MorphNet Predicts Cell Morphology from Single-Cell Gene Expression

Hojae Lee, Joshua D. Welch
bioRxiv (2022)


Gene expression and morphology both play a key role in determining the types and functions of cells, but the relationship between molecular and morphological features is largely uncharacterized. We present MorphNet, a computational approach that can draw pictures of a cell’s morphology from its gene expression profile. Our approach leverages paired morphology and molecular data to train a neural network that can predict nuclear or whole-cell morphology from gene expression. We employ state-of-the-art data augmentation techniques that allow training using as few as 103 images. We find that MorphNet can generate novel, realistic morphological images that retain the complex relationship between gene expression and cell appearance. We then train MorphNet to generate nuclear morphology from gene expression using brain-wide MERFISH data. In addition, we show that MorphNet can generate neuron morphologies with realistic axonal and dendritic structures. MorphNet generalizes to unseen brain regions, allowing prediction of neuron morphologies across the entire mouse isocortex and even non-cortical regions. We show that MorphNet performs meaningful latent space interpolation, allowing prediction of the effects of gene expression variation on morphology. Finally, we provide a web server that allows users to predict neuron morphologies for their own scRNA-seq data. MorphNet represents a powerful new approach for linking gene expression and morphology.