Simulating neural spike trains

Simulating neural spike trains

I recently taught a 3 hour class as part of the Quantitative Biology Workshop at MIT.  I spent the first couple of hours teaching students (undergraduates) how to simulate neural spike trains given the firing rate of a neuron.  The goal of the first part of the class was to implement a Poisson spike generator and plot spike trains as a raster plot.  Students told me they really liked the class, so I thought I would share this with everyone.

The goal of this tutorial is to understand how neurons encode a stimulus.  We will simulate spike trains using MATLAB and visualize spiking activity by making raster plots.  Here is an example.

02_Raster_BaselineAndStimUnderstanding the raster plot. In neuroscience, the words firing and spiking commonly refer to action potentials generated by a neuron.  We can monitor neural activity, and therefore spikes, by placing electrodes close to a neuron in the brain.  Neuroscientists commonly place electrodes in live animals (such as mice / rats / monkeys) to monitor neural activity in region(s) of the brain that are of interest to them while the animal is performing a task, e.g. dots moving on a screen are commonly used to study how the brain processes movement.  In the above plot (known as a raster plot), each black bar represents one spike or action potential and each row of black bars (or a spike train) represents spiking activity of a neuron over a period of 1.5 s.  This plot is a simulation of spiking activity of a neuron in an area MT (that encodes motion information) of a monkey.  This neuron has a baseline firing rate of 6 Hz (meaning, the neuron fires 6 action potentials per second on average, when there is no stimulus) and increases its firing rate to 30 Hz when the monkey sees a stimulus moving at 45 degrees from horizontal (now the neuron fires 30 spikes per second on average!).  Simulating spike trains like the ones in the raster plot above requires only one piece of information: the firing rate of the neuron.

Neural encoding.  In this tutorial, we are interested in transforming information from the real world into spike trains.  This transformation is call neural encoding.  In other words, the term neural encoding refers to representing information from the physical world (such as direction of a moving stimulus) in the activity of a neuron (such as its firing rate).  Encoding information in the firing rate of a neuron is commonly called rate coding.

One abstraction that I have found extremely useful while thinking about how the brain encodes information is to think of two spaces – the physical space and neural space.  The physical space consists of physical properties of objects, such as direction, shape, speed, color, loudness, etc.  Neural space consists of properties of a neuron / group of neurons / dendrites (depending on the context) such as firing rate (most commonly encountered), morphology, receptor density, number of spines, local field potential amplitude, etc.

Neural encoding is merely a transformation from physical space to neural space.  For example, the neuron in our plot above represents a quantity from physical space (direction of a moving stimulus) as a quantity in neural space (firing rate).  To understand the encoding of this neuron, we use the concept a tuning curve.

03_Neuron3_tuningCurveA tuning curve is merely a graphical representation of neural encoding.  On the X axis, we represent the quantity from physical space and on the Y axis we represent the quantity from neural space.  In this example, we plot the neural quantity firing rate (Y axis) as a function of the physical quantity direction of motion (X axis), and each neuron will have its own tuning curve.  Neuroscientists determine the tuning curve of a neuron by presenting the animal with various stimuli (varying along the X axis) while recording the activity of that neuron.

If you recall, we can simulate spike trains merely by knowing the rate at which a neuron should fire.  So how does a tuning curve relate to simulating spike trains?  Tuning curves translate physical quantities (direction of motion) into firing rate.  Continuing our example, say we were asked to simulate a spike train for an MT neuron (whose tuning curve is given above) to a stimulus moving at a direction of 90 degrees.  From the tuning curve, we see that this neuron will fire at about 45 Hz to the given stimulus.  Now, we know the firing rate, and we can simulate spike trains in response to the given stimulus.

Hopefully, by now, you have understood the motivation behind simulating spike trains.  Now let’s take a moment to reflect on how to simulate spike trains.  We will then worry about how to implement that concept in MATLAB.

