Abstract
Single-particle tracking (SPT) allows researchers to observe the motion of individual molecules within living cells. However, when observing long trajectories (exceeding 100 points), particles often switch their mode of motion over time, making global model fitting misleading. This paper proposes a non-parametric sequential algorithm to detect temporal "change points"—the moments when a particle transitions between Brownian motion, subdiffusion, and superdiffusion. The method utilizes a local window statistic that remains independent of specific mathematical models, effectively controlling false detections in purely Brownian paths. Through Monte Carlo studies and application to mRNA complex (mRNP) trajectories in neuronal dendrites, the authors demonstrate that this approach offers superior computational speed and reliability compared to existing parametric methods.
Introduction
The interior of a living cell is a crowded, dynamic environment where molecules rarely follow a single physical law for the duration of their journey. A protein or mRNA complex might drift aimlessly via Brownian motion, become confined in a dense neighborhood (subdiffusion), or be actively transported along microtubule tracks by molecular motors (superdiffusion).
While Single-Particle Tracking (SPT) provides the raw data for these paths, traditional analysis tools often assume a stationary mode of motion. If a particle switches dynamics mid-trajectory, applying a single model to the entire path results in inaccurate physical parameters. To address this, researchers need robust methods to identify "change points"—the precise timestamps where the underlying physics of motion shifts.
Methodology
The Sequential Sliding Window
The proposed algorithm employs a "sliding window" strategy to analyze trajectories locally. For any given point in a trajectory, the method compares a "backward" window (the path immediately preceding the point) and a "forward" window (the path immediately following it).
- The Test Statistic ($T_n$): For each window of size $k$, the algorithm calculates a statistic based on the maximum displacement of the particle from its starting position, normalized by an estimate of the local diffusion coefficient.
- Three-Decision Classification: Based on the value of $T_n$, the motion within the window is categorized as subdiffusive (low displacement), superdiffusive (high displacement), or Brownian (intermediate displacement).
- Change Point Identification: A potential change point is flagged whenever the classification of the "forward" window differs from the "backward" window.
- Clustering and Refinement: Because stochastic noise can cause multiple consecutive points to be flagged, the algorithm clusters these detections and identifies the point of maximum difference within the cluster as the final estimated change point.
Non-Parametric Advantage
Unlike many existing tools, this method is non-parametric. It does not require the user to specify a particular model for subdiffusion (such as the Ornstein-Uhlenbeck process). Instead, it relies on the fundamental scaling properties of the displacement, making it more versatile across different biological contexts.
Results and Performance
Computational Efficiency
The most significant advantage of this sequential approach is its speed. In benchmarking tests, the proposed algorithm processed trajectories in less than 0.05 seconds. In comparison, state-of-the-art parametric methods using Maximum A Posteriori (MAP) or Markov Chain Monte Carlo (MCMC) estimations required between 11 and 115 seconds per trajectory. This massive reduction in overhead makes the algorithm suitable for high-throughput analysis of the thousands of trajectories typically generated in modern SPT experiments.
Accuracy in Low-Signal Regimes
The algorithm was tested against simulated scenarios involving transitions between Brownian motion and directed drift. It proved particularly effective in "low-signal" regimes—where the drift velocity is small ($v \le 0.8$). In these cases, traditional parametric methods often fail to detect any change in motion, whereas the non-parametric window statistic successfully identifies the transitions.
Biological Application: mRNP Transport
The utility of the method was demonstrated on real-world data involving messenger RNA complexes (mRNPs) in neuronal dendrites. The algorithm successfully segmented these trajectories, identifying transitions to active transport (superdiffusion). Notably, it also identified subdiffusive segments—periods where the mRNP was likely "trapped" or confined—that previous analyses had incorrectly categorized as simple Brownian motion.
Discussion and Future Work
While the algorithm is highly efficient, its performance is sensitive to the choice of the window size ($k$). A large window provides better statistical power for detecting subtle changes but may fail to resolve two change points occurring in rapid succession. Conversely, a small window offers high temporal resolution but is more susceptible to stochastic noise.
Future developments aim to implement a multi-scale approach. By automatically selecting or combining multiple window sizes, the algorithm could potentially capture both rapid transitions and gradual shifts in dynamics without manual parameter tuning. This would further solidify the tool's role in the automated, large-scale analysis of intracellular transport.
Summary updated