Exploring Segmentation-Free Methods in Spatial Transcriptomics: FIXTURE and Sainsc

helixhorizonsoluti
Apr 3
3 min read

After introducing platforms, tissue imaging with cell segmentation, and TensionMap’s mechanical insights, we’re now shifting gears to explore segmentation-free methods. These innovative approaches bypass the need for predefined cell boundaries, offering a fresh perspective on analyzing high-resolution spatial data. At Helix Horizon, we’re excited to leverage tools like FIXTURE and Sainsc to empower researchers with flexible, scalable solutions.

FIXTURE: Pixel Inference vs Sainsc: Density Clustering

What Are Segmentation-Free Methods?

Traditional spatial gene expression data analysis often relies on segmenting tissue images into individual cells to assign gene expression profiles. However, this can falter when cell boundaries are unclear, nuclei are absent (e.g., erythrocytes), or imaging data is incomplete. Segmentation-free methods sidestep these challenges by analyzing gene expression at the pixel level, preserving resolution and capturing cellular heterogeneity without assuming cell outlines. This is especially powerful for submicron-resolution datasets where transcript density is sparse but biologically rich.

FIXTURE: Scalable Pixel-Level Inference

FICTURE (published in 2024) is designed for large-scale, submicron-resolution spatial transcriptomics. It uses a factor-based approach to infer cell types and tissue domains without segmentation. Here’s how it works:

Workflow: FIXTURE collapses raw transcript data into a hexagonal grid, then applies Latent Dirichlet Allocation (LDA) to identify expression factors (e.g., cell-type signatures). These factors are mapped back to individual pixels, assigning probabilistic cell-type labels at native resolution.
Strengths: It scales to large tissue areas (>5mm²) and broad gene panels (>1,000 genes), handling billions of transcripts efficiently. It’s flexible, accepting external single-cell RNA-seq references or unsupervised clustering inputs.
Key Feature: By avoiding segmentation, FIXTURE retains fine spatial details and adapts to diverse tissue types, even where cell boundaries are ambiguous.

Sainsc: Density-Based Cell-Type Assignment

Sainsc (introduced in 2024) takes a different tack, using a segmentation-free, density-driven model inspired by SSAM. It’s tailored for transcriptome-wide, nanometer-resolution data. Here’s the breakdown:

Workflow: Sainsc applies Kernel Density Estimation (KDE) to model spatial gene expression as a continuous density. Local maxima of expression are clustered (e.g., via Leiden algorithm) to define cell-type signatures, which are then assigned across the tissue without cell borders.
Strengths: It excels at detecting rare or small cells (e.g., erythrocytes in spleen) missed by segmentation-based methods. It’s computationally efficient, with Python and Julia implementations for rapid analysis.
Key Feature: Sainsc provides an assignment score map, offering confidence levels for cell-type calls, enhancing interpretability.

Comparing FIXTURE and Sainsc

While both tools avoid segmentation, their approaches differ:

Scalability: FIXTURE shines with massive datasets, processing entire tissue sections quickly. Sainsc, while efficient, is optimized for interactive, exploratory analysis rather than sheer scale.
Resolution: FIXTURE’s pixel-level inference preserves submicron detail via grid-based factorization. Sainsc’s KDE smooths expression into densities, potentially softening ultra-fine features but boosting sensitivity to low-expression cells.
Flexibility: FIXTURE integrates external references easily, ideal for cross-study comparisons. Sainsc leans on unsupervised clustering, better suited for discovery in novel tissues.
Output: FIXTURE delivers probabilistic cell-type maps; Sainsc adds confidence scores, aiding validation.

Why It Matters for Your Research

Segmentation-free methods like FIXTURE and Sainsc unlock new possibilities—capturing cell types and tissue domains that traditional methods miss, all while handling the complexity of modern spatial data. At Helix Horizon, we can implement these tools to fit your project, whether you need FIXTURE’s scalability or Sainsc’s sensitivity. Our bioinformatics services ensure you get the most from your spatial data.