Inter spike interval (ISI).  The term ISI refers to the time interval between two spikes.  For a neuron firing at 8 Hz, the average time interval between two spikes is 125 ms.  If we draw one spike at every 125 ms, have we simulated the activity of a neuron?  In theory, yes, but the firing pattern of this neuron does not resemble that of a real neuron.  When we look at firing from a real neuron and measure the ISIs, they look something like this.

ISI_pictureWe can use a concept in probability theory known as the Poisson process to simulate spike trains that have characteristics close to real neurons. In your own time, look up Poisson process and Poisson spike generator. For now, we will use a simple formula (derived from the math that we skipped just now) to generate spikes. Here is an excerpt from Dayan and Abbott, p30 explaining how to simulate spikes.

Spike sequences can be simulated by using some estimate of the firing rate, fr, predicted from knowledge of the stimulus, to drive a Poisson process. A simple procedure for generating spikes in a computer program is based on the fact that the estimated probability of firing a spike during a short interval of duration dt is fr*dt. The program progresses through time in small steps of size dt and generates, at each time step, a random number x chosen uniformly in the range between 0 and 1. If x < fr*dt at that time step, a spike is fired; otherwise it is not.

In this technique, we are marching through time, in small steps of size dt (typically set to 1 ms) and deciding if there should be a spike at this time point.  An alternative (and quicker) technique is to decide when the next spike is going to be and I for this, I recommend you read this excellent post by Jeff Preshing.  In fact, I implemented this technique in the PulsePod.  For the purposes of this MATLAB exercise, I will stick to the first technique.

Spikes in MATLAB!

Finally, we can get down to business.  Before you open MATLAB and start typing, please make sure you recall basic MATLAB concepts such as the difference between a command window and a the editor, matrix representation of data, commonly used data types (int, char, logical, etc), MATLAB path and the difference between a MATLAB script and a function.

Represent a spike train in MATLAB.  In this tutorial, we will represent spike trains in MATLAB matrices.  Let each element of a matrix represent a time interval of 1 ms.  If there is a spike in this time interval, then we set the value of the element to 1, else we set it to zero.  Open up your MATLAB command window, type the following command-

myFirstSpikeTrain = [0 1 0 0 0 1 0 0 1 0];

If this spike train starts at time 0 and each element represents a 1 ms bin, then there are 3 spikes in this train, one between 1 and 2 ms, one between 5 and 6 ms, one between 8 and 9 ms.  By looking at this 10 ms spike train alone, estimate the firing rate of this neuron (Ans: 300 Hz).

Generate a “realistic” spike train in MATLAB.  Read the quotation from Dayan and Abbott once more.  We can break it down into three steps

  1. Compute the product of fr*dt. Let’s simulate a 10 ms long spike train for a neuron firing at 100 Hz.  Therefore, fr = 100 Hz, dt = 1 ms and fr*dt = 0.1 (remember that Hz is the inverse of s and 1 ms is 1/1000 s and fr*dt is therefore dimensionless).
  2. Generate uniformly distributed random numbers between 0 and 1.  To generate random numbers, use the MATLAB function rand()rand(1, 10) will generate 10 random numbers.
  3. Compare each random number to fr*dt.  If the product is < fr*dt, then there is a spike! We can summarize these three steps into the code below.
fr = 100; % Hz
dt = 1/1000; % s
nBins = 10; % 10 ms spike train
myPoissonSpikeTrain = rand(1, nBins) < fr*dt;

And I got the following output (You'll obviously get something different, all zeroes is not uncommon)

myPoissonSpikeTrain =
0 1 0 0 0 0 0 0 1 0

Repeat this simulation 20 times.  We could just type the commands 20 times. But it would be better to do all of it at once.  For that, we will represent the data in a 2D matrix.  Each row of the matrix will represent one simulation. Here is the code.

fr = 100; % Hz
dt = 1/1000; % s
nBins = 10; % 10 ms spike train
nTrials = 20; % number of simulations
spikeMat = rand(nTrials, nBins) < fr*dt;

