Poster #40: Xingyan Liu

Geometry-guided phenotyping reveals interpretable spatial programs of Ebola infection beyond viral abundance

Liu X, BSPH¹; Ma R, PhD¹ ¹Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA, USA.

Optical pooled screening (OPS) links genetic perturbations to high-content single-cell imaging, enabling analysis of viral accumulation and host-gene effects on spatial organization. Ebola virus replication involves subcellular events, including inclusion-body formation, diffuse cytoplasmic localization, and budding, yet linking these phenotypes to perturbations at genome scale remains challenging. In an Ebola OPS study (Carlson et al., 2025), infection phenotypes were summarized using marker intensities and a deep embedding fine-tuned to four expert-defined VP35 localization phases. Although powerful, latent embeddings do not specify which spatial properties change, while discrete phases may compress continuous or independent variation. We developed a geometry-guided, interpretable representation built on intrinsic cell morphology rather than microscope coordinates. Each segmented cell is mapped to radial and angular coordinates defined relative to its distance-transform center, nuclear mask, and boundary, making measurements comparable across cell size, shape, and orientation. Across six viral and host channels, the representation captures radial and nuclear localization, angular organization and polarity, puncta organization, membrane enrichment, and morphology. We applied this framework to 38.7 million quality-controlled cells and ~20,000 targeted genes. We constructed unsupervised phenotypic axes and supervised axes aligned with 3,810 expert-annotated cells. Nuclear DNA content showed that abundance axes were largely independent of cell-cycle variation, whereas the angular axis retained some dependence. Viral abundance captured substantial phase information but was insufficient: abundance-based axes distinguished later phase transitions (ROC-AUC 0.79–0.88) but were near chance for the earliest transition (ROC-AUC 0.52–0.54, versus 0.85 using the geometry-guided representation), corresponding to initial inclusion-body formation. Beyond abundance, geometry captured nuclear displacement, angular organization of viral protein and RNA, JUN nuclear translocation, and radial VP35 localization. Gene-level profiles linked these phenotypes to perturbations involving vesicle organization, oxidative phosphorylation, ribosome biogenesis, and protein deneddylation. Morphology-aware geometry provides a framework for interpretable phenotyping and genetic discovery from large-scale optical pooled screens.