Discussion: meetings/2026-02-24

[jeromelecoq]

Material for the meeting;
Here is the latest version of the figure 2

Link to the experimental google sheet :
https://docs.google.com/spreadsheets/d/1wAeloFJgvRjrseoVeNm4YQd8BezGWRon-Z-b1iJAz9c/edit?usp=sharing

[jeromelecoq]

Here is a proposed poll structure based on our discussion today :

# Google Form draft — Predictive Processing OpenScope Analysis Priorities

## Form description
We’re prioritizing community analyses for the OpenScope Predictive Processing dataset, based on the **Analysis Plan** in arXiv:2504.09614.

This poll is meant to:
- Pick **what to do first**
- Identify **who wants to contribute**

Please answer based on what you can realistically start soon on incoming data.

---

## Section 1 — About you

1. **Name** (Short answer) *Required*
2. **Email** (Short answer) *Required*
3. **Time you can commit (next 8–12 weeks)** (Multiple choice) *Required*
   - <2 hrs/week
   - 2–5 hrs/week
   - 5–10 hrs/week
   - 10+ hrs/week
4. **Modalities you are experienced working with** (Checkboxes)
   - Neuropixels / spiking
   - Calcium imaging (2p)
   - Voltage imaging
   - Glutamate imaging
   - Behavior / running / closed-loop variables
   - Computational / neuronal models

---

## Section 2 — Choose priorities (what we schedule first)

### Analysis options (use these as the answer choices in the next questions)

Below, the choices are organized under the paper’s **Analysis Plan** headers and use the paper’s numbering.

### I. What kind of information is encoded by mismatch responses?

**I-1 — Analysis #1: Single-neuron mismatch detection (ETA/PSTH + bootstrap)**
- For each neuron × mismatch stimulus: construct ETA (imaging) or PSTH (Neuropixels) and test significance against the appropriate control per session (bootstrap; e.g., p<0.01).

**I-2 — Analysis #2: Novel vs control comparison (multiplicative/additive/subtractive/no-effect)**
- Scatter plot novel vs control responses + linear fit (slope/offset interpretations).

**I-3 — Analysis #3: Compare different mismatch stimuli (Sessions 1–2)**
- Compare relative responses across mismatch stimulus identities; test expected relationships (e.g., orientation change magnitude), fraction-of-neurons + shuffle tests; depth/subregion effects.

**I-4 — Analysis #4: Decoding novelty per se vs mismatch identity (Sessions 1–2)**
- Population decoding / information metrics (including mutual information); compare novelty-per-se vs individual mismatch identity; performance vs #neurons.

### II. Distinguish between two categories of prediction made by neurons

**II-1 — Analysis #1: Closed-loop vs open-loop vs control comparisons (Session 2)**
- For the same mismatch stimulus: compare closed-loop vs open-loop and control vs open-loop to distinguish detailed predictions vs deviation-from-ensemble (adaptation).

**II-2 — Analysis #2: Decoding/info for mismatch identity vs novelty (Session 2)**
- Use population decoders (e.g., linear/SVM) and/or mutual information to quantify encoding of individual mismatch identity vs novelty per se; test differences across closed-loop/open-loop/control.

**II-3 — Analysis #3: Emergence of prediction signals (single neuron + population; within session & across days)**
- Trial-by-trial response changes; model fits (exponential/linear); bootstrap validation.
- Population latent analysis (e.g., TCA), manifold/geometry shifts; cross-day changes.

### III. Mismatch responses across different types of predictions

**III-1 — Analysis #1: Map locations of mismatch neurons (2p + Neuropixels; region/layer/cell type)**
- Map density/% of mismatch neurons by region/layer/cell type; timing variability and peak latencies; functional clustering; coding subspace geometry.

**III-2 — Analysis #2: Same-neuron comparison — sensorimotor (Session 2) vs temporal sequence (Session 3)**
- Compare mismatch strengths and patterns across these prediction types; scatter plots to test common vs distinct circuitry.

