Topologically-based parameter inference for agent-based model selection from spatiotemporal cellular data

This paper introduces TOPAZ, a novel computational pipeline that integrates topological data analysis with Bayesian inference methods to effectively infer parameters and select the most biologically plausible agent-based models from spatiotemporal single-cell imaging data.

The Gist

Imagine you are watching a massive, chaotic dance floor filled with thousands of people (cells). From a bird's-eye view, you see them swirling, clustering, and moving in streams. You want to know: What are the invisible rules they are following to create this beautiful chaos?

Are they just bumping into each other and bouncing away? Or are they also holding hands and trying to face the same direction as their neighbors?

This is the challenge scientists face when studying cell movement. They have amazing cameras that can track every single cell, but turning that video into a clear set of "rules" is incredibly hard.

This paper introduces a new tool called TOPAZ (which stands for TOpologically-based Parameter inference for Agent-based model optimiZation). Think of TOPAZ as a super-smart detective that helps figure out the hidden rules of the cellular dance floor.

Here is how it works, broken down into simple steps:

1. The Problem: Too Much Data, Too Many Guesses

Scientists have two main tools, but both have flaws:

• The "Agent-Based Model" (ABM): This is like a video game simulator. You write down rules (e.g., "if a cell gets too close to another, it pushes away") and watch the simulation run. The problem? There are so many possible rules and settings that it's like trying to find a specific needle in a haystack by guessing randomly.
• Topological Data Analysis (TDA): This is a way of looking at the shape of the data. Instead of looking at individual cells, it looks at the "holes" and "loops" in the crowd. It's like looking at a cloud and saying, "That looks like a bunny," without worrying about the individual water droplets. It's great at describing what the shape is, but it doesn't tell you why it formed that way.

2. The Solution: TOPAZ (The Detective)

TOPAZ combines these two tools into a single pipeline. It acts like a detective who uses a special magnifying glass (TDA) to compare the real crime scene (the actual cell video) with different suspect theories (the computer simulations).

Here is the step-by-step process of the investigation:

• Step 1: Taking a "Topological Snapshot"
  Imagine taking a photo of the dance floor and turning it into a wireframe map. This map highlights the shapes: where are the big empty circles? Where are the tight clusters? TOPAZ creates a "Crocker Plot," which is essentially a heatmap showing how these shapes change over time. It's like turning a complex movie into a simple, color-coded scorecard.

• Step 2: The "Guess and Check" (ABC)
  The detective runs thousands of simulations with different rules.

  • Scenario A: Maybe the cells just push and pull.
  • Scenario B: Maybe they push, pull, and try to align their direction like a school of fish.
    TOPAZ compares the "scorecard" from the real video with the scorecards from the simulations. If a simulation's scorecard looks almost identical to the real one, that simulation gets a "thumbs up."

• Step 3: The "Simplicity Test" (Model Selection)
  This is the most clever part. Imagine you have two theories:

  1. Theory A: Cells push and pull. (Simple, 2 rules).
  2. Theory B: Cells push, pull, and align. (Complex, 3 rules).

  If Theory B only explains the data slightly better than Theory A, TOPAZ says, "No thanks, that extra rule isn't worth the complexity." It uses a mathematical rule called BIC (Bayesian Information Criterion) to penalize overly complicated theories. It only picks the complex theory if the data demands it.

3. The Experiment: The Fibroblast Dance

The authors tested TOPAZ using data from fibroblasts (a type of cell involved in wound healing).

• They created two versions of a simulation: one where cells just bumped into each other, and one where they also tried to align their direction (like a flock of birds).
• They fed the "alignment" simulation data into TOPAZ and asked, "Which rules created this?"
• The Result: TOPAZ correctly identified that the "alignment" rule was necessary. It realized that the simple "bump and push" model couldn't explain the specific shapes and streams seen in the data.

The Big Picture Takeaway

Think of TOPAZ as a translator.

• On one side, you have biologists with high-tech cameras showing complex cell movements.
• On the other side, you have mathematicians with abstract equations trying to describe those movements.

TOPAZ translates the visual "shape" of the cell movement into a specific set of mathematical rules. It tells us not just that the cells are moving in a certain way, but exactly which biological interactions (like pushing, pulling, or aligning) are responsible for it.

Why does this matter?
If we can figure out the rules of how cells move, we can better understand how wounds heal, how tumors spread, or how tissues form. TOPAZ gives scientists a reliable way to test their theories against real-world data, ensuring they aren't just making up complicated stories that don't actually fit the facts.

Technical Summary

1. Problem Statement

The paper addresses a critical challenge in computational biology: translating high-resolution, spatiotemporal single-cell imaging data into mechanistic insights about cell population dynamics.

• The Gap: While Agent-Based Models (ABMs) offer a flexible, bottom-up framework for simulating emergent collective behaviors (e.g., cell migration, swarming), they are often difficult to calibrate, sensitive to parameter choices, and hard to validate against complex data. Conversely, Topological Data Analysis (TDA) provides robust, scale-invariant descriptors of spatial organization (e.g., connectivity, holes, boundaries) but lacks inherent mechanistic interpretability.
• The Specific Challenge: Existing methods often use TDA for descriptive insights or parameter inference within a single model. There is a lack of unified pipelines that can rigorously select between competing mechanistic hypotheses (e.g., does a specific interaction rule like "alignment" actually exist in the data?) using topological summaries.
• Goal: Develop a computational pipeline that integrates TDA, Approximate Bayesian Computation (ABC), and Bayesian Model Selection to infer biologically plausible parameters and distinguish between competing ABMs based on spatiotemporal cellular data.

