TL;DR
This paper presents a novel probabilistic framework and computational approach to reconstruct the full 3D trajectories of particles moving on a closed surface (like a bacterial cell wall) when only a small, censored portion of that surface is observed. This is crucial for understanding biological processes where molecules can enter and exit the field of view, enabling more accurate analysis of dynamics and spatial organization.
What Problem Does This Paper Address?
Many biological processes involve molecules moving on curved surfaces, such as proteins on the membrane of a bacterial cell. Observing these dynamics often relies on microscopy techniques that only capture a limited region of the cell (a "region of observation" or ROO). This means particles can disappear from view and then reappear later, making it challenging to track their complete trajectories and understand their overall movement patterns. Existing methods struggle with these "re-entrance" events and large unobserved regions. This paper specifically addresses the problem of connecting partial 2D observations of particle trajectories on a cylindrical surface (like a rod-shaped bacterium) to reconstruct their full 3D paths, accounting for particles entering and exiting the observed area.
Methods
The core of the paper is a probabilistic framework that models particle movement as Brownian motion with drift on a cylindrical surface. The key innovation lies in how they handle the censored observations. They formulate the problem as finding the most likely configuration of connections between "output" tracklets (particles leaving the observed region) and "input" tracklets (particles entering the observed region). This is achieved by:
- Probabilistic Modeling: Using Poisson processes for particle birth and death events, and Brownian motion with drift for particle movement.
- Likelihood Maximization: Defining a likelihood function that quantifies the probability of a given configuration of connections.
- Integer Linear Programming: Transforming the optimization problem of maximizing the likelihood into a solvable integer linear programming problem.
- Parameter Estimation: Developing consistent estimators for key parameters like birth rate ($\lambda$), death rate ($\tau_d$), drift velocity ($\mathbf{v}$), and diffusion variance ($\Sigma$) from the observed data.
This approach is data-driven and aims to avoid hidden parameters.
Datasets
The paper primarily uses simulated datasets to evaluate their method. They generate synthetic trajectories mimicking MreB protein dynamics on a cylindrical surface, varying parameters like the ratio of observed to unobserved regions, drift, diffusion, and birth/death rates. They also mention testing their method on a real TIRFM dataset of MreB aggregates in Bacillus subtilis (though detailed results for this are in an appendix).
Key Insights and Ideas
- Bridging the Gap: The most significant insight is the development of a principled probabilistic framework to connect fragmented trajectories across unobserved regions, effectively "filling in the blanks" of particle movement.
- Re-entrance Events are Key: Explicitly addressing how particles can leave and re-enter the field of view is critical for accurate trajectory reconstruction on closed surfaces.
- Linear Programming for Connections: Formulating the complex connection problem as a linear optimization problem makes it computationally tractable.
- Parameter Estimation is Crucial: The accuracy of the trajectory reconstruction heavily relies on accurately estimating the underlying dynamics parameters (drift, diffusion, birth/death rates).
Key Results
- The proposed method demonstrates good performance in reconstructing particle trajectories on a cylindrical surface, even with a significant portion of the surface unobserved.
- The Adjusted Rand Index (ARI) is used to quantify the accuracy of the connection results, showing high ARI values (often > 0.7, and > 0.9 in some cases) when using estimated parameters.
- The method is shown to be robust to variations in drift, diffusion, and particle density.
- The parameter estimation procedures for birth rate ($\tau_\alpha$) and death rate ($\tau_d$) are shown to be unbiased and perform reasonably well, even with shorter observation times (e.g., 5-minute movies).
- The study provides insights into the number of rotations particles make around the cylinder, showing that the distributions of estimated and theoretical rotations are similar.
Limitations and Future Work
- Homogeneity Assumption: The current model assumes homogeneity in time and space for birth and death rates, and constant speed and noise. The authors acknowledge that real biological systems might deviate from this.
- Age Dependency: The assumption of a memoryless lifetime (exponential distribution) is a simplification. Future work could explore how death rates might depend on particle "age," though this is complicated by partial observation.
- Spatial Dependence: While the current model assumes constant death rates, future work could investigate how these rates might depend on position on the surface.
- Experimental Design Guidance: The simulation study is suggested to be useful for designing future experiments, particularly in determining optimal observation times ($T_s$).
- Real Data Application: The authors plan to investigate more real TIRFM datasets and explore more complex scenarios, such as particles born or dying non-uniformly on the membrane.
Summary updated