Unbiased Phenotyping of CRISPR-Edited Tumors Using Perturb-map and Whole Transcriptome Analysis

The heterogeneity of tumor phenotypes in patients, and the lack of a holistic definition of tumor subtypes, are major barriers to achieving clinical success. Many clinical programs have been halted for failing to refine the patient population that would benefit from treatment. Our OCTO suite of machine learning models, trained on multi-modal data generated from 1000s of patient samples, provides a unified and unbiased classification of patient subtypes that does not solely rely on genotypes or expression of specific biomarkers. Though OCTO enables us to identify target populations for active clinical programs, predicting response for preclinical assets is still an unmet need.In contrast to patient tumors, traditional preclinical models are by design homogeneous, and often do not represent an actual patient populations. Using a scaled in vivo CRISPR perturbation platform, Perturb-map, we have generated 600+ mouse model variants of NSCLC. Each tissue section from Perturb-map contains hundreds of individual tumors, each carrying a specific genetic perturbation, spatially resolved with a molecular barcode. We built a multi-modal (H&E, mIF, transcriptomics) dataset profiling 100,000s of individual tumors. A subset of samples was profiled using the mouse Whole Transcriptome (WTx) panel (22,000-plex) using the Cosmx Spatial Molecular Imager, providing the first data set coupling high-throughput CRISPR perturbations in vivo with unbiased phenotyping at subcellular resolution.Leveraging the unique combination of throughput from Perturb-map and depth of phenotyping from the WTx panel, we mapped genetic perturbations to their impact on tumor development at an unprecedented scale. Comprehensive phenotyping of hundreds of CRISPR-edited tumors revealed the impact of each gene perturbation on relevant signaling pathways within tumor cells and on the composition of the tumor microenvironment, including deep characterization of the immune compartment. Specifically, we found significant overlap among the molecular footprints of known tumor drivers in NSCLC (including EGFR, KRAS and STK11 mutations) in preclinical models and in patients. We also linked tumor phenotypes with their sensitivity to specific treatments, including to immune checkpoint blockade.Altogether, we present a unique dataset generated by profiling hundreds of NSCLC preclinical model variants, each carrying a specific genetic perturbation, using a comprehensive Whole Transcriptome (WTx) panel. A subset of mouse tumor archetypes presented characteristics of specific patient tumor subtypes, and could be used to evaluate population-specific drug responses. Citation Format: Maxime Dhainaut, Liang Zhang, Michela Meister, Anastasia Mavropoulos, Andrew Benz, Isabel Lee, Sierra McKinzie, Daniel Millman, Derrick Tran, Dulce Ovando, Lucas Cavalcante, Keith Mitchell, Eshed Margalit, Dexter Antonio, Abhishaike Mahajan, Aaron Salm, Eric Siefkas, Francis Fernandez, Hargita Kaplan, Harry Winters, Jake Schmidt, Joy Tea, Jeremie Decalf, Lisa Ryan, Maede Zolanvari, Nicole Snell, Rob Schiemann, Rodney Collins, Ryan Huang, Yubin Xie, Yves Falanga, Shafique Virani, Lacey Padron, Emily Corse, ShanShan He, Daniel Bear, Joseph M. Beechem, Ron W. Alfa. Unbiased phenotyping of CRISPR-edited tumors with Perturb-map and subcellular imaging of the mouse whole transcriptome (22,000-plex) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6755.