DANDI:000231

For a subset (44/115) of the behavioral sessions that we report here, we also recorded neuronal activity in identified columns of the barrel cortex

Methods, Neural recordings during behavior, paragraph 1

44 files are named with ecephys, and the same 44 hold the extracellular array recording and a non-empty units table.

includes 88.9 hours of behavioral video at 200 frames per second and 33.8 hours of 64-channel neural recording at 30 kHz

Background & Summary, paragraph 5

Reading (b), the main Open Ephys recordings under sourcedata, gives 33.770 h over 44 sessions, within 0.05 h of 33.8; each session has 64 channel files of equal length. The record headers of CH1 show a second, shorter segment in 7 sessions (0.48 to 0.51 s at the start of 6 files, and in KF134 2018-02-10 a separate recording, number 0, of 39.73 s before the main one, number 1); these 42.8 s are left out. Nothing printed shows they are the test recordings the paper mentions, which it says were a few seconds long, while six of the seven are 14 or 15 records under the main segment's recording number; leaving them out is a choice of this attempt. The whole files give 33.782 h, also within, so the outcome does not depend on it. Reading (a), the span of the NWB recordings, gives 33.656 h (-0.42%). The NWB files hold 56 to 64 channels.

How the test differed from the paper’s method

• The samples of each Open Ephys file are computed from its size (1024-byte header, 2070-byte records of 1024 samples). To leave out segments apart from the main recording, such as the test recordings the Data Records mention ("In some cases the Open Ephys file includes test recordings that were just a few seconds long"), segments are found from the first and last record headers and a bisection on the record timestamps of channel CH1, which splits each file where the timestamps break; the longest contiguous segment is taken as the main recording.

Shape discrimination mice were trained to lick right for convex shapes and left for concave shapes. Shape detection mice were trained to lick right for either shape (concave or convex) and to lick left on catch trials when no shape was presented at all

Methods, Two behavioral tasks, paragraph 1

The task assigned by the rule matches the experiment description in all 115 sessions (30 detection, 85 discrimination). Over trials to analyse with Table 6 stimuli, 21,752 of 21,752 follow the rule, in every session; all 32,350 trials follow it, including 1,533 with flatter shapes given the rule of their shape. The metadata task is taken from the experiment description, and from the keywords only where the description names none, which never happens; the keywords name "shape discrimination" in every file, detection sessions included. The task given by "nothing" alone also agrees with the description in all 115 sessions.

How the test differed from the paper’s method

• The files hold stimuli "flatter convex" and "flatter concave", which Table 6 does not list. The rule is judged on the Table 6 stimuli, and the flatter shapes are shown apart with the rule of their shape.

'nothing' was only used in the detection task, not in the discrimination task

Data Records, NWB file, trials table, Table 6 (stimulus)

The experiment description names the task in every file. All 3,749 "nothing" trials are in the 30 detection sessions; none of the 85 discrimination sessions has one.

How the test differed from the paper’s method

• The criterion allows the task from the experiment description or the keywords. The description is used. The keywords include "shape discrimination" in all 115 files, the 30 detection sessions included, so they cannot name the task.

No mice were trained on both tasks

Methods, Two behavioral tasks, paragraph 1

Each of the 15 mice has sessions of one task only, 10 discrimination and 5 detection, by the rule and by the experiment description.

The shape stopped at one of three different positions (close, medium, or far)

Methods, Two behavioral tasks, paragraph 2; Table 6 (servo_position)

Trials with a shape take exactly close (9,564), medium (9,494) and far (9,543), and all 115 sessions hold the three. Nothing trials also carry a servo position.

Only the first lick (left or right) during the response window determined the trial outcome (correct or incorrect)

Methods, Two behavioral tasks, paragraph 4; Table 6 (outcome, choice)

On all 32,350 trials outcome is correct exactly when choice equals rewarded_side, error when it is the other side, and spoil exactly when choice is nogo; the 46 direct-delivery trials are consistent too.

Even in fully trained mice, we found it best to begin every session with 45 trials of Forced Alternation, before switching to random trials

Methods, Session control and trial types to be dropped from analysis, paragraph 2

In the 110 sessions whose table holds trials numbered 0 to 44, all 4,229 such trials are non-random and the first random trial is number 45. The other 5 tables begin after trial 44; their first random trials are 61, 76, 88, 102 and 124.

