Abstract 639: Rapidly evaluating cancer dependencies by label-free imaging of zero-passage primary cells

• Published in: Cancer research (Chicago, Ill.) Ill.), 2022-06, Vol.82 (12_Supplement), p.639-639
• Main Authors: Al-Jazrawe, Mushriq, Molnar, Csaba, Rindtorff, Niklas, Blum, Steven, Colgan, William, Alimova, Maria, Misek, Sean, Rios, Carmen, Tseng, Moony, McFarland, James M., Regev, Aviv, Cimini, Beth A., Carpenter, Anne E., Bass, Adam, Klempner, Samuel J., Boehm, Jesse
• Format: Article
• Language: English
• ISSN: 1538-7445; 1538-7445
• DOI: 10.1158/1538-7445.AM2022-639

Successful mapping of cancer dependencies requires conducting genetic and pharmacological screens in a diversity of cell models. However, existing model development approaches require long periods of culture time during which evolutionary pressures reduce heterogeneity. It also remains difficult to create long-term models of many cancers, greatly limiting the share of patient samples that can be studied. These limitations make it challenging to create large and diverse datasets for the discovery and validation of biomarker-vulnerability relationships. To enable high-throughput genetic and pharmacological screens in primary, including nondividing, cells without an intermediate model generation step, we are developing a label-free imaging-based platform of early living tissue perturbation. Here, we describe our efforts to establish an ex vivo cell-preservation and imaging system for malignant ascites from patients with advanced gastroesophageal (GE) cancer, whose prognosis remains poor and there is an urgent need for rapid evidence-based therapeutic discovery. First, we optimized a workflow to acquire and perturb cells within 24 hours of sample collection. We observed that incorporating microenvironmental factors by mixing ascites fluid with organoid media extends the preservation of cellular composition and viability. We next hypothesized that label-free microscopy can be a substitute for fluorescence-based labels which fade over time in live-cell imaging. To test this, we created a dataset consisting of over 1.0M cells from 14 samples (10 unique patients). Before seeding, we added fluorescence labels to annotate cell type and viability during imaging. We then trained predictive models to infer cell annotations based on brightfield morphological features only and observed an overall accuracy of 92% and 82% for cell identity and viability, respectively. Prediction of tumor cell fraction from label-free images alone showed strong correlation with tumor fractions estimated by flow cytometry. We also integrated single-cell RNA sequencing data to generate a candidate panel of 28 compounds which are predicted to exhibit antitumor activity via different mechanisms that are of relevance to our study cohort. We observe that label-free inference of compound activity showed strong correlation (R2 > 0.8) with fluorescent-based methods. We are now expanding the scale of our rapid screens and utilizing transcriptomics to link the molecular profile with response. Since ou