P04.06.B AUTOMATED TUMOR MICROTUBE TRACING FOR HIGH-THROUGHPUT ANALYSIS OF MORPHOLOGICAL NETWORK CONNECTIVITY IN GLIOBLASTOMA

Abstract BACKGROUND Cellular tumor networks in glioblastoma are based on ultralong cellular membrane protrusions, known as tumor microtubes (TMs). Tumor networks are directly associated with resistance to cytotoxic therapies. Targeting this malignant connectivity has demonstrated promise in preclinical studies and is currently being explored in early clinical trials. TM length and the extent of TM-based cell-cell contacts serve as surrogate markers of morphological network connectivity, making them highly valuable readouts for high-throughput drug screening. However, existing image analysis tools struggle to accurately delineate individual TMs within densely interconnected multicellular networks. In this study, we aimed at developing two fully automated image analysis pipelines to enable precise TM identification and comprehensive assessment of morphological network connectivity. MATERIAL AND METHODS We utilized five primary glioblastoma cell populations that were lentivirally transduced with ZsGreen. These cells were seeded into well plates and incubated for 144 hours. Fluorescence images were captured every six hours using live-cell imaging microscopy. To analyze TM lengths, we developed two fully automated image processing pipelines. These pipelines include pre-processing, thresholding, speckle removal, nuclei selection, skeletonization, tracing measurement, and nuclear counting. One pipeline was implemented using the scikit-image library in Python, while the other utilized built-in tools in ImageJ. The results were compared to manual measurements obtained using the Simple Neurite Tracer (SNT) tool in ImageJ. RESULTS Both automated pipelines demonstrated high accuracy when compared to manual TM tracing. Spearman’s correlation coefficient indicated a strong agreement between manual and automated results, with values ranging from 0.89 to 0.95 across the five primary glioblastoma cell populations. Correlation coefficients were highest at earlier time points, when the tumor network density was lower. Notably, the median correlation coefficient was slightly higher for the Python-based pipeline compared to the ImageJ-based pipeline (0.91 vs. 0.89). CONCLUSION We have developed simple, effective, and fully automated tools for processing data from high-throughput imaging experiments. These pipelines provide an efficient approach for analyzing TM-based network connectivity and hold potential to facilitate preclinical research focused on identifying therapeutic candidates targeting tumor networks.