Since the camera ran at 200 frames per second, frame F in the video is defined as occurring at exactly time F/200 in the common video timebase

Methods, Alignment of multiple data streams, paragraph 4

All processed and pose timestamps of 460 whisker series are within 9e-13 s of the 0.005 s grid, with no step shorter than one frame (252,019,206 of 252,170,504 steps are one frame, the rest gaps). Contact starts and stops are on the grid and equal frame_start x 0.005.

How the test differed from the paper’s method

• For each whisker the angle timestamps are read in full; the other six processed series are checked to have the same length and the same first and last time, not read in full.

these timestamps will in general contain gaps, for instance during the 133 ms blackout synchronization pulse at the start of each trial. This is because we did not interpolate over gaps longer than 4 frames (20 ms)

Data Records, NWB file, processing group processed_whisker_position, last paragraph

No run of interpolated frames is longer than 4 frames. In every session at least 98.4% of trials have a gap at the trial start within 10% of 133 ms for each of C1 to C3 (C0 at least 90.3%). The gap (missing time) is a median 0.135 s, 27 frames, and 0.135 to 0.140 s on 90% of trials.

This is null on interpolated frames

Data Records, processed_whisker_position, Table 3 (bend)

Bend is NaN on all 368,927 interpolated frames and on no non-interpolated frame.

The sampling rate is nearly 30 kHz, but not exactly, because it has been converted to the common video timebase used throughout the NWB file

Data Records, NWB file, acquisition group extracellular array recording, last paragraph

The 44 recordings are stored as a starting time and a rate, so spacing is exactly regular; rates are 29,999.9354 to 30,000.0199 Hz, within 0.0003% of 30 kHz, and none is exactly 30 kHz.

How the test differed from the paper’s method

• The files hold no timestamps array, contrary to Table 2; the rate and regular spacing are taken from the stored rate, which equals (samples minus one) over the span.

This is calculated by taking the depth of the electrode in the electrodes table which is indexed by the field "neuron_channel" for this unit ... This specifies the laminar location of the unit, based on its depth

Data Records, NWB file, units table, Table 5

All 999 units have the depth of the electrodes row their neuron_channel indexes, and electrode ids equal row indexes. The layer depth ranges do not overlap and run 1 (10 to 127 um), 2/3, 4, 5a, 5b, 6, 6b (1372 to 1424 um); the files spell L1 to L6 this way, with sublayers. Each unit's layer is the layer in its electrode's location string for all 999 units once case is ignored; the 446 that differ as spelled are 5a, 5b and 6b against 5A, 5B and 6B.

How the test differed from the paper’s method

• The layer names of the files ("1", "2/3", "4", "5a", "5b", "6", "6b") are read as L1 to L6, with 5a and 5b in L5 and 6b in L6.

we ordered the channels in the Neurodata Without Borders (NWB) file in order of the anatomical location of the corresponding recording site, from superficial to deepest. We excluded broken channels from the NWB file

Methods, Ordering of neural data channels; Methods, Neural recordings during behavior, paragraph 4; Data Records, electrodes table

In all 44 sessions y increases strictly in multiples of 20 um (26 sessions have a larger step at a broken channel), the table has 56 to 64 rows, as many as the data columns, and the ElectricalSeries rows are 0 to n-1. All 2,747 location strings hold the five items, and y minus 20 um times the anatomical index including broken channels is constant in each session, so a full probe spans 1260 um between end sites.

Its "location" field specifies the barrel-related cortical column in which the recording device was inserted during this session

Data Records, electrodes table, Table 4 (electrode_group); Methods, Neural recordings during behavior, paragraphs 2 and 3

Each of the 44 sessions has one electrode group naming a column; C3 13, C1 10, C2 10, D2 3, C4 2, D1 2, and one each of B2, B3, E1 and E2.

In this dataset, there is only one contiguous chunk of data included per session, so the length of this list is one

Data Records, NWB file, units table, Table 5 (obs_intervals, spike_times)

All 999 units have one interval holding all their spikes. It starts at the first sample time and ends one sample period after the last sample time, within the criterion's one sample.

