Topological Data Analysis of Spatial Protein Expression in Multiplexed Spatial Proteomics Studies

This paper introduces TOASTER, a novel topological data analysis method that bypasses error-prone cell segmentation and phenotyping to directly associate continuous spatial protein expression patterns with patient outcomes, demonstrating improved statistical power and robustness in multiplexed spatial proteomics studies.

The Gist

Imagine you are trying to understand the layout of a bustling city by looking at a giant, high-resolution photograph of it.

The Old Way: Counting People in Houses
Traditionally, scientists analyzing tissue samples (like a slice of a tumor) have tried to do this by first drawing a box around every single "house" (cell) in the photo. They then try to guess what kind of person lives in each house (is it a T-cell? A B-cell?). Once they've labeled every house, they look at the neighborhood map to see if the arrangement of these "people" predicts how sick a patient is.

The Problem:
This approach is like trying to understand a city by only looking at the front doors.

1. It's messy: Sometimes houses are squished together, or the photo is blurry. Drawing perfect boxes around them is hard and prone to mistakes.
2. It throws away data: If a person is standing in the garden (outside the house boundary), the old method ignores them. But in biology, that "garden" space is full of important chemical signals.
3. It misses the vibe: It focuses on who is there, but ignores how much of a signal they are sending.

The New Way: TOASTER (The "City Vibe" Detector)
The authors of this paper, Sarah and Michael, invented a new tool called TOASTER. Instead of trying to draw boxes around cells, TOASTER looks at the entire image as a continuous landscape of light and color.

Think of the protein expression in a tissue sample like a mountain range made of light.

• High protein levels = Tall, bright peaks.
• Low protein levels = Deep, dark valleys.

How TOASTER Works (The "Flood" Analogy):
TOASTER uses a technique called Topological Data Analysis. Imagine you are slowly pouring water into this mountain range (the image).

1. The Flood (Filtration): As the water level rises, new islands (connected groups of high protein) appear out of the water. These are "births."
2. The Lakes (Loops): As the water rises further, some islands might get surrounded by water, creating lakes. Eventually, the water fills the lakes, and they "die."
3. The History Book: TOASTER doesn't care about the individual houses. It keeps a running diary (called a "Topological Event History") of exactly when these islands were born and when the lakes were filled in.

The "Event History" as a Fingerprint:
Every patient's tissue sample gets its own unique "flood diary."

• If a patient is responding well to cancer treatment, their "mountain range" might have fewer, larger islands (meaning the immune cells are clustered together in big, strong groups).
• If a patient is not responding, the islands might be tiny and scattered everywhere.

The Test:
TOASTER takes these "flood diaries" from all the patients and asks: "Do the diaries of the patients who got better look different from the diaries of the patients who didn't?"

It uses three different ways to compare these diaries:

1. The Shape Matcher: Looks at the overall curve of the diary.
2. The Grid Checker: Checks specific points on the water level to see where the groups differ most.
3. The Pattern Finder: Uses math to find hidden patterns in the differences.

Why This is a Big Deal:

• No Boxes Needed: It skips the messy step of trying to draw lines around cells. It works even if the tissue sample has a tear or a hole in it (like a ripped map), because it looks at the whole picture.
• More Power: In their tests, TOASTER was better at spotting the difference between sick and healthy patients than the old methods.
• Real World Success: They tested it on Triple-Negative Breast Cancer. They found that patients who responded to immunotherapy had a very specific "topological fingerprint"—their immune cells (CD3, CD4, CD8) were clustered in a way that the old methods missed.

In a Nutshell:
Instead of trying to count and label every single cell (which is like trying to count every grain of sand on a beach), TOASTER looks at the shape of the beach itself. It realizes that the shape of the protein landscape holds the secret to whether a patient will beat their cancer, even if we can't perfectly see every single cell.

Technical Summary

1. Problem Statement

Multiplexed spatial proteomics platforms (e.g., MIBI, IMC, multiplexed IHC) generate high-resolution images capturing the spatial distribution of proteins in tissue. Standard analysis pipelines rely on a complex pre-processing workflow involving:

1. Segmentation: Identifying cell boundaries.
2. Phenotyping: Assigning cell type labels based on protein markers.
3. Point Process Modeling: Treating cells as points with "marks" (types) to test associations with patient outcomes.

Limitations of Current Approaches:

• Error Propagation: Segmentation and phenotyping are prone to errors due to overlapping cells, uneven morphologies, or tissue tears. These errors are not accounted for in downstream statistical tests.
• Information Loss: Protein expression outside predicted cell boundaries is discarded, losing valuable tissue context.
• Discretization Bias: Focusing solely on cell types ignores the continuous quantitative levels of protein expression, which may be biologically significant.
• Data Artifacts: Existing methods struggle with gaps or tears in tissue samples common during handling.

There is a lack of statistical methods capable of directly associating continuous spatial protein expression with patient-level outcomes without relying on imperfect cell segmentation.

