Skip to main content
2025 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session E

N6 - Pseudospacetime Analysis to Extract Temporal Variance in Spatial Data

Saturday, October 11, 2025
10:00 AM - 11:00 AM PDT
Location: Exhibit Hall F, G & H

Introduction: : Tissues self-organize, repair, and maintain multicellular structures over time. Dynamic techniques, like intravital microscopy, capture asynchronous, individual cell behaviors simultaneously, suggesting a single snapshot can reveal diverse dynamic states; yet they lack multicellular context and molecular depth. In contrast, current spatial-omics technologies generate deep molecular profiles spatially but ignore temporal features. Inspired by pseudotime algorithms in static RNA analysis, we will develop the first "pseudospacetime" framework to infer temporal dynamics from static spatial-omics snapshots (Fig. 1a-c). To analyze temporal changes of neighborhoods, we will leverage currently “thrown-out variance” in static data, akin to reconstructing a synchronized exercise routine from “variance” in movement of 1,000 individuals performing out of sync.

Currently only spatial gene expression is used in other pseudospacetime calculations, not cellular neighborhoods. For the first time, using neighborhood count matrices analogous to single-cell RNA count matrices, we adapt the underlying computational approaches of Slingshot, Scanpy’s Diffusion Pseudotime (DPT) and Partition-based Graph Abstraction (PAGA), and MOSCOT. To compare accuracy, we validate these approaches against synthetic dynamics, comparing reconstructed sequences of neighborhood events to ground-truth dynamics. We further validated our approach with our CODEX datasets where we captured T cell transitions over time. Our approach is a generalist approach for spatial-omics datasets and enables a context-aware representation of cellular dynamics in space, with implications for understanding spatially encoded progression states in health and disease.

Materials and

Methods: : Our pseudospacetime approach employs three trajectory inference algorithms: Scanpy DPT and PAGA, MOSCOT, and Slingshot. These were selected to represent distinct pseudospacetime inference architectures: MOSCOT applies optimal transport, PAGA models graph-based cluster connectivity, and Slingshot fits principal curves through the data (Fig. 1c).

We applied the pseudospacetime algorithms to two datasets: (1) a synthetic dataset we generated and (2) a CODEX dataset from an antigen-specific, syngeneic murine melanoma model of adoptive T cell therapy, with tissue samples harvested at distinct timepoints (Fig. 2a). Using our previously established cellular neighborhood algorithm, we defined seven CD8+ T cell-centric neighborhoods which then served as input for pseudospacetime inference. To generate the synthetic data, we simulated a linear trajectory of six neighborhood states by defining prototype neighborhood vectors and adding tunable Gaussian noise to model variability and transitions between states (Fig. 1d).

All pseudospacetime algorithms were applied to spatial embeddings of cell neighborhoods. For MOSCOT, both the neighborhood embeddings and harvest days were used as input. We set low epsilon values to generate sparse cell mappings and normalized the cost matrices to stabilize training and improve lineage inference. For Slingshot, DPT, and PAGA, inputs were limited to the first 10 principal components to preserve local variance and avoid over-smoothing during diffusion map construction.

We evaluated algorithm performance on the synthetic dataset by assessing how the inferred pseudospacetime reflected the known neighborhood ordering. For the real dataset, we assessed biological interpretability by exploring spatial patterns of pseudospacetime states and changes in neighborhood composition.

Results, Conclusions, and Discussions:: To test whether spatial context alone could support trajectory inference, we applied our pseudospacetime framework to a synthetic dataset with known linear transitions. UMAP embeddings of neighborhood vectors preserved spatial structure (Fig. 1e) and pseudospacetime value assignments correlated with pre-established temporal arrangements (Fig. 1f-h, j-k).

To assess whether we could recover the ordering of major cellular states, we applied PAGA, Slingshot, and MOSCOT without providing temporal labels. All three methods successfully reconstructed the correct sequence of neighborhood states (Fig. 1m, o-q), Notably, MOSCOT requires the number of inferred states to match the number of timepoints, so the output was truncated with this constraint. These findings show that spatial neighborhood vectors contain sufficient structure to recover pseudospacetime orderings and existing tools can be adapted to spatial datasets with appropriate modifications.

We next applied Scanpy’s algorithms to our murine melanoma CODEX dataset (Fig. 2a). Using our T cell–centric neighborhood algorithm (Fig. 2b), we identified distinct environments across multiple timepoints, enabling us to trace temporal transitions (Fig. 2c-e). These transitions align with key tumor states (immune-infiltrated, inflammation, and tumor outgrowth), mirroring pseudospacetime trends: early pseudospacetime was enriched for tumor populations (e.g., Tumor, Tumor PDL1+ MHCII+, Tumor Ki67+), while late pseudospacetime was dominated by CD8+ T cells and macrophages (Fig. 2g-i) Pseudospacetime bins followed a cyclical pattern, with early bins appearing at days 0 and 12 (Fig. 2j). Mapping pseudospacetime back to the tissue revealed distinct values aligned with specific microenvironments, demonstrating the framework’s ability to capture tumor-immune dynamics (Fig. 2k).

Finally, we applied PAGA, MOSCOT, and Slingshot to our CODEX dataset–all captured consistent trajectories. Early pseudospacetime was enriched for inflammatory and vascular features, and later pseudospacetime showed immune cell recruitment (Fig. 3b, h-j). These trajectories reflected known tumor evolution stages and revealed how multiple dynamic processes can co-occur within a single tissue (Fig. 3d-e). Each algorithm contributed distinct strengths: PAGA enabled interpretable lineages, MOSCOT modeled complex trends and cycles, and Slingshot smoothed transitions.

Our approach provides a powerful, underutilized strategy for reconstructing biological trajectories from spatial context, with potential for application to diverse datasets and further algorithmic exploration.

Acknowledgements and/or References (Optional): :