Discussion: meetings/2025-06-17

[hbelski]

Hi all,

Please post any content that you wish to present and discuss on next Tuesday's meeting here.

Thanks,

Hannah

[jeromelecoq]

Here is what I am planning to discuss : https://docs.google.com/presentation/d/1WFPHhe7XP8DtXoCHj_VGkGm4ukBfVv69UaC4gMWOlio/edit?usp=sharing

[Sarruedi]

Maybe we could compute cumulative distributions of aROC values across ROIs at each subsample size to track how discriminability evolves as more trials are included. This would give us a sense of how reliable the separation between standard and oddball becomes across the population of ROIs. We can then look at what trial count results in aROC values exceeding a chosen threshold that works for most ROIs, helping us find a compromise that balances discriminability and the number ROIs.

[lrudelt]

I think I don't understand the concept of aROC here, do you refer to average rate of change? Does that mean the difference in mean between standard and an oddball in this case?

[Sarruedi]

When I said aROC, I didn’t mean “average rate of change”. Instead I meant area under the ROC curve, which tells you how well you can tell two distributions apart.

The idea is this: you take the ΔF/F for each trial, group them into standard and oddball, and use aROC to measure how separable those two distributions are.

Then you subsample different numbers of trials (e.g., 5, 10, 15…) from each condition, repeat that a bunch of times, and calculate the aROC each time. That gives you a curve showing how discriminability improves as you add more trials for each distribution (for each ROI).

It’s essentially a way to estimate how reliably an ROI distinguishes the stimuli depending on how much data you have. Does this make sense?

[lrudelt]

Thanks for the clarification! I was confused about the area under the ROC curve because I was missing a classification model. So one provides the ΔF/F for each trial, and then tries to classify them into the conditions that are compared (i.e., this was an oddball trial, or a standard trial, using a binary classifier)? Then the aROC measures how well that classifier performs and that the data allow to discriminate between the conditions? Sorry, I am absolutely not familiar with this approach.
If this a common way to test the discriminability of two sets of data then this sounds like a good idea.

[Sarruedi]

Exactly, you got it! The nice thing about the aROC approach is that it doesn't require you to assume any specific shape for the underlying distributions. So even if your ΔF/F values aren't normally distributed you can still get a measure of how well you can tell the conditions apart.

Here's an example in which it's been applied to distributions of spike counts in Ephys across different sizes of time bins. The context is different but the visualizations may be useful to understand: https://elifesciences.org/articles/34044

[Sarruedi]

Here’s an initial look at the analysis of orientation tuning across the two tuning blocks. The focus is on comparing preferred orientation and orientation selectivity across synaptic inputs. The goal is simply to gain a clear understanding of the actual data before delving into more complex analyses, such as changes in spatial selectivity across layers or population structure:
Slap2_Overview_Analysis_15062025_OriTuning.pptx

[jeromelecoq]

These orientations tuning plots are very useful ! I am curious to hear your reading of these

[Dedalus9]

Edited on 6/25 to fix bugs in analysis code. All figures have been changed in addition to adding a few more.

Acknowledgments and a Note to Collaborators

This post may interest @JJYma79. This analysis was done prior to their posting, but I think there is overlap between our analyses. For distinguishing odd-blocks and receptive field-blocks, it builds on code written by @maierav.

In this analysis, I sometimes speculate about the interpretation of figures. These speculations are intended as hypotheses that can be followed up on in future analyses.

Introduction

To preface, when analyzing orientation tuning, I believe we should remain agnostic to whether any synapse prefers any one orientation. In Figure 1, for instance, it is evident that the orientation tuning vector of certain synapses exhibits preferences for multiple orientations and after the odd-block they can also change their preferences for multiple orientations. This suggests that the effect of the odd-block on the orientation tunings of individual synapses may involve more than simply making them respectively exhibit greater preference for the unexpected stimulus and diminished preference for the expected one. Rather, it may well be the case that preferences for unexpected stimuli over diminished ones are encoded by population of synapses rather than by individual ones. It is this intuition that I explore in the analysis below.

Brief Overview of Metrics and Preprocessing

Computation of the Orientation Tuning Vector

The orientation tuning vector of each ROI was computed by averaging the response of each ROI to each orientation of the visual stimulus for the 0.5s following its presentation.