2. Methodology: TOASTER

The authors propose TOASTER (Test Of Association between Spatial protein expression and clinical Traits-of-intERest), a method that bypasses segmentation and phenotyping. It uses Topological Data Analysis (TDA) to characterize the geometric structure of protein expression directly from raw image data.

Core Concept: Topological Event History

TOASTER treats protein intensity as a random field. It uses a filtration process (iteratively thresholding intensity levels) to track the "birth" of topological features:

• Degree-0 Homology: Connected components (clusters of high intensity).
• Degree-1 Homology: Loops (rings of high intensity surrounding low intensity).

The method summarizes these births using an adaptation of the Nelson-Aalen cumulative hazard function, termed the "Topological Event History" ($\hat{A}(t)$).

• Univariate Case: Tracks the birth of local minima in a single protein's intensity field.
• Bivariate Case: Tracks the birth of shared local minima where two proteins simultaneously exhibit low intensity (or high intensity, depending on the filtration direction) at the same spatial location. This captures colocalization or interaction between two markers.
• Robustness: The algorithm naturally handles missing data (holes/tears) by adjusting the "at-risk" set of pixels, excluding missing neighbors from the calculation.

Association Testing Strategies

Once the topological event history curves are generated for each patient, TOASTER tests their association with clinical outcomes (binary, continuous, or censored survival) using three distinct approaches:

1. Functional Data Analysis (FPCA): Decomposes the curves into functional principal components (FPCs) and uses them as covariates in a Cox proportional hazards model (for survival) or logistic regression (for binary outcomes).
2. Gridwise Testing: Evaluates the association at discrete time points (quantiles) along the curve. P-values from each grid point are combined using the Cauchy combination test, which is robust to dependent p-values.
3. Kernel Association Testing: Constructs a distance matrix between patients based on the Euclidean distance of their curves. This is converted into a kernel matrix and tested against the outcome using a score test based on residuals from a null model.

3. Key Contributions

• Segmentation-Free Analysis: TOASTER is the first method to directly test associations between continuous spatial protein expression and clinical outcomes, eliminating reliance on error-prone cell segmentation.
• Topological Summarization: It introduces the "Topological Event History" (Nelson-Aalen adaptation) as a robust summary statistic for spatial proteomics, capturing both connected components and loops.
• Handling Artifacts: The method is inherently robust to tissue tears and gaps, a common issue in spatial proteomics that often invalidates cell-based approaches.
• Flexible Framework: It offers three distinct statistical testing frameworks (Functional, Gridwise, Kernel) to accommodate different data structures and outcome types.
• Software Implementation: The authors provide an open-source R package (TOASTER) for implementation.

4. Results

Simulation Studies

The authors benchmarked TOASTER against existing methods (DenVar for univariate, DIMPLE for bivariate) using Gaussian random fields.

• Power: TOASTER demonstrated significantly higher power (approx. 0.88–0.91) compared to DenVar (0.49) and slightly higher power than DIMPLE in various scenarios, particularly when proteins were simulated independently but from different parametric models.
• Type I Error: TOASTER successfully controlled Type I error at the nominal 0.05 level across all three testing strategies. In contrast, DenVar was found to be overly conservative.
• Robustness: In simulations with induced "holes" (missing data), TOASTER maintained high power and error control, whereas competitors struggled.

Real-World Application: Triple-Negative Breast Cancer (NeoTRIP Study)

TOASTER was applied to Imaging Mass Cytometry (IMC) data from 106 patients with triple-negative breast cancer treated with combination chemotherapy and immunotherapy.

• Goal: Determine if the topological structure of immune markers (CD3, CD4, CD8, CD20) predicted Pathologic Complete Response (pCR) vs. Recurrent Disease (RD).
• Findings:
  • Patients achieving pCR exhibited fewer births of connected components during filtration compared to those with RD, suggesting protein expression was more concentrated in clusters in responders.
  • Strongest Associations: The joint expression of CD3+CD4 and CD3+CD20 showed the most distinct topological differences between responders and non-responders (p-values < 0.005 across models).
  • Method Comparison: Different testing strategies yielded varying significance levels for specific markers (e.g., Gridwise and Kernel tests were significant for CD3, while Functional was not), highlighting the value of using multiple approaches to capture different aspects of the data.

5. Significance

This paper represents a paradigm shift in spatial proteomics analysis. By moving away from cell-centric models to continuous field-based topological analysis, TOASTER addresses critical limitations in current workflows:

1. Accuracy: It avoids the propagation of segmentation errors.
2. Completeness: It utilizes the full spatial signal, including intercellular regions.
3. Clinical Utility: It successfully identified topological signatures of immunotherapy response in breast cancer that might be missed by traditional cell-counting methods.

The work establishes a new statistical framework for spatial biology, demonstrating that the geometric arrangement of protein expression (topology) contains predictive information about patient outcomes that is independent of, and potentially superior to, discrete cell phenotyping.