Spatial Trajectory Inference

The foundational algorithm assigns each cell a pseudotime value representing its progress along a biological continuum. Unlike non-spatial methods, the ordering is constrained by a spatial neighborhood graph, ensuring that transitions between states are physically proximal. This prevents biologically implausible leaps across tissue compartments. The process typically involves constructing a graph where edges connect spatially adjacent cells, then finding a minimum spanning tree or principal curve through the high-dimensional expression space that respects these spatial adjacencies.

Instead of a static ordering, this approach learns a velocity vector for each cell, predicting its future transcriptional state. Techniques like RNA velocity are adapted to the spatial context, modeling the rate of change in gene expression as a function of both the cell's current state and its local neighborhood. The resulting vector field visualizes the flow of differentiation across the tissue, identifying sources (progenitor zones), sinks (terminally differentiated regions), and potential attractor states.

This probabilistic framework models the coupling between cells in different spatial regions or time points. Optimal transport calculates the most efficient way to map the distribution of cells from an earlier state to a later state, minimizing a cost function that incorporates both transcriptional dissimilarity and physical distance. The resulting transition matrix quantifies the probability of a cell at location x differentiating into a cell type at location y, providing a rigorous, mass-preserving model of spatiotemporal dynamics.

Advanced methods borrow concepts from differential geometry to analyze the shape of the inferred trajectory manifold. By calculating the curvature and torsion of the principal curve through the spatial-expression space, researchers can identify sharp bends that represent critical fate decision points. High curvature regions often correspond to branching points where a progenitor population diverges into two distinct lineages, providing a quantitative geometric signature of cell fate commitment within the tissue architecture.

This characteristic integrates cell-cell communication directly into the trajectory model. The algorithm weights the probability of a differentiation transition based on the co-expression of cognate ligand-receptor pairs between neighboring cells. A cell's pseudotime progression is thus influenced by the signaling molecules it receives from its niche. This moves beyond descriptive trajectory inference to a mechanistic model where the tissue microenvironment actively shapes the differentiation path.

Persistent homology, a tool from topological data analysis, is used to identify robust, multiscale features of the spatial trajectory. This method tracks the birth and death of connected components, loops, and voids in the spatial-expression data as a resolution parameter changes. It can detect cyclic trajectories (e.g., the cell cycle in a germinal center) or bifurcating structures that are invariant to noise, providing a coordinate-free summary of the dynamic process's global architecture.

A foundational statistical measure quantifying the degree to which a gene's expression at one location depends on values at neighboring locations. Positive autocorrelation indicates clustering of similar expression levels, while negative autocorrelation suggests a chessboard-like pattern. Moran's I is the canonical metric, ranging from -1 (dispersion) to +1 (clustering). Trajectory inference algorithms exploit autocorrelation to identify smooth expression gradients along developmental paths.

A graph data structure where nodes represent cells or spatial spots, and edges connect proximate neighbors based on a distance threshold or k-nearest neighbors (k-NN). This graph is the computational backbone for trajectory inference, enabling message passing in spatial graph neural networks and defining the topology over which pseudotime is computed. Edge weights often encode inverse distance or similarity.

A computational assignment of a scalar value to each cell representing its position along a continuous biological process, such as differentiation or cell cycle progression. Unlike real time, pseudotime is inferred from transcriptomic similarity and, in spatial contexts, constrained by physical proximity. Algorithms like Monocle and Slingshot use minimum spanning trees or principal curves to order cells, with spatial variants adding a distance penalty to prevent discontinuous jumps.

A probabilistic framework that models tissue organization as a sequence of hidden states (e.g., anatomical zones) with spatial dependencies. Each state emits gene expression values according to a state-specific distribution. The Viterbi algorithm decodes the most likely state path across the tissue. Trajectory inference can be framed as a state-transition problem where cells move through a defined sequence of spatial domains.

A computational method that disentangles mixed gene expression signals from multi-cellular spots into constituent cell-type proportions. This is critical preprocessing for trajectory inference on lower-resolution technologies like Visium, where a single spot may contain 1-10 cells. Algorithms like RCTD and SPOTlight use reference single-cell data to estimate cell-type composition, enabling trajectory reconstruction on deconvolved cell-type maps.

The computational alignment of multiple tissue sections or spatial datasets into a common coordinate system. This enables cross-modality integration—for example, aligning spatial transcriptomics with histological images or serial sections. Rigid, affine, and non-linear transformations are applied to correct for tissue deformation. Registered data allows trajectory inference to span multiple sections, reconstructing 3D developmental gradients.