GRN-INFORMED CELLFLOW: ENHANCING CELL STATE TRAJECTORY INFERENCE WITH BIOLOGICAL REGULATORY NETWORKS

Abstract

This paper presents a study on integrating gene regulatory network (GRN) information into computational models for reconstructing cellular trajectories, with a particular focus on enhancing the CellFlow framework. Gene regulatory networks describe interconnected systems of genes and regulatory elements — including transcription factors and signaling pathways — that coordinate gene expression and ensure proper regulation of processes such as cell identity maintenance, lineage differentiation, and adaptive responses to environmental changes. Mapping these interactions provides a foundation for understanding how gene activity patterns shape cellular behavior and transitions between states. In this work, we introduce GRN-informed CellFlow, an extension of the original CellFlow model that explicitly incorporates regulatory dependencies between genes. Unlike the baseline approach, which treats genes as independent features, the proposed method integrates known gene–gene relationships to guide trajectory reconstruction. To achieve this, we constructed a GRN matrix using zebrafish cell data: transcription factors were identified via UniProtKB annotations, while their interactions with target genes were inferred using correlation analysis and the GRNBoost2 algorithm. The resulting network Laplacian was employed as a regularizer during model training, enabling CellFlow to account for structured dependencies between genes. Experimental results showed that GRN integration slightly worsens the loss function compared to the classical CellFlow configuration, yet improves the biological interpretability of reconstructed trajectories. These findings highlight the potential of combining structured network information with algorithmic approaches to cell trajectory inference.