ODDBALL Analysis (for SLAP2 data)

[maierav]

This thread if for anyone interested in analyzing the SLAP2 data for oddball responses / predictive processing.

ORIENT YOURSELF ABOUT THE EXPERIMENTAL SCOPE AND METHODOLOGY
#76

WHAT HAS BEEN DONE ALREADY?
We have basic python code provided by Jérôme Lecoq, Carter Peene, and Alex Maier to plot the oddball responses of individual ROIs (dendritic spines) based on the glutamate signal (see next section).

EXISTING CODE TO BUILD ON:
https://colab.research.google.com/drive/17-pAT6-xZXFM5mz3ccUt1W7w3Yaxazh7?usp=sharing

WHAT CAN I DO AS A NEXT STEP?
That's entirely up to you! Here are some ideas:

-compare if and how ROIs with overlapping receptive fields share orientation tuning or oddball responses

-examine temporal dynamics, such as trial-by-trial correlations

• create population statistics

• examine whether oddball responses are static or drift over the course of the presentations

• compare trial mean versus median

• convert the oddball responses into %-change form baseline

• examine the relationship between different oddball responses (there were 3)

• examine the relationship between oddball responses and response magnitude of each ROI

• search the literature for what others have found out about oddball responses and open questions

• extend the analysis to more of the imaging sessions (right now, we only analyze one of them)

• anything else you can think of or find interesting

DISCUSS YOUR IDEAS AND RESULTS
Share your plans, figures, and codes here to coordinate with others:
#79

I STILL DO NOT KNOW OR FEEL UNSURE ABOUT WHAT TO DO
Brainstorm with others on here:
#78

[lrudelt]

Contextual modulation of synaptic inputs during oddball

I did a preliminary analysis of the SLAP2 data to investigate contextual modulation of synaptic inputs during oddballs.

Analysis code: https://colab.research.google.com/drive/1rbVXa0x8xn0Mz18ejZoc6Ald6f6R6dDq?usp=sharing
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.

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). See slides for plots and more details.

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

[lrudelt]

One example of suppression in the standard direction

Example of a suppressed (off?) response in the standard direction (0 degrees). There is also some suppression in the 45 degree direction, potentially because of the small angular distance to the standard stimulus.

[lrudelt]

Interesting example of changes in stimulus tuning after oddball block

This ROI is interesting, because there is quite an interesting tuning (on response to 90 and 45 degrees, off response to 0 and 45 degrees) appearing after the oddball block, thus only showing in the control condition of the second half of the recording (direction tuning block). This is most likely not a confound of bleaching, because the response should be weaker in the second control block.

[lrudelt]

Few significant modulated responses in DMD 2 or any imaging plane of the other data set

Few significant cases of suppression and enhancement in DMD 2 of the same data set as above. Especially, these could be confounds of bleaching.

Few significant cases of suppression and enhancement in DMD 1 of data set 2025-04-24.

Same for DMD 2...

[maierav]

Great job. Super interesting. Did you sub-sample the presentations so that they match in number (n) between the orientation tuning and oddball blocks before comparison?

[lrudelt]

Great point! Roughly, for 45 and 90 there are about twice as many samples in the control compared to oddball block because it is split in half. However, now with access to deltaF/F I will take the entire oddball block, such that the number of trials is the same.

The number of trials for 0 degrees are already subsampled to match the control condition.

[Sarruedi]

I reanalyzed oddball-evoked activity using %ΔF (rather than z-scored values), as discussed here. The stimulus-evoked response for each ROI was calculated as the %ΔF during the stimulus window (0–0.343 s) minus the baseline (–0.343–0 s).

The scatter plots below show the evoked responses to the standard stimulus (x-axis) versus each oddball type (y-axis) for all ROIs.
Each point is color-coded based on statistical comparison:
Red: Significantly facilitated by the oddball (paired t-test, p < 0.05)
Blue: Significantly suppressed
Gray: No significant change

This approach allows us to:

• Compare oddball sensitivity across test recordings (DMD1 vs DMD2),
• Examine the range and symmetry of evoked responses across the population,
• Quantify the fraction of oddball-sensitive ROIs,
• Identify candidate ROIs for follow-up analyses

Here are the current plots:
sub-794237/sub-794237_ses-20250508T145040_image+ophys.nwb

sub-794237/sub-794237_ses-20250424T142019_image+ophys.nwb

sub-794237/sub-794237_ses-20250403T103830_image+ophys.nwb

The next step is to relate this functional oddball modulation to potential experience-dependent changes in tuning properties or RF selectivity. Eventually we can map these properties back on the the location along the dendrites...

[maierav]

This would also be interesting to redo using @Irudelt 's approach above, contrasting the oddball presentation with the presentations of the exact same gratings (45 and 90 degrees orientation) during the orientation tuning blocks. This way we can dissociate stimulus feature differences from the oddball-context. In the same vein, we might want to compare 0 degree orientation between the two blocks to investigate whether there is an effect of "expectation"/"predictability" when it turns into a standard stimulus.

