DANDI:000294 · Encore Science

Claims

From the paper 10.1162/neco_a_01672

Reproducing Buccino et al. (2024) from dandiset 000294 10

Notebook reproduce_buccino_2024.ipynb

Reproduced9

ReproducedResults 3.3, paragraph 1; Methods 5.2, last paragraph, and 5.6.5
The experimental data are simultaneous patch-clamp and HD-MEA recordings of two cultured neurons, each stimulated with the eCode protocols repeated in four to six runs of about 200 s.
We acquired paired patch-clamp and extracellular signals from two cells of cortical cultures in embryonic rats ... The eCode protocols were repeated during four to six runs (each run lasted approximately 200 s)
Results 3.3, paragraph 1; Methods 5.2, last paragraph, and 5.6.5
Two NWB files, one neuron each; cell 1 (sub-3148, the code's cell1_211006_3148) has 5 runs and cell 2 (sub-3436, cell1_211011_3436) 4. The median run lasts 195.8 s from the first to the last patch sample (195.7 to 195.8 s, the same span as each run's MEA series). The runs are those the code used; cell 2's runs 3 to 6 are stored as run1 to run4, and the descriptions of its MEA series keep the original numbers.
ReproducedMethods 5.2
Each run applies the eCode protocols with the sweep numbers, step durations and amplitudes (in percent of the estimated rheobase) that the Methods list.
IDthresh: The purpose of this protocol was to finely screen the current levels to correctly reestimate Irheo before processing. It consisted of 21 sweeps with a 270 ms square pulse of increasing amplitudes from 50% to 130% with step increments of 4%
Methods 5.2
In all 9 runs, all eight protocols are present with 21, 2, 11, 11, 6, 4, 4 and 1 sweeps (IDthres, firepattern, IV, IDrest, APWaveform, HyperDepol, sAHP, PosCheops). Step durations measured on the timestamps are within 0.1 ms of 270, 3600, 3000, 1350, 50, 450 and 270, and 250, 225 and 450 ms; the PosCheops ramps rise in 3925 to 3975, 1965 to 1988 and 1307 to 1322 ms measured from 5 pA above holding. Against the IDthres scale (120.4 to 120.8 pA for cell 1, 301.2 to 301.7 pA for cell 2) every amplitude is within 1.2 percentage points of its stated percentage. IV has 11 steps from -140 to 60% (reading a); reading (b), 9 steps to 20%, does not match the sweep count.
How the test differed from the paper’s method
- The 0% IV step of reading (a) has no relative tolerance; 10% of the IV increment (2% of the scale) was used. Its measured level is within 2% of holding in every run.
- APWaveform's sweep count (6) and PosCheops' (1) are not stated in the Methods; 6 is the number of amplitudes from 200 to 350% by 30%.
ReproducedMethods 5.6.5, paragraph 1; Methods 5.2question to the reader
The holding current was set so that the read-out membrane potential was about -70 mV, which is about -84 mV after liquid junction potential correction.
the holding current was manually adjusted to obtain a resting membrane potential of around -70 mV (note that with LJP correction, this corresponds to -84 mV)
Methods 5.6.5, paragraph 1; Methods 5.2
The median pre-stimulus membrane potential is -69.54 mV for cell 1 (given as measured) and -69.12 mV for cell 2, over 300 and 240 sweeps, within -70 +/- 7 mV (reading a). Run medians range from -71.0 to -66.6 mV. Neither cell is within -84 +/- 8.4 mV (reading b); the files do not say whether the voltages are corrected for the junction potential, and reading (b) failing does not show that they are not. The window before the code's ton gives the same values to within 0.1 mV. Whether the criterion's window should end at the code's ton or at the onset in the files is asked of a reader (questions/000294.yaml).
How the test differed from the paper’s method
- The pre-stimulus window ends at the first change of the injected current found in each sweep (350 ms for most protocols), which in these files is 50 ms before the code's ton (400 ms for IDrest); the window ending at the code's ton is shown as a variant.
ReproducedMethods 5.6.5, paragraph 1
Only cells whose extracellular spikes were at least 50 uV were kept, so both recorded cells show extracellular spikes of 50 uV or more.
We discarded cells with extracellular spikes lower than 50 μV
