Bridging Gene Regulatory Networks and Causal Representation Learning in Single-Cell Genomics Data

Understanding gene regulatory mechanisms is key to advancing our capacity to interpret and manipulate cellular physiology, with significant implications for bioengineering and precision medicine. Two major computational paradigms, namely gene regulatory network (GRN) reconstruction and causal representation learning (CRL), offer distinct perspectives on transcriptional regulation. GRN methods focus on capturing detailed, fine-scale interactions among genes and transcription factors, whereas CRL seeks to identify a small set of latent variables that drive gene expression, providing a coarser but potentially more generalizable representation. In this work, we propose methods that incorporate GRN-derived structures into CRL models, guiding their training and enriching their biological interpretability. Computational experiments on scPerturb-seq datasets demonstrate that GRNs and CRL can work in concert, yielding biologically interpretable latent representations without sacrificing predictive performance.