Coordinating pilot analysis of SLAP2 data

[jeromelecoq]

Hi all,
I am starting a thread here to help organize the analysis of slap2 data, especially the 3 experiment we have collected so far. I felt a forum post would be easier and perhaps more welcoming.

Today we discussed having 3 subthreads:

• Analysis of receptive fields block : The experiments we ran so far included an approximately 10 min long receptive field block. We recorded synapses from proximal and apical dendrites in V1. A simple analysis questions:

1. Are receptive fields aligned across both dendritic compartments? Some models have proposed that apical dendrites received distinct inputs from proximal dendrites.
2. Are receptive fields clustered in space? An analysis of spatial location (in the brain) against screen location would directly test this.

• Analysis of orientation tuning : We have 2 orientation tuning block. One before the oddball block and one after the oddball block.

1. Are orientation tuning equally clustered across synapses? A similar spatial analysis would test this.
2. The two orientation blocks could be used to look at the stability of inputs across the session. After the oddball block, we should expect a change in representation of vertical bars, given the number of presentations of those during the oddball block.

• Analysis of oddball block: We present a vertical orientation repetitively and then disturb it with 4 types of oddballs : One halt, One omission, and 2 alternative orientations.

1. Do synapses detect those oddballs? It is really an open question!
2. If they do, are those different depending on the oddball types.
3. Are those mismatch response stable or do they extinguish quickly?
4. Are those oddball responses in line with the orientation tuning responses ? Especially the two alternative orientation that are presented earlier and later in the experiment.

Perhaps, some could sign up to tackling those questions?

[Sarruedi]

I would love to team up with others to start digging into some of the questions.

Can you maybe clarify what you mean when you say: Are receptive fields clustered in space? An analysis of spatial location (in the brain) against screen location would directly test this.

I'm not sure I understand what you mean when you say spatial location in the brain. What reference points do we have available in the data?

[jeromelecoq]

Right. Good point.

So every ROI is a single synapse. In the data there are 2 planes: One apical and one proximal dendrite.
One plane is deep at the somatic layers and the other is very superficial.
The X,Y location of the ROI is given to you by the ROI center of mass.
The Z location is something we can give you directly but grossly you can assume one plane was in the layer I and the other in the layer 2/3

For every ROI, you can measure the receptive field, that will give you the location of the synapse response onto the screen.
We have not compared the physical location of the synapses in the brain tissue to their response location onto the screen. This is what I am referring to.

[koweiss]

This is what I stated during the meeting that we would be interested in taking a first stab at. Once this spatial information is available, along with the location of the soma, we will jump right into the analysis.
One specific question we will tackle (to add to the above): If and when an oddball response takes place at a subset of the imaged synapses, to what extent these responses correlate in time and in space? if such correlation exists, does it change as a function of the proximal and distal planes?

[jeromelecoq]

I think a proxy to the spatial information is already available. Each trace is associated with an ROI mask. This mask is rather localized so I would really start to use that as a measure of relative position. Later on we can add the dendritic tree once it is segmented.

[rcpeene]

Yeah, the image masks are stored together as a 3D array, where the dimensions are roi, x_pos and y_pos. Taking each ROIs mask from this gives the full field 2D mask of the ROI. From there you could find the x,y pair of the maximum or median activity for that ROI to get a point in 2D space relative to the other ROIs. This approach doesn't include any metric of real distance though.

[Sarruedi]

Thanks for the clarification. Just one more question regarding the GluSnFR used in the test experiments as I'm not yet familiar with these sensors. How do we know that the signal reports the activity of single synapses? Do the receptors spatially cluster at spines? Just wondering how the expression is targeted... Is the construct specifically targeting postsynaptic compartments?

[seyedolmohadesin]

The indicator we are currently using (iGluSnFR4.v9601) is described in the preprint.

From Kaspar:
"iGluSnFR4 labels the entire membrane, and reports extracellular glutamate. However, it turns out that this is quite specific to synaptic partners, because glutamate is buffered and taken up at each synapse by glia. We do not measure receptor activation.
We know that the glutamate signals are specific because:

1. when we label axons, the signals are highly correlated to the spiking of the cell
2. signals occur preferentially at the tips of spines
3. We have modeled the process of glutamate and release and diffusion, with known AMPA and NMDA receptor activations to guide the models, using molecular diffusion models."

