How we watch neurons being activated
A large portion of my research consists in measuring the activity of neurons to understand how they respond to external signals and communicate with each other. The currency of the brain is electrical activity: when a neuron is active, it “fires” an action potential. An action potential is a depolarization that is conducted through a neuron’s cable (axon) to a neuron it signals to. Traditionally, neuroscientists have used electrophysiology to probe the firing pattern of neurons. This is a very precise technique, but it is limited to the analysis of a handful of neurons at one time and it is very time consuming. A method that has been developed more recently solves these issues. It takes advantage of the fact that when a neuron is depolarized (active), calcium flows into it. We can monitor calcium dynamics inside the neuron by using dyes that fluoresce differently depending on whether they bind calcium. These dyes can be introduced to certain types of cells through genetic methods.
The larval zebrafish is a great model system to probe neural activity with calcium imaging: its brain is transparent enabling the experimenter to monitor neural activity in the awake behaving animal.
Currently, I am interested in how a few neurons forming a small circuit communicate with each other to encode stimuli from the environment and compute a response resulting in behavior. The behavior I chose to look at is the escape behavior because it is one of the most basic and important behaviors for the fish. The brain is such a complex entity, with many layers of regulation, that we don’t even understand some basic principles governing behavioral response at the individual cell level. The beauty of my system is that I can look very precisely into the neuronal computation, control the input to the system (the external stimuli) and monitor the output (the behavior).
This is usually the rough method I use to analyze calcium imaging data:
- Load image stacks into MatLab
- Denoise images using a Wiener filter so that the registering process in step 3 works best. I made sure denoising the image does not alter the change in fluorescence computation.
- Register images in X-Y with respect to a reference image: pad each image and then translocate it in X-Y. Find the translocated image with the strongest correlation to the reference image
- Manually pick ROIs (I only look at a few neurons and this is more accurate than any software computation)
- Compute the change in fluorescence (Δf/f = raw fluorescence – baseline fluo) / (baseline fluo) for each ROI at each timepoint
- Find whether there was a response for each ROI in each trial using thresholds and shape finding
- Find the amplitude of the response and standardize it: (ampl. – median(ampl.)) / interquartile range
- Compare response amplitude changes and probability of response changes across different conditions
I work with a population of neurons consisting of approximately 12 cells that form anatomically distinct groups. I wanted to ask whether these groups were also functionally distinct. That is, whether their response to stimuli could be clustered into groups that mirror their anatomical differentiation or not. To do this, I applied correlation and clustering algorithms.
Individual neurons (colored shapes) vary in their activity in response to stimuli
The variables I used for analysis include the normalized responses traces (fluorescence as a function of time after the stimulus), normalized probability of response and response amplitude. I reasoned that two factors were most important to determine functional correlation between neurons: response probability, and response dynamics, which encompasses shape (latency to start, latency to peak, on and off response speed) and amplitude.
Correlation of the neurons’ responses in the 4 clusters
Here I correlated the shape of the responses (normalized for amplitude). This shows that clusters 2 and 3 (the lower clusters anatomically) are very correlated with themselves and with each other. Clusters 1 and 4 (the upper clusters anatomically) follow the same pattern but the correlation is smaller.
I was happy to see this. Then I did k-means and hierarchical clustering to find that activity in the neurons fell quite well within the four anatomical clusters. By looking at the traces again, I saw that the difference in the shapes of the responses between the lower and upper clusters is due to a difference in the start of the responses. This meant that the lower and upper clusters varied in the timing of their responses, which was exciting. As a skeptical scientist, I thought about the ways in which this could be wrong, and sadly I found that the timing difference was due to be an artifact! It turns out that the imaging laser scans the image from bottom to top, and this takes 1/8 sec, which is a large amount of time with respect to the timing of neuronal activity. When I analyze one image, it is actually not a snapshot over the whole image, but there is a difference of 0.125 sec between the bottom and top of the image. So it appeared as though the bottom cluster responded after the other but in fact, it was because it was scanned first, and before it had time to respond to the stimulus. Lesson learned…
So there wasn’t a real difference between the top and bottom clusters in terms of response shape (normalized for amplitude). I did find however that the amplitude of the responses between the left and right clusters were different. The biological basis of this difference is explored in my article (currently under review).