Write the poissonSpikeGen function in MATLAB.  Packaging code into functions is very useful if we want to use a piece of code repeatedly.  Now let’s write a function that takes three inputs – the firing rate fr, length of the simulation tSim and number of trials nTrials. We have introduced a new variable here – length of simulation.  This is simply another way of representing nBins.  The number of bins is simply tSim/dt, where we have chosen dt to be 1 ms.  Your function should look something like this.

function [spikeMat, tVec] = poissonSpikeGen(fr, tSim, nTrials)
dt = 1/1000; % s
nBins = floor(tSim/dt);
spikeMat = rand(nTrials, nBins) < fr*dt;
tVec = 0:dt:tSim-dt;

This function (remember that the file name of a MATLAB function should be identical to the function name!) gives two outputs – the spike matrix spikeMat and a time vector tVec, a vector describing the time stamps for each column of the 2D matrix spikeMat.  In this vector, we represent the beginning time point, meaning that the interval between 10 and 11 ms will get a time stamp of 10.  Since each time stamp is 1 ms long, the increment is dt.  The last time stamp will start at tSim – dt representing the interval between tSim – dt and tSim.

Let’s take a moment to see why this function is useful.  Suppose you are asked to simulate spike trains from a neuron firing at 30 Hz for a period of 1 s and repeat your simulation 20 times, once you have saved the poissonSpikeGen function in the MATLAB path, you can simply complete your simulation with one command in your MATLAB command window. Go on then, clear your MATLAB workspace and type the following command at the command prompt.

[spikeMat, tVec] = poissonSpikeGen(30, 1, 20);

Plot the spike matrix.  Now, it would be nice if we can visualize this data!  I’ve written a simple function plotRaster.  Copy the code and save this function in your MATLAB path.

function [] = plotRaster(spikeMat, tVec)
hold all;
for trialCount = 1:size(spikeMat,1)
    spikePos = tVec(spikeMat(trialCount, :));
    for spikeCount = 1:length(spikePos)
        plot([spikePos(spikeCount) spikePos(spikeCount)], ...
            [trialCount-0.4 trialCount+0.4], 'k');
    end
end
ylim([0 size(spikeMat, 1)+1]);

This code simply cycles through each trial, finds all the spike times in that trial, and draws a black line for each spike.  Once you have saved this function, use it to visualize the spike matrix, then label the axes. Plot the X axis in ms.

[spikeMat, tVec] = poissonSpikeGen(30, 1, 20);
plotRaster(spikeMat, tVec*1000);
xlabel('Time (ms)');
ylabel('Trial Number');

When I ran the above code, I got the following output.

01_raster_30Hz

Yay! that’s it!! Quite simple right? Let’s celebrate by doing another problem.

A neuron in area MT of a monkey has a baseline firing rate of 6 Hz.  Upon presentation of a moving stimulus, the firing rate of this neuron increases to 30 Hz.  Make a raster plot representing the firing activity of this neuron. Show baseline for 500 ms.  The stimulus was presented for 1 second.  Your plot should look something like this.02_Raster_BaselineAndStimThis is the plot we set out to achieve in the beginning of this tutorial!  Good job.  Here is the code I used to make this plot.

% simulate the baseline period
[spikeMat_base, tVec_base] = poissonSpikeGen(6, 0.5, 20);
tVec_base = (tVec_base - tVec_base(end))*1000 - 1;

% simulate the stimulus period
[spikeMat_stim, tVec_stim] = poissonSpikeGen(30, 1, 20);
tVec_stim = tVec_stim*1000;

% put the baseline and stimulus periods together
spikeMat = [spikeMat_base spikeMat_stim];
tVec = [tVec_base tVec_stim];

% plot the raster and mark stimulus onset
plotRaster(spikeMat, tVec);
hold all;
plot([0 0], [0 size(spikeMat, 1)+1]);

% label the axes
xlabel('Time (ms)');
ylabel('Trial number');

I want to keep going, teach me more!