[koweiss]

@seyedolmohadesin @jeromelecoq
Is the spiking of the cell receiving these synaptic inputs recorded/imaged?

[jeromelecoq]

Not currently.

[jeromelecoq]

I think one way to get started is :
1/ Pick one of the questions listed above.
2/ Say so here.
3/ explain how you are going to answer the question here. Note that we can collectively help further refine your analysis plan. So don't hesitate to ask questions here.
4/ Share linked code on google collab or similar on how you tried to answer the question here.
5/ Perhaps join the webinar in 2 weeks where we will discuss progress. I am hoping those that did analysis can discuss with their code.

I think later on we can create Github tickets and organize ourselves better but we need to get started in a simpler way. Perhaps this forum is the simplest way,

[maierav]

Here is a first stab at the oddball analysis. This likely is still buggy, and I would appreciate if someone could help with checking the code and hunt down bugs. Also, any improvements or additions are welcome. And please feel free to use this code as a starting point for your own analysis:

https://colab.research.google.com/drive/17-pAT6-xZXFM5mz3ccUt1W7w3Yaxazh7?usp=sharing

standard is the regular, repeated stimulus (0 deg drifting grating)
oddballs are labeled by their orientation (45 and 90) as well as static for the static oddball

[maierav]

I did a quick check for a variance effect:

The greater variance for the greater response is expected, so there does not seem to be a systematic oddball effect.

[Sarruedi]

Here's a first attempt regarding: Mean Response Correlation based on sorting the ROIs as a function of 'distance'

DMD1

DMD2

Still tweaking code as I switched to spyder, will upload asap. In a nutshell: I first extracted the spatial coordinates of each ROI by computing the centroid of its binary mask from the image segmentation data. I then calculated the pairwise Euclidean distances between all ROI centroids to construct a spatial distance matrix. For the selected stimulus condition, I retrieved the trial-wise fluorescence traces and computed the average response over trials for each ROI. Using these trial-averaged traces, I calculated the Pearson correlation coefficient between every pair of ROIs to create a similarity matrix, which reflects how similar the temporal dynamics of responses were across ROIs. I then converted this into a dissimilarity matrix (1 – correlation), flattened both the spatial distance and dissimilarity matrices, and performed a Spearman correlation analysis to quantify the relationship between spatial distance and functional similarity. Plotted is the mean response correlation with error bars (SEM) as a function of distance, to check how functional similarity decays with spatial separation across the field of view.

I haven't sub selected the ROIs based on additional criteria - response amplitude, tuning preferences etc. that are of interest at this stage.

[jeromelecoq]

Apparently this U curve (correlation that has some sort of minimum at a distance) is something that we have seen in other neurons with other stims. Unclear why but this seems consistent.

[maierav]

Very surprising for someone who knows little about synaptic inputs like me. Higher correlation with increasing distance? And such drastic differences in overall correlation that cannot be explained by SNR...

[Sarruedi]

I wonder whether it is due to some spatial clustering of shared inputs in particular zones. I think it would be good to do a sanity check and look at the response strength of the ROI pairs with low correlations versus high correlations, just to make sure that the low correlations are not due to a lack of a response etc.

[maierav]

Here's a first result on orientation tuning (single ROI), before and after the oddball block:

[jeromelecoq]

Interesting. I looked at the test code that is linked at the end of the experiment notes (validate_nwb_slap2.py) script) and found:

Try to get DfOverF data first (preferred), fallback to Fluorescence

    if 'DfOverF' in ophys.data_interfaces:
        data_interface = ophys.data_interfaces['DfOverF']
        trace_type = 'DfOverF'
    elif 'Fluorescence' in ophys.data_interfaces:
        data_interface = ophys.data_interfaces['Fluorescence']
        trace_type = 'Fluorescence'

So As Carter mentioned during the presentation, we made a mistake and packaged DF as DF/F. I understood that Kaspar and Maedeh have been using DF more readily for their analysis. I am not quite sure yet as to the full rationale. I think they wanted to relate with SNR more directly. Obviously we would need to normalize across experiments. Perhaps you could compute the DF/F from the trace?

