Abstract 3703: Cell identity revealed by precise cell cycle state mapping links data modalities

Abstract Purpose: Despite advances in single-cell technologies, integrating multimodal data to capture both the genetic and phenotypic dynamics of tumor cells in their microenvironment remains a challenge. Approach: We leverage variations in cell cycle state as a natural barcode to link data across different modalities. We have demonstrated the feasibility of this method in a gastric cancer cell line (NCI-N87) using two widely different technologies: single-cell RNA sequencing and live-cell imaging. Our approach begins with the transfection ofFUCCI probes into cells, followed by acquisition of 3D images using a Leica microscope. These images are then utilized to train a label-freeU-Net-based prediction model (FNet) to infer subcellular organization, specifically for the nucleus, mitochondria, and cytoplasm. FNet achieved correlation coefficients of 0.76, 0.70, and 0.78 for these compartments, respectively. Segmentation masks generated by CellPose enabled cell statistics calculations. We further applied deep learning methods, including a customized convolutional neural network (CNN)and a truncated ResNet50, to distinguish between cells in G1, S and G2M phases of the cell cycle. The CNN achieved an accuracy of 90.67% when using the nucleus alone and 91.63% when using all three compartments (nucleus, mitochondria, and cytoplasm). Deep features were extracted by removing the classification layer and flattening the convolutional layer feature maps. These features were then employed for cell cycle pseudotime inference. When projecting the features onto pseudotime, we observed a strong correlation between the inferred pseudotime and FUCCI-derived cell cycle states (Pearson r =-0.899, p-value < 1E-6). This finding demonstrates the ability to assign continuous cell cycle states directly from imaging data. Comparing imaging- and sequencing-derived cell cycle pseudotime obtained for the same cell line, we assigned 449 imaged cells to their corresponding 693 sequenced cells, achieving an average resolution of 3.4 cells per sequencing profile and 2.4 cells per imaging profile. Conclusion: This integration has opened new possibilities for linking genotype (from sequencing) and phenotype (from imaging), providing a forward-looking approach to genotype-phenotype mapping at single-cell resolution. Assigning the transcriptome of a sequenced cell to its closest living relative (still actively growing and expanding), will enable tracing the cellular trajectory in real-time, enabling a deeper understanding of how tumor cells adapt to their microenvironment and evolve over time. Citation Format: Saeed S. Alahmari, Andrew Schultz, Jordan Albrecht, Vural Tagal, Zaid Siddiqui, Sandhya Prabhakaran, Issam El Naqa, Alexander Anderson, Laura Heiser, Noemi Andor. Cell identity revealed by precise cell cycle state mapping links data modalities [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3703.