MorphCell learns interpretable shape representations of 3D cellular morphology

Abstract

Three-dimensional (3D) cell morphology provides a measurable phenotype of cellular state and function, yet learning transferable and interpretable representations across biological systems and imaging modalities remains challenging. Here we introduce MorphCell, a self-supervised framework for learning 3D cellular morphology representations from cell-surface point clouds. MorphCell uses cross-view reconstruction and spherical self-reconstruction to capture global and local geometric features and reconstruct complete 3D surfaces. Pretrained on diverse non-biological 3D geometries, MorphCell transfers to biological datasets without task-specific fine-tuning, providing a common representation across imaging modalities and biological contexts. We demonstrate this transferability across confocal microscopy, X-ray microscopy and volume electron microscopy, spanning red blood cells, wheat root cells and colon epithelial cells. MorphCell captured biologically relevant 3D shape variation and outperformed baseline representations in multiple classification tasks. Separating physical scale from the learned shape representation further revealed context-dependent contributions of shape and scale: shape alone outperformed shape-scale fusion for red blood cell morphotype classification, whereas combining shape with the nuclear-to-cytoplasmic volume ratio increased balanced accuracy for cancerous versus non-cancerous colon cells from 62.7% to 92.8%. MorphCell therefore provides a general framework for representing, reconstructing and interpreting 3D morphology, enabling the discovery of biologically meaningful morphological features beyond conventional measurements of scale.