Methods 5.6.5, paragraph 1
With the template method on the run with the most channels after denoising (cell 1 run 4, 561 channels; cell 2 run 1, 545), the largest channel reaches a negative peak of 217.0 uV (peak-to-peak 274.0 uV) for cell 1 and 159.2 uV (192.2 uV) for cell 2; the smaller, cell 2's, is given as measured. Every run of both cells is above 50 uV under both readings (negative peaks 159.2 to 218.0 uV). 125 to 217 patch spikes per run outside PosCheops, 92 to 98% of waveforms kept by the 2 SD rule. The code's mean template is higher by up to 13.6 uV (negative peak) and 13.7 uV (peak-to-peak). The largest channel of cell 1 is the channel the export notebook plots (ch885); for cell 2 that notebook plots ch384, adjacent to the largest channel found here (156.8 uV). The patch and MEA clocks stay aligned within sweeps; all 1616 spikes have their MEA trough 0.15 to 0.70 ms before the patch peak, none at the +/- 2 ms edge of the search, with medians of -0.35 to -0.25 ms in every window of the sweep up to 4.5 s, and per-run slopes of -0.003 to 0.023 ms per s of sweep, against the 20.4 ms per s a clock mismatch of 20 against 20.41 kHz would give.
How the test differed from the paper’s method
- The band-pass filter is applied forward and backward (sosfiltfilt), the default of the code's filtering library; the Methods give only the order and cut-offs. A standard choice that cannot move amplitudes of 159 uV and more below 50 uV.
ReproducedResults 3.3, paragraph 2; Figure 6data in NeuroMorpho.org Buccino_Hierlemann_Cell1 and Buccino_Hierlemann_Cell2; GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/<cell>/morphology_corrected.swc
In both cells the axon initial segment does not leave the soma but a dendrite (an axon-bearing dendrite).
Moreover, for both cells, the AIS did not originate directly from the soma but from one of the dendrites, which is termed an axon-bearing dendrite (ABD
Results 3.3, paragraph 2; Figure 6
In the repository's morphology_corrected.swc of both cells, the single AIS path (type 4, 184 and 192 points) starts from a dendrite point (type 3), 42.9 and 39.1 um from the soma point, and the axon (type 2) starts from the AIS. The NeuroMorpho.org copies were not used.
ReproducedResults 3.2, paragraph 4data in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais (model) and notebooks/ground-truth/1-generate-ecode-protocols-and-features-hay_ais.ipynb
The ground-truth model's rheobase, estimated with a 270 ms step, is 150 pA.
All current amplitudes were computed with respect to the neuron's rheobase current (estimated using a 270 ms step). For our groundtruth model, the rheobase current was 150 pA
Results 3.2, paragraph 4
With the ground-truth notebook's series (from 0.15 nA by 0.005 nA), the first step, 150 pA, fires one spike in the 270 ms step. Because that series starts at the stated value, a series from 0.10 nA was also run; the model first fires at 140 pA, also within 150 +/- 15 pA. The code's own compute_rheobase_for_model on the LFPy cell model gives the same 150 and 140 pA. The repository's IDrest 300% step of 0.45 nA implies the same 150 pA scale.
How the test differed from the paper’s method
- Simulated with BluePyOpt 1.14.26 (NEURON 9) on the repository's model definition, in place of BluePyOpt 1.13.88; the main series uses its NEURON simulator, and the code's compute_rheobase_for_model on its LFPy simulator, as the notebook uses, is shown and agrees.
ReproducedFigure 3B; Results 3.2, paragraph 2rests on an assumptiondata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais and notebooks/ground-truth/7-AIS-test.ipynb
In the ground-truth model the action potential starts in the distal AIS and travels back toward the soma.
The AP is initiated at the distal end of the AIS and travels back toward the soma
Figure 3B; Results 3.2, paragraph 2