[maierav]

I find that super interesting. When we look at the spatial structure of orientation tuning (working on that), we might want to take that into account. In other words, we might find that it is not the Euclidian distance that matters but something more "functional". What I mean by that is that synapses that are far apart but share overlapping RF's might also be closer in tuning. That might indicate common inputs.

Do you feel comfortable sharing your code?

Format does not matter. If it's on GitHub, I can convert it to a Colab notebook.

[Sarruedi]

I haven't subselected any ROIs at this stage. I was initially hoping to select the ones with a significant response above baseline, which is why I was looking into the deltaF/F versus deltaF, but for now I just took them all. Maybe it's easiest to normalize via zscore or min/max to subselect ROIs with a response above or different from baseline at the moment. In addition, I didn't do any proper analysis to identify the center of the RF - simply took the position with max dF - so this could also add noise.

Happy to share my code. Will try to upload it tomorrow when I get the chance. Just wanna double check the distance calculation, to make sure it's correct. The idea of comparing with the orientation tuning is great.

Adding some relevant papers here: https://pmc.ncbi.nlm.nih.gov/articles/PMC5533220/

https://pubmed.ncbi.nlm.nih.gov/20428163/

[KasparP]

> I have a question about the data. In the notebook, it says that the information is stored under a key called 'DfOverF', but it’s actually just ΔF and not normalized. Could you clarify why this is the case? It’s a bit confusing to interpret the responses without having properly normalized ΔF/F data.

Regarding ΔF vs ΔF/F: both have their uses. When combining signals across synapses, dF is more appropriate. But normalizing to baseline is important to compensate for bleaching etc; for many analyses we aggregate dF and F0 separately then compute dFF for the analysis at the end. There is no single 'Properly Normalized' signal for every analysis! In the original data files from which these NWB files are generated, DF and DFF (and F0) are all included.

[Sarruedi]

Thanks for the explanation Kaspar. It would be great to include to F0 and/or DFF in the nwb file the future.

[Sarruedi]

Hi Alex, I uploaded the demo code on github:

[SLAP2_DemoScript_SR_AM.py]
https://github.com/AllenNeuralDynamics/openscope-community-predictive-processing/blob/main/code/data-access/SLAP2_DemoScript_SR_AM.py

It’s messy and quite hacky in places (especially the further you go along in the code), but I wanted to share it with you in case you have time to take a look and want to expand on the orientation-RF-spatial analysis we discussed.

I’d also really appreciate any feedback, especially if you spot any bugs or have suggestions for improvements.

[Sarruedi]

Trying to get a sense of the modulation of dF

Subselecting ROIs (not very stringent at all as zscores generally quite low)

Correlation across ROIs

[Sarruedi]

Hi Nicholas, thanks for putting together the visualization. It’s looking really exciting!

It would be great to include a color bar alongside the plot will help everyone see exactly what values correspond to each hue. Does your current scale spans the “zero-change” ROIs as well? It’d be great to get a sense of the fraction of ROIs changing by also visualizing unchanged ones.

My other suggestion is to use a more colorblind-friendly diverging map: A diverging palette like bwr or seismic would work really nicely. These colormaps clearly highlight positive versus negative deviations in line with your current plots but also keep the neutral (no-change) midpoint distinct.

[Dedalus9]

@jeromelecoq I updated my post to include plots from additional experiments. I also included the static stimulus type, although it is not present in the pre- or post- blocks (blocks before the odd-ball block and after it). It seems like we're seeing a similar effect. Also, according to the meta-data that comes along with the NWB files DMD1 represents proximal dendrites and DMD2, apical dendrites: does that sound correct?

@Sarruedi Thank you! I edited the plots to incorporate your suggestions. The green/red scale did span the "zero-change" ROIs, but I've changed the color scheme to seismic to better visualize those "zero-change" ROIs

[Dedalus9]

Question Addressed: I am also now looking at spatial clustering of orientation tuning and how that changes with presentation of the oddball block.