**III-3 — Analysis #3: Same-neuron comparison — oddball (Session 1) vs sequence (Session 3)**
- Compare magnitudes/patterns across sessions (including ANOVA-style structure described in plan); scatter plots.

**III-4 — Analysis #4: Inhibitory interneuron analyses (sessions 1–3; emphasis on Session 2)**
- Compare inhibitory activation patterns; closed-loop vs open-loop differences; extend to sessions 1 and 3.

**III-5 — Analysis #5: Temporal mismatch analysis (Session 4)**
- Test timing-based mismatch effects (baseline/evoked under control vs temporally deviant stimuli); interneuron contributions to predictive timing.

**III-6 — Analysis #6: Test various prediction models across session types (modeling-heavy)**
- Within-session and across-day changes; autocorr/cross-corr; self-supervised representations; model comparison with cross-validation and cross-lab validation.

---

5. **Suggestions / edits to the analysis list above** (Paragraph)
Examples: split/merge items, clarify wording, add a missing analysis.

---

## Section 3 — Rank analyses (science vs personal)

6. **Science priority: rank the analyses (most important science-first → least)** (Ranking / ranked list) *Required*
- Choices: I-1…III-6

7. **Personal preference: rank the analyses you’d most like to contribute to (most → least)** (Ranking / ranked list) *Required*
- Choices: I-1…III-6

---

## Section 4 — Publication bundles (read first)

### Publication bundle options (built from the analysis items above)

**B1 — Novelty coding taxonomy (Hypothesis I)**
- Includes: I-1 to I-4

**B2 — Detailed prediction vs ensemble deviation (Hypothesis II)**
- Includes: II-1 + II-2 (and related pieces of I-4)

**B3 — Emergence/learning of prediction signals**
- Includes: II-3

**B4 — Common vs distinct mechanisms across mismatch types (Hypothesis III)**
- Includes: III-1 to III-5

**B5 — Integrative modeling / model comparison**
- Includes: III-6

8. **Suggestions / edits to the publication bundle list above** (Paragraph)
Examples: missing bundle theme, bundles that should be split/merged.

---

## Section 5 — Rank publication bundles (science vs personal)

9. **Science priority: rank the publication bundles (most important science-first → least)** (Ranking / ranked list)
- Choices: B1–B5

10. **Personal preference: rank the publication bundles you’d most like to contribute to (most → least)** (Ranking / ranked list)
- Choices: B1–B5

---

## Section 6 — Final comments

11. **Any comments / constraints we should consider when scheduling?** (Paragraph)
Examples: data access limitations, tooling needs, coordination needs across labs, figures you think are “must-haves” early.

[jeromelecoq]

@koweiss @maierav @Sarruedi maybe make comment on this poll. It is constructed from the paper analysis plan which is quite detailed already.

[maierav]

Looks great. Thank you for all that effort, Jerome!

Just one question of clarification: "Section 4 — Publication bundles" - does this refer to grouping the analyses outlined above into discrete papers?

[koweiss]

@jeromelecoq Thanks for putting this together, Jerome.
One general question I have: has the analysis done so far for different sessions/conditions been sufficient to refine any of the hypotheses listed (or that were in the original paper)? I'm asking because testing various predictions of these hypotheses might require different ways to analyze the same data which should be taken into account. Ideally, for people to commit time and effort to one (or more) type of analysis, each model prediction should be clearly spelled out along with each analysis proposed so people can measure success/milestone completion. Additionally, it would be nice to have each analysis tied to elements of Figure 2 above. Currently, this seems to be only listed under "III-6 — Analysis #6: Test various prediction models across session types (modeling-heavy)".

[Nina1337]

Thank you Jerome for the detailed work. I am happy it is time for analytics now. My question is if you are interested in machine learning models with post hoc xai method as this is not described here but my labs traditional way of working.

