Cellular morphology emerges from polygenic, distributed transcriptional variation

Abstract

Height and disease risk are canonical polygenic traits, but whether an analogous architecture governs cellular phenotypes remains unclear. Here, we show that cellular morphology behaves as a polygenic systems phenotype by integrating cross-modal modeling, perturbation profiling, and human genetic variation. An adversarially aligned autoencoder trained on four perturbation datasets predicted morphology from gene expression and generalized without retraining to matched RNA-seq and Cell Painting profiles from 100 genetically diverse iPSC donors. Seventeen morphological features were strongly predicted, particularly those capturing organelle distribution and cytoskeletal architecture. Predictive signal was broadly distributed across genes, with weak individual gene-morphology correlations. Independent CRISPR perturbations supported model-prioritized genes including TIAM1, RAB31, and ABCC5, while genetic analyses identified variant-gene-morphology associations involving PDHX, SLC11A2, and MAPKAPK5. These findings extend polygenic models of complex traits to cellular architecture and provide a framework for interpreting cross-modal prediction in functional genomics.