Context: I z-score normalized each ROI's entire time-course using only its pre-block ("baseline") frames to set that ROI's baseline mean to zero and its baseline variability to zero so that any deviation in the odd- or post-blocks is expressed in its units of its own baseline, i.e., before its exposure to the odd block. In the pre/post block and plane (i.e., DMD1 and DMD2), I then extracted peri-stimulus windows (-0.5s to 0.5s) for each stimulus presentation of interest.

Figures: I measured how each ROI's orientation tuning pattern aligns with its immediate dendritic neighbors before and after the oddball block. For each ROI, I first computed the mean response in the 0-0.5s window following each orientation (0°, 45°, 90°, 120°, ...) in both the pre- and post-blocks, yielding two tuning vectors per ROI. Using the Euclidean distances between ROI centroids, I then identified its k = 10 nearest neighbors and calculated the cosine similarity between its tuning vector and each neighbor's -- averaging those values to get one "local tuning similarity" score per ROI which is, in effect, a metric of the similarity of that ROI's orientation tuning with its neighbors. Finally, I overlaid those scores onto the spatial masks to visualize how local tuning coherence changes from pre to post. I also computed the difference between the post "local tuning similarity" score per ROI and the pre "local tuning similarity" score per ROI to get a visual for how "local tuning similarity" changes. I also overlaid those differences onto the spatial masks.

Interpretation:
For the average cosine similarity scores:

• A positive value means the ROI's tuning vector (as I defined it) is similar in direction to its neighbors, indicating that it responds similarly across orientations.
• A value of 0 means the ROI's tuning vector is orthogonal to its neighbors. That is, there is no linear relationship in their tuning with its neighbors. In other words, its response to different orientations are uncorrelated with that of its neighbors.
• A negative value means that the ROI's tuning vector is diametrically opposed to that of its neighbors. That is, the response of that ROI across orientations is anti-correlated with its neighbors.

For the difference in average cosine similarity scores for each ROI between the post- and pre- blocks:

• A positive difference means increased local tuning similarity after the odd-ball block
• A zero difference means no net change in local tuning similarity after the odd-ball block
• A negative difference means decreased local tuning similarity after the odd-ball block

Observation: It would seem that after presentation of the odd-ball block, there are spatially distinct clusters of neighboring dendritic synapses that either increase or decrease in local tuning coherence.

Next steps: I would be curious to see how smooth these differences are and how quickly these clusters emerge during the odd-ball block. I would also like to see how spatial orientation tuning clusters overlap with deviant (i.e., oddball) tuned clusters (e.g., static, 45°, 90°)

sub-794237_ses-20250403T103830

sub-794237_ses-20250424T142019

sub-794237_ses-20250508T145040

[jeromelecoq]

I had a quick discussion on this analysis with Kaspar.
Important caveat that we will probably have to discuss in depth next week: There IS bleaching in those long experiments focused on single cells. The signal is strong and stays along but we need to be careful in the interpretation as bleaching will create a decreasing baseline which will be accompanied with smaller transients.
In addition, deeper planes come with lower signal levels that create a different detection floor for events, compare to more superficial planes.

The right approach is probably to use a DF/F measure with a local F0 that is a slow running average. DF/F will inherently normalize for the changing concentration of the fluorophore over the course of the experiment. I recognize that you currently have DF packaged in those experiments so it could be confounding the analysis. We should repackage the data to provide those more complete traces for all.

We could also develop a number of documentation pages on the website to help develop these careful analysis. There are a number of control analysis that we could do to protect ourselves from these interpretations caveats.

[Oweiss-Lab]

Understood. I think it is important to understand these issues and ensure the data is corrected for these effects (or controlled for) before we embark on extensive analysis and comparisons.

[farznaj]

To avoid overlap, what analyses should i have my students work on. i see the list above ... but any preferences?

[maierav]

Please feel free to keep working on the oddball analyses or orientation tuning analyses. There is still lots to do beyond the simple population averages (e.g., studying individual ROI's). I won't be able to get back to these analyses for the next month or so.

