Understanding changes in gene regulatory networks, known as rewiring, is a challenge. Methods addressing this problem require multi-omics input data, which is costly to generate for model organisms, and simply not available for non-model organisms. We propose a new approach to detect rewiring events between cell types, that uses only single-cell RNA sequencing data. The approach consists of training machine learning models to predict the expression of target genes from the expression of transcription factors. Then a local (i.e., per cell), model-agnostic explainability method, SHAP, determines the importance of each transcription factor for each target gene. We consider that these transcription factor importances capture the gene network state of a cell. A new space is created describing cells in terms of regulatory network associations between genes. We compute in this space the statistically different associations between cell types, resulting in a set of marker regulatory associations. Several visualizations help the selection of putative rewiring events between GRN states in different cell types. The proposed workflow is applied to two datasets, demonstrating the potential for detecting rewiring. The implementation of our workflow follows the conventions of the scverse initiative and allows easy integration with Scanpy and other packages for single-cell analysis.