GDIS-Bio: A Generalized Dynamical Instability Framework for Localizing Transcriptional State Transitions in Single-Cell Trajectories

Background/Objectives: Single-cell RNA sequencing enables reconstruction of developmental trajectories, but identifying regions of major transcriptional state reorganization remains challenging, particularly when no explicit dynamical model is available. We developed GDIS-Bio, a framework that applies the Generalized Dynamical Instability Score (GDIS) to pseudotemporally ordered single-cell state spaces to localize transition-associated instability. Methods: GDIS-Bio was evaluated in pancreatic endocrine differentiation (GSE114412) and externally validated in human induced pluripotent stem cell-derived cardiac differentiation (GSE175634). Biological transition landmarks and primary analytical settings were defined independently of GDIS. Robustness was assessed across alternative PCA dimensions and sliding-window configurations, and localization was benchmarked against total variance, Gaussian differential entropy, pseudotemporal step distance, and lag-1 pseudotemporal autocorrelation using structure-preserving null models and individual-level paired tests. Results: In GSE114412, GDIS peaks localized reproducibly near the NEUROG3-early transition across both endocrine lineages and differentiation replicates, with significant localization under circular-shift null testing. In GSE175634, significant localization was retained for MES→CMES, PROG→CM, and PROG→CF transitions without dataset-specific retuning. GDIS was not universally superior to conventional metrics; variance, entropy, and step distance were competitive or better in several settings, whereas GDIS showed its clearest comparative advantage over lag-1 pseudotemporal autocorrelation in the terminal cardiac transitions. Conclusions: GDIS-Bio provides a transferable, model-independent framework for localizing transition-associated dynamical instability in single-cell trajectories and is best interpreted as a complementary integrative measure rather than a universally superior early-warning statistic.