Run as 7-AIS-test.ipynb runs it (the repository's IDrest_300 protocol with the AIS recordings on the LFPy cell model, get_peak_cutout over 12 spikes, first and last left out). The code's cut-outs have 0.05 ms samples. At the sample the distal and middle AIS tie under both readings (-0.35 ms at -20 mV, -0.25 ms at the peak, from the somatic peak sample), and the soma follows 0.15 and 0.25 ms later. Between samples (linear crossing, parabolic peak) the distal AIS leads the middle by 0.0042 and 0.0202 ms. A fixed-step run (dt 0.005 ms) of the whole protocol times each of the 12 spikes; the distal AIS leads the middle by 0.0051 to 0.0070 ms at the crossing and 0.020 to 0.025 ms at the peak, and the middle leads the soma by 0.162 to 0.168 and 0.23 to 0.24 ms, in 12 of 12 spikes. The order distal, middle, soma holds under both readings at every resolution that separates the two AIS points.
How the test differed from the paper’s method
- Assumed. The paper does not say how the initiation site is read from Figure 3B (Results 3.2, "The AP is initiated at the distal end of the AIS and travels back toward the soma"), nor at what resolution. The times are read between the 0.05 ms samples of the code's cut-outs and checked on a fixed-step run at 0.005 ms, because a 0.05 ms grid cannot order two points 15 um apart that the simulation separates by 5 to 7 us; read at the sample, the two AIS points tie under both readings and the claim's strict order would not be met.
- Simulated with BluePyOpt 1.14.26 in place of 1.13.88; the code's protocol, recordings and get_peak_cutout run unchanged on its LFPy simulator. The fixed-step check uses its NEURON simulator (dt 0.005 ms, a standard choice).
ReproducedFigure 3A; Results 3.2, paragraph 3data in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais/fitting
On the simulated 80-electrode MEA, the largest extracellular signal is on the electrode closest to the AIS.
The red trace depicts the signal with the largest amplitude, which appears on the electrode closest to the AIS
Figure 3A; Results 3.2, paragraph 3
The template recomputed with the code (IDrest 300%, calculate_eap with get_extra_kwargs, the code's planar probe, LFPy cell model) matches the repository's template.npy with correlation 0.99992 and an RMS difference of 1.3% of the template; with the probe moved by 2 or 5 um the difference rises to 6.6 to 16.1%, so the file was simulated in the frame of the LFPyCellModel. In that frame the AIS segment midpoints run from (11.8, -1.0) to (13.8, -28.3) um and its midpoint is (12.6, -13.7) um. The nearest electrode is channel 43 at (25, -25) um, 16.8 um away against 39.3 um for the next, and it is the nearest electrode to every AIS segment. Channel 43 has the largest peak-to-peak and the largest negative peak of the repository's template. The second-largest peak-to-peak, 0.939 of the largest, is channel 24 at (-25, 25) um, 54.0 um from the AIS midpoint; moving the probe by 2 um in x makes channel 24 the largest, so the largest channel depends on the placement, which the match to the file fixes.
How the test differed from the paper’s method
- The AIS is placed with the segment midpoints of the instantiated LFPyCellModel (BluePyOpt 1.14.26, LFPy 2.3.7), where LFPy puts the soma midpoint at the origin. In the first round it was placed from NEURON's 3D points relative to the mean soma point, which puts the AIS midpoint at (1.7, -23.4) um, about 14 um off; that placement also gives channel 43 as nearest (23.3 um, next 26.8 um) and is shown as a variant.
ReproducedTable 1; Results 3.2, paragraph 6data in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais/fitting/efeatures
For the ground-truth model, the soma strategy uses 76 intracellular features; all adds 11 extracellular features (87), sections adds 33 (109), and single adds 77 (153).
A total of 76 intracellular features were extracted from these selected protocols ... the all strategy ... adds 11 extracellular features (total = 87). The sections strategy added 33 (three sections times 11 extracellular features), and the single strategy added 77 features
Table 1; Results 3.2, paragraph 6
The code's create_evaluator for hay_ais has 76 intracellular objectives for every strategy and 0, 11, 33 and 77 extracellular ones, totals 76, 87, 109 and 153; the feature files features_BPO_<strategy>.json give the same counts. The Methods' 79 is the count of the six training protocols' features in features.json before the code drops the three sag features of IV -20%.

