Stability of motor representations after paralysis (DANDI:000147)

Claims

From the paper Stability of motor representations after paralysis

Reproducing Guan, Aflalo et al. (2022) from dandiset 000147 14

Notebook reproduce_guan_2022.ipynb

Not reproducedFigure 1-figure supplement 6b
A diagonal LDA classifier could not classify the cue location from neural activity in the attempted-movement window.
Diagonal LDA classifies the 12 cue locations at 11.5% against a chance level of 8.3% (binomial p = 1e-12), above all 500 label shuffles across trials (highest 10.0%) and all 500 shuffles within finger (highest 10.2%). The authors' classifier gives 11.2%, and removing each finger's mean activity first gives 11.7%. Five of the ten sessions are above chance at p < 0.05. With folds that never split a screen the accuracy is 11.1%, above all 200 shuffles of location within screens (highest 9.8%), and with folds that never split a run it is 11.3%. Every screen cues each location once; location and position within the screen are weakly associated over all trials (chi-square p = 0.037). The cued finger is classified at 93.4% from the same features. With folds that never split a run, 11.3% is above all 200 shuffles of location within runs (highest 9.8%). Position within the screen is itself decoded at 10.5%, above all 200 shuffles within screens (highest 9.5%), but after removing each position's mean activity location is still classified at 11.1%.
How the test differed from the paper’s method
- Eight stratified folds within each session; the paper does not say how this classifier was cross-validated, and the authors' code uses eight folds for its classifiers.
- The paper shows a confusion matrix without an accuracy or a test, so the comparison is with its statement that the location could not be classified.
ReproducedResults, "Intracortical recordings during finger flexion"
Over 10 sessions the recordings yielded 95.8 (SD 6.7) sorted neurons and 408 (SD 40.8) trials of the main finger task per session.
The ten files hold 95.8 units per session (SD 6.7) and 408.0 trials per session (SD 40.8), 4080 trials in all, as reported.
ReproducedFigure 1b, Figure 1-figure supplement 3, Results
Online BCI classification of the cued finger was 86% accurate (SD 4% across sessions, weighted by trials) over 4016 trials in 10 sessions, against a chance level of 17%. No-Go trials of the first session are excluded because the classifier could not output No-Go that day.
The recorded decoder outputs match the cued finger on 85.96% of the 4016 scored trials, with a trial-weighted SD across sessions of 3.6%. Shuffling the outputs gives 16.9%, the paper's chance level of 17%.
ReproducedResults, "Representational structure did not trend toward task optimum"; Figure 1b
Most (70%) of the online classification errors were middle-ring or ring-pinky confusions.
Of the 379 online errors between two fingers, 69.7% confuse middle with ring or ring with pinky. Counting errors that involve No-Go, the fraction is 46.8% of 564 errors, or 59.2% of the 446 errors on finger trials.
How the test differed from the paper’s method
- The paper does not say which errors form the denominator. Only errors between two fingers, leaving out No-Go, give the paper's 70%.
ReproducedFigure 1-figure supplement 4a, Results, Methods "Finger classification"
Offline classification of the cued finger from sorted firing rates in the window 0.5 to 1.5 s after the cue, with diagonal LDA and leave-one-out cross-validation within each session, was 92% accurate (SD 2% across sessions) over 4080 trials.
Diagonal LDA with leave-one-out cross-validation within sessions gives 93.4% (weighted SD 1.6%) over 4080 trials. The classifier and eight stratified folds of the authors' code give 92.4% (SD 1.7%).
How the test differed from the paper’s method
- The paper's methods say leave-one-out; the authors' code uses eight folds, a mean-count cutoff of 0.5, z-scoring and LDA with identity covariance. Both are reported, and the measured value is the leave-one-out one.
ReproducedResults, Figure 1-figure supplement 5, Methods "Single-unit tuning to finger flexion"
