Fig. 1: Illustration of the CausCell framework.
From: Causal disentanglement for single-cell representations and controllable counterfactual generation
a The framework of CausCell, which consists of a causal disentanglement module and a diffusion-based generative module. In the causal disentanglement module, encoding gene expression data into exogenous embeddings, processed through an SCM layer to learn causal relationships, yielding endogenous embeddings. Concept labels are predicted via supervised constraints. In the diffusion-based generative model, employing a denoising diffusion model to generate realistic gene expression profiles, conditioned on latent concepts via cross-attention mechanisms at each step. b The schematic diagram of the disentanglement representation for the cells in CausCell. Each cell can be represented as multi-concept embeddings. c The schematic diagram of counterfactual generation in CausCell. CausCell can generate new single-cell omics data for the unseen concept combinations. Credit: all of these concept-related icons in (b, c), except for cell icons, https://smart.servier.com/.