Partially reproduced1

Partially reproducedResults 3.1, paragraph 2; Methods 5.6.2 and 5.6.5question to the reader
The HD-MEA recorded more than 500 electrodes near the pipette simultaneously, out of at most 1024, at 20 kHz, and the patch signal was digitized at 20 kHz.
The patch-clamp and the HD-MEA systems were synchronized so that we could simultaneously acquire the patch-clamp readout and the extracellular signals from more 500 electrodes in proximity of the patch pipette
Results 3.1, paragraph 2; Methods 5.6.2 and 5.6.5
The MEA recorded 678 (cell 1) and 675 (cell 2) electrodes, above 500 and at most 1024, at 20000 Hz in every run. The patch timestamps are at 20000 Hz only for IDthres; firepattern, IV, IDrest, HyperDepol, sAHP and PosCheops are at 20408.2 Hz (49 us intervals) and APWaveform at 125000 Hz (8 us), neither 20 nor 200 kHz. The intervals are exact within each sweep and the step durations measured on them are the nominal ones, so they are the sampling of the deposited data. The electrode and MEA parts hold; the patch rate, with tolerance 0, does not. The criterion has no partial clause; partially_reproduced is used because the claim holds in part (electrodes and MEA rate) and fails in part (patch rate). Whether APWaveform at 125 kHz fits the paper's "typically 200 kHz" is asked of a reader (questions/000294.yaml); it cannot change the outcome, since six protocols at 20.41 kHz already fail.

