💻 bioinformatics

Characterizing dynamic tissue architectures by identifying cell-type-specific spatiotemporal gene programs with stGP

The paper introduces stGP, a statistical framework that decomposes spatiotemporal transcriptomic data into temporal and spatial components to identify interpretable, cell-type-specific gene programs, thereby revealing how dynamic tissue architectures and localized cellular responses evolve across time and anatomical niches.

Original authors: Yu, B., Tan, Z., Wan, X., Wang, H., Yang, C.

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yu, B., Tan, Z., Wan, X., Wang, H., Yang, C.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a bustling city where the buildings (cells) are constantly changing their interior design (gene programs) based on two things: when they were built or renovated (time/age) and where they are located in the city (space/neighborhood).

Sometimes, a building changes its design because it's getting old. Other times, it changes because it's next to a park or a busy street. The big problem for scientists has been: How do we tell if a building's new look is because of its age, or because of its neighborhood?

This paper introduces a new tool called stGP (spatiotemporal Gene Programs) to solve this puzzle. Here is how it works, using simple analogies:

1. The Problem: The "Mixed-Up" City

In the past, scientists looked at these cellular cities using snapshots (spatial transcriptomics). They could see the buildings and their designs, but they couldn't easily separate the "age" factor from the "location" factor.

The Old Way: Imagine trying to listen to a choir where everyone is singing different songs at different volumes. You hear a jumble of noise and can't tell who is singing what, or why.
The Challenge: If a specific type of cell (like a brain cell) changes its behavior as a person gets older, is that change happening everywhere at once? Or is it only happening in a specific corner of the brain? Existing tools were like a blurry camera that couldn't focus on both the time and the place at the same time.

2. The Solution: stGP (The "Smart Filter")

The authors built stGP, which acts like a sophisticated, smart filter for these cellular cities. It breaks down the data into two distinct layers:

The "Time Track" (Temporal Component): This tracks how a specific type of cell changes its "song" (gene program) as time passes, regardless of where it is. It's like a timeline showing how a building's style evolves from 1920 to 2020.
The "Map Layer" (Spatial Component): This tracks how that same cell type behaves differently depending on its specific address. It's like seeing that buildings in the "downtown" district look different from those in the "suburbs," even if they are the same age.

The Magic: stGP keeps the "identity" of the gene program consistent (the same set of genes) but separates when it is active from where it is active. It's like taking that blurry choir recording and using software to isolate the bass line (time) from the melody (space), so you can hear exactly what each section is doing.

3. What They Discovered (The "City Tours")

The team tested stGP on three different "cities" (datasets) and found some fascinating things:

Tour 1: The Aging Human Brain (The DLPFC)
- The Scene: They looked at the prefrontal cortex of people aged 15 to 87.
- The Finding: They found that the "layers" of the brain (like the floors of a skyscraper) stay very stable over time. The buildings in the "top floor" (Layer 2/3) stayed in the top floor, and the "bottom floor" (Layer 5/6) stayed at the bottom, even as the people got older.
- The Twist: However, the interior design of these layers did change with age. Some layers got "quieter" (less active) as people aged, while others had a "mid-life crisis" (peaked in middle age and then declined). stGP was the only tool that could clearly see both the stable location and the changing age-related activity.
Tour 2: The Aging Mouse Brain (The White Matter)
- The Scene: They looked at microglia (the brain's immune cells) in aging mice.
- The Finding: They discovered a specific "immune program" that acts like a fire alarm. In young mice, this alarm is barely on. But in old mice, it goes off loudly, specifically in the white matter (the brain's wiring).
- The Connection: This alarm wasn't just because the mice were old; it was triggered because these immune cells were hanging out near T-cells (another type of immune cell). stGP showed that the "fire alarm" was a localized event happening in specific neighborhoods, not a global brain-wide panic.
Tour 3: The Injured Mouse Kidney (The Construction Site)
- The Scene: They watched kidneys heal after a simulated injury over several weeks.
- The Finding: They saw a very clear, repeatable "construction schedule."
  1. Phase 1: Immediate emergency response (acute injury).
  2. Phase 2: A transition to repair (inflammation to healing).
  3. Phase 3: Long-term remodeling (scar tissue formation).
- The Proof: They ran the analysis on the left kidney and the right kidney separately. stGP found the exact same schedule in both, proving the tool is reliable and robust. It also showed that the "repair crew" (immune cells) and the "construction crew" (fibroblasts) were working together in specific zones, not randomly.

4. Why This Matters

Think of stGP as a time-space decoder. Before, scientists had to guess if a cell's behavior was due to age or location. Now, stGP can say: "This gene program is the same everywhere, but it gets louder as you get older, and it only gets really loud in the white matter of the brain."

This allows scientists to:

Distinguish between a city that is just getting old versus a city that is being rebuilt in a specific neighborhood.
Identify "aging hotspots" where things go wrong faster than elsewhere.
See how different cell types interact in specific neighborhoods (niches) as time passes.

In short, stGP takes the messy, tangled web of "when" and "where" in our bodies and untangles it, giving us a clear map of how our tissues change over time and space.