[I'll post the orientation tuning code on here later today or tomorrow.]

[maierav]

I added a primer for new student to get started at:

#76 (comment)

please feel free to edit.

[Sarruedi]

@KasparP @seyedolmohadesin Regarding our discussion about deltaF versus deltaF/F yesterday. I would like to clarify the steps for proceeding with the analysis to make sure we are on the same page on how to process the signals.

To quantify ROI responses I changed from zscore to % change in fluorescence (%ΔF) relative to a common baseline, followed by an estimate of variability across trials based on our discussion.

Here is an outline of the approach and some example traces:

Step1. Average Trace Across Trials:
For each ROI, compute the trial-averaged fluorescence trace across all trials, yielding a single time series per condition.

Step2. Baseline Calculation:
Calculate the average baseline by taking the mean fluorescence during the pre-stimulus period (e.g 0.343s before the stimOnset)

Step3. Percent Change Conversion:
Converted the average trace to percent change by subtracting the average baseline and dividing by it, then multiplying by 100 to get the %ΔF relative to baseline at each timepoint.

Step4. Error Estimation: To estimate the variability, convert each individual trial’s trace to percent change using the same average baseline as the denominator. SEM across trials at each timepoint by taking the standard deviation of the percent traces and dividing by the square root of the number of trials.

This results in traces like these:

Is this similar to your approach, or how exactly do you calculate the responses in your current analysis? The normalization step that I used should ensure comparability across ROIs with different baseline expression levels. Maybe you have a different way to account for this?

[lrudelt]

Hi Sarah, thanks for sharing! I have one question: Since the fluorescence signal is in dF, it can take negative values, right? How do you account for this when computing percentage changes? E.g., think of an ROI where the baseline ends up being negative, what is the interpretation of percentage change in this case? Or am I getting something wrong here?

[Sarruedi]

I agree that % change can be misleading and cause issues with the baseline estimate. Since we didn't have access to F, I simply used the metric to serve as a rough index of how much an ROI’s activity deviates from its own baseline. It's definitely not ideal in terms of calculating evoked activity due to potential flip signs, so I ended up analyzing the %ΔF time without baseline subtraction yesterday. I have to double check if I uploaded these plots yesterday. Btw, since we now have access to both DeltaF and F, I most likely switch to deltaF/F and redo the analysis. Does this make sense?

[lrudelt]

Yes, I will also rerun my analyses now with deltaF/F :)

[lrudelt]

Contextual modulation of synaptic inputs during oddball

Hey, to coordinate our efforts I wanted to post here a short summary of my preliminary analysis of Oddball responses in the SLAP2 data. I will also present a short summary in the next meeting. Please find a more detailed post in the discussion post for next meeting #86.

Google Collab code link will follow shortly, the code is based on code by @maierav, @jeromelecoq and @Sarruedi.
The analysis is related to the analysis by @Dedalus9 presented in #89 .

My focus was on investigating contextual modulation of synaptic inputs for expected versus unexpected sensory stimuli (e.g., similar to Hamm et al. 2021, or Furutachi et al., 2024 for somatic responses).
Thus, the idea was to do a similar analysis to panels I and J of Fig 1 in Hamm et al. 2021 (see below), but for individual ROIs, reflecting individual synapses. The difference to the plots by @Sarruedi and @maierav is the comparison of oddball responses to a random control condition (the direction selectivity block), with the aim to disentangle response modulation from stimulus tuning.

Precise question addressed: Do we find evidence for the suppression of synaptic responses for expected inputs and enhancement for unexpected inputs? What is the spatial distribution of these responses (basal vs apical, or even branch specific?).

Preliminary results: Significant (average) modulation of synaptic inputs through stimulus context is quite rare, with few ROIs showing suppression of the standard stimulus, and very few ROIs showing enhancement of unexpected stimuli (45 and 90 degree orientation oddballs). Such modulations were mainly found in the 2025-05-08 experiment for image plain DMD1 (corresponding to synapses onto proximal dendrites).

