Agenda

Round-table update on analysis for the Predictive processing data release manuscript

Please submit 2 slides maximum prior to the meeting at the discussion link at the bottom of this meeting page. Any analysis is welcome, including from new participants.

At the onset of the meeting, we will review the number of presentations to go over as well as introduce any new comer to the meeting.

Meeting Recording

.

Meeting Notes

SLAP 2 Glutamate Signal-To-Noise Analysis: Ido presented a quality-control analysis of SLAP 2 glutamate recordings, and Jerome and Maedeh agreed that the work could support the data-release paper after aligning the SNR metric and regenerating the figures.

Analysis Scope: Ido analyzed SLAP 2 glutamate recordings because a corresponding quality-control panel was missing from the existing figure. The analysis examined recording planes, proximal and apical dendrites, individual synapse traces, detected glutamate-release events, and signal-to-noise groupings.

Event Detection: Ido estimated the noise level using mean absolute deviation and classified detected peaks by their number of standard deviations above the noise floor. Synapses were then grouped with two-dimensional k-means using the proportions of events below two standard deviations and above four standard deviations; top-left points represented higher-SNR synapses and bottom-right points lower-SNR synapses.

Session And Compartment Effects: Ido observed that glutamate-plus-calcium sessions generally appeared to have higher SNR than glutamate-only sessions and that more recent recordings contained more high-SNR synapses. Maedeh noted that expression quality, microscope maintenance, ROI selection, imaging rate, and more uniform sampling could all contribute to these differences; she also cautioned that dendritic location may not be the primary explanation.

Metric Alignment: Ido had used different SNR definitions for glutamate-only and glutamate-plus-calcium sessions to emphasize their apparent separation. Jerome said the metric should be the same across conditions, while Sarah suggested that a common continuous or discrete visualization could eventually be selected across modalities.

Data Release Action: Maedeh agreed to share the existing SNR function and example-trace approach with Ido. Jerome asked Ido to submit the analysis and reproducible code as a pull request; after the metrics are aligned across modalities, the team will merge the analyses into a common figure for the data-release paper.

Neuropixels Mismatch Response Analysis: Roberto examined population responses to motor halt, omission, and orientation oddballs, while the group identified running state, control blocks, timing, anatomical location, and cell type as necessary factors for deeper analysis.

Initial Response Patterns: Roberto plotted PSTHs from Neuropixels sensorimotor sessions across isocortex, hippocampus, thalamus, and other regions. Orientation changes produced the strongest response overall, omissions produced smaller responses, and motor halts were usually nearly silent except in a small number of cases.

Behavioral Context: Jerome asked whether the analysis separated mice that were running from mice that were stationary. Roberto had initially applied standard unit-QC filters only, not behavioral sorting. Jerome explained that the motor halt should be interpreted in the context of whether the mouse was moving because the grating was coupled to disk movement; Roberto agreed to divide the analysis by running and standing state.

Stimulus Design: Jerome clarified that the grating normally drifted according to disk speed. A motor halt uncoupled the grating from the disk, an omission made the display gray for approximately 250 milliseconds, and orientation oddballs changed the grating orientation. Samuel noted that halt and omission are brief transient disruptions whereas orientation is a renewed drifting stimulus, so comparisons should account for response timing and appropriate control blocks.

Further Stratification: Roberto quantified mean firing-rate changes and found larger omission responses in the motor-trained cohort, while the orientation effect remained strongest. He also found that layer 2/3 units showed a clearer cohort difference than pooled units and indicated that analyses by layer, neuron type, and individual unit should be expanded.

Required Follow-Up: Sarah recommended scatter plots comparing each unit's standard-condition and oddball activity to estimate the fraction and magnitude of affected cells rather than relying only on population averages. Jerome asked Roberto to compare mismatch responses with control blocks and inspect behavioral context; Roberto agreed that the exploratory analysis needed more work and would share the code and associated figure.

Interactive PSTH Exploration: David presented an interactive Neuropixels viewer for examining mismatch and control responses under multiple stimulus contexts, and Jerome, Roberto, and Sarah supported merging its exploratory and teaching value into the data-release manuscript.

Viewer Functionality: David's viewer allows users to select oddball, sensorimotor sequence, and duration experiments; choose mismatch events; filter by area, probe, unit type, and QC status; and display raw firing rate, baseline-subtracted or z-scored responses, individual-unit traces, and heat maps.

Unit Sorting: The viewer supports sorting units by depth, time to first peak, response magnitude, or Rastermap ordering. Mismatch and control blocks can be displayed using the same unit order so users can visually compare responses within a selected population.