It comprises four data interfaces, each of which is an ndx_pose.pose.PoseEstimation containing pose tracking data for one of the four whiskers that we tracked. In turn, each PoseEstimation comprises eight PoseEstimationSeries ... which is the same for all markers within a session but can vary across sessions

Data Records, NWB file, processing group raw_whisker_position

All 115 sessions hold four PoseEstimation objects with eight series of shape (N, 2) each, and one conversion value per session; it varies across sessions from 0.0403 to 0.0524 mm per pixel.

but this whisker rarely made contact with the shapes in most mice

Methods, Two behavioral tasks, paragraph 5

Over trials to analyse, C0 makes fewer contacts than each of C1, C2 and C3 in 10 of 15 mice (C0 fraction 0.0% to 7.3%); in 219CR, 234CR, 245CR, 255CR and KF119 it makes 20% to 49% of contacts.

How the test differed from the paper’s method

• A whisker with no contacts table in a session (C0 in 21 sessions, C3 in 8) is counted as making no contacts there.

We have also added optional fields containing the location of each contact, expressed both as a Cartesian location within the frame and also as the angle of the whisker at the time of the contact

Data Records, NWB file, processing group identified_whisker_contacts

All 431 contacts tables (290,345 rows) have start_time, stop_time, tip_x, tip_y, base_x, base_y and angle, and no row starts after it stops. 29 whisker-sessions have no table (C0 in 21, C3 in 8), read as whiskers that made no contact, as the criterion's wording allows.

processing A group containing a ProcessingModule named "behavior". This includes processed data about behavioral events, whisker position, and whisker contacts

Data Records, Annotated data in the NWB file, Table 1 and the sections after it

Under the naming of Table 1 and UsageNotes.ipynb (one module "behavior"), all 115 files hold left and right licks and rewards, the four raw pose and processed position interfaces with the seven Table 3 series, contacts, the 14 Table 6 columns and a video link that resolves to a deposited video; the 44 ecephys files hold the recording, units with the Table 5 columns, and electrodes with the Table 4 columns (y, location and one electrode group per session, x and z null). The group names of the Data Records sections are used in no file. 29 files hold three contacts tables, not the four the Data Records describe.

At the top level, there is one folder per subject ... Video files are stored in a directory with the same name as the corresponding NWB file. Finally, we also provide the raw neural data files in a directory called "sourcedata"

Data Records, Directory structure of the entire dataset

All 115 NWB files are in their subject's folder and named as described (71 behavior+image, 44 behavior+ecephys+image), each with one video in a folder of its name. sourcedata holds a folder for each of the 44 ecephys sessions and no other, each with Open Ephys files and Kilosort/Phy outputs.

Importantly, the frame rate of this video file is spuriously set to 30 frames per second, even though the true frame rate is 200 frames per second

Data Records, Raw data, video files; Methods, High-speed videography and whisker tracking, paragraph 1

All 115 videos are Matroska, H.264, at a nominal 30 frames/s; 111 are 640 x 550 and 4 are 640 x 600. Frames over 200 exceed the last whisker time by 0.005 to 0.065 s (1 to 13 frames).

We purchased breeder mice from the C57BL/6 J strain ... Mice began behavioral training between postnatal days 90 and 180 ... We used males and females arbitrarily and in roughly equal proportion

Methods, Two behavioral tasks, paragraph 7

All 15 mice are C57BL/6J (spelled C57BL6/J), 10 female and 5 male (binomial p = 0.302, minority one third, so both sex readings hold), 197 to 436 days old at their first session.

The size of the dataset ... 15 mice, 88.9 hours, 115 sessions, 18,514 trials, 63,979,800 frames

Methods, High-speed videography and whisker tracking, paragraph 2; Background & Summary, paragraph 5

Mice (15 subject folders) and sessions (115 NWB files) match. The video frames match exactly under reading (a), 63,979,800 frames or 88.86 h; reading (b), the longest raw pose series of each session, gives 63,082,162 (-1.40%). The trials match under neither reading, 32,350 rows (+74.7%) under (a) and 22,905 with ignore_trial False (+23.7%) under (b), both outside 10%, so the criterion gives not reproduced. Other counts were tried for information (random trials 24,776; random and not spoiled 24,590; ignore_trial False with Table 6 stimuli 21,752; ignore_trial False in sessions without optogenetics 17,287); none gives 18,514.

