Abstract LB170: Artificial intelligence-assisted morphology-based detection and enrichment of malignant effusion tumor cells as a method for molecular profiling

Abstract Patients with metastatic breast cancer (MBC) can develop malignant effusions (MEs) when excess fluid and tumor cells accumulate in the pleural or peritoneal space. MEs are associated with significant morbidity and mortality. Tropism of tumor cells in these spaces is poorly understood and tumor cells residing in these fluids are difficult to study with traditional methods due to low abundance. To overcome this limitation and understand the pathobiology of MEs in MBCs, we applied a label-free method to isolate tumor cells based on morphology using an artificial intelligence (AI) model This method enabled isolation of live cells for downstream molecular analysis and characterization of morphological heterogeneity. We collected ME samples from 9 MBC patients with cytology-confirmed MEs, including 3 triple negative and 6 hormone receptor-positive/HER2-negative breast cancers. ME cells were imaged and isolated using a Deepcell AI model trained on a separate cohort of MEs to isolate and enrich tumor cells based on deep learning representations of morphological features. Quantitative morphological features were then extracted from cell images to perform cluster analysis and visualized by dimensionality reduction projections. In parallel, flow cytometry (FC) was used to quantify % EpCAM+/Claudin4+ tumor cells in ME samples. Copy number (CN) analysis was performed before and after enrichment with as few as 200 cells from each sample. Tumor cell fraction (TCF) estimates were calculated from the AI classifier, FC, and CN analyses. CN profiling of malignant cells sorted by the AI classifier revealed genomic alterations common in breast cancer in 7 of 9 patients. We observed increased amplitude of genomic alterations in tumor-enriched vs. unenriched samples, indicating successful tumor cell enrichment in 14 out of 16 samples. We observed a significant correlation in TCF between the AI classifier and FC with a Spearman test (rho=0.568, p0.008). In some cases, the AI classifier reported higher TCF than FC. This could indicate that the AI model detected tumor cells missed by antibody labeling or false positives from morphologically similar cell types, such as macrophages and mesothelial cells. Remarkably, analysis of serial ME samples from the same patient revealed morphological shifts in cell populations which were consistent with changes in TCF by CN analysis. We demonstrate the feasibility of label-free identification and isolation of tumor cells in MEs using an AI classifier. CN profiling of isolated tumor cells verified enrichment from low initial frequencies. Future work will include expression analysis of ME tumor and immune cells, and analysis of organoids derived from MEs. Combining molecular and morphological profiling of MEs will enable new insights into their pathobiology and inform targeted treatment strategies for patients with this condition. Citation Format: Mark J. Magbanua, Laura Huppert, Janifer Cruz, Jackson Goudreau, Shaopu Zhou, Keyi Yin, Pavel Stejskal, Ryan Carelli, Julie Kim, Linda Hsie, Maggie Wang, Zhouyang Lian, Michael Phelan, Manisha Ray, Chassidy Johnson, Amy L. Delson, Ron Balassanian, Jennifer Rosenbluth, Laura van't Veer, Michelle Melisko, Hope S. Rugo, Mahyar Salek, Maddison Masaeli, Hani Goodarzi. Artificial intelligence-assisted morphology-based detection and enrichment of malignant effusion tumor cells as a method for molecular profiling [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 2 (Clinical Trials and Late-Breaking Research); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(8_Suppl):Abstract nr LB170.