Not tested in a notebook 13

1 metadata insufficient, 12 too large to analyse

Blocked13

BlockedMethods 5.6.3, paragraph 2; Results 3.3, paragraphs 1 and 2· Metadata insufficient
The cultured cells were recorded between days in vitro 18 and 24.
All experiments were conducted between days in vitro (DIVs) 18 and 24
Methods 5.6.3, paragraph 2; Results 3.3, paragraphs 1 and 2
The NWB subject has only subject_id, species and description ("Rat cortical embryonic cell culture"); neither the files nor the dandiset's metadata give an age, a date of birth or plating, or days in vitro. The session dates (2021-10-06 and 2021-10-11) are 5 days apart, which fits either reading but does not give the age.
BlockedFigure 4A; Results 3.2, paragraph 7· Too large to analysedata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais, scripts and notebooks/ground-truth
On the ground-truth model, soma, all and sections give good intracellular fitness and single worse; single gives the best extracellular fitness and soma the worst.
While the soma, all, and sections strategies appeared to yield low scores (i.e., better fitness) for intracellular features, the single strategy produced solutions with worse intracellular fitness. Conversely, the single strategy yielded the best performance with respect to extracellular features, while the worst performance was achieved by using the soma strategy
Figure 4A; Results 3.2, paragraph 7
Needs 10 optimizations per strategy of 600 generations of 20 offspring (480,000 evaluations for the four strategies). In the run recorded, one evaluation of the six training protocols on the released ground-truth model takes 12.2 s intracellularly (NEURON simulator) and 20.5 s with the extracellular signals (the code's evaluator for the all strategy on the LFPy simulator), so at least 2457 CPU hours, 614 hours on the 4 cores of the runner, against a 30-minute notebook limit and a 5-hour run. The timing varies with the runner's load; the run reviewed in round 2 printed 8.0 and 14.3 s and 1699 CPU hours. No optimization results are in the repository. The code's evaluators do build and run in the locked environment (table1-feature-counts), so time is the only obstacle.
BlockedResults 3.3, paragraph 3· Too large to analysedata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais, scripts and notebooks/ground-truth
Over the ten seeds, the ground-truth fits reach intracellular scores of 62.8 +/- 13.6 for soma and 59.3 +/- 17.0 for all.
(intracellular score: GT fitting: 62.8 ± 13.6 for soma and 59.3 ± 17.0 for all
Results 3.3, paragraph 3
Needs the ground-truth optimizations; see fig4a-strategy-balance for the estimate.
BlockedFigure 4C; Results 3.2, paragraph 8· Too large to analysedata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais, scripts and notebooks/ground-truth
In validation, only the single strategy reproduces that the ground-truth model fires no spike in the depolarizing phase of HyperDepol -160% and one at HyperDepol -40%; soma, all and sections fire at -160%.
the soma, all, and sections strategies failed to reproduce the absence of spikes observed in the depolarization phase (second row: HyperDepol (−160%). Only the single ... strategy correctly did not generate a spike for HyperDepol (−160%) and did generate one for HyperDepol (−40%)
Figure 4C; Results 3.2, paragraph 8
Needs the best solution of each strategy's optimizations; see fig4a-strategy-balance. The ground-truth part is computed. In the validation run of the code's evaluator the ground truth fires no spike in the 700 to 970 ms depolarizing phase of HyperDepol -160% (and -120%) and one at -40% (and -80%), as the criterion expects of it.
BlockedFigure 4B, left; Results 3.2, paragraph 8· Too large to analysedata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais, scripts and notebooks/ground-truth
On validation protocols (86 intracellular features), the best solutions of the four strategies do not differ in intracellular feature scores.
No significant differences were ... detected across strategies (post hoc Mann-Whitney tests with Holm p-value correction)
Figure 4B, left; Results 3.2, paragraph 8
Needs the best solution of each strategy's optimizations; see fig4a-strategy-balance. The validation evaluator of 4-optimization-performance.ipynb has 110 objectives and 86 distinct feature names, the 86 features the claim states.
BlockedFigure 4B, right; Results 3.2, paragraph 8· Too large to analysedata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais, scripts and notebooks/ground-truth
On extracellular features computed on all 80 electrodes (796 features), all three extracellular strategies score significantly better than the soma strategy.
All strategies making use of extracellular data significantly outperformed the soma strategy (post hoc Mann-Whitney tests with Holm p-value correction: soma VS all - p < 10−5, soma VS sections - p < 10−5, soma VS single - p < 10−5
Figure 4B, right; Results 3.2, paragraph 8
What the paper reports
- soma against all: post hoc Mann-Whitney test, p < 10⁻⁵, Holm-corrected (n = 796 extracellular features per strategy)
- soma against sections: post hoc Mann-Whitney test, p < 10⁻⁵, Holm-corrected (n = 796 extracellular features per strategy)
- soma against single: post hoc Mann-Whitney test, p < 10⁻⁵, Holm-corrected (n = 796 extracellular features per strategy)
Needs the best solution of each strategy's optimizations; see fig4a-strategy-balance. The extracellular features of the ground truth's firepattern 120% template, computed as 4-optimization-performance.ipynb does, are 11 kinds on 80 electrodes, 880, of which 800 are finite (recovery_slope is not finite on any electrode); the claim states 796, and 800 is within its tolerance of 80.