Most (89%) neurons were significantly tuned to at least one finger movement, by a two-tailed t-test of each finger's firing rate in the 0.5 to 1.5 s window against No-Go trials, p < 0.05 after FDR correction, with neurons below 0.1 Hz excluded and neurons pooled across sessions.
With FDR correction over all units and fingers pooled, 87.9% of 958 units are tuned to at least one finger. FDR within each session gives 87.5%, within each unit 83.9%, and no correction 91.7%. No unit fell below the 0.1 Hz cutoff, whether the rate is taken in the window over all trials (lowest 0.113 Hz) or over the whole recording (lowest 0.147 Hz).
How the test differed from the paper’s method
- The FDR family is not stated in the paper; the measured value pools all units and fingers.
- Student's t-test with equal variances; the paper does not say which form of the t-test.
ReproducedFigure 2d and 2e, Results
In the session-averaged representational dissimilarity matrix (cross-validated Mahalanobis distance, window 0.5 to 1.5 s, No-Go excluded), the thumb is distinct, while the middle, ring and pinky fingers are close.
In the session-averaged RDM, max-scaled, the distances among middle, ring and pinky are 0.21 to 0.48 and every distance involving the thumb is 0.64 or more. This holds in 9 of the 10 sessions on their own.
How the test differed from the paper’s method
- The noise covariance uses OAS shrinkage, as in the authors' code; the paper names Ledoit-Wolf.
ReproducedFigure 2f, Resultsuses data from outside the dandiset
The measured RDMs match the able-bodied PC-IP fMRI model better than the unstructured (pairwise-equidistant) model, by whitened unbiased cosine similarity (WUC), p = 5.7e-5.
Mean WUC is 0.881 with the able-bodied PC-IP model and 0.781 with the unstructured model. The PC-IP model is higher in all 10 sessions; the bootstrap 95% interval of the difference is 0.074 to 0.127, and the bootstrap t-test gives p = 4.4e-5 (paper 5.7e-5).
How the test differed from the paper’s method
- The model RDMs are read from the authors' code repository at the commit archived with the paper, not from the dandiset.
- The p-value comes from 1000 resamples of the sessions (mean difference over its bootstrap SD, nine degrees of freedom); the paper used rsatoolbox's bootstrap t-test.
ReproducedFigure 2-figure supplement 5a, Resultsuses data from outside the dandiset
The measured RDMs match the able-bodied motor cortex (MC) fMRI model better than the able-bodied PC-IP fMRI model (p = 1.9e-6), and the WUC with the MC model is close to the noise ceiling.
Mean WUC with the MC model is 0.969, higher than with the PC-IP model in all 10 sessions (difference 0.088, bootstrap 95% interval 0.060 to 0.112, p = 7.9e-5; paper 1.9e-6). It is 0.015 below the lower noise ceiling of 0.984, and at or above its own session's ceiling in 1 of 10 sessions.
How the test differed from the paper’s method
- The model RDMs are read from the authors' code repository, as for fig2f-able-bodied-over-unstructured.
- The criterion's margin of 0.05 below the noise ceiling is a choice; the paper says only "close to".
ReproducedResults, "Representational structure did not trend toward task optimum"
The finger RDMs were consistent across sessions, with an average pairwise correlation between session RDMs of r = 0.90 (SD 0.04, min 0.83, max 0.99).
The 45 pairs of session RDMs correlate at a mean Pearson r of 0.928 (SD 0.041, min 0.817, max 0.990), against 0.90 (SD 0.04, min 0.83, max 0.99). The mean is 0.028 above the paper's, inside the tolerance of 0.03.
ReproducedFigure 3a, Resultsuses data from outside the dandiset
Over sessions the model fit did not trend from the able-bodied model toward the unstructured model (session by model interaction t(6) = 0.50, one-tailed p = 0.32).
The session by model interaction is t(16) = 1.47, one-tailed p = 0.080 toward the unstructured model (paper t(6) = 0.50, p = 0.32). The WUC with the PC-IP model alone declines across sessions (slope -0.009 per session, p = 0.038); the slopes for the unstructured and MC models are not significant (p = 0.20 and 0.23). In other specifications the trend is also not significant. The same t read with 6 degrees of freedom gives p = 0.096; the difference between the unstructured and PC-IP fits regressed on session gives t(8) = 1.30, p = 0.116; with the MC model in place of PC-IP the slope is negative (p = 0.84).
