CytoBridge turns sparse spatial transcriptomics snapshots into virtual tissues, writing cell–cell interaction into the dynamics

A time series of spatial transcriptomics can now be joined into one continuous population evolution rather than a set of disconnected sections. CytoBridge places each cell's gene-expression state and spatial coordinate in a single joint space, learns how cell state, position and population size change over time, and treats cell–cell interaction as an explicit term in the dynamics. Earlier spatial and spatiotemporal transcriptomics methods mainly aligned, coupled or reconstructed distributions across time points, such as stVCR, MOSCOT, Spateo and STORIES; they describe population evolution under prescribed transport or diffusion principles and do not explicitly model how local cell–cell interactions shape multicellular dynamics. On five spatial datasets (chicken embryonic heart, mouse organogenesis MOSTA, axolotl brain regeneration ARTISTA, Alzheimer's disease progression in TgCRND8 mice, and zebrafish embryogenesis), one intermediate time point at a time was withheld from fitting and the predicted distribution compared with the held-out observation; across 33 combinations of held-out stage and evaluation space, CytoBridge had the lowest Sliced-W2 in 12, more than any other method or control.

The model splits each cell's velocity into two parts: input from neighboring cells' signals, and the cell's own regulatory progression plus unmeasured environmental influences. Interaction direction and strength can then vary with time and spatial position instead of being fixed as an analytic kernel. The interaction term is parameterized by an attention-style equivariant graph neural network whose edges are set jointly by spatial distance and ligand–receptor pairing priors and rebuilt as cells move and change state; population growth or loss is described by a separate growth field, and stochastic fluctuations by a time-varying diffusion scale. On a simulated benchmark built on the MOSTA mouse organogenesis atlas with known velocity fields, growth rates and interaction networks, the inferred full velocity field agreed with the generator field at an aggregate cosine similarity of 0.977 in physical space and 0.952 in gene-expression space, attention strengths correlated with generator values at a mean Spearman correlation of 0.80, and growth values at Pearson r = 0.92; on a separate independent finite-range attractive particle simulator, the radial attraction profile was recovered at Pearson r = 0.876.

The demonstrated scope is leave-one-stage interpolation and interaction reconstruction on five spatial transcriptomics datasets, plus comparisons against lineage-tracing fate proportions and new-RNA directions withheld from training in two non-spatial single-cell time courses; these data are observational, and the inferred interactions and simulated perturbation responses are model predictions.

The authors note that the simulated benchmark's generator is itself a CytoBridge model sharing its structural assumptions, so it serves only as a controlled recovery test; each section samples one plane of a three-dimensional tissue, and sparse temporal sampling or alignment error limits recovery of intermediate populations, as in the AD dataset where the interval of more than a year between 5.7 and 17.9 months limits the temporal resolution of inferred changes; separating interaction from intrinsic context requires ligand–receptor and neighborhood features that vary enough among comparable cells; and because the data are observational, the inferred interactions and simulated perturbation responses still need perturbation experiments to test causal roles.