Fig. 1: Overview of the MorphDiff framework.

a In this study, the multimodal dataset consists of the morphology images with five channels collected using CP (Cell Painting), and the gene expressions collected using L1000 assays. CellProfiler is then used to segment the cells and extract CellProfiler features at the single-cell level. Morphology images and gene expression together characterize the cell morphology responses to specific perturbations. The scale bar is 20 μm. b MorphDiff is composed of two main components: Morphology VAE (MVAE) and Latent Diffusion Model (LDM). The MVAE encoder encodes the multi-channel cell morphology images into latent representation, and the decoder reconstructs the original cell morphology images based on the latent representation. LDM is trained to denoise from Gaussian random noise Z_T to morphology latent representation Z₀, recursively conditioned on L1000 gene expression. The scale bar is 20 μm. c MorphDiff can be applied in two ways to generate cell morphology images with perturbations: L1000 gene expression to cell morphology generation (G2I, Gene to Image) and perturbed L1000 gene expression combined with control morphology images to perturbed cell morphology images generation (I2I, Image To Image). I2I is implemented with SDEdit⁶² without re-training. The scale bar is 20 μm. d Illustration of the downstream applications of MorphDiff, including prediction on unseen perturbations, feature analysis with CellProfiler, as well as morphology-based MOA retrieval with DeepProfiler. DMSO stands for Dimethyl sulfoxide, which is considered as control group without perturbation. The scale bar is 20 μm. a, b, c, d are created in BioRender. Group, A. (2025) https://BioRender.com/nu1zlqw.