Virtual dynamics of spatiotemporal single-cell interactions

Reconstructing the complex dynamics of cells and tissues over space and time is a critical challenge in building virtual cells and tissues. However, spatial omics experiments only provide limited snapshots of this process. To address this, we present CytoBridge, a computational method that reconstructs continuous virtual cell dynamics from discrete spatial transcriptomic data.

CytoBridge begins with a joint transcriptomic-spatial space and learns stochastic trajectories of cell states, positions, and population sizes. It does this through an unbalanced mean-field Schrodinger bridge formulation that explicitly incorporates cell-cell interactions as a dynamic driver. A time-varying interaction graph, constructed from spatial proximity and ligand-receptor priors, decomposes velocity into two components.

The first is an interaction component, estimated from observed neighborhoods and ligand-receptor features. The second is an intrinsic-context component, capturing intrinsic regulatory processes and unmeasured contextual influences.

CytoBridge was applied to five datasets across three spatial transcriptomics platforms, covering development, regeneration, and neurodegeneration. It accurately generated tissue dynamics at unmeasured time points, followed ligand-receptor signaling through measured and unmeasured stages, and supported hypothesis-generating in silico perturbation.

CytoBridge offers a general method for transforming heterogeneous spatiotemporal transcriptomic measurements into continuously evolving virtual tissues.