CellSpecks: A Software for Automated Detection and Analysis of Calcium

Transcription

1 Biophysical Journal, Volume 115 Supplemental Information CellSpecks: A Software for Automated Detection and Analysis of Calcium Channels in Live Cells Syed Islamuddin Shah, Martin Smith, Divya Swaminathan, Ian Parker, Ghanim Ullah, and Angelo Demuro

2 Supplementary Information Text 1 CellSpecks: A Software for Automated Detection and Analysis of Calcium Channels in Live Cells Syed Islamuddin Shah, Martin Smith, Divya Swaminathan, Ian Parker, Ghanim Ullah, and Angelo Demuro A Comment on the Algorithm of CellSpecks For automated detection and analysis of hundreds, sometimes thousands of functional Ca 2+ channels imaged simultaneously at the single channel and millisecond resolution, we developed a Java-based fluorescent detection software called CellSpecks. Our goal is to ease the handling of a large data set, routinely generated during TIRFM, and facilitate the detection of channels with low open probability and brief open time, otherwise challenging and time consuming when done manually. CellSpecks opens files by using either the Bio-Formats library for Metamorph stack files (.stk movie file) or.tiff sequential image files. During the opening of the image files, the onedimensional data array produced by these libraries for each frame is converted into a twodimensional array and the resulting 2D frames are stored as a stack object in memory called Original stack. After cloning (copying) the Original stack, the program generates Blur (smooth), Noise, and Signal stacks (after performing noise-subtraction) as described in the main text. For each of these tasks, CellSpecks analyzes each pixel independently using its intensity values over time employing multithreading capability of java to speed up processing of the multi gigabyte stacks. We generate as many threads as there are pixels in a movie frame and the time trace of each pixel is processed by the thread allocated to it in order to perform various tasks including noise detection, signal isolation, and event detection and attribution. Three primary methods for managing efficient thread pooling were examined both experimentally in the context of the program and in a simulated benchmarking environment. We found that the CPU bound performance differences between the three methods (operation nesting, semaphore thread-count mediation, and JVM thread-count mediation) are less important than ensuring efficient RAM usage and using the solution that most resembles the problem. Channel Locations

3 After constructing a list of events, CellSpecks performs a series of steps to generate a list of channel locations. In order to locate channels, as a first pass, a two-dimensional image g(x, y) is generated after summing the intensity values of the Signal stack over time for each pixel and normalizing by the highest intensity value recorded over the entire stack. The local maxima of a pixel (x,y) with respect to a 2x2 neighboring grid in g(x, y) is interpreted as likely channel location and added to a list of temporary channel objects. At this stage channel objects have probable channel locations but no events associated with them. The map g(x, y ) is biased towards high fluorescence as some channel openings are brief or have small amplitudes, and it is possible that g(x, y) does not capture all channel locations. These remaining channels are included as explained below. We next use the Eventlist to dynamically link events to channel locations and using event position and intensity attributes revised and add channel locations to the channel object. This process takes into account that as the coordinates of a channel are dynamically generated from their constituent events, events early in the stack can determine location of channels to which nearby larger, later events may be assigned. The first part of this channel assignment procedure is to determine the nearest neighboring channel for each event. Events are linked to channels one by one, either to the closest known channel, if there exists one with which it overlaps, or to a new channel location which is then added to the channel list. Whether or not the event is added to the existing channels list depends on two criteria: The first criteria consists of a binary overlap verification that checks whether or not the coordinates of the event and channel in consideration when rounded to the nearest pixel in space are equal. Event locations are calculated from their constituent EventPartStack using a mean of the EventPartStack coordinates weighted by intensity, as given in Equation 1. e X = 2 ' % & ( & &40 & )*+ ( ( & ) & Є [0, 2] 2 ' & ( & &40)*+ (( & ) & Є [0, 2], (1) where X is the x-location of the event and x 6 is x-coordinate of the i th pixel of event e. I 6 and p 6 are the relative (obtained by sum of intensities over time divided by global maximum) and absolute intensities of the pixels in the event and P is the number of EventPartStacks for the event. The X coordinate for the location of a channel (C) is similarly calculated using a mean of the coordinates of the channel s constituent events weighted by each event s total intensity, as in Equation 2.

