🧬 biology

spAlignDE unifies cross-sample and cross-modal spatial alignment with mismatch-aware differential expression

spAlignDE is a computational framework that unifies cross-sample and cross-modal spatial alignment using image-like continuous fields and shooting-based diffeomorphic registration, while introducing a mismatch-aware differential expression test to reduce false signals caused by imperfect alignment.

Original authors: Dongyuan Song, Su Xu, Yu Wang, Lingzhen Meng, Aarya Dalal, Yuxin Yin

Published 2026-11-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dongyuan Song, Su Xu, Yu Wang, Lingzhen Meng, Aarya Dalal, Yuxin Yin

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 trying to compare two maps of the same city, but one was drawn from a helicopter, the other from a bicycle, and both were folded, stretched, and torn in different ways. The streets might be in the same relative positions, but the angles are wrong, the scales don't match, and some neighborhoods are missing entirely. This is the daily reality for scientists studying the inside of living tissues. They use powerful tools to measure genes, proteins, and chemicals while keeping track of exactly where those molecules are located within a slice of tissue. This field, known as spatial omics, allows researchers to see how the body's architecture changes during aging or disease. However, to learn anything meaningful, they must be able to line up these different tissue slices perfectly. If they cannot match a specific spot in a healthy kidney to the exact same spot in an injured one, or align a gene map with a standard anatomical atlas, any comparison they make is built on shaky ground.

For years, researchers have struggled with this alignment problem. Existing methods often require the two samples to share the exact same molecular features to line them up, which is impossible when comparing a gene map to a simple photograph of tissue or a standard brain atlas. Even when alignment is attempted, the process is rarely perfect. Small errors can cause a piece of tissue from one sample to be forced onto a completely different part of another sample. When scientists then try to measure differences in gene activity between these misaligned spots, they often mistake these geometric errors for real biological changes, leading to false conclusions about what is happening inside the body.

A team of researchers has now introduced a new computational framework called spAlignDE that solves these problems by treating tissue structures like continuous images rather than just lists of data points. Instead of looking for matching genes to connect two samples, the system identifies the shapes and patterns of the tissue itself—such as the layers of the brain or the compartments of a kidney—and uses those shapes as a guide to line everything up. This approach allows the researchers to align different types of data, such as gene expression maps, chemical accessibility maps, and standard microscope images, without needing them to share any common molecular features. The system works by converting these tissue structures into smooth, continuous fields, similar to how a digital image is made of pixels, and then mathematically warping one field to fit the other with high precision. This method is fast enough to handle massive datasets containing over a million cells in just a few minutes, a task that would take other methods much longer or fail to complete entirely.

Crucially, the researchers recognized that even the best alignment is never perfect. There are always tiny mismatches where a spot in one sample does not correspond exactly to a spot in the other. Rather than ignoring these errors or pretending the alignment is flawless, spAlignDE includes a built-in safety mechanism. It calculates a risk score for every location, estimating how likely it is that the alignment is slightly off. If a location has a high risk of mismatch, the system automatically increases the uncertainty of the statistical test performed there. This means that if the data is shaky, the system becomes more conservative, reducing the chance of falsely claiming a gene is active when it is not. This "mismatch-aware" approach ensures that the final results are reliable, filtering out false signals that would otherwise confuse the scientific picture.

The team tested this new system against twelve other existing methods using a variety of real-world datasets, including mouse brains at different ages and kidney sections from healthy and injured animals. In every test, spAlignDE outperformed the competition, achieving the highest accuracy in matching anatomical regions and preserving the natural patterns of gene expression. It successfully aligned gene maps from different technologies, matched brain sections to a standard three-dimensional atlas, and even connected gene data with chemical accessibility maps from a different type of sequencing. In one large-scale test involving twenty sections of a mouse brain containing over 1.45 million cells, the system completed the alignment in under three minutes while maintaining high precision.

When applied to real biological questions, the system revealed details that were previously hidden. In a study of the aging mouse brain, the researchers used spAlignDE to track how gene activity changed over time. They found that a specific gene, known to decline with age, did not fade away uniformly across the entire brain. Instead, the decline was concentrated in a specific arch-shaped region of white matter, a pattern that would have been missed if the researchers had only looked at the brain as a whole. Similarly, they observed that immune cells called T cells moved closer to specific types of brain cells as the animals aged, a change in neighborhood organization that was only visible once the tissue slices were perfectly aligned.

In another application, the researchers compared healthy and injured kidney sections. By aligning the two samples and accounting for potential mismatches, they identified exactly where specific genes increased or decreased in response to injury. They found that the changes were not uniform across the organ; some genes went up in the outer cortex while others went down in the inner medulla. This level of detail allowed them to link specific biological processes, such as inflammation or metabolic changes, to the precise anatomical locations where they occurred. Without the ability to align the samples accurately and account for the uncertainty of that alignment, these localized patterns would have been lost in the noise of the data.

The work demonstrates that accurate comparison of biological tissues requires more than just matching data points; it requires a deep understanding of the physical structures those points represent and a statistical honesty about the limits of that matching. By unifying the alignment of different data types with a rigorous method for handling errors, spAlignDE provides a new standard for how scientists can compare tissues across time, conditions, and technologies. It allows researchers to see the body not just as a collection of molecules, but as a dynamic, structured landscape where every change has a specific place and a specific meaning.

