scTIDE: Deciphering Critical Transitions Through Cell‐Perturbed Manifold Graphs and Optimal Transport Conditional Flow Matching

A tipping point marks the threshold or critical state where a biological system shifts from one stable state to another. Deciphering critical transitions and their associated signaling molecules is essential for elucidating complex biological processes and for enabling timely interventions to avert or postpone catastrophic deteriorations. However, existing critical-state detection methods rely mainly on Euclidean-space statistics, which may overlook nonlinear dynamical behavior among molecules and distribution-based molecular patterns, leading to limited robustness and performance in high-dimensional, sparse, and noisy single-cell data. In this study, we introduce single-cell Tipping-point Identification via Distributional Embedding (scTIDE), a framework that integrates manifold-based graph representations with optimal-transport conditional flow matching (OT-CFM) to capture intrinsic topological structure and identify critical transitions at the individual-cell level. Specifically, for a given cell, scTIDE quantifies distributional differences between a distribution derived from the reference manifold graph and a perturbed distribution inferred from the cell-perturbed manifold graph using OT-CFM, thereby identifying critical stages and key signaling molecules. The reliability and effectiveness of our model are demonstrated through synthetic models and eight distinct single-cell datasets, where it outperforms existing methods. Moreover, scTIDE reveals possible critical transitions for unseen cells and visualizes the intricate biological progression.