How I Addressed the Impact Photobleaching

This analysis was conducted on datasets normalized by $F_0$. Moreover, because photobleaching may complicate interpretation of the amplitude of the averaged response of each ROI to the stimulus for the 0.5s following its presentation, I chose to focus my analysis instead on the comparison of the shape of the orientation tuning vector of each ROI regardless of whether one or another ROI may exhibit in its recorded signal more or less amplitude. To do this, I performed I implemented these preprocessing steps:

1. ROI Selection by Pre-Block Tuning Spread (Top 25%)

To focus my analysis on synapses that exhibited clear tuning prior to the presentation of the odd-block, I selected ROIs on the basis of their "tuning spread" in the pre-block. I computed that spread as follows:

$$\Delta_i = \max_j x_{i,j} - min_j x_{i,j}$$

where $x_{i,j}$ is ROI $i$'s response to orientation $j$. I then kept only those ROIs whose $$\Delta_i$$ lay in the top 25% of the spread distribution. This helps ensure that (1) we analyze only ROIs with clear orientation preferences and (2) subsequence min-max scaling does not disproportionately amplify noise. One caveat is that if an ROI loses all orientation preferences despite having exhibited clear preferences in the pre-block, then we'd expect subsequent min-max scaling to amplify noise rather than any true orientation tuning.

2. Min-max scaling

I min-max scaled the orientation tuning vector of each ROI as follows:

Let $x_{i,j}$ be the response of ROI $i$ to orientation $j$. We define the minimum and maximum of each that ROI's orientation vector as,

$$m_i = \min_{1 \leq k \leq m} x_{i,k}$$
$$M_i = \max_{1 \leq k \leq m} x_{i,k}$$
where $k$ corresponds to the index for each orientation, $m_i$ to the minimum response of ROI $i$ across all $m$ orientations, and $M_i$ to the maximum response of ROI $i$ across those same orientations.

Then the normalized response for that ROI, $\tilde{x_{i,j}}$ can be defined as,

$$\tilde{x_{i,j}} = \frac{x_{i,j} - m_i}{M_i - m_i}$$
with $i = 1, ..., N$, $j = 1,..., m$. This normalization step rescales that ROI $i$'s entire tuning curve into the $[0,1]$ range.

Figures

Figure 1: Radar Plots of Pre- vs Post-Block Orientation Tuning for Preprocessed ROIs

Figure 1.1: DMD1

Figure 1.2: DMD2

Take-away: Orientation tuning vectors vary between ROIs. Some ROIs exhibit a preference for multiple orientations rather than just one orientation. Some ROIs change their preference for multiple orientations after the presentation of the odd-block. Why some ROIs change their preference for multiple orientations after the presentation of the odd-block rather than for just the deviant or standard stimuli admits of multiple answers. First, it may well be the case that the odd-block was not long enough to create lasting changes in the tuning preferences of synapses. Second, synapses may be more or less tuned to discriminate deviant from standard stimuli, but this won't be evident by analyzing the tuning vectors of individual ROIs but will only become evident after analyzing the tuning vectors of populations of ROIs or synapses along the dendritic branches. Finally, tuning changes for multiple orientations may reflect different features encoded by the same synapse, e.g., uncertainty. In any case, a population level analysis may be revealing.

Figure 2: The presentation of the odd-ball block changes the orientation tuning vectors of ROIs in a (seemingly) structured way

Because orientation tuning vectors seem to vary their preferences for multiple orientations after the presentation of the odd-block, I sought next to determine how orientation preferences in the pre-block were changed by the odd-block. These were a few questions that came to mind:

1. Did some ROIs increase their preference for particular orientations and others for multiple orientations?
2. For those that increased their preference for multiple orientations, which orientations did they prefer more?
3. Did the pre-block orientation preferences of an ROI impact how they would change their orientation after the odd-block's presentation?

To answer these questions (or approach an answer to them), I conducted a principal components analysis (PCA). To do this, I constructed a cross-covariance matrix, each row of that matrix corresponding to how orientation tuning the pre-tuning block varied with orientation tuning in the post-block across ROIs. I then derived PC components from this cross-covariance matrix.

