Quantifying responsive neurons to mismatch stimuli in neuropixel (& other modalities) by a statistical criterion we set

[davidgwyrick]

Creating a discussion platform here to expand the point Stefan Mihalas raised in the meeting on 09.08.26 about establishing some set of statistical criteria about what constitutes a significant response to a particular mismatch signal for a given unit/roi. This could be part of the data product release as a starting point to looking at the data, as Stefan correctly pointed out, many researchers don't want to calculate these things themselves. Although I think more do than don't.

I see a few issues that we also ran across in the V1 Deep Dive project.

1. Given that there are so many different mismatch & control stimuli presented, a potential data product that gives a particular units responsiveness would be multiplexed. e.g. it might respond strongly to the 90 degree mismatch in the standard oddball block, but not the 45 degree mismatch. Or ROIs that were recorded across sessions in the mescope & slap2 modalities might be context specific & only respond to a particular context block like the sensorimotor block.
2. Compared to the V1DD, we have the additional advantage/complexity that control stimuli were recorded. A unit/roi might elicit a significant response during a mismatch event relative to non-mismatch events, but when compared to the control response it is not significantly different. Is the latter definition of significance the one we would report or both?
3. Defining the window over which to calculate such metrics like responsiveness or mismatch index is also a choice we have to make for each stimuli. We would need to think a bit about the duration mismatch response, since it's an extended response. And the sensorimotor stimulus set, as it depends on the behavior of the animal as well.
4. This table would contain many columns per unit/roi and we would need to define what ultimately is a responsive unit.

All of these are surmountable and could be explained in the text as part of the data release. I think this definitely falls into analysis territory but everyone who is going to use this data will have to do this, so standardizing it during data release might be worth our efforts here. The full data would be released but this would be an additional derived data product that could be a nice supplement people could choose to utilize. It could also complicate things if we don't defend our choices well. In any case, it would also add another tick mark to this interactive figure "Show only responsive units" , that may be a nice way to round out the static figure (which I haven't remade yet, sorry Jerome). If we want to do this, we can iterate the set of criteria we want to start out with and see how things look in terms of how many responsive units remain & whether there are any interesting mismatch signals in the data (& in which area) in the first place.

[maierav]

That screenshot looks awesome!

Additional considerations: Do we correct for baseline? How? Delta? %-change? Z-score? What matters more to post-synaptic neurons, absolute firing rates or relative changes thereof? What kernel should be used to convolve the binary spike trains in order to derive "analog" (SDF) spiking time series? How can averages over animals be justified? Mean vs. median, SEM and CI, parametric vs. non-parametric tests - is all variance is Gaussian?

[davidgwyrick]

Hey Alex, thanks for the questions. Just to clarify, this is just for 1 mouse for now. Doing significance testing for each unit/roi across all of the different stimuli is the first step before we think about how to look across mice. For that, I do think looking at baseline subtracted, stimulus evoked activity is one way of approaching it. Of course we have to define this baseline. For some stimuli, like the sequence block, it's not clear what a good baseline is since there is not interstimulus interval. Comparing absolute firing rates for a mismatch event vs a repeat event or vs a control event runs the problem of firing rate drift over the course of the experiment (e.g. firing rates tend to increase the longer an animal is head-fixed on a rig) & faster timescale behavioral effects (e.g. comparing trials where the animal is running vs at rest).

These are the questions we have to address though if we want to define which units are responsive to a particular mismatch stimuli as Stefan suggested. And as you can see, there are always a lot of choices one has to make.

[davidgwyrick]

For those that are following this, claude and I wrote a markdown file that describes the responsiveness criterion I came up with per mismatch context. There are many decisions I made here, and would like feedback from the group, particularly for the duration mismatch and sensorimotor mismatch contexts. Since I was able to quantify responsiveness against the unit's own expectation (Q1) and against the matched control block (Q2), I can set 2 criterion for significance. In the interactive figure 10 of branch /edit/neuropixel-mismatch-responsiveness I added a new filter to show only responsive units. See screenshot below.

https://github.com/AllenNeuralDynamics/openscope_p3_data_release_paper/blob/edit/neuropixel-mismatch-responsiveness/docs/mismatch-responsiveness-definitions.md