BlockedFigure 4D; Results 3.2, paragraph 8· Too large to analysedata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais, scripts and notebooks/ground-truth
The extracellular strategies fit the ground-truth template better on almost all electrodes, while the soma solution fails to capture the extracellular signals on many channels.
It is evident at first glance that the three extracellular strategies achieve a better fit for almost all electrodes, while the solution from the soma strategy ... failed to capture the complex dynamics of extracellular signals in many recording channels
Figure 4D; Results 3.2, paragraph 8
Needs the best solution of each strategy's optimizations; see fig4a-strategy-balance.
BlockedFigure 5; Results 3.2, paragraph 9· Too large to analysedata in GitHub alejoe91/multimodalfitting at commit 24e8e61, cell_models/hay_ais and notebooks/ground-truth/5-AP-distribution.ipynb
Across the neuron, the all strategy gives lower AP waveform errors than the soma strategy at every tested location, and all extracellular strategies beat soma at every dendritic location.
For all tested locations, the all strategy produces lower errors than the strategy. For all dendritic locations (both apical and basal), all extracellular strategies (all, single, sections) outperform the soma strategy
Figure 5; Results 3.2, paragraph 9
Needs the best solution of each strategy's optimizations; see fig4a-strategy-balance.
BlockedFigures 7A and 8A; Results 3.3, paragraph 3· Too large to analyseuses data from outside the dataset
For both experimental cells, the soma strategy gives better intracellular scores and the all strategy better extracellular scores.
For both cell models, the application of the soma strategy yielded better intracellular scores, while the use of the all strategy resulted in better extracellular fitting
Figures 7A and 8A; Results 3.3, paragraph 3
Needs 10 optimizations of 600 generations of 20 offspring for soma and all on each cell, 480,000 evaluations; at the ground-truth model's cost in the run recorded (12.2 s per intracellular evaluation, 20.5 s with the extracellular signals), at least 2182 CPU hours, 546 hours on the 4 cores of the runner (1487 CPU hours in the run reviewed in round 2). The dataset holds what the fits need from the recordings (the patch sweeps of every protocol and the MEA data from which templates of 159 to 218 uV are computed here), and the linked runs are those the code used (recordings-two-cells-runs).
BlockedResults 3.3, paragraph 3· Too large to analyseuses data from outside the dataset
Intracellular scores are 123.9 +/- 11.2 (soma) and 149.2 +/- 47.2 (all, one non-converging seed left out) for cell 1, and 152.8 +/- 9.22 (soma) and 177.6 +/- 20.4 (all) for cell 2, almost twice those of the ground-truth fits.
cell 1 fitting: 123.9 ± 11.2 for soma and 149.2 ± 47.2 (excluding one seed that did not converge) for all; cell 2 fitting: 152.8 ± 9.22 for soma and 177.6 ± 20.4 for all
Results 3.3, paragraph 3
Needs the experimental optimizations; see fig78a-soma-intra-all-extra.
BlockedFigures 7B and 8B; Results 3.3, paragraphs 3 and 6· Too large to analyseuses data from outside the dataset
The experimental cell models of both strategies reproduce the main features of the spiking patterns of the validation protocols, except the soma model's PosCheops 300% response in cell 2, so including extracellular features does not keep the optimization from finding acceptable models.
the models fitted on the experimental data could reproduce the main features of the intracellular spiking patterns of the validation protocols (panels B, with the exception of the PosCheops (300%) response for the soma strategy in cell 2
Figures 7B and 8B; Results 3.3, paragraphs 3 and 6
Depends on fig78a-soma-intra-all-extra, blocked for the cost of the optimizations. The reference of reading (a) is computed. Over runs, the recorded spike counts per sweep are, for cell 1 and cell 2, firepattern 12 to 15 and 0 to 1 (120%), 33 to 37 and 14 to 30 (200%); HyperDepol 0 in every sweep and run; sAHP 1 and 0 to 2, 2 to 3 and 3 to 5, 3 to 4 and 7 to 8, 4 and 8 to 10 (150 to 300%); PosCheops 76 to 90 and 52 to 80.
BlockedFigures 7C and 8C, and 7A and 8A right; Results 3.3, paragraph 4· Too large to analyseuses data from outside the dataset
The simulated templates are qualitatively similar to the recorded ones, with mismatches mainly in cell 2, whose fits have higher extracellular scores than cell 1.
The worse fit of cell 2 with respect to cell 1 is also ... reflected by the overall higher extracellular scores
Figures 7C and 8C, and 7A and 8A right; Results 3.3, paragraph 4
Depends on fig78a-soma-intra-all-extra, blocked for the cost of the optimizations.
BlockedFigure 10 legend; Results 3.3, paragraph 5· Too large to analyseuses data from outside the dataset
In the cell 2 all model, the template mismatch is explained by too small a somatic and too large an AIS contribution; the recorded positive upstroke near the basal dendrite coincides with the somatic negative current peak, and the initial negative phase with the AIS peak.
The upstroke of the experimental EAP close to the basal dendrite ... seems to coincide with the somatic negative peak ... The initial negative phase instead is time-locked to the AIS negative peak
Figure 10 legend; Results 3.3, paragraph 5
Depends on fig78a-soma-intra-all-extra, blocked for the cost of the optimizations.