Important caveats:

• The current stimulus design only allows to observe a given stimulus either as expected, or unexpected, respectively, which does not allow to study the relation between suppression and enhancement for the same stimulus (i.e., we miss one condition as in the plot by Hamm et al.).
• We are currently lacking controls for the static and contrast oddballs, so these have been omitted in the analysis.
• Due to photobleaching, we have to be careful in the interpretation of the current results. To mitigate this effect some extent, I split the data set in two halves, and compared the first half of oddball responses for each stimulus orientation to its control in the first direction selectivity block (previous to the oddball block), and the second half of the oddball block to the second direction selectivity block (immediately after the oddball block). This reduces the effect of bleaching between the compared conditions. Also, this way we can be quite certain that enhanced responses in the first oddball half, as well as suppressed responses in the second oddball half are not caused by bleaching.
• Statistical testing is still preliminary, e.g., proper multiple-comparison correction is not yet implemented

Figure from Hamm et al., PNAS 2021

[JJYma79]

Exploring Dimensionality Reduction and Clustering for Joint DMD1&DMD2 Analysis- Seeking Feedback

I've been experimenting with a dimensionality reduction and clustering pipeline for our ΔF/F time series data from both DMD1 and DMD2 (716 synapses total, analyzed together, not separately now), and I'd be grateful for any opinions and suggestions:

My current workflow involves three steps, but I'm not sure whether this approach is appropriate currently.

First, I applied SVD to reduce our 716 traces to the top 3 principal components and used K-means clustering (k=3) on these 716 traces in the reduced PC space (Figure 1). This seems to reveal three potential functional clusters that show distinct activity patterns across different phases of a trial (pre-stimulus vs. post-stimulus periods). I'm wondering if this actually reflects meaningful spines subpopulations with different temporal response profiles, or if it's just an artifact of dimensionality reduction.

Second, the SVD explained variance analysis (Figure 2) suggests the first component captures >30% of variance, with diminishing returns after ~6 components. The elbow around components 3-5 guided my choice of dimensions, but I'm curious if others have found different optimal ranges for similar neural data.

Third, I tried hierarchical clustering on the top 30 PC traces (Figure 3) to see if certain temporal patterns group together. Interestingly, the most dominant PCs do appear to show very distinct patterns (e.g., PC1, PC2, PC3 cluster separately), potentially indicating that different principal components capture fundamentally different aspects of synaptic dynamics, but I'm uncertain about the biological interpretation.

Questions: Does this multi-level approach (synaptic clustering + temporal module clustering) make sense for current exploration? Are there alternative methods you'd recommend? I'm particularly interested in how to validate these clusters against our receptive field, orientation tuning, and oddball detection data. Any thoughts or suggestions would be greatly appreciated!

[JJYma79]

Google Collab code link will follow as soon as possible, the code is based on code by @maierav, @jeromelecoq and @Sarruedi.

[Dedalus9]

Hi @JJYma79, I think this is a very good start, but I think to make this more interpretable it would help to visualize the loadings of each PC component. I imagine, also, that you reduced the dimensionality of the dataset along ROIs, using a matrix where each row refers to a stimulus presentation and each column to an ROI. If that is the case, what values correspond to each ROI? Also, I wonder if it might also be informative to color-code this trajectory by the type of stimulus presented at the time. It would seem as if the "PCA" space of this matrix does have some structure but interpreting what drives that structure would be crucial for interpreting what the trajectory in this PCA space means.

I have pursued a similar analysis but focused on orientation tuning vectors and how they vary once the odd-block is presented. Once I post it I will link to it here: it may give you some helpful directions.

[JJYma79]

Hi @Dedalus9,
Thank you so much for the thoughtful feedback!

Yes, I did reduce the dimensionality along the ROI dimension

I really appreciate your suggestions about visualizing the PC loadings and color-coding the trajectory by stimulus type, and I'll work on implementing both of these visualizations to make the analysis more interpretable and post the updated results and code soon, looking forward to both your results and next discussion!

Thanks again for the valuable guidance!

[Dedalus9]

Of course! Also, the link to the PCA analysis I mentioned: #92 (comment)

Perhaps it'll give some inspiration