📊 statistics

JASPER: Joint Bayesian Analysis of Spatial Expression via Regression

JASPER is a novel Bayesian framework that improves the detection of spatially varying genes in transcriptomic data by jointly modeling inter-gene correlations and complex spatial patterns through spatial basis function regression, thereby outperforming existing methods in accuracy and biological interpretability.

Original authors: Pritam Dey, Rajarshi Guhaniyogi, Yang Ni, Bani K. Mallick

Published 2026-04-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Pritam Dey, Rajarshi Guhaniyogi, Yang Ni, Bani K. Mallick

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery inside a bustling, crowded city. This city is a piece of human tissue (like a slice of breast cancer tissue), and the "citizens" are genes.

In the past, scientists could only take a "smoothie" of the whole city, blend all the citizens together, and ask, "Is this gene active?" This told them if a gene was working, but not where.

Spatially Resolved Transcriptomics (SRT) is like giving the detective a high-tech map. Now, they can see exactly which genes are active in which specific neighborhood. The goal is to find the "Spatially Varying Genes" (SVGs)—the genes that act like local landmarks, turning on only in specific districts (like the "downtown" of a tumor) and turning off in others.

However, finding these landmarks is tricky. The current tools used by detectives have two big problems:

They work alone: They check every gene one by one, ignoring the fact that genes often work in teams (co-expression).
They are rigid: They assume the city's layout is perfectly smooth and predictable (like a grid), but real biology is messy, with weird shapes and sudden changes.

Enter JASPER.

What is JASPER?

Think of JASPER as a super-smart, team-oriented detective who uses a flexible, custom-made map.

Here is how it works, using simple analogies:

1. The "Team Player" Approach (Joint Modeling)

Old Way: Imagine checking if a specific streetlight is on by looking at it in isolation. If the light flickers, you might think it's broken, or you might miss it because you didn't notice the other lights on the same block are flickering too.
JASPER's Way: JASPER knows that genes are like a choir. If one singer (gene) is off-key, it might be because the whole choir is changing the song. JASPER listens to the whole choir at once. By understanding how genes "talk" to each other, it can spot the ones that are truly special (spatially varying) much more accurately, ignoring the noise.

2. The "Flexible Map" (Spatial Basis Functions)

Old Way: Imagine trying to draw a map of a city using only a ruler and a protractor. You assume every street is straight and every block is a perfect square. If the city has a winding river or a jagged coastline, your map fails. Most old methods use "rigid kernels" (like that ruler) that assume biological patterns are smooth and predictable.
JASPER's Way: JASPER uses Lego blocks (called spatial basis functions) to build the map. It doesn't care if the pattern is a straight line, a circle, or a jagged lightning bolt. It can snap these blocks together to fit any shape the data takes. This means it doesn't get confused by the messy, complex reality of real tissue.

3. Counting the "People" Directly (Negative Binomial Model)

Old Way: Some methods try to "normalize" the data first, which is like trying to count people in a room by first washing everyone's clothes and then guessing how many people were there based on the laundry. This can distort the truth.
JASPER's Way: JASPER counts the actual people (gene expression counts) directly, accounting for the fact that some rooms are naturally crowded and some are empty. It uses a math tool (Negative Binomial) that understands that in biology, "crowdedness" varies wildly and isn't always a perfect average.

Why Does This Matter? (The Results)

The paper tested JASPER on real human breast cancer data and mouse brain data. Here is what happened:

Better Detective Work: JASPER found more "landmarks" (SVGs) than the old tools, and it made fewer mistakes (fewer false alarms).
Biological Truth: The genes JASPER found were more likely to be the "real deal." When checked against known medical databases, JASPER's list included famous cancer markers (like COX6C and EFNA1) that the other tools missed or ignored.
Understanding the Disease: Because JASPER found the right genes, it helped scientists understand how the cancer spreads. It highlighted that the cancer cells are interacting with their "neighborhood" (the extracellular matrix) to help them invade other areas. This is crucial for finding new drugs.

The Bottom Line

If existing methods are like trying to navigate a complex, winding city with a rigid, pre-printed map, JASPER is like having a GPS that learns the city's layout in real-time, listens to the traffic reports from all the cars (genes) at once, and guides you straight to the most important locations.

It turns a messy, confusing biological puzzle into a clear picture, helping doctors and scientists understand diseases like cancer at a much deeper level.

1. Problem Statement

Spatially Resolved Transcriptomics (SRT) technologies allow for the measurement of gene expression within the specific spatial context of tissue architecture. A primary goal in analyzing SRT data is identifying Spatially Varying Genes (SVGs)—genes whose expression levels change significantly across spatial locations.

The paper identifies three critical limitations in existing SVG detection methods (e.g., SpatialDE, SPARK, SOMDE):

Ignoring Gene Co-expression: Current methods typically analyze genes one-by-one (gene-by-gene screening). They fail to account for the fact that genes are often co-regulated and correlated, leading to inflated false positive and false negative rates.
Rigid Spatial Kernels: Most prominent methods rely on pre-specified spatial covariance kernels (e.g., Gaussian processes with Matérn or squared-exponential kernels). These assume stationarity, which is frequently violated in complex biological tissues where spatial patterns are non-stationary and diverse.
Normalization Artifacts: Many methods require pre-normalization of count data (e.g., library size normalization), which can introduce spurious spatial effects or distort the natural mean-variance relationship inherent in count data.

2. Methodology: JASPER

The authors propose JASPER, a fully Bayesian framework that addresses these limitations through Joint Bayesian Analysis via Regression.