Findings about the data

What attempts found about the dataset itself, for its owner and the archive.

Metadata errorThe dandiset's metadata name the paper by its title but link it to doi:10.1162/neco_a_01676, which Crossref gives as "Energy Complexity of Convolutional Neural Networks". The paper is doi:10.1162/neco_a_01672.
Metadata errorThe CurrentClampSeries are stated in volts with conversion 1 but hold millivolts (-100 to +46), and the CurrentClampStimulusSeries are stated in amperes with conversion 1 but hold picoamperes (holding levels of 12 and -37, steps up to about 1000).
OtherThe patch timestamps are at 20.41 kHz (49 us) for every protocol except IDthres (20 kHz) and APWaveform (125 kHz, 8 us). The paper states digitization at 20 kHz, and APWaveform at high rates, typically 200 kHz. The intervals are regular and agree with the nominal step durations.
Metadata errorThe icephys sequential recordings table leaves out the first sweep of every protocol after IDthres in each run and every PosCheops sweep (35 of 300 recordings for cell 1, 28 of 240 for cell 2), so the grouping tables do not describe the protocols. The protocol_name and run columns of the intracellular recordings table are complete.
Description missingNo culture age, plating date or days in vitro is given in the files or the dandiset's metadata; the subject has only an id, a species and a description.
Description missingThe electrodes table's x, y and z are empty (NaN), location is "unknown" and impedance -1; electrode positions are only in the custom columns rel_x and rel_y, without a description of their frame.
Selection not documentedCell 2's file holds runs 3 to 6 of the recording, renamed run1 to run4; only the descriptions of its MEA series and the paper's export notebook say so. Runs 1 and 2 of cell 2 are not deposited, and the dandiset does not say why.
OtherIn every run 117 to 132 of the 675 to 678 MEA channels are constant for at least one of ten one-second segments (the channels the paper's code removes before computing templates); the deposited series keep them, with nothing marking them.
OtherIn the deposited sweeps the stimulus starts 50 ms earlier than the timings of the paper's feature code (350 ms against ton 400 ms for IDrest, 450 against 500 for firepattern), so the code's timings cannot be applied to the files as they are.
Paper inconsistentMethods 5.2 describe IV as 11 steps from -140 to 20% by 20%, which is 9 steps; the data hold 11 steps from -140 to 60%.
Paper inconsistentMethods 5.4 give 79 intracellular features for optimization and Table 1 gives 76; the code's feature files have 79 before it drops three sag features of IV -20%, and 76 after.
Methods not availableFor fig3b-ap-initiates-distal-ais, the paper does not say how the initiation site is read from Figure 3B or at what time resolution. The code's cut-outs (get_peak_cutout) are on a 0.05 ms grid, on which the distal and middle AIS tie at the -20 mV crossing and at the peak; they are ordered only between samples or in a finer simulation (5 to 7 us at the crossing, 20 to 25 us at the peak). Question for the authors - was the order read from a finer simulation than the plotted cut-outs?

Questions

Where the author or a verifier believed the claims file misread the paper, they asked, citing the paper, and a reader who had not seen the data answered from the paper alone, amending the claims file where it was wrong. The author then ran the claim as answered. Attempts before protocol pilot-11 also asked here for method details the claims file did not give.

openasked 2026-10-30about mea-electrodes-and-rates
The criterion allows the APWaveform patch series only at 20 or 200 kHz. The Methods say "high sampling rates (typically 200 kHz)", which reads as a range of high rates, not one value. Should any rate well above 20 kHz (the deposited APWaveform sweeps are at 125 kHz, 8 us) count as matching for APWaveform?
Passage. APWaveform: This protocol was defined to precisely record one to two action potentials at high sampling rates (typically 200 kHz). (Methods 5.2); Whole-cell current-clamp recordings were low-pass-filtered at 10 kHz and digitized at 20 kHz. (Methods 5.6.5)
Why it matters. Under the claims file APWaveform at 125 kHz fails; under the wider reading it passes. The outcome of this attempt does not change, since six other protocols are at 20.41 kHz (49 us) and already fail the 20 kHz requirement with tolerance 0, but the APWaveform part would. The independent verifier raised it (round 1, problem 3).
No reader has answered it yet.
openasked 2026-10-30about holding-potential
The criterion reads the membrane potential "before the stimulus onset of each sweep (the code's ton, e.g. 400 ms for IDrest)". The Methods give the eCode protocols without the delay before the step. Should the pre-stimulus window end at the stimulus onset, wherever it falls, with the code's ton only an example, or at the code's ton?
Passage. the holding current was manually adjusted to obtain a resting membrane potential of around -70 mV (note that with LJP correction, this corresponds to -84 mV) (Methods 5.6.5)
Why it matters. In the deposited sweeps the stimulus starts 50 ms before the code's ton (350 ms against 400 ms for IDrest), so a window ending at ton includes 50 ms of the step. The outcome of this attempt does not change, since both windows give the same medians to within 0.1 mV (-69.54 and -69.48 mV for cell 1, -69.12 and -69.04 mV for cell 2), but a reading of the criterion that takes ton literally would measure part of the response. The independent verifier raised it (rounds 1 and 2).
No reader has answered it yet.