Technical Summary of spAlignDE

Problem Statement
Spatial omics technologies enable the measurement of molecular profiles while preserving tissue architecture, facilitating comparisons across samples and modalities (e.g., transcriptomics, epigenomics, histology). However, these analyses are confounded by spatial heterogeneity, including differences in orientation, scale, sectioning angle, and local deformation. Existing spatial alignment methods face four primary limitations: (1) high computational demands that hinder scalability to large, cell-resolved datasets; (2) the potential for overly flexible transformations to distort local neighborhoods or force unsupported correspondences; (3) difficulty in cross-modal alignment when datasets lack directly comparable molecular features; and (4) a lack of rigorous statistical frameworks for location-resolved differential expression (DE) that account for residual alignment mismatch. Treating estimated correspondences as exact can underestimate uncertainty and inflate false discovery rates in downstream DE analysis.

Methodology
The authors propose spAlignDE, an integrated framework that unifies cross-sample and cross-modal spatial alignment with mismatch-aware local differential expression analysis. The framework consists of two complementary components:

Structure-Guided Spatial Alignment:
- Representation: spAlignDE converts modality-specific tissue structures (from transcriptomics, ATAC-seq, histology, or anatomical atlases) into comparable, image-like continuous fields. For cross-sample alignment, it uses joint clustering of batch-corrected gene expression and spatial neighborhoods to identify shared structures. For cross-modal alignment, it identifies structures independently within each modality and pairs them based on geometric compatibility without requiring shared molecular features.
- Registration: The method employs shooting-based large-deformation diffeomorphic metric mapping (S-LDDMM). Unlike standard variational LDDMM that optimizes time-dependent velocity fields, S-LDDMM estimates a single initial momentum field, significantly improving scalability and reducing memory usage.
- Handling Mismatch: To accommodate missing tissue or partial overlap, spAlignDE adapts a probabilistic missing-data formulation. It fits an expectation-maximization mixture model with one matched component and two unmatched appearance components. The posterior probability of the matched component ( $W_M(x)$ ) weights each location during registration, ensuring that locations without valid counterparts contribute less to the transformation estimation.
- Workflow: The process involves global pre-alignment (using centroid matching, mask overlap, or interactive adjustment), structure-pair identification based on a composite geometric score (Dice, signed-distance-field, Chamfer, area, and thickness), and final S-LDDMM refinement.
Mismatch-Aware Local Differential Expression:
- Testing Framework: After alignment, spAlignDE constructs a shared testing grid. It estimates local expression means using kernel-weighted neighborhoods of aligned observations.
- Variance Inflation: Recognizing that nominally matched locations may still differ in anatomy or cell composition (residual mismatch), the method estimates a local mismatch-risk score ( $R^i_{loc}$ ). This score is derived from discrepancies in the local distributions of putatively stable genes and local sampling density discordance between samples.
- Statistical Model: The method inflates the variance of local expression contrasts based on the estimated mismatch risk. The working model is $\hat{\Delta}_{ig} = \gamma_g + z_i^\top \beta_g + \theta_{ig} + \epsilon_{ig}$ , where the variance of the error term $\epsilon_{ig}$ is scaled by a factor dependent on $R^i_{loc}$ . This reduces inferential precision in poorly supported regions without altering the estimated effect direction.
- Output: The framework provides location-specific P-values, adjusted q-values, and identifies connected significant regions.

Key Contributions and Results

Cross-Sample Alignment Performance: In benchmarks against 12 existing methods using paired MERFISH mouse brain data, Visium kidney data, and a large MERFISH aging cohort (20 sections, 1.45 million cells), spAlignDE achieved the highest agreement in spatial gene expression patterns and anatomical annotations. It demonstrated superior scalability, completing the alignment of 20 brain sections in 2.43 minutes with peak GPU memory usage below 0.1 GB.
Cross-Modal Alignment: spAlignDE successfully aligned spatial transcriptomics with histology (H&E and Nissl stains), the Allen Mouse Brain Common Coordinate Framework (CCFv3), and spatial ATAC-seq data. In the atlas alignment benchmark, spAlignDE outperformed the task-specific method xIV-LDDMM, leaving only 0.67% of cells unlabeled compared to 5.81% for xIV-LDDMM, and showed higher consistency between transferred atlas labels and expression-based cell type assignments.
Mismatch-Aware Inference: Semi-synthetic simulations demonstrated that the mismatch-aware DE test significantly improved false discovery control (FDP) across various alignment methods compared to "naïve" testing that assumes exact correspondence. The method reduced FDP from levels as high as 0.389 to 0.178 for certain methods, with near-zero median FDP for spAlignDE, while maintaining statistical power.
Real-World Applications:
- Aging Mouse Brain: The framework localized age-associated down-regulation of Gamt to an arch-shaped white-matter region along the corpus callosum and identified increased proximity and spatial overlap of T-cells with oligodendrocytes, microglia, and ependymal cells in older brains.
- Injured Kidney: In ischemia–reperfusion injury models, spAlignDE resolved spatially restricted expression differences (e.g., increases in Cbr1 and Cd44, decreases in Myo5a) and identified compartment-specific functional enrichments (e.g., metabolic processes in the cortex, extracellular matrix organization at the interface) that were obscured by compartment-level aggregation.

Significance and Claims
The paper claims that spAlignDE provides the first rigorous, location-resolved statistical comparison framework that explicitly accounts for residual alignment mismatch. By unifying diverse alignment modalities through structure-guided continuous fields and integrating mismatch risk into the variance of statistical tests, spAlignDE addresses the critical gap between spatial registration and reliable differential expression analysis. The authors assert that this approach allows for the detection of localized biological changes that would otherwise be masked by global summaries or confounded by alignment errors, offering a scalable and robust solution for the growing field of spatial multi-omics. The method is implemented as an open-source Python package to facilitate reproducibility and adoption.

Technical Summary of spAlignDE

More like this