Cega: a single particle segmentation algorithm to identify moving particles in a noisy system

Abstract

Improvements to particle tracking algorithms are required to effectively analyze the motility of biological molecules in complex or noisy systems. A typical single particle tracking (SPT) algorithm detects particle coordinates for trajectory assembly. However, particle detection filters fail for data sets with low signal-to-noise levels. When tracking molecular motors in complex systems, standard techniques often fail to separate the fluorescent signatures of moving particles from background signal. We developed an approach to analyze the motility of kinesin motor proteins moving along the microtubule cytoskeleton of extracted neurons using the Kullback-Leibler divergence to identify regions where there are significant differences between models of moving particles and background signal. We tested our software on both simulated and experimental data and found a noticeable improvement in SPT capability and a higher identification rate of motors as compared with current methods. This algorithm, called Cega, for "find the object," produces data amenable to conventional blob detection techniques that can then be used to obtain coordinates for downstream SPT processing. We anticipate that this algorithm will be useful for those interested in tracking moving particles in complex in vitro or in vivo environments.

FIGURE 1:

Cega workflow. (A) Example time sequence taken from image sequence of kinesin motors moving within an axonal compartment, as well as corresponding tracks calculated after Cega detection. (B) Diagram of Cega steps leading to spot detection. Font between each step indicates values used to process axonal and dendritic data. (C) Image sequence of time points in A after processing with each of Cega’s steps. Image stills represent how data are manipulated at each step. Green arrows indicate moving spots, while red arrows indicate positions of stationary spots. Dim-colored spots in the KL divergence images represent locations of spot detection. Spots in the LoG images represent detected spots, colored by appearance over time. Spots corresponding to the same track are connected with the same colored lines.

FIGURE 2:

Analysis of Cega spot detection. (A, B) Jaccard index values calculated for simulated spots detected with Cega, median or minimum background subtraction methods, or standard methods, as the SNR increases. (C, D) Recall rate for detection methods as SNR increases. Simulations were run 100 times and resulted in a SD of <0.0045.

FIGURE 3:

Analysis of tracking after Cega or median background subtraction methods. Simulated data using axonal background signal was used where mean photon emissions were set to 200 photons, which corresponds to a SNR of 2.8. (A, B) Kymographs of tracks determined from simulated particles within axonal and dendritic compartments. Although the same number of particles were simulated in the axonal and dendritic movies, the dendritic movie was smaller in size, resulting in a higher density of particles. In the merge kymograph and zoom-in area, cyan indicates locations where only simulated tracks and Cega detected tracks overlap, whereas yellow indicates where only simulated tracks and median background-subtracted detected tracks overlap. Magenta tracks are where only Cega and median background-subtracted tracks overlap. Jaccard indices and recall rates for tracks determined after Cega filtering and median background subtraction are listed below each corresponding kymograph.

FIGURE 4:

Analysis of Cega performance on original data set. (A, B) Kymographs of tracks determined from kinesin motors moving on microtubule arrays within axonal and dendritic compartments. In the merge kymograph, yellow areas indicate where Cega and median background-subtracted detected tracks overlap.