Notebooks

Every registered entry that uses this dataset. An entry can hold several notebooks, and can use more than one dataset.

Reproduction Patch-clamp and HD-MEA recordings of two cultured neurons (Buccino et al. 2024)

Tests the findings of Buccino et al. (2024) on dandiset 000294, which holds simultaneous patch-clamp and high-density microelectrode array recordings of two cultured rat cortical neurons stimulated with the eCode protocols. It checks the runs, the stimulus protocols, the holding potential, the extracellular spike amplitude and the recording rates, and, from the paper's code repository, the reconstructed morphologies, the ground-truth model's rheobase, AP initiation and template, and the feature counts. The model fits are too large to run here. All files

Reproducing Buccino et al. (2024) from dandiset 000294
reproduce_buccino_2024.ipynbExecuted, with outputs · Source

Verification

Round 1
2026-10-30Agreed with 23 of 23 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The handover of the first run, the executed notebook with the conclusions and summary removed, the helper modules and the claims file, with the paper under paper/. The verifier also read the dandiset's metadata in the archive.
Round 2
2026-10-30Agreed with 23 of 23 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The full executed notebook of the second run, with its conclusions and summary, the helper modules, the claims file, the README and this record, with the paper under paper/.

Method differences

How the tests differed from the paper’s method or data, and the claims each applies to.

The 0% IV step of reading (a) has no relative tolerance; 10% of the IV increment (2% of the scale) was used. Its measured level is within 2% of holding in every run.ecode-protocol-structure
APWaveform's sweep count (6) and PosCheops' (1) are not stated in the Methods; 6 is the number of amplitudes from 200 to 350% by 30%.ecode-protocol-structure
The pre-stimulus window ends at the first change of the injected current found in each sweep (350 ms for most protocols), which in these files is 50 ms before the code's ton (400 ms for IDrest); the window ending at the code's ton is shown as a variant.holding-potential
The band-pass filter is applied forward and backward (sosfiltfilt), the default of the code's filtering library; the Methods give only the order and cut-offs. A standard choice that cannot move amplitudes of 159 uV and more below 50 uV.extracellular-spikes-above-50uv
Simulated with BluePyOpt 1.14.26 (NEURON 9) on the repository's model definition, in place of BluePyOpt 1.13.88; the main series uses its NEURON simulator, and the code's compute_rheobase_for_model on its LFPy simulator, as the notebook uses, is shown and agrees.gt-rheobase
Assumed. The paper does not say how the initiation site is read from Figure 3B (Results 3.2, "The AP is initiated at the distal end of the AIS and travels back toward the soma"), nor at what resolution. The times are read between the 0.05 ms samples of the code's cut-outs and checked on a fixed-step run at 0.005 ms, because a 0.05 ms grid cannot order two points 15 um apart that the simulation separates by 5 to 7 us; read at the sample, the two AIS points tie under both readings and the claim's strict order would not be met.fig3b-ap-initiates-distal-ais
Simulated with BluePyOpt 1.14.26 in place of 1.13.88; the code's protocol, recordings and get_peak_cutout run unchanged on its LFPy simulator. The fixed-step check uses its NEURON simulator (dt 0.005 ms, a standard choice).fig3b-ap-initiates-distal-ais
The AIS is placed with the segment midpoints of the instantiated LFPyCellModel (BluePyOpt 1.14.26, LFPy 2.3.7), where LFPy puts the soma midpoint at the origin. In the first round it was placed from NEURON's 3D points relative to the mean soma point, which puts the AIS midpoint at (1.7, -23.4) um, about 14 um off; that placement also gives channel 43 as nearest (23.3 um, next 26.8 um) and is shown as a variant.fig3a-largest-channel-at-ais

Provenance

Attempt record: 000294-buccino-2024-01
Paper: 10.1162/neco_a_01672
Date: 2026-10-29
Made by: An agent, run by CatalystNeuro
Dataset: DANDI:000294
Version: 0.260528.0854 (doi:10.48324/dandi.000294/0.260528.0854)
Run on: GitHub Actions
Harness: Claude Code
Model: claude-opus-5-5
Protocol: pilot-12

Transcripts of the agent runs

Each run’s prompt, messages and tool calls, with long output cut and the text of the paper left out, and its final report.

← DANDI:000289 DANDI:000336 →