GSET: Gene Signals from Embedding Trajectories select compact contexts for mechanistic GRN inference

Mechanistic inference of gene regulatory networks from temporal transcriptomic data can become computationally prohibitive at transcriptome scale, making gene selection an important determinant of which regulatory relationships can be recovered. However, gene selection for GRN inference commonly relies on static expression or co-expression structure, rather than prioritizing dynamically informative genes whose temporal patterns capture how transcriptional programs change and relate to one another. We introduce GSET (Gene Signals from Embedding Trajectories), an upstream retrieval framework that extracts dynamic information from temporal gene embeddings using two complementary signals: an optimal transport residual, which captures gene-specific deviations from global embedding movement, and a KL divergence, which detects changes in local co-expression neighborhoods. Temporal gene-embedding similarity identifies static transcriptional programs, whereas vectorial GSET signals represent coordinated temporal changes among genes, complementing this static structure. Lagged comparisons of these signals further prioritize temporal relationships between programs and candidate bridge genes associated with their transitions. Integrated with CardamomOT, GSET combines static program cores with dynamically retrieved bridge genes into compact, program-informed gene sets for mechanistic GRN inference. In synthetic benchmarks, at equal gene budget these sets recover more of the regulatory links between programs than static selections of the same size, and inference on the edges within the set is comparable to that of the full-gene model on the same gene pairs. On a time course of mouse fibroblasts reprogrammed into induced pluripotent stem cells, GSET traces the transition from the fibroblast program to the cell-cycle program. By providing a dynamic retrieval layer upstream of GRN inference, GSET nominates bridge genes as candidate regulators of cell-state transitions, which, once refined by a mechanistic model such as CardamomOT and confirmed by perturbation experiments, could open the way to controlling these transitions.