I applaud you if you just said that, so here are some exercises for you.  I’m going to give you some exercises and what your output should look like.

In this unit, we will simulate activity of 8 neurons in area MT of a monkey, whose direction tuning properties are given to you.  Download the dataset tuning.mat and load it into your MATLAB workspace.  The variable tuningMat contains tuning curves of 8 neurons recorded from area MT.  Each tuning curve describes each neuron’s firing rate as a function of direction of stimulus motion.  For each neuron, assume that the baseline firing rate is the lowest firing rate described by the tuning curve.

 (1)  Plot the tuning curve of neuron 3, and label the axes. It should look like this.03_Neuron3_tuningCurve

(2)  Plot all 8 tuning curves and use MATLAB’s plot tools to explore the tuning curves. Write down the tuning direction of each neuron. Your window should look something like this.

All8TuningCurves

  • Which neuron would you expect to respond best to a stimulus moving at 90 deg?
  • Which neuron would you expect to have the maximum firing rate in response to a stimulus moving at an angle of 225 degrees?

(3)  Simulate firing in each of the 8 neurons (500 ms baseline) to a single presentation of a 1 s long stimulus moving at an angle of 90 degrees.  Which neuron would you expect to respond best to this stimulus?

04_8MTneurons_90DegStimulus

(4)  Repeat the stimulation for a stimulus moving at an angle of 270 degrees.

05_8MTneurons_270DegStimulusCan you infer the direction of the stimulus by looking at the firing patterns of these 8 neurons?

(5)  Simulate firing in all 8 neurons over 30 trials for a 1 s long stimulus moving at an angle of 180 degrees.  It should look something like this.

06_rastersFromAllNeurons_90DegStimulus

Hope you enjoyed doing these exercises.  Happy simulations!

Picture from the cerebellum

Purkinje cell

Cerebellum is one of the regions in the brain where neurons are beautifully organized. Here is a picture from the cerebellum of a transgenic mouse expressing ChR2 (ChannelRhodopsin-2) and eYFP (Yellow Fluorescent Protein) in the basket cells of the cerebellum. The Purkinje cell was filled with an alexa dye. The picture above was taken as part of IBRO-KIST summer school 2012 in the laboratory of George Augustine. Recently, one of these pictures was chosen as one of the winners of the 2012 Koch Institute Image Awards here at MIT. In this video, I’m talking a little bit about the experiment and how the image was captured.

Visualize networks in your EEG data

In my previous post, I wrote about the Phase Locking Value statistic that can be used to quantify neural synchrony between two electrodes and how it can be a proxy for connectivity. Once you have quantified changes in connectivity using such measures (others include cross spectrum, coherence), it is good to visualize your results. The connected topoplot function is just the right thing for the job.

Below is a visual summary of what the script can do. In this figure, I have randomly connected ~10% of all possible channel pairs. It is possible to either specify or skip connection strengths. Detailed help and demo are available with the script.

Connected topoplot illustration

Now, an illustration from real data. You can use the PLV statistic to isolate stimulus-induced transient networks in your EEG data. Below is typically how your results would look like – a network within a time window along with an associated connectivity pattern for the network.

PLVresult-Illustration

In the above example, the subjects viewed a stimulus at time 0 and there are two prominent transient networks that could be isolated – one within the first 300 ms of viewing the stimulus and second between 300 and 600 ms of viewing the stimulus.  Plots below the connected topoplots are an average PLV time course of the whole network that is visualized. Therefore, in the first 300 ms, the isolated network de-synchronizes and between 300 ms and 600 ms, (a different) isolated network synchronizes.

Good luck with your analyses!