Technical Summary: Characterizing Dynamic Tissue Architectures with stGP

Problem Statement
Cellular gene programs unfold over biological time within spatially organized tissues, where the same cell type can activate distinct molecular programs depending on its location, developmental stage, or local multicellular niche. While spatiotemporal transcriptomics enables in situ measurement of these processes, analyzing such data presents a significant challenge: temporal progression (e.g., aging or injury recovery), anatomical organization within tissue sections, and local niche-associated variation occur simultaneously across distinct, often unaligned, tissue sections. Existing methods struggle to disentangle whether a gene program's activity is driven by temporal changes, spatial structure, or their interaction. Current approaches, such as PCA, NMF, and spatially aware methods like SpatialPCA, Popari, or MEFISTO, often fail to provide a unified framework that decomposes cell-type-specific program activity into separable temporal trajectories and spatial effects while preserving the interpretable molecular identity (gene loadings) of each program.

Methodology: The stGP Framework
To address these challenges, the authors present spatiotemporal Gene Programs (stGP), a statistical framework designed for multi-sample spatiotemporal transcriptomic studies. stGP operates within a target cell type and models gene expression as a small number of latent gene programs.

Model Structure: stGP represents the expression matrix $Y$ $Y$ for a sample as a decomposition into a baseline effect, a spatiotemporal cell embedding matrix $H$ $H$ , and a gene loading matrix $W$ $W$ .
- Gene Loadings ( $W$ ): Defined as non-negative vectors on a simplex, these represent interpretable, weighted gene sets shared across all samples. This ensures the molecular identity of a program remains consistent.
- Cell Embeddings ( $H$ ): The activity of each cell for a specific program is decomposed into two components:
  1. Temporal Effect ( $\alpha$ ): A sample-level effect shared by all cells of that type from an individual at a specific biological age/stage. This is modeled using a Gaussian process (GP) prior smoothed over the temporal covariate.
  2. Spatial Effect ( $b$ ): A cell-level heterogeneous effect varying within a tissue section. This is modeled using a separate GP prior defined on pairwise Euclidean distances within the section, allowing for local spatial variation without requiring cross-section registration.
Inference: The model estimates gene loadings, cell embeddings, and variance components (quantifying the relative contribution of time vs. space) using a blockwise backfitting algorithm. The algorithm alternates between updating program activities, gene loadings on the simplex, and variance components via marginal log-likelihood maximization. The number of programs can be automatically determined during model selection.
Computational Efficiency: By exploiting the block-diagonal structure of within-section spatial kernels and low-rank temporal terms, stGP avoids the $O(N^3)$ complexity of standard GP inversion, achieving $O(\sum N_t^2 + T^3)$ complexity per update.

Key Results
The authors evaluated stGP through comprehensive simulations and three real-world spatiotemporal transcriptomic datasets:

Simulation Studies: stGP outperformed baseline methods (PCA, NMF, SpatialPCA, Popari, STAMP, MEFISTO) in recovering latent gene programs, detecting true genes, and reconstructing cell embeddings. Crucially, stGP was the only method capable of accurately recovering programs driven purely by temporal effects or purely by spatial effects, correctly disentangling the two sources of variation.
Human Dorsolateral Prefrontal Cortex (DLPFC) Aging: Applied to MERFISH data from donors aged 15–87, stGP resolved layer-specific excitatory neuron architecture in unaligned sections more accurately than baselines. It identified stable laminar structures (L2/3, L4, L5/6) while revealing age-dependent attenuation of specific transcriptional states. Furthermore, stGP uncovered an oligodendrocyte program (stGP4) associated with a shared multicellular niche (MCN2) in deep white matter, showing that aging remodels the activity and deployment of pre-existing niches rather than creating entirely new ones.
Aging Mouse Brain: In a MERFISH dataset of mouse brains, stGP identified a late-life microglial program (stGP2) characterized by MHC-I antigen presentation and interferon-responsive genes. This program showed a progressive increase with age and became spatially enriched in white-matter tracts (corpus callosum and anterior commissure). Analysis revealed that this spatial activation was strongly coupled to local T-cell neighborhoods, distinguishing it from general anatomical region effects.
Mouse Kidney Injury-Repair: Applied to a Xenium time-course dataset following bilateral ischemia–reperfusion injury, stGP identified reproducible, stage-specific immune and epithelial programs across bilateral kidney replicates. It successfully separated acute injury responses from later inflammatory-to-repair transitions and persistent immune-stromal remodeling, mapping these programs to specific cellular neighborhoods and injury zones.

Significance and Claims
The paper claims that stGP provides an effective and robust framework for dissecting dynamic tissue architecture by converting spatiotemporal transcriptomic data into interpretable temporal trajectories and spatially resolved models. Its primary significance lies in its ability to:

Decouple Variation: Explicitly separate shared temporal responses from stable anatomical structures and localized spatial remodeling.
Preserve Interpretability: Maintain consistent, non-negative gene loadings across samples, allowing programs to be defined by clear molecular signatures.
Reveal Contextual Dynamics: Uncover how cell-type-specific programs are coordinated across biological time and spatial microenvironments, such as identifying "aging hotspots" or "niche-coupled" states that are missed by methods treating time and space as a single confounded factor.

The authors position stGP as a versatile tool for studying biological processes where molecular state, spatial organization, and biological time are tightly intertwined, offering insights into preserved tissue architecture versus dynamic remodeling.

1. The Problem: The "Mixed-Up" City

2. The Solution: stGP (The "Smart Filter")

3. What They Discovered (The "City Tours")

4. Why This Matters

More like this