How the test differed from the paper’s method

• Video frame counts are read from each file's duration tag at the nominal 30 frames/s, not by decoding 2 TB of video. Every duration is a whole number of seconds; the container duration equals the tag in all 115 files. No video stream reports its own frame count (0 of 115), so the tag is the only count the headers give.
• Reading (b) takes the longest of the four whiskers' raw pose series in each session, as the criterion's N_frames differs between whiskers.

choice_time (float) The time of the choice lick, which is the first lick after the response window opened. On spoiled trials with no lick, this value is 45 seconds after the response window opened

Data Records, NWB file, trials table, Table 6 (choice_time)

On 32,085 non-spoiled trials, the first lick at or after the window opening is within 5 ms of choice_time on 77.6% (78.0% allowing a lick 5 ms early), below 99% in every session (lowest 52.5%); its side matches choice on 99.1%. On the 7,185 trials that do not match, choice_time is 0 to 5 ms after the window opening on 7,183 of them, a lick in the 100 ms before the opening precedes it on 84.8%, and a lick lies within 5 ms of choice_time on only 5.7%. This suggests that a lick already under way when the window opens is taken as the choice at the opening, which the definition does not say. Over all 32,085 non-spoiled trials choice_time is 0 to 5 ms after the opening on 7,848 (24.5%), 7,183 of them mismatched. On all 265 spoiled trials choice_time is NaN, not 45 s after the window.

ignore_trial (bool) This is True if 'direct_delivery' is True, or 'stim_is_random' is False, or outcome is 'spoil', or 'optogenetic' is True, or the trial was otherwise munged

Data Records, NWB file, trials table, Table 6 (ignore_trial)

All 46 direct-delivery, 7,574 non-random and 2,482 optogenetic trials have ignore_trial True, but 160 of the 265 spoiled trials (random, not optogenetic, in 50 sessions) have it False. ignore_trial equals "stim_is_random False or optogenetic" on every trial, and no trial is ignored without one of the four flags (no munged trials are marked).

stop_time (float) Stop time of the trial. (Equal to the start time of the next trial.) ... The index of this DataFrame is called "id" and is identical to the field "trial"

Data Records, NWB file, trials table, Table 6 and the paragraph after it

id equals trial on all 32,350 rows. stop_time equals the next start_time within 1 ms on 32,230 of 32,235 consecutive pairs; the 5 others, in 4 sessions, are all where the trial numbers skip, with gaps of 7.0 to 68.0 s. The criterion asks for every row, so it is not reproduced. Read as the next trial run, the statement holds on all 32,230 pairs whose trial numbers are consecutive (variant). The paper mentions only trials dropped before the video began, at the start of a table. Whether the criterion should compare the next trial run is asked of a reader (questions/000231.yaml); the outcome is tested on the claims file as it is.

single units had to pass all of the following objective criteria ... ≤5% of the inter-spike intervals less than 3 ms; ≤1.5% change per minute in spike amplitude; ≤20% of the recording at <5% of the mean firing rate

Technical Validation, Spike sorting

The ISI part holds for all 999 units (largest fraction of intervals below 3 ms 4.93%). The firing-rate part fails under both readings, so the criterion gives not reproduced; with 60 s bins 6 units exceed 20% (largest 23.3%), with one bin per trial 78 do (largest 40.2%); at a 30% bound, 0 and 6. Leaving out the 16 units of KF134 2018-02-09, whose neural clock the data check did not establish, changes neither count. Amplitude criteria, outside the outcome as the criterion says. 893 of 999 units match their Kilosort cluster by id with equal spike counts, and their NWB spike times equal the Kilosort samples over the rate plus a constant (spread 2.3e-12 s); the other 106 have more Kilosort spikes than NWB spikes. None of the 893 changes amplitude by more than 1.5% per minute (largest 1.47%, median 0.33%). The criteria on amplitudes below the detection threshold and below 10 uV, and on common-mode artefacts, were not run.

How the test differed from the paper’s method

• In reading (a) the last, partial 60 s bin of each unit's interval is left out (a standard choice).
• The change per minute in spike amplitude is measured, for information only, as the slope of a least-squares line of Kilosort amplitude against time over the mean amplitude; the paper does not say how it measured it. Kilosort amplitudes are template scaling factors, not microvolts, and the paper gives no detection threshold, so the other amplitude criteria cannot be run.

we required that every whisker be labeled in ≥95% of the frames, that ≤2% of the contact events contained even a single frame with a missing label

