2354P - Computational pathology-derived features capture varied epithelial-mesenchymal transition (EMT) states

Background

EMT is a process through which epithelial cells acquire mesenchymal-like traits, such as increased migratory capacity, which facilitate cancer cell metastasis. Here we propose an interpretable, scalable, machine learning (ML)-based method to assess EMT status directly from H&E-stained images. We extracted human interpretable features (HIFs) quantifying cancer cell nuclear morphology and the tumor microenvironment from breast cancer H&E images and confirmed concordance with a known ground-truth EMT expression signature.

Methods

Using a breast cancer-specific EMT signature gene set, we calculated quantile normalized EMT scores for each patient in TCGA-BRCA with RNA-seq data (n = 751). Convolutional neural network models trained using H&E-stained whole slide images were used to extract a set of HIFs across available slides (n = 775). An ordinary least squares regression was run for each HIF against EMT score with tumor purity as a covariate. FDR correction was applied to all p-values.

Results

Lower EMT scores were associated with epithelial attributes, such as high nuclei circularity, while high EMT scores were associated with mesenchymal attributes, such as high nuclei eccentricity and greater variation in nuclei perimeter. Epithelial-type tumors were associated with increased immature stromal content and cancer cell proportion. Mesenchymal-type tumors were associated with increased mature stromal content and stromal cell proportion. Table: 2354P

Conclusions

ML model-generated HIFs effectively capture cancer cell morphology reflective of EMT status, including nuclei eccentricity and circularity. The relationships observed in these results are concordant with previous studies showing increased EMT in tumors with greater extracellular matrix stiffness and large numbers of cancer associated fibroblasts. Overall these results reflect significant agreement between EMT scores and HIFs, indicating potential for quantification of EMT states directly from H&E images.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

PathAI, Inc.

Funding

PathAI, Inc.

Disclosure

C. Kirkup, S. Vilchez, S. Srinivasan, M. Lin, J. Abel, M. Drage, J. Conway, A. Khosla, A. Taylor-Weiner: Financial Interests, Full or part-time Employment: PathAI, Inc. All other authors have declared no conflicts of interest.