Figure 2.1: Cross-Covariance Matrices of DMD1 + DMD2

Figure 2.2: DMD1, PC Loadings from PCA of Cross-Covariance Matrix

Figure 2.3: DMD2, PC Loadings from PCA of Cross-Covariance Matrix

Figure 2.2 + 2.3 Description: This figure shows how of the total cross-covariance is captured by the first four PC components in DMD1and DMD2. It also shows the loading values for each PC component. Each loading value shows the weight that a given grating angle contributes to the corresponding PC of the pre-vs-post cross-covariance matrix. These PCs are mutually orthogonal, uncorrelated axes that reveal the dominant, structured ways in which tuning vectors shift following the odd-ball block.

Take-away: For both DMD1 and DMD2, the PC loadings of the first PC component exhibit clear, non-random loading patterns, indicating that the odd-ball block may drive synaptic tuning along a low-dimensional manifold in orientation space. In other words, rather than a single uniform gain change, ROIs remap their preferences in several independent directions. I will probe this hypothesis further in future analyses.

Figure 3: Rank-1 Reconstruction of Pre-/Post-Tuning Block Cross-Covariance Matrix from Individual Principle Components

Figure 3.1: DMD1

Figure 3.2: DMD2

Figure Description: Each panel shows the matrix:

$$\lambda_i u_i u_i^T$$

for principle component $u_i$, with rows corresponding to pre-block orientations and columns to post-block orientations. Above each heatmap I list the eigenvalue $\lambda_{i}$ and the fraction of total cross-covariance it explains. PC1 captures the dominant, population-wide shift in the shape of orientation tunings.

Figure 4: Pre 45°/90°/0° vs. Post-block Tuning After Rank-2 Reconstruction from PC components

Figure 4.1: DMD1, Rank-2 Reconstruction of the Cross-Covariance Matrix

Figure 4.2: DMD2, Rank-2 Reconstruction of the Cross-Covariance Matrix

Figures 4.3 - 4.8 Descriptions: Each panel plots every ROI's rank-2 reconstructed tuning: the centered pre-block response (x_axis) at either 45°, 90°, or 0° versus the centered post-block response at the indicated orientation (y-axis) where the original data have been approximated using only the first two principal components of the pre-post cross-covariance matrix. The dashed gray line is the identity (perfect stability), the solid orange line is the zero-intercept OLS fit (with 95 % CI shaded), and the annotated slope quantifies how strongly the top-2 covariance modes preserve (slope≈1), attenuate (0<slope<1), or invert (slope<0) tuning at each orientation when viewed through this low-dimensional reconstruction.

Note on Reconstruction By reconstructing each ROI's tuning curves from the pre- and post- tuning blocks using only the top two covariance modes, we are in effect projecting them onto that subspace spanned by the population's dominant cross-covariance modes. In other words, if an ROI's pre-block/post-block orientation tuning exhibits features orthogonal to these modes, they are filtered out; if they exhibit features aligned with them, they are retained. For this reason, reconstructing them using the population's dominant cross-covariance, we can see how each ROI's tuning changes along this manifold (if at all).

Figure 4.3: DMD1, Pre 45° vs. Post-block Tuning After Rank-2 Reconstruction from PC components

Figure 4.4: DMD1, Pre 90° vs. Post-block Tuning After Rank-2 Reconstruction from PC components

Figure 4.5: DMD1, Pre 0° vs. Post-block Tuning After Rank-2 Reconstruction from PC components

Figure 4.6: DMD2, Pre 45° vs. Post-block Tuning After Rank-2 Reconstruction from PC components

Figure 4.7: DMD2, Pre 90° vs. Post-block Tuning After Rank-2 Reconstruction from PC components

Figure 4.8: DMD2, Pre 0° vs. Post-block Tuning After Rank-2 Reconstruction from PC components

Figure 5: Pre 45°/90°/0° vs. Post-block Tuning Without Rank-2 Reconstruction from PC components

Below are scatter plots with the original data, i.e., the data that has not been reconstructed with the first two principal components of the cross-covariance matrix.