[Sarruedi]

Great point. I'm already working on that. My plan is to first sort ROIs based on their orientation‐tuning preferences from the initial tuning block. Once we know which cells prefer 45° versus 90° (and 0°), we can directly compare their responses when those same gratings appear as oddballs versus as standard stimuli. That should let us tease apart pure feature tuning from any expectation/predictability effects. I’ll share the analysis once it's ready

[Dedalus9]

@Sarruedi I wonder if it would help to distinguish whether synapses indeed prefer just one orientation. In my analysis, for instance, I found synapses tuned for multiple orientations: some were tuned to mid-range angles (45 - 90) and others to angles further from the mid-range (0 or 158). So, if you were to sort ROIs based on orientation tuning preferences, perhaps we should distinguish synapses that exhibit just one dominant orientation preference from those that exhibit multiple orientation preferences, as these latter synapses may represent higher-level features spanning multiple orientation preferences.

As to teasing apart pure feature tuning from the effects of the odd-ball block (i.e., expectation/predictability effects), it could be important to look for more complex effects than whether the odd-ball block changes an ROI's preference for just one orientation (e.g., the 45-, 90-, or 0- degree orientations). After all, the odd-ball block may change an ROI's preferences for multiple orientations (e.g., mid-range angles v. "edge angles"). In this latter case, one could argue that the odd-ball block causes tuning changes for higher-level features that span multiple orientation preferences. I address this in part in a recent (and very preliminary) analysis where I compare orientation tunings in the pre- and post- tuning blocks, showing variation across multiple orientation preferences for several ROIs after presentation of the odd-ball block: #92 (comment). This analysis isn't perfect of course -- I will be refining it -- but I think it touches on the topics discussed in this thread.

[Sarruedi]

Hi @Dedalus9, I absolutely agree. That's a really important point. I've been looking into this as well, not just in terms of the preferred orientation, but also considering changes in circular variance and the hierarchical clustering of the responses to capture shifts in tuning sharpness and potential broadening across multiple orientations. I won't be presenting tomorrow, so we have enough time to discuss other issues. How about we try to prepare a summary combining both of our approaches for the week after?

[Dedalus9]

@Sarruedi Circular variance does make sense. The orientation degrees of the gradient spans 0 to 158 and increase in increments of 23 degrees (except for the 158 which is 1 degree greater than then its neighboring orientation, 157 -- I'm not sure why this discrepancy is present @jeromelecoq). Since a visual grating that is 180 degrees would look like one that's 0 degrees, computing the circular variance seems reasonable.

I could prepare a slide that engages explicitly with your work and include this in my presentation. I would be interested to see your work on hierarchical clustering.

[JJYma79]

Analysis Plan

Hi guys, I am considering trying to understand whether at the spine level, it also follows the cortex level compute principle, for example, whether there exists unexpected stimuli that cause large activity based on @maierav @Sarruedi @jeromelecoq @lrudelt @Dedalus9 ’s analysis, if this plan aligns with our current exploration objectives @jeromelecoq.

Specifically, I currently divide this prediction error computation principle into the following 5 steps and subsequent analysis to explore this question:

1. Which of the four types of oddball causes greater activity (current results seem to be static, not sure whether the current result makes sense to our experimental design)

And I will continue to explore subsequent questions based on the current dPCA analysis strategy, which is well-suited for separating task-relevant neural dynamics from mixed population responses(I guess?), allowing us to isolate condition-specific variance (oddball vs. standard stimuli) from time-dependent variance and interaction effects.

2. In our Pre orientation tuning, oddball, post orientation tuning - whether the oddball activity is greater than pre- and post-activity

3. If the hypothesis in 2 holds, whether at the current spine level it presents a bottom-up pattern, or whether there exists a decrease in glutamate levels

4. Whether at the spine level there exists preferential coding of prediction error on a plane-by-plane basis, or more preference for participation in prediction maintenance

5. What are the temporal characteristics of glutamate signals in prediction error generation

Please feel free to point out any misunderstandings or inappropriate usage on my part - I'm always open to feedback and corrections.

The patterns are still not clear for the figure4, will be updated later

[JJYma79]

Hey @Dedalus9 , apologies for the late reply. It's been busy during weekdays and I’m just now getting a chance to revisit this and share some progress.
Yes, that is exactly my intention. I'm currently using dPCA to disentangle variance components associated with time, stimulus identity, and their interaction.
Since this is my first time applying dPCA to real data, I’m intentionally starting with a coarse implementation, despite the availability of excellent methodological descriptions in previous work (links above), to better understand what this method can and cannot reveal currently.
These are the main references I’m currently following:

https://www.nature.com/articles/s41467-024-51750-7

https://elifesciences.org/articles/10989

The new updated analysis supports what was seen in the earlier rough analysis. Time-related dynamics remain consistent across conditions, confirming stable trial progression. Strong responses to static and deviant stimuli support the earlier observation that unexpected inputs drive distinct neural dynamics. Clearer separation in CI space further supports the idea of prediction error–like responses

Here are more explicit marginalization setups:

I marginalize over stimulus conditions to extract time-related variance (i.e., condition-invariant trial progression), marginalize over time to isolate stimulus-specific responses, and include the interaction term to capture condition-dependent temporal modulations—potentially reflecting mismatch or prediction error signals. The goal is to uncover condition-specific neural trajectories in a low-dimensional space while preserving temporal structure, and to test whether unexpected stimuli (e.g., static or rotated gratings) elicit separable population dynamics compared to standard inputs. (These three figures are for DMD2)

Fig. 1 – Stimulus-related dPC1 across conditions

dPC1 projections vary considerably across conditions, especially for the deviant (90° and 45°), static, and blank stimuli.