2. Methodology: The TOPAZ Pipeline

The authors introduce TOPAZ (TOpologically-based Parameter inference for Agent-based model optimiZation), a five-step computational framework:

1. Topological Feature Extraction (TDA):

   • Input: Spatiotemporal point clouds of cell trajectories $(x, y, \theta)$, where $\theta$ is the cell orientation.
   • Technique: Persistent Homology (PH) is computed using the Vietoris-Rips filtration.
   • Output: Crocker plots (Contour Realization Of Computed K-dimensional hole Evolution in the Rips complex). These are matrices summarizing the evolution of Betti numbers (specifically $\beta_0$ for connected components and $\beta_1$ for loops) over time and proximity scales. These serve as the summary statistics.

2. Dimensionality Reduction (Optional but Recommended):

   • Techniques like t-SNE or PCA are used to visualize the high-dimensional Crocker matrices, aiding in the identification of distinct topological signatures associated with different parameter sets.

3. Parameter Inference (ABC & AABC):

   • Approximate Bayesian Computation (ABC): A rejection-sampling algorithm is used. Simulations are run with parameters drawn from a prior distribution. The resulting Crocker matrices are compared to the "observed" (ground truth) data using a distance metric (Sum of Squared Errors, SSE). Samples with SSE below a tolerance threshold are accepted to form the posterior distribution.
   • Approximate Approximate Bayesian Computation (AABC): To reduce computational cost, AABC generates additional samples by interpolating between existing simulation results using kernel density estimation (Epanechnikov kernel) on the summary statistic space, avoiding the need for new, expensive forward simulations.

4. Distributional Assessment:

   • Before model selection, statistical tests (PERMANOVA, Energy Distance, and Maximum Mean Discrepancy) are applied to ensure that the posterior predictive distributions of the summary statistics are statistically distinguishable across candidate models.

5. Model Selection (BIC):

   • The Bayesian Information Criterion (BIC) is calculated for each candidate model.
   • Formula: $BIC = -2 \ln(L) + k \ln(n)$, where $L$ is the likelihood (derived from the SSE of the top 1% of ABC samples), $k$ is the number of parameters, and $n$ is the number of data points.
   • The model with the lowest BIC score is selected as the optimal model, balancing goodness-of-fit with model complexity (penalizing unnecessary parameters).

3. Key Contributions

• Novel Pipeline (TOPAZ): The first framework to explicitly couple TDA-derived summary statistics with ABC/AABC for model selection (distinguishing between hypotheses) rather than just parameter inference.
• Mechanistic Extension of D'Orsogna Model: The authors extend the standard D'Orsogna ABM (Model$_{DO}$), which includes attraction and repulsion forces, by adding a directional alignment interaction (Model$_{AL}$). This creates a testbed for distinguishing between simple repulsion/attraction dynamics and those involving alignment (fluidization).
• Integration of Topology and Statistics: Demonstrates how topological invariants (Betti numbers) can serve as effective summary statistics for likelihood-free inference in complex, stochastic biological systems.
• Open-Source Implementation: The authors provide a publicly available GitHub repository with the full code, simulation scripts, and notebooks.

4. Results

The framework was validated using synthetic data generated from collective fibroblast movement simulations.

• Parameter Recovery: TOPAZ successfully recovered ground-truth parameters for both Model$_{DO}$ (2 parameters: relative attraction strength $C$ and range $L$) and Model$_{AL}$ (3 parameters: $C$, $L$, and alignment strength $W$). The median of the AABC posterior distributions closely matched the true values.
• Model Discrimination:
  • When data was generated from Model$_{AL}$ (with alignment), the pipeline correctly selected Model$_{AL}$ as the optimal model (lowest BIC) over Model$_{DO}$.
  • When data was generated from Model$_{DO}$ (no alignment), the pipeline correctly selected Model$_{DO}$, rejecting the more complex Model$_{AL}$ despite the extra parameter.
• Visual Validation: t-SNE visualizations of the Crocker plots clearly separated the topological signatures of aligned vs. non-aligned cell populations, confirming that the topological features were sensitive to the mechanistic difference.
• Robustness: The use of PERMANOVA and Energy Distance confirmed that the posterior predictive distributions were statistically distinct between the two models, validating the BIC selection.

5. Significance and Future Directions

• Mechanistic Insight: TOPAZ provides a rigorous, data-driven method to determine which biological mechanisms (e.g., alignment vs. no alignment) are necessary to explain observed cell behaviors, moving beyond descriptive modeling.
• Scalability: The framework is designed to be extensible. While currently applied to cell tracking, it can incorporate spatial transcriptomics or proteomics data to link intracellular signaling states with population-level dynamics.
• Handling Noise: The use of TDA (which is robust to noise) combined with Bayesian model selection offers a promising approach for analyzing noisy, high-dimensional experimental data where traditional likelihood-based methods fail.
• Future Work: The authors plan to apply TOPAZ to real experimental data, incorporate more complex interaction rules, and explore more data-efficient inference algorithms (e.g., Sequential Monte Carlo) for higher-dimensional parameter spaces.

In summary, this paper presents a powerful, extensible tool for computational biologists to move from "simulating what happens" to "inferring why it happens" by rigorously testing competing mechanistic hypotheses against spatiotemporal single-cell data.