4 C X = ; < &40(: & (%) & (= & (>)) ) )*+ = & > & Є [0, ;] ; < & (= & (>)) &40 )*+ = & + & Є [0, ;]. (2) Where E is the number of events in the channel and e i (x) is calculated from equation 1 for i th event and v i (e i (x)) represents the total intensity of the i th event calculated as the sum of the intensity measured for all the pixels (P) constituting event e i in space and time, i.e. A v 6 e 6 ). (3) Equations equivalent to 1 and 2 are used to calculate the y coordinates e(y) of each event and C(y) of each channel respectively. The second criteria for event-channel association takes into account the relative intensity of event e at channel location C calculated as I D e(x, t), e(y, t), C(x), C(y) = OPQ ( IJKL= M4K(=N IJKL= 2 &40 G & (:,H) G & :,H M4K(=N & Є 0, 2 ), (4) where P is the total number of EventPartStacks that constitute event e, the numerator sums intensity of all EventPartStacks over the duration of an event and the denominator is the maximum summed intensity of a constituent EventPartStack. This relative intensity calculation is used to determine an overlap likelihood using the following equation: o e x, e y, C x, C(y) = G T :(%,U),:(V,U),H(%),H(V) H(%)W:(%) X H(V)W:(V). (5) If o e x, e y, C x, C(y) > 0.1, then event is added to channel C otherwise a new channel with current event is added in the channels list. This second approach is an arbitrary but consistent threshold: the relative intensity of an event at the center of the channel multiplied by the inverse of distance between the event and channel centers must be greater than 0.1. As the distance between the center of the channel and the center of the event must be at most 2, (opposite corners of the same pixel), this threshold limits the intensity of the event at the location of the channel center to at least ( a.b ) (or 14.1%) of the maximum intensity in that event. The b required relative intensity can be less than 14.1% if the distance between the channel and event centers are less than 2. But these centers are determined by an intensity weighted coordinate calculation anyway, so it is extremely unlikely that a relative intensity of less than 14.1% will occur, given that these coordinates are calculated at sub-pixel resolutions.

5 Finally, channels containing no events (those channels added using the weight array but found to not accurately represent real channel locations) are removed. At the end of this process, channel locations are adjusted (using equation 4) for the actual events they contain and any channels found to be within the same pixel are combined. Synthetic data set To validate the accuracy of the algorithm, we generated a synthetic data set of 50 channels randomly distributed on a grid of 128 x 128 pixels for a total duration of 2 seconds in a sequence of 1000 tiff images. We deliberately randomized the opening and closing of these channels as well as their open and close dwell-times to mimic experimental conditions and to test the robustness of CellSpecks. First, we randomly select the number of open events for each channel between 5 and 15. Second, open times for these events are randomly generated anywhere between 11 and 50 frames. To keep channel flux tractable, fluorescence of zero (intensity units) represents a closed channel whereas an open channel was assigned a fluorescence of 200 (in line with the frequently observed intensity values when the channel is open). To make the channel signal realistic, noise derived from normal distribution was superimposed on each channel s time-trace as well as all other pixels in the image frame. The mean and standard deviation was adjusted to give a signal-to-noise ratio ranging from 5 to 40. Also in line with experimental observations, fluorescence from an open channel s location is allowed to spread to the nearest neighbor pixels. The code generating this data is available from authors upon request.

6 Flowchart 1: Noise detection and subtraction algorithm.

7 Signal Stack - SS Initialize EventPartStack (for each pixel in space and time of SS) Link EventPartStack elements to non-zero intensity pixels (left, right, up, down, previous, and next links) Generate Eventlist by adding to it isolated blocks of EventPartStack (screen for small events with intensity < 50 or duration < 10 frames) Event location (e(x), e(y)) calculation (Eq. 1) Channel List with associated events Flowchart 2: Event detection and association of events with channels algorithm.