Best, Nina

[Alishamsnia]

Hey,
I hope you are doing well. I wanted to share the latest updates of the analysis, also based on the proposed analysis roadmap. To recap all the analysis I have done, I am attaching all of them to this message. I am also tagging Dr. Najafi @farznaj, @najafi-laboratory.
1. Preprocessing
I applied preprocessing. Our preprocessing pipeline consisted of four sequential steps designed to isolate physiological signals from noise while minimizing distortion:

a. 60 Hz Notch Filter: First, we applied a zero-phase 60 Hz notch filter (using signal.filtfilt) to eliminate electrical power-line noise without disrupting gamma-band activity or broadband neural structures.

b. High-Pass Filter: Next, we removed slow signal drifts using a high-pass Butterworth filter with a 0.5 Hz cutoff; this filter was specifically chosen for its maximally flat passband, which prevents artificial ringing and ensures stable, continuous neural frequencies.

c. Common Average Referencing (CAR): Third, we implemented Common Average Referencing (CAR) to subtract non-neural environmental and motion interference shared globally across the recording channels.

d. Artifact Detection: Finally, we conducted artifact detection using a conservative z-score threshold of 5σ. This approach targets extreme, non-physiological amplitudes to minimize false positives, allowing us to accurately identify and reject both unreliable channels and globally contaminated time points.

The attached files are step by step effect of the preprocessing.




2. Alignment
Then I aligned each of the LFP on the stimulus time stamp (provided in the dataset) just to make sure everything makes sense:
alignment_report_820459_2025-11-12_PREPROCESSED.pdf

3. Spectrogram Analysis
Then for the follow-up analysis I applied spectrogram analysis for each channel based on the depth and different frequency bands (Alpha, Beta, etc). The overall objective of this depth-resolved analysis is to characterize how distinct neural oscillations and stimulus-evoked responses are distributed across cortical layers using high-density Neuropixels recordings.
To achieve this, the methodology employs four complementary visualizations per brain region:

• Time-Frequency Spectrogram: The first figure provides a broad view of dynamic spectral power shifts across all frequencies relative to stimulus onset, divided into deep, middle, and superficial layers to capture global state changes.

• Power Mapping: The second figure maps the power of specific canonical frequency bands (such as Delta, Theta, or Gamma) directly onto the physical channel array, offering high-resolution spatial localization to pinpoint exactly which anatomical depths generate distinct rhythms.

• Baseline Power quantification: The third figure isolates and quantifies the absolute baseline power of these frequency bands across the depth categories, establishing the underlying rhythmic architecture of the tissue layer by layer.

• Power Change Tracking: Finally, the fourth figure tracks the percentage of power change over time for each specific band, effectively measuring event-related synchronization and desynchronization to reveal the precise temporal dynamics of how sensory stimuli actively modulate these layered neural circuits.
  spectrogram_depth_analysis (1).pdf

4. Receptive Field and Orientation Tuning Analysis
Finally, based on the provided analysis roadmap, I performed receptive field and orientation tuning analysis.

A. Receptive Field Mapping
This analysis estimates the visual receptive fields (RFs) of neurons recorded with Neuropixels probes using the RF mapping block of the OpenScope dataset.

• Dataset & Selection: In this session there were 1215 RF mapping trials and 3172 spike-sorted units, of which 817 passed quality control filters and were used for analysis.

• Stimulus Presentation: During the RF mapping block, a visual stimulus was presented at 81 spatial positions arranged in a 9×9 grid, with X positions and Y positions ranging from −40 to 40 (in steps of 10). These X and Y values represent locations of the visual stimulus on the monitor in visual space, not electrode positions in the brain.

• Evoked Response Construction: For each neuron, spikes were aligned to stimulus onset and firing activity during the stimulus response window was compared to a baseline period before the stimulus. The difference between stimulus-driven firing and baseline firing provides the evoked response.

• RF Map Interpretation: By averaging the evoked response across trials for each stimulus location, we obtain a receptive field map, where the X axis represents horizontal stimulus position and the Y axis represents vertical stimulus position in the visual field. Each pixel in the RF map therefore indicates how strongly a neuron responds when the stimulus appears at that particular spatial location. The color scale represents the baseline-subtracted firing rate (Hz), where positive values indicate increased firing and negative values indicate suppression relative to baseline.