Figure 5.1: DMD1, Pre 45° vs. Post-block Tuning Without Rank-2 Reconstruction from PC components

Figure 5.2: DMD1, Pre 90° vs. Post-block Tuning Without Rank-2 Reconstruction from PC components

Figure 5.3: DMD1, Pre 0° vs. Post-block Tuning Without Rank-2 Reconstruction from PC components

Figure 5.4: DMD2, Pre 45° vs. Post-block Tuning Without Rank-2 Reconstruction from PC components

Figure 5.5: DMD2, Pre 90° vs. Post-block Tuning Without Rank-2 Reconstruction from PC components

Figure 5.6: DMD2, Pre 0° vs. Post-block Tuning Without Rank-2 Reconstruction from PC components

Conclusion and possible future directions

Tentative Conclusion

The odd-block seems to induce tuning changes not as a uniform gain shift but instead along a low-dimensional manifold in orientation space, suggesting that the odd-ball stimulus may induce changes along a small number of orthogonal, uncorrelated dimensions at the population level rather than through high-dimensional reshaping (i.e., for particular orientations). Moreover, what changes occur seems to depend on what tuning curve that ROI exhibited in the pre- orientation tuning block, e.g., an ROI that strongly prefers the 0-degree orientation won't exhibit strong preference for 90-degree orientation after the odd-block's presentation.

Future Directions I would like to pursue / that can be pursued

1. Changes across orientations may be better explained if we accounted for non-linear or multi-way interactions among orientation preferences. It may therefore be worthwhile to explore the application ICA or tensor-PCA. I have, however, some preliminary results from applying ICA and have generated similar results as above, suggesting that there may not be non-linear or multi-way interactions among orientation preferences. Further analysis would be necessary, however.

2. Map each ROI’s position and dendritic branch morphology to its loading on PC 1 vs. PC 2, testing whether spatial clustering underlies specific change modes. We could also link reconstructed tuning shifts to somatic firing or calcium activity to assess how dendritic remapping impacts output.

3. Analyze time-resolved tuning dynamics to see how quickly and along which axes ROIs enter/exit the cross-covariance manifold during stimulus blocks. This is related to the analysis of @JJYma79.

4. Develop an empirical statistical model (e.g., a fully Bayesian hierarchical model) to distinguish group-level effects from ROI-to-ROI variability around group-level effects. This would compensate for a deficiency in the above analysis: it only captures population level variability. Also, I discussed statistical modeling with @lrudelt, so this suggestion is related to that discussion. We could make a discussion thread dedicated to this if this would be of interest to anyone else.

[Sarruedi]

Thanks a lot for uploading your analysis @Dedalus9. Just to preface today's meeting, while I get the idea of staying agnostic about whether any given synapse prefers a specific orientation, I’d actually argue the opposite for today's meeting. I think tuning preferences and how these can shift with experience is a fascinating question. I'm not pre-selecting ROIs with strong tuning either, because I think it's an important question to ask whether synapses that start off with weak or no clear tuning can become biased through something like repeated exposure to the standard stimulus. That kind of plasticity tells us potentially interesting things about experience-dependent plasticity at the level of synaptic integration.

[Dedalus9]

Slide Presentation:
Oweiss_Lab_Nicholas_Rodriguez_6_17_Presentation.pptx

Note: will edit to include a slide on analysis caveats, which I mainly referred to verbally rather than in writing during my presentation

[Dedalus9]

@Sarruedi I agree that it may well be the case that synapses that exhibit no clear tuning may develop clear tuning after the odd-block's presentation. So, we could perhaps divide the dataset into two subsets: those ROIs that exhibit clear tuning and those that don't, evaluating then how their tuning changes if in the odd-block they exhibit clear tuning.

I also need to investigate whether there are ROIs that exhibit clear tuning but no longer exhibit clear tuning after the odd-ball block. If that is indeed the case, that would a significant caveat for my analysis because I assume that ROIs which exhibit clear tuning in the pre- orientation tuning block retain clear tuning in the post- orientation tuning block. If this is not the case, then the min-max scaling step I perform on the orientation tuning vectors of ROIs in the post- orientation tuning block would be problematic: it would amplify noise giving the impression of clear tuning when there isn't.

[Sarruedi]

@maierav @Dedalus9 it would be great to rerun your analysis on the data once the orientation/direction issue we discussed yesterday is fixed. Ideally we fix the original script so it's not an issue in the future. Just to make sure we are all on the same page

[Dedalus9]

Sounds good. I am also working on developing python classes that will make it easier to load data including multiple experimental sessions. I can provide this later this week if anyone would find this interesting. If it seems helpful, we could push it to this GitHub repository for others to use.

@jeromelecoq What do you think?

[Dedalus9]

I have rerun my analysis. I am editing figures @Sarruedi. I will include a link to a GoogleColab notebook with my code as a comment to my initial post in this thread.