Synora: vector-based boundary detection for spatial omics

Synora is a modality-agnostic computational framework that robustly identifies tumor-stroma boundaries in spatial omics data by introducing a novel "orientedness" metric to distinguish true compartment interfaces from random cell infiltration using only cell coordinates and binary annotations.

The Gist

Imagine you are looking at a bustling city from a drone. In this city, there are two distinct groups: the "Residents" (healthy tissue) and the "Invaders" (cancer cells).

For a long time, scientists have been trying to draw a perfect line on a map to separate the Residents from the Invaders. This line is called the tumor boundary. It's a crucial place because that's where the two groups talk to each other, exchange signals, and decide whether the disease will spread or be stopped.

However, drawing this line is incredibly hard. Sometimes, the Invaders sneak into the Resident neighborhoods, and sometimes the Residents wander into the Invader zones. Traditional maps often get confused, thinking a chaotic mix of people is a border, when it's actually just a messy crowd.

Enter Synora. Think of Synora as a new, super-smart GPS system for these biological cities.

The Problem with Old Maps

Previous tools tried to find the border by just counting how many different types of people were in a neighborhood.

• The Flaw: If you have a mix of Residents and Invaders, the old tools say, "Aha! This is a border!"
• The Reality: Sometimes, that mix is just a chaotic crowd where everyone is jumbled together randomly (like a mosh pit). Other times, it's a true border where the two groups are standing on opposite sides of a fence, facing each other. The old tools couldn't tell the difference between a border and a mosh pit.

The Synora Solution: "The Direction Detector"

Synora introduces a new concept called "Orientedness."

Imagine you are standing in the middle of a crowd.

• Scenario A (The Mosh Pit): People are pushing you from all sides randomly. There is no clear direction. Synora says, "This is just a mix, not a border."
• Scenario B (The True Border): You are standing on a line. All the Invaders are pushing you from the left, and all the Residents are pushing you from the right. There is a clear direction to the pressure.

Synora measures this "directional push." It doesn't just ask, "Are there different people here?" It asks, "Are the different people pushing from opposite sides?"

How It Works (The Three Steps)

1. The Scan: Synora looks at every cell and its neighbors. It checks if the neighbors are a mix of types (like a salad) and, more importantly, if that mix is organized in a specific direction.
2. The Score: It gives every cell a "Boundary Score." If a cell feels the directional push of two opposing groups, it gets a high score and is marked as part of the border. If it's just in a random mix, it gets a low score.
3. The Map: Once the border is drawn, Synora can tell you exactly how far any cell is from the edge. Is it deep inside the tumor? Deep in the healthy tissue? Or right on the front line?

Why This Matters

The researchers tested Synora on real data from 15 different cancer types (like breast, lung, and colon cancer) and even on protein imaging data.

• It's a Universal Tool: It works on almost any type of spatial data, whether it's looking at genes (transcriptomics) or proteins. You don't need complex genetic data; just the location of the cells and a simple label saying "Tumor" or "Not Tumor" is enough.
• It Finds Hidden Secrets: Because it draws the border so precisely, it found new "neighborhoods" of cells that other methods missed. For example, in some patients, it found that specific immune cells were gathering right at the border in a way that predicted how the disease would behave.
• It's Tough: Even if the data is messy (like if 50% of the cells are missing or the border is very jagged), Synora still finds the line.

The Big Picture

Think of Synora as the difference between looking at a blurry photo of a crowd and seeing a high-definition map that clearly shows the fence line.

By understanding exactly where the tumor ends and the healthy tissue begins, doctors and scientists can better understand how cancer invades, how the immune system fights back, and where to aim new treatments. It turns a fuzzy, confusing edge into a clear, measurable line, helping us decode the language of the tumor's neighborhood.

Technical Summary

1. Problem Statement

Spatial omics technologies (e.g., Visium HD, CODEX, Xenium) now allow for single-cell resolution mapping of tissue microenvironments. A critical biological feature in cancer is the tumor–stroma boundary, a dynamic interface where malignant and non-malignant cells interact to drive invasion, immune modulation, and therapeutic resistance.

However, existing computational methods struggle to accurately detect these boundaries because:

• Ambiguity in Heterogeneity: Both true tissue interfaces and regions of random immune infiltration exhibit high cellular diversity (mixed cell types). Traditional neighborhood-based clustering or frequency-based metrics cannot distinguish between a structured interface (segregated cell types) and disorganized infiltration (intermixed cell types).
• Data Requirements: Image-based segmentation requires high-quality histology and is computationally heavy; Copy Number Variation (CNV) methods are limited to assays where genomic alterations can be inferred.
• Lack of Directionality: Existing methods often treat spatial neighborhoods as scalar compositions, ignoring the directional organization of cells relative to the interface.

2. Methodology: The Synora Framework

Synora is a modality-agnostic computational framework that identifies tumor boundaries using only cell coordinates and binary tumor/non-tumor annotations. It introduces a novel metric called "orientedness" to resolve the ambiguity between structured boundaries and random infiltration.