How the test differed from the paper’s method
- Ordinary least squares on the twenty session-by-model WUC values, with sixteen residual degrees of freedom. The paper's t(6) implies a different model or degrees of freedom, which it does not describe, and its code for this test is not in the repository.
- Three specifications are shown, since the paper's t(6) is not explained.
- No Bayes factor is computed.
ReproducedFigure 3b, Results, Methods "Measuring changes in the representational structure"
The distances between the most confused finger pairs (middle-ring and ring-pinky) did not increase across sessions (one-tailed p > 0.99) or across runs within sessions (one-tailed p = 0.67).
For the middle-ring and ring-pinky distances of 85 per-run RDMs, the session term is t(8) = -3.56 (paper -4.5) and the run term t(83) = -1.05 (paper -0.45). In all six estimates the coefficients are negative and the one-tailed p for an increase is 0.73 or more.
How the test differed from the paper’s method
- The paper's code for this regression is not in its repository. The regression follows the methods' formula on per-run RDMs, with degrees of freedom of sessions minus two for the session term and runs minus two for the run term, as the paper's t(8) and t(82) suggest.
- The dandiset has 85 runs; the paper's t(82) implies 84.
- Per-run RDMs use five folds of eight trials each, so a fold's noise covariance rests on one or two trials per finger and inflates the distances (mean 5.06 against 0.68 for session RDMs). A variant with the session's noise covariance (mean 0.76) is reported.
- No Bayes factors are computed.
ReproducedFigure 3-figure supplement 1, Results
Across all finger pairs the inter-finger distances did not increase; they decreased slightly across sessions (t(8) = -4.0, two-tailed p = 0.004) and across runs within sessions (t(82) = -2.4, p = 0.019), with very small effect sizes (Cohen's f2 of 0.008 and 0.005).
Over all ten pairs the session and run coefficients are negative in all six estimates, and Cohen's f2 is 0.006 for session and 0.012 for run in the main regression, and 0.017 and 0.013 with the session's noise covariance (paper 0.008 and 0.005). In the main regression the session decrease is t(8) = -2.28, p = 0.052, and the run decrease t(83) = -3.12, p = 0.0025. With the session's noise covariance they are t(8) = -3.76, p = 0.0055, and t(83) = -3.29, p = 0.0015, against the paper's t(8) = -4.0, p = 0.004, and t(82) = -2.4, p = 0.019.
How the test differed from the paper’s method
- The same regression and choices as fig3b-confused-pairs-no-increase.
- The session decrease is significant at 0.05 only in the variant with the session's noise covariance; the pre-registered criterion asks for negative coefficients and small effect sizes, not significance. The per-run fold covariance inflates the distances of the main regression, so the variant with the session's covariance is the better estimate of the session term.
BlockedFigure 1-figure supplement 6a· Metadata insufficient
Linear regression could not decode the cue's screen coordinates from neural activity in the attempted-movement window; cross-validated r2 was close to 0 in every session.
The trials table gives the cue location as an index from 1 to 12 on the grid of three rows and four columns, described only as a screen location index, with no map to the grid or to screen coordinates. As a check, the coordinates were regressed on spike counts under the two plain orders of the index. Ordinary least squares gives a cross-validated r2 below 0 for both coordinates in all ten sessions under both orders. Ridge regression gives r2 up to 0.097 for x and 0.104 for y, with both coordinates at or below 0.05 in 5 (row order) or 6 (column order) of the ten sessions. The result depends on the regression as well as on the unknown layout.
How the test differed from the paper’s method
- The check assumes the index runs along the rows or down the columns; neither is documented.
- The paper does not say whether its regression was regularised, so both ordinary least squares and ridge regression are shown, each with eight unshuffled folds per session.