• Key Metrics: The RF center is defined as the spatial location with the strongest response in the smoothed RF map, and the peak RF response is the magnitude of that maximum response.

• Statistical Significance: To determine whether a neuron is visually responsive, we statistically compare firing rates during the stimulus period to baseline firing rates; this produces a visual responsiveness p-value, where neurons with p < 0.05 are considered visually responsive.

• Population Results: Across the population, we observed that most neurons have small peak RF responses (generally below ~5 Hz) while a smaller subset shows stronger visual responses. The p-value distribution indicates that 72 neurons are significantly visually responsive in this dataset.

• Cortical Depth Variation: We also examined how RF strength varies with cortical depth by plotting peak RF response versus neuron depth (X axis: probe depth in micrometers, Y axis: peak RF response in Hz). This allows us to visualize how visually responsive neurons are distributed across cortical layers.
  Overall, this RF analysis provides three key metrics for each neuron: the location of its receptive field center in visual space, the strength of its visual response, and whether the neuron is statistically responsive to visual stimuli, which together form the baseline characterization needed before performing more advanced analyses such as orientation tuning or mismatch response analysis.
  rf_maps_selected_units.pdf

B. Orientation Tuning
This analysis measures orientation tuning of neurons recorded with Neuropixels probes using the Sequence Mismatch block of the OpenScope experiment.

• Dataset & Selection: In this session there were 6240 stimulus presentations and 3172 recorded units, of which 817 units passed quality control filters and were used for the tuning analysis.

• Trial Filtering: The sequence experiment contains several trial types (standard, omissions, mismatch conditions, etc.), but for estimating orientation tuning we only used standard trials to avoid contamination from prediction or mismatch effects. This resulted in 4852 trials used for tuning.

• Stimulus Details: The stimuli in this block were presented with three orientations: 0°, 45°, and 90°, which correspond to different orientations of the visual grating stimulus.

• Evoked Response Construction: For each neuron, spikes were aligned to stimulus onset and firing activity during the stimulus response window was compared to a baseline period before stimulus onset. The difference between stimulus firing and baseline firing gives the evoked response, which represents how strongly a neuron responds to that orientation relative to its spontaneous activity. The orientation tuning curve is constructed by averaging this evoked response across trials for each orientation.

• Interpreting the Figure: In the orientation tuning figure, the X axis shows the stimulus orientation in degrees, while the Y axis shows the baseline-subtracted firing rate in Hertz (Evoked − baseline rate). Each point represents the mean neural response to that orientation, and the error bars represent variability across trials.

• Quantitative Metrics: Several quantitative metrics are computed to summarize tuning. The preferred orientation is the orientation that produces the strongest response, while the Orientation Selectivity Index (OSI) measures how selective the neuron is to a particular orientation; OSI ranges from 0 (no selectivity) to 1 (strong selectivity for a single orientation).

• Statistical Significance: Statistical significance of orientation tuning is assessed using a one-way ANOVA comparing responses across orientations, producing an orientation p-value. If the p-value is small, it indicates that the neuron responds differently to different orientations and is therefore orientation tuned.

• Example Unit: For example, in Unit 9 the neuron shows a strong preference for 90° orientation, with OSI = 1.0, indicating very strong selectivity, and ANOVA p ≈ 2.4 × 10⁻¹⁵, indicating highly significant tuning.

• Population Results: Across the population of neurons analyzed, 252 units were classified as orientation tuned using the criteria p < 0.05 and OSI > 0.2, meaning these neurons show reliable orientation-dependent responses.

The orientation tuning analysis therefore provides key information about feature selectivity of visual cortical neurons, identifying which neurons encode specific orientations of visual stimuli and quantifying the strength and statistical reliability of that tuning.

6. Reference Materials
For your reference I added my notebook. There are more detailed metrics for each analysis. You can scroll down the notebook to the last cells for the RF and orientation tuning analysis:
intro_to_ephys_nwbs.ipynb