I think this component currently successfully captures stimulus-specific variance. The deviant (orange) and static (red) conditions evoke the strongest divergence from the standard (blue), consistent with novelty or prediction error responses.

Fig. 2 – Time-related dPC1 across conditions
All conditions exhibit highly similar time-evolving dynamics along this component.

Indicates a robust, shared temporal structure across stimuli — likely reflecting general trial progression.

Fig. 3 – Stimulus × Time interaction component
Clear divergence emerges post-stimulus between conditions, particularly for static and 90°.

Suggests condition-specific temporal modulation; potentially reflects violation detection mechanisms in the oddball structure.

While these three figures revealed condition-specific and shared temporal components in 1D projections, they do not capture how neural states evolve as full trajectories over time. Therefore, we next examined the data in a 3D condition-invariant (CI) space to better characterize the geometry and temporal evolution of neural population responses.

Fig. 4 – CI trajectories decomposed into time-related components
Each condition forms a closed loop trajectory in the CI space, with visible divergence around the stimulus time.

I think these results reflect temporally evolving, condition-specific dynamics; CI space provides meaningful low-dimensional visualization.

Fig. 5 – CI trajectories decomposed into interaction-related components
DMD2 (orange) shows more variability across conditions than DMD1 (blue).

I think these figures suggest DMD2 captures more stimulus-specific modulation; may encode deviance or sensory surprise.

Fig. 6 – CI trajectories decomposed into stimulus-related components
Neural trajectories differ substantially across stimulus conditions, with non-standard stimuli (e.g., 90°, static, blank) producing more irregular and divergent paths in CI space.

These might suggest increased trajectory tangling under deviant or unexpected stimuli, potentially reflecting stronger sensory drive, mismatch processing, or altered internal states.

Fig. 7 – Overlaid DMD1 vs. DMD2 trajectories
Almost identical across conditions, likely time-driven.

Clear divergence by condition, indicating that it carries discriminative stimulus-specific information.

Next Steps for current part:

1. Quantify trajectory differences 1) Procrustes distance to compare overall trajectory geometry 2) Euclidean distance between trajectories at key time points (like stimulus onset) 3) Trajectory tangling index to assess dynamical smoothness
2. CI axis classification 1) Train a simple LDA or logistic regression classifier in CI space to test condition separability. 2) Cross-validation classification accuracy to identify informative dimensions.
3. Time-aligned averaging & trajectory centers: project trial-averaged trajectories into CI space, compute center-of-mass shifts across conditions.

Of course, the current results still require further exploration and validation. I’ll continue reviewing related work, and I welcome any suggestions or feedback you may have.

If you don’t mind, would that be possible to provide a more detailed explanation together additional references and context later? I’m thinking this might be a more careful or robust way to do it. And the code still contains a lot of redundant parts, and I plan to share a cleaner version on Sunday or early next week when I have time. I’d also be grateful if you’re willing to take part or help correct any mistakes I’ve made.

[JJYma79]

And for the second question @Dedalus9, I defenitelly and really think your analysis is impressive. I’m just trying to explore the data using some complementary approaches, which I guess could open up more angles for discussion. I saw your new analyses, they look super cool, and I’ll go through them more carefully on Sunday and reach out with questions if that’s okay. I really appreciate what you've done so far!

[Dedalus9]

@JJYma79 if you have time, perhaps you could share the code or write in latex the steps you took to get the covariance matrices needed for PCA? Alternatively, you could provide the pseudo-code. I tried to replicate your results, but the PCA trajectories I got don't look nearly as clean as yours: I may be looking at metrics completely different than from what you're looking at, however.

[JJYma79]

Hi Nicholas @Dedalus9,
Apologies for the delayed responses again. I've just cleaned up the code to make it clearer to read, but my updates might come a bit slowly. Please feel free to email me anytime as well if anything is unclear; my email is [email protected].

Here is the link to the code: https://colab.research.google.com/drive/1UPuZQQH0ChFb968yulbJjtWA59NZCFxP?usp=sharing

[Dedalus9]

@JJYma79 Hi Yujie, I see where my misunderstanding was. I had thought you were computing dPCA to see how dendritic synapses evolved over trials. In your analysis, however, you were computing dPCA with respect to a dataset that first averages out trials. In effect, you are looking at the average response of each synapse to each stimulus at each time regardless of trial order. I think this analysis won't be useful unless you compare it with the pre- orientation tuning block or post- orientation tuning block because otherwise you can't tell the effect of the odd-ball block. Alternatively, you could not average trial index initially: you could preserve that trial dimension as its own factor in dPCA to see how synaptic responses change over the odd-ball block.