Not tested in a notebook 6

6 data missing from the dandiset

BlockedFigure 1-figure supplement 4b, Results· Data missing from the dandiset
In the calibration task, offline cross-validated finger classification was 96% accurate (SD 3%) over 530 trials in nine sessions.
The dandiset holds only the main finger task. The calibration task, recorded at the start of each day, is not in any of the ten files.
BlockedFigure 1-figure supplement 6c· Data missing from the dandiset
Fingers could not be classified from neural activity during the control task, in which the participant viewed the cues without attempting movements.
The no-movement control task is not in the dandiset; every file is described as the main task and contains only its trials.
BlockedResults, Figure 2-figure supplement 4· Data missing from the dandisetuses data from outside the dandiset
Compared with the RDMs of individual able-bodied fMRI participants, participant NS's average PC-IP RDM was statistically typical (permutation test, p = 0.55).
The test needs the RDMs of the 36 individual able-bodied fMRI participants. Neither the dandiset nor the authors' code repository holds them; the repository has only the group means.
BlockedFigure 4e, Figure 4-figure supplement 2· Data missing from the dandisetuses data from outside the dandiset
In a time-resolved decomposition of the RDM into muscle and somatotopy component models, the muscle model emerged 170 ms before the somatotopy model (two-sided Wilcoxon signed-rank test across sessions, p = 0.002).
The decomposition needs the muscle and somatotopy model RDMs of Ejaz et al. (2015). They are not in the dandiset, and the authors' code repository loads them from a models folder that holds only the two fMRI models and the unstructured model at the archived commit.
BlockedFigure 4-figure supplement 1, Results· Data missing from the dandisetuses data from outside the dandiset
Among combinations of the hand-usage, muscle and somatotopy models, the hand usage plus somatotopy and muscle plus somatotopy combinations matched the data best.
The hand-usage, muscle and somatotopy model RDMs are not in the dandiset or the authors' code repository.
BlockedFigure 4-figure supplement 4· Data missing from the dandiset
The able-bodied match and the representational dynamics hold when only well-isolated units (lowest third of L-ratio) are used.
The units table has spike times, observation intervals, electrodes and electrode groups only, the same in every file. There are no waveforms, features or L-ratio values from which to select well-isolated units, and the representational dynamics part also needs the missing component models.

Findings about the data

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

Data missingThe dandiset holds only the main finger task. The calibration task (nine sessions) and the no-movement control task (six sessions) analysed in the paper are not included.
Data missingThe model RDMs that the paper's main comparisons rest on are not in the dandiset. The two able-bodied fMRI group RDMs and the unstructured model are in the authors' code repository; the muscle, hand-usage and somatotopy models and the individual fMRI participants' RDMs are in neither.
Data missingThe units have no waveforms or sorting quality metrics, so the paper's selection of well-isolated units by L-ratio cannot be repeated.
OtherSpike times are not in ascending order within units. In nearly every unit the spikes of the first seconds of each run are stored before the rest of the previous run (590 to 1016 backward jumps per file). No trial's analysed range meets an affected span, so sorted spike times give the same counts, but tools that assume ascending spike times could be misled near run boundaries.
OtherThe cue location is an index from 1 to 12, described as the screen location index of the cue, with no map to the grid of three rows and four columns or to screen coordinates. The trials table's Run and TrialNumber columns read "no description", and the start time has only the NWB default description. Decoding the finger over time is consistent with a start time at the cue onset, but does not pin it to the cue within a few hundred milliseconds.
OtherElectrode coordinates and impedances are missing (NaN), and the participant is identified as P1 in the files and as NS in the paper.
OtherThe paper says the first session had 84 No-Go trials. The file has 64, and the paper's own total of 4016 scored trials (4080 minus the first session's No-Go trials) implies 64.
Selection not documentedThe two published versions hold the same ten files. The earlier version, which the authors' download script pins, names the bioRxiv preprint; the later one names the eLife paper. Neither says which of the paper's tasks the data cover.

Notebooks

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

Reproduction Finger representations in human posterior parietal cortex after paralysis (Guan et al. 2022)

Tests the findings of Guan, Aflalo et al. (2022) that dandiset 000147 can support. It checks online and offline finger decoding, single-unit tuning and whether the cue location can be decoded, builds the cross-validated finger RDMs of the ten sessions, compares them with the able-bodied fMRI models from the authors' repository, and tests whether the representation changed across sessions and runs. The calibration and control tasks, the component model RDMs and sorting quality metrics are not in the dandiset, so the claims that need them are blocked. All files

Reproducing Guan, Aflalo et al. (2022) from dandiset 000147
reproduce_guan_2022.ipynbExecuted, with outputs · Source

Verification

Round 1
2026-10-13Agreed with 20 of 20 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The first executed notebook with the conclusion cells and the summary removed, the helper module, the claims file and the paper's full text (PMC9555862). For the blocked claims the verifier checked the archive's asset list and metadata and the authors' repository.
Round 2
2026-10-13Agreed with 20 of 20 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 conclusion cells and summary, the helper module, the README and this attempt record, against the verifier's round-1 points.

