstream effector β-catenin.…”