Baseline Handling: David emphasized that baseline selection must be handled carefully because an automated choice such as the preceding two seconds can include the prior stimulus. Different stimulus designs required different baselines, including a shorter preceding interval or a trial farther back for the duration condition; he warned that AI-generated analysis code must be checked for such assumptions.

Manuscript Role: Jerome described the viewer as a didactic figure that teaches readers how raw data, baseline normalization, and control comparisons reveal an oddball response. Sarah agreed that the first purpose should be helping readers understand the data and analysis choices rather than presenting a single definitive oddball metric; example responsive neurons could be added to demonstrate that mismatch-sensitive cells exist.

Implementation Plan: David said he had a branch ready for a pull request and had precomputed matrices for the interactive display. Jerome asked the team to merge the contribution so others could test it and provide feedback. Carter noted that the visualization requires data snapshots, and Jerome confirmed that the precomputed material can be committed to the repository.

Electrophysiology Quality Control: Sarah presented work from Marina and Yuheng characterizing Neuropixels recording quality, probe stability, depth profiles, waveform classes, and anatomical sampling, with follow-up checks requested for depth labeling and unit classification.

SNR Stability: Sarah reported SNR distributions for units already classified as good using established quality metrics. An SNR threshold around three or four would generally retain reliable units, and comparisons of the first and last sessions showed that recordings were broadly stable, although the final session often shifted slightly toward lower SNR.

Probe Depth Profiles: The team examined firing-rate distributions across acute probe insertions and probes A–F. Sarah noted stereotyped insertion and brain-surface locations across animals, while Severine questioned units appearing beyond the expected depth range and recommended checking channel locations against the CCF and identifying channels outside the brain.

Noise Verification: Severine reported that units at negative or otherwise invalid depths had previously been checked against the CCF and were all classified as noise. Sarah said she would verify whether the presented plots truly contained only single units or whether a filtering criterion had been missed.

Waveform Classification: Marina used WaveMAP to classify waveform shapes and visualize electrophysiological properties such as firing rate, amplitude, waveform measures, and SNR across brain areas. The analysis was intentionally permissive in its number of classes because recordings covered prefrontal and visual cortex, hippocampus, thalamus, and striatum.

Cell-Type Validation: Jerome suggested comparing waveform classes with SST tags as a sanity check and examining responses to the zebra stimulus, which was designed to separate inhibitory populations such as VIP, SST, and PV cells by their spatiotemporal and orientation-tuning properties. Sarah agreed to investigate, while she and Severine noted that the identity of a high-firing or narrow-spike class still required checking.

Data Access: Sarah said Marina had downloaded the relevant files and locally stored unit metrics to avoid repeatedly streaming the data. Jerome asked for the file size so the team could determine whether the metrics could be committed to the repository; Sarah agreed to confirm the size and details.

Quality Control And Analysis Ordering: Samuel raised the risk that strict QC could exclude biologically important but variable neurons, and Sarah and Stefan distinguished exploratory discovery from formal QC, with Stefan agreeing to contribute a manuscript paragraph requiring QC before inferential analysis.

Potential Selection Bias: Samuel asked how the team could avoid quality thresholds being dominated by large numbers of neurons that are irrelevant to a particular computation and accidentally rejecting neurons carrying meaningful oddball-related signals, especially if those neurons have lower or more variable SNR.

Two Analysis Directions: Sarah distinguished the goals of the data-release paper from exploratory science. For release-level QC, she would report SNR distributions, unit counts, and recording reliability to establish dataset quality; for discovery, she would first use tools such as David's and Roberto's viewers to find responsive neurons and then inspect their properties and cell types.

Strict QC Sequence: Stefan recommended a strict ordering in which quality control is completed before analysis of interesting effects. He warned that low-SNR units, long-tailed distributions, and other statistical artifacts could create spurious oddball-like responses if researchers selected interesting units before filtering.

Manuscript Contribution: Jerome asked Stefan to turn this recommendation into a paragraph for the manuscript. Stefan agreed, and Jerome said he would follow up by email with guidance on where to begin.

Next Meeting Analysis Contributions: Jerome closed the session by inviting additional modality analyses, and Antonio agreed to present his prepared work at the following meeting after posting it in advance.

Next Presentation: Antonio said he had prepared an analysis but had not posted it to the discussion thread and would present it the following week if possible.

Submission Process: Jerome encouraged anyone working on two-photon data to post analyses for presentation. He said the team would repeat the same round-table format next week, with participants posting materials ahead of time and the meeting information added to the website.

💬 Join the discussion for this page on GitHub