Phase Locking Value

Phase Locking Value

Phase Locking Value (PLV) is a statistic that can be used to investigate task-induced changes in long range synchronization of neural activity from EEG data. This method is introduced in Lachaux et al., (1999). I have implemented the computation of this statistic in MATLAB and you can download it from the link below. In this post, I will talk a little bit about how PLV can be useful when analyzing EEG data. Note that this statistic may also be used to quantify LFP (local field potential) data recordings to study synchrony at a finer scale compared to EEG.

Download MATLAB script

It is useful to keep in mind a few properties of the PLV statistic before diving in further.

  1. PLV statistic is a time course. This means that for every time point in your EEG data, you can extract a measure of connectivity. Therefore, this quantity can be used to observe transient changes in connectivity without pre-defining a time window of analysis.
  2. One PLV time course for every electrode pair. For example, if data is recorded from 28 EEG electrodes, there are 378 possible PLV time courses.  
  3. One PLV time course is computed over multiple trials. In practice, I have observed that this metric is suitable for datasets with a large number of trials. It is good to have a few hundred trials for each experimental condition for which long range synchronization is to be quantified using PLV.

The PLV statistic can be argued to be a proxy for connectivity. Intuitively, if the EEG signal in two channels (electrodes) during an experimental condition rises and falls together more than a baseline value, then there is more synchronization or loosely speaking, enhanced connectivity between these two electrodes. If it is less than the baseline value, there is desynchronization or loosely speaking, decreased connectivity between the two electrodes. Note that this metric does not care about the co-variation in the power of the EEG signal between two electrodes.

The schematic below illustrates how PLV is computed in this implementation. Explanation for each step follows.

Computing the Phase Locking Value - an illustration

Filtering. The EEG data is first filtered in the desired frequency band of interest, for example, gamma band, 35-45 Hz. In practice, it is best to use an FIR (finite impulse response) filter to filter such data in comparison to IIR (infinite impulse response) filters. Loosely speaking, FIR filters filter the signal in time domain and IIR filters operate in the frequency domain. Put simply, an FIR filters computes the value of the filtered signal at a time point from the values of the previous and future points. How far the FIR filter looks is knows as the order of the FIR filter.  For EEG signals, a useful rule of thumb is to ‘look at’ about 4 to 5 cycles of the desired EEG rhythm. As an example, consider gamma rhythm (35-45 Hz). Here, one cycle is about 25 ms and I would set the order of the filter to be 100 ms. If the sampling rate is 500 Hz, then the filter order would be 50 data points.

Hilbert transform. This step is used to quantify rising and falling of EEG data. Hilbert transform of a signal can be used to compute the instantaneous amplitude as well as the instantaneous phase. We ignore the former and use the phase (φ) for PLV computation. φ is a value between –π and π. Think of the filtered EEG signal as a series of crests and troughs. A value of π indicates the peak of a crest and a –π indicates the bottom of a trough. 

PLV metric. Remember that there is a phase time course φ for every EEG electrode. Now consider the phase time courses of two electrodes. The difference between these two time courses (Δφ) quantifies locking between the phases of these two electrodes. If an experimental stimulus influences signal in two electrodes to rise and fall together or with a certain lag, then Δφ will be consistent between trials. If there is absolutely no relationship between when the signal in these two electrodes will rise and fall, then Δφ will be random. All we need to do now is to quantify the randomness in Δφ. This can be done using a little trick from complex number theory. Please see the equation in the analysis schematic to understand this trick.

Normalization. This step is performed to make the PLV metric useful in practice. Note that due to its definition, PLV is always a value between 0 and 1; 0 signifying purely random rise and fall whereas a value of 1 signifies that one signal perfectly follows the other. In practice, the value of a PLV obtained is going to vary very little over time and the absolute PLV is not what we are interested in. We are interested in knowing whether an experimental stimulus induced a change in PLV. We are interesting in answering questions such as “was there increased connectivity between the frontal and parietal electrodes when the subject performed a task?” . To find this, we should test if PLV after stimulus is significantly greater than the PLV before stimulus. To simplify things, we can use the pre-stimulus period as a baseline and perform a z-transform normalization.

Good luck!