Technical Validation, Inclusion criteria

The contact part fails under every reading; up to 7.7% of a whisker's contacts in a session hold an interpolated or absent frame, and 16 sessions exceed 2% for C1 to C3 (27 for C0 to C3). The frame part holds for C1 to C3 (at least 99.35% of processed frames not interpolated, at least 97.6% of video frames with a pose) but not for C0 (93.1% and 89.7% at worst). Measured values are under whiskers (i) and frames (a). The stop frame belongs to the contact, duration_frames equals frame_stop - frame_start + 1 on all 290,345 contacts and stop_time equals frame_stop x 0.005 s; with stops taken as exclusive the largest fraction is still 7.7%. The expected value is per whisker and session. With contacts pooled over the whiskers of a session (variant), the largest fraction is 2.007% for C0 to C3 and 2.008% for C1 to C3, both in KM91 2016-12-16 (54 of 2,691 and 54 of 2,689 contacts), just above 2%; 1 session (i) or 2 sessions (ii) exceed 2%. It would be 1.94% with exclusive stops, which the files rule out. Pooled, the result rests on one contact (53 of 2,691 would be 1.97%) and on counting a frame absent from the processed series as missing, as the criterion says. Whether the 2% bound is per whisker or pooled over a session's contacts is asked of a reader (questions/000231.yaml); the outcome is tested on the claims file as it is.

How the test differed from the paper’s method

• A contact covers its frames from start to stop inclusive, as duration_frames = frame_stop - frame_start + 1 and stop_time = frame_stop x 0.005 s in the files (shown in the notebook).

which began exactly 2 seconds after the linear servo motor started moving the shape toward the mouse ... reached its final position between −0.8 s and −0.4 s, depending on whether the final position was close, medium, or far

Methods, Two behavioral tasks, paragraph 3

No motor start or stop time is in the files; the trials table's only servo column is servo_position. For information, the window opens a median 3.023 s after start_time (5th to 95th percentile 3.020 to 3.443 s), the same for each position.

Correct licks were rewarded with water and incorrect licks were punished with a timeout, typically 9 seconds long

Methods, Two behavioral tasks, paragraph 4

The criterion's two readings of the timeout disagree, and it blocks the claim in that case. Over trials to analyse, the median of stop_time minus choice_time is 12.21 s on error trials (reading a, outside 9 +- 0.9 s) and 3.88 s on correct trials, so the difference is 8.34 s (reading b, within); per session the difference is 7.04 to 8.54 s. The reward part also fails as computed, with a reward on 8.5% of the 5,130 error trials over both sides. It holds on the left (reward on 100% of 8,917 correct trials and on none of 2,497 error trials) but not on the right, where the reward series is identical to the right lick series in every file and a "reward" falls on 16.6% of error trials; there it counts licks, so right rewards cannot be tested, a defect of the deposit (findings). Which interval the paper calls the timeout is a question for the paper's authors (findings).

In contrast the neural and video timebase were synchronized very well by a linear fit. Residuals were never more than a single frame (5 ms)

Technical Validation, Alignment data

The pulse in the neural stream is on an analog input of Open Ephys. The 44 .openephys indexes list 360 ADC and 270 AUX channel files besides the 2,880 CH files, but the dandiset holds only 2,816 CH files (64 per session) and all_channels.events holds only its 1024-byte header. The video side would also need the brightness of every frame of 2 TB of video.

We observed that the behavioral timebase was a poor fit to the other two timebases and that its residuals varied slowly but predictably over the session

Technical Validation, Alignment data

All times in the NWB files are in the video timebase, and the dandiset's other files are videos, Open Ephys files and Kilosort/Phy outputs; its 44 .log files are all named phy.log, one in each of 44 sorting folders under sourcedata (printed), so they are spike-sorting logs. No Arduino log or behavioral timebase is deposited. The paper does not say where the Arduino times are. The README of the original release on Zenodo (ref. 13, doi:10.5281/zenodo.4743837) lists a per-session behavioral_logfile; asked whether the claim should use it, a reader answered that the paper never says what that release holds, so the claim keeps data dataset and the claims file is unchanged (questions/000231.yaml). The answer changed nothing in the test; the outcome stays blocked, data_missing, as before. The Zenodo release was not read.

We also excluded sessions with fewer than 90 trials to analyze

Technical Validation, Inclusion criteria

The readings disagree, and the criterion blocks the claim in that case. Under (a), ignore_trial False, one session (228CR, 2019-07-15) has 84 trials; under (b), random and not spoiled, it has 101 and every session has at least 90. Which trials the paper counts is a question for the paper's authors (findings).