Method differences

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

The paper does not say which errors form the denominator. Only errors between two fingers, leaving out No-Go, give the paper's 70%.fig1b-errors-middle-ring-pinky
The paper's methods say leave-one-out; the authors' code uses eight folds, a mean-count cutoff of 0.5, z-scoring and LDA with identity covariance. Both are reported, and the measured value is the leave-one-out one.fig1s4a-offline-accuracy
The FDR family is not stated in the paper; the measured value pools all units and fingers.fig1s5-tuned-fraction
Student's t-test with equal variances; the paper does not say which form of the t-test.fig1s5-tuned-fraction
The check assumes the index runs along the rows or down the columns; neither is documented.fig1s6a-gaze-regression
The paper does not say whether its regression was regularised, so both ordinary least squares and ridge regression are shown, each with eight unshuffled folds per session.fig1s6a-gaze-regression
Eight stratified folds within each session; the paper does not say how this classifier was cross-validated, and the authors' code uses eight folds for its classifiers.fig1s6b-gaze-classification
The paper shows a confusion matrix without an accuracy or a test, so the comparison is with its statement that the location could not be classified.fig1s6b-gaze-classification
The noise covariance uses OAS shrinkage, as in the authors' code; the paper names Ledoit-Wolf.fig2d-rdm-structure
The model RDMs are read from the authors' code repository at the commit archived with the paper, not from the dandiset.fig2f-able-bodied-over-unstructured
The p-value comes from 1000 resamples of the sessions (mean difference over its bootstrap SD, nine degrees of freedom); the paper used rsatoolbox's bootstrap t-test.fig2f-able-bodied-over-unstructured
The model RDMs are read from the authors' code repository, as for fig2f-able-bodied-over-unstructured.fig2s5-mc-over-pcip
The criterion's margin of 0.05 below the noise ceiling is a choice; the paper says only "close to".fig2s5-mc-over-pcip
Ordinary least squares on the twenty session-by-model WUC values, with sixteen residual degrees of freedom. The paper's t(6) implies a different model or degrees of freedom, which it does not describe, and its code for this test is not in the repository.fig3a-no-trend-to-unstructured
Three specifications are shown, since the paper's t(6) is not explained.fig3a-no-trend-to-unstructured
No Bayes factor is computed.fig3a-no-trend-to-unstructured
The paper's code for this regression is not in its repository. The regression follows the methods' formula on per-run RDMs, with degrees of freedom of sessions minus two for the session term and runs minus two for the run term, as the paper's t(8) and t(82) suggest.fig3b-confused-pairs-no-increase
The dandiset has 85 runs; the paper's t(82) implies 84.fig3b-confused-pairs-no-increase
Per-run RDMs use five folds of eight trials each, so a fold's noise covariance rests on one or two trials per finger and inflates the distances (mean 5.06 against 0.68 for session RDMs). A variant with the session's noise covariance (mean 0.76) is reported.fig3b-confused-pairs-no-increase
No Bayes factors are computed.fig3b-confused-pairs-no-increase
The same regression and choices as fig3b-confused-pairs-no-increase.fig3s1-all-pairs-small-decrease
The session decrease is significant at 0.05 only in the variant with the session's noise covariance; the pre-registered criterion asks for negative coefficients and small effect sizes, not significance. The per-run fold covariance inflates the distances of the main regression, so the variant with the session's covariance is the better estimate of the session term.fig3s1-all-pairs-small-decrease

Provenance

Attempt record: 000147-guan-2022-01
Paper: Stability of motor representations after paralysis
Date: 2026-10-13
Made by: An agent, run by CatalystNeuro
Dandiset version: 0.221122.2256
Harness: Claude Code
Model: claude-opus-5-5
Protocol: pilot-2

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.

← Dandiset 000130 Dandiset 000207 →