Dissecting cell state transitions by causal inference

Abstract

Trajectory inference methods are frequently used for cell fate analysis, however, most of them are similarity-based and lack an understanding of the causality underlying differentiation processes. Here, we present CIBER, a Causal Inference-Based framework for the Evaluation of feature effects and the Reconstruction of cellular differentiation networks. CIBER provides a novel paradigm for dissecting cell state transitions other than trajectory inference and differential analysis. It is a versatile framework that can be applied to various types of data, including transcriptomic, epigenomic and microarray data. It can identify both known and potential cell-lineage structures with minimal prior knowledge. By integrating the CIBER-learned network with structural causal model and applying in silico perturbation as inventions, we generated an effect matrix that quantifies the impact of different features on each differentiation branch. Using this effect matrix, CIBER can identify crucial features involved in haematopoiesis, even if these features show no significant difference in expression between lineages. Moreover, CIBER can predict novel regulation associations and provide insight into the potential mechanism underlying the influence of transcription factors on biological processes. To validate CIBERs capabilities, we conducted in vivo experiments on Bcl11b, a non-differentially expressed transcription factor identified by CIBER. Our results demonstrate that dysfunction of Bcl11b leads to a bias towards myeloid lineage differentiation at the expense of lymphoid lineage, which is consistent with our predictions.