A. Statistical Model

JASPER models the data using a hierarchical structure:

Observation Model (Negative Binomial):
Instead of normalizing data first, JASPER models raw gene expression counts ( $C_j(s_i)$ ) directly using a Negative Binomial distribution. This accounts for overdispersion and zero-inflation common in SRT data. The mean is linked to a latent Gaussian variable $Y_j(s_i)$ via a log-link function, adjusted for library size ( $N_i$ ).
$C_j(s_i) \mid Y_j(s_i) \sim \text{NB}(\mu_j(s_i), N_i)$
Latent Spatial Regression:
The latent variables $Y(s_i)$ are modeled using Spatial Basis Function Regression:
$Y(s_i) = \alpha + B^\top X(s_i) + \epsilon(s_i)$
- $X(s_i)$ : A vector of spatial basis functions (e.g., cubic B-splines) evaluated at location $s_i$ . This avoids the need for a pre-specified covariance kernel, allowing flexible modeling of complex, non-stationary patterns.
- $\Sigma$ : A covariance matrix capturing gene co-expression patterns across the $p$ genes.
- $\epsilon(s_i)$ : Independent noise.
Variable Selection (Spike-and-Slab Priors):
To identify SVGs, JASPER employs a multivariate spike-and-slab prior on the regression coefficient vectors ( $\beta_j$ ) for each gene.
- Indicator Variable ( $\gamma_j$ ): $\gamma_j = 1$ if gene $j$ is spatially varying; $\gamma_j = 0$ otherwise.
- Prior: If $\gamma_j=0$ , $\beta_j$ is fixed at 0 (spike). If $\gamma_j=1$ , $\beta_j$ follows a multivariate normal distribution (slab).
- This approach treats SVG detection as a variable selection problem, borrowing strength across correlated genes.

B. Inference Algorithm

Posterior Inference: The authors use a Gibbs sampler with Polya-Gamma data augmentation to sample the latent Gaussian variables and a collapsed Gibbs sampling scheme for the indicator variables $\gamma$ .
Selection Criterion: Genes are selected as SVGs if their Posterior Probability of Inclusion (PPI) exceeds a threshold chosen to control the posterior expected False Discovery Rate (peFDR) at 0.05.
Scalability: For very large datasets, a simplified "JASPER-normalized" version based on Gaussian assumptions is proposed.

3. Key Contributions

Joint Modeling: Unlike existing methods, JASPER simultaneously models multiple genes, explicitly capturing co-expression structures. This allows the method to "borrow information" across genes, improving detection accuracy, especially when genes are highly correlated.
Kernel-Free Flexibility: By using spatial basis functions (B-splines) rather than fixed Gaussian process kernels, JASPER can adapt to diverse and non-stationary spatial patterns without the risk of kernel misspecification.
Direct Count Modeling: The use of Negative Binomial distributions on raw counts eliminates the need for pre-normalization, preserving the statistical properties of the data and avoiding artifacts introduced by normalization techniques.
Bayesian Variable Selection: The framework integrates variable selection directly into the spatial regression model, providing a principled way to identify SVGs while accounting for uncertainty.

4. Results

A. Simulation Studies

Co-expression: In simulations with correlated genes, JASPER significantly outperformed SPARK and SpatialDE, particularly when spatial effects were weak and gene correlations were high.
Kernel Misspecification: When data was generated using a periodic exponential sine squared kernel (violating standard kernel assumptions), kernel-based methods (SpatialDE, SOMDE) had low true positive rates (~43-63%). JASPER achieved a 93.5% true positive rate, demonstrating robustness to kernel misspecification.
Sparsity: JASPER's performance improved as the proportion of SVGs increased, leveraging the correlation structure to better learn underlying patterns.

B. Real-World Applications

The method was applied to three distinct datasets:

Human Breast Cancer (10x Visium):
- JASPER detected 672 SVGs.
- Biological Validation: Genes uniquely detected by JASPER (e.g., COX6C, EFNA1) showed higher overlap with known cancer gene databases (Harmonizome, COSMIC) and higher Moran's I statistics (spatial autocorrelation) than competing methods.
- Enrichment: JASPER-unique genes were enriched in pathways critical to breast cancer, such as "cell adhesion molecule binding" and "ECM-receptor interaction," which were missed or less prominent in other methods.
- Replication: 89.8% of SVGs found in a secondary ST dataset were also found by JASPER in the Visium dataset (vs. ~61% for SPARK).
Mouse Visual Cortex (STARmap):
- JASPER detected 53 SVGs, capturing all 13 genes highlighted in the original study (vs. 11 for SPARK, 3 for SpatialDE).
- It successfully identified genes with sparse, clustered high-expression patterns that other methods missed (e.g., Gad1).
- 35.3% of JASPER's unique detections were associated with visual areas in the Harmonizome database, outperforming competitors.
Mouse Olfactory Bulb (ST):
- JASPER identified 713 SVGs and detected all 10 genes highlighted in the original study.
- It showed superior overlap with known olfactory-associated genes (22.5%) compared to SPARK (17.3%) and SpatialDE (5.7%).

5. Significance

JASPER represents a significant advancement in spatial transcriptomics analysis by shifting the paradigm from independent gene screening to joint Bayesian modeling. Its ability to handle complex, non-stationary spatial patterns without rigid kernel assumptions, combined with its capacity to leverage gene co-expression, makes it a more robust and biologically interpretable tool. The method not only improves statistical power (reducing false negatives/positives) but also uncovers biologically relevant pathways and therapeutic targets that existing methods fail to detect, particularly in complex disease microenvironments like breast cancer.