The workflow consists of three main components:

A. Core Metrics

Synora decomposes boundary evidence into two complementary components:

1. Mixedness ($M$): Measures local neighborhood heterogeneity (similar to the Gini–Simpson index). It quantifies whether a cell's neighbors contain a mix of tumor and non-tumor types. High mixedness occurs in both true boundaries and infiltrated regions.
2. Orientedness ($O$): A novel vector-based metric that quantifies directional spatial bias.
   • It calculates the annotation-weighted vector sum of neighbors around a central cell.
   • At a true boundary, tumor cells cluster on one side and stromal cells on the other, creating a strong directional vector.
   • In infiltrated regions, cell types are randomly intermingled, resulting in a net vector near zero.
   • Edge Correction: The algorithm includes a geometric correction to prevent cells near the imaging field boundary from being spuriously classified as boundaries due to missing neighbors on one side.

B. Boundary Score

The final BoundaryScore is computed as the product of Mixedness and Orientedness:
$$\text{BoundaryScore} = M \times O$$
This ensures that only neighborhoods that are both mixed (heterogeneous) and directionally structured (segregated) are identified as boundaries.

C. Three-Step Workflow

1. Boundary Identification: Cells are classified into three compartments (Boundary, Nest, Outside) based on the BoundaryScore.
2. Distance Stratification: A signed distance-to-boundary metric is calculated for every cell (positive for tumor nests, negative for stroma), enabling continuous spatial profiling.
3. Shape Metrics: The framework calculates morphological descriptors (e.g., boundary-to-nest ratio, fractal dimension, solidity) to quantify interface abundance and irregularity.

3. Key Contributions

• Novel Metric (Orientedness): The introduction of a vector-based metric that distinguishes structured interfaces from random infiltration, a capability lacking in previous scalar-based methods.
• Modality-Agnostic Design: Requires minimal input (coordinates + coarse labels), making it applicable across transcriptomics, proteomics, and metabolomics without needing specific molecular features or CNV data.
• Robustness: The method is designed to handle realistic experimental perturbations, including missing data and complex tissue architectures.
• Open Source: The Synora R package is publicly available, standardizing interface detection across the field.

4. Results

A. Synthetic Validation

• Orthogonality: In synthetic datasets generated with Perlin noise, orientedness increased as cell distributions transitioned from random mixing to spatial segregation, while mixedness remained constant.
• Robustness: Synora maintained high performance (AUPRC > 0.72) under significant perturbations:
  • Infiltration Noise: Up to 25% of tumor cells reassigned to non-tumor.
  • Missing Data: Up to 50% of cells randomly removed.
  • Complexity: Boundaries with varying fractal complexity.
• Comparison: Traditional mixedness-only approaches failed significantly under infiltration noise (AUPRC dropped to ~0.4), whereas Synora remained robust.

B. Application to Spatial Transcriptomics (Visium HD)

• Datasets: Analyzed 15 datasets across 6 cancer types (BRCA, CRC, LUAD, OC, PCa, PDAC).
• Gene Signatures: Identified consistent boundary-enriched genes across cancers, including ECM remodeling genes (COL1A1, FN1) and immune/chemokine genes (CXCL9, CXCL14).
• Cancer-Specificity: Revealed distinct pathway enrichments per cancer type (e.g., Angiogenesis in PCa/LUAD; TGF-β in PCa; Chemokine signaling in CRC).
• Cellular Gradients: Confirmed conserved tumor-fibroblast segregation but highlighted cancer-specific immune and vascular distributions near the boundary (e.g., neutrophil enrichment in CRC stroma; endothelial peaks in PDAC).

C. Application to Multiplexed Protein Imaging (CODEX)

• Validation: Applied to a Colorectal Cancer (CRC) dataset, showing superior precision in boundary detection compared to original manual annotations.
• Discovery: Incorporating Synora-derived features (BoundaryScore and distance) into neighborhood clustering revealed novel cellular neighborhoods (CNs) missed by frequency-based approaches:
  • CN3: A DC–Treg–macrophage-enriched tumor boundary neighborhood.
  • CN2, CN6, CN7: Novel T-cell enriched niches.
• Clinical Correlation: Identified differential abundance of these neighborhoods between clinical groups (Diffuse Inflammatory Infiltration vs. Crohn's-like lymphoid reaction), linking boundary architecture to specific immune phenotypes (e.g., Th2 programs and granulocyte recruitment in DII).

5. Significance

• Standardization: Synora provides a standardized, quantitative framework for defining tissue interfaces, enabling cross-study and cross-platform comparisons.
• Biological Insight: By distinguishing structured boundaries from random infiltration, the tool reveals "hidden" biological programs active at the tumor–stroma interface that are otherwise obscured by noise.
• Clinical Potential: The ability to detect boundary-associated cellular neighborhoods offers new biomarkers for understanding immune evasion and therapeutic response.
• Future Directions: While currently 2D, the vector-based logic is extensible to 3D spatial datasets. The authors note that future work should benchmark against ground-truth histopathology and extend to 3D volumetric data.

In summary, Synora transforms spatial boundary detection from a composition-based problem to a directional-organization problem, offering a robust, low-input solution to a critical challenge in spatial biology.