Claims 34
From the paper DREDge: robust motion correction for high-density extracellular recordings across species · claims file · checked 2026-10-21: ready
- edf3a-ap-dredge-slow-trend
In the fast-motion macaque recording, DREDge's spike-based (AP) motion estimate follows the drift visible in the unregistered spike raster.
a Spike raster before registration, with AP DREDge displacement estimate overlaid in red.
Extended Data Fig. 3a legend; Methods, Tracking fast motion from spikes using clustering and splinesHow it will be judged
The legend shows the estimate over the raster and states no result in words; that it follows the slow drift is read from the Methods ("First, slower timescale motion is estimated using DREDge"). Run the pipeline (opening comment, steps 1 and 2) on the whole recording. The paper shows this only as a figure, so the measure is the paper's own registration quality metric, the mean template correlation (Methods, Spike registration quality metrics) on a 1 um by 1 s raster of mean spike amplitudes, bins outside the probe after correction masked. Reproduced when, under readings (a) and (b) of step 2 and (i) and (ii) of step 1, the mean template correlation after DREDge correction exceeds that of the unregistered raster; not reproduced when it is lower under all; methods_underspecified when the readings disagree. The size of the estimated displacement (range and s.d. in um) and the plot of the trace over the raster are reported. That a small improvement may reflect little drift in an anesthetized, paralyzed animal is noted, not judged.
- edf3b-hdbscan-clusters
After AP DREDge registration, high-amplitude spikes form clusters that HDBSCAN isolates using their registered positions and amplitudes.
Colored spikes represent high amplitude spikes that were clustered by HDBSCAN using their registered positions and amplitudes.
Extended Data Fig. 3b legend; Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 1How it will be judged
Pipeline step 3, HDBSCAN on the DREDge-registered spikes, under each combination of the open readings named there (amplitude cut, coordinates, min_cluster_size) and of registration readings (a) and (b). The legend shows several colored clusters and gives no count, so the claim holds when HDBSCAN returns at least two clusters other than noise. Reproduced when every combination does; not reproduced when none does; methods_underspecified otherwise. The number of clusters and spikes per cluster are reported for each.
- edf3c-spline-subsecond
A smoothing spline fit to the centered positions of the clustered spikes captures sub-second displacement left after AP DREDge registration.
Left, centered positions of the clustered spikes (blue) with the spline fit, that captures the sub-second displacement, overlaid in red.
Extended Data Fig. 3c legend; Methods, Tracking fast motion from spikes using clustering and splinesHow it will be judged
Pipeline step 3 under each combination of its open readings. The paper gives no number and no test for this panel. The control is borrowed from another panel of the paper, the "randomly permuted or 'scrambled'" motion estimate of Fig. 2e (legend), which the paper uses for ISI violations, not for this figure, and whose permutation unit it does not give. For each combination, the pooled within-cluster s.d. of corrected depth of the clustered spikes is computed after correction by the spline and after correction by the same spline trace with its time bins randomly permuted. The permutation bin (0.4 s, the knot spacing), the 100 permutations with a fixed seed and the one-sided 5% level are choices of this file, not of the paper. Reproduced when, in every combination, the spline-corrected s.d. is below the 5th percentile of the scrambled s.d. values; not reproduced when in none; methods_underspecified otherwise. The spline trace's range and s.d. (um) and its power spectrum are reported, as is the noted circularity that the spline is fit to the spikes it is evaluated on.
- edf3de-spline-improves-registration
Correcting the AP-registered spikes with the spline estimate of sub-second displacement gives a more stable spike raster and tighter spike localizations than AP DREDge registration alone.
d Top, AP registered spike raster. Bottom, AP registered spike raster corrected using the spline estimate of sub-second displacement. e Localizations of detected spikes after AP DREDge registration (left), and spline correction (right).
Extended Data Fig. 3d and 3e legends; Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 1How it will be judged
The legends show the comparison and state no result in words; the improvement is read from the Results, which call the spline estimate "a faster timescale AP-based estimate". Two measures, one per panel, both from the paper. Panel d, the mean template correlation (Methods, Spike registration quality metrics) of all detected spikes, on the 1 um by 1 s raster of mean amplitudes, after AP DREDge correction and after AP DREDge plus spline correction (the spline displacement subtracted from each spike's registered depth at its time). Panel e, the pooled within-cluster s.d. of corrected depth over the HDBSCAN clusters, before and after spline correction; the paper names no measure for panel e, so this one is the file's choice, and since the spline is fit to minimize these residuals it is reported with that circularity noted. The 1 s bins of the template correlation are the Methods' and cannot show motion faster than 1 s, so the panel d measure is also reported at 0.1 s bins as a variant that does not enter the outcome. Reproduced when both measures improve (higher correlation, lower s.d.) under every combination of the open readings of steps 2 and 3; not reproduced when neither improves under any; partially reproduced when one measure improves under every combination and the other does not; methods_underspecified when the readings disagree. The size of each change is reported.
- fig2b-ap-misses-fast-motion
In a human recording with large fast motion, DREDge's AP-based estimate captures the slow trend of the motion but not the faster motion from heartbeats and other sub-second brain motion.
DREDge’s AP-based motion estimation was able to capture the slow trend of the true motion, but not the faster motion due to heartbeats and other sub-second brain motions
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 1; Fig. 2bHow it will be judged
Judged on the recording the paper names. Not tested on this dandiset, since the paper places Fig. 2b in a human recording.
- fig2b-spline-lfp-overlap
In the Fig. 2b human recording, the spline-based AP motion trace and the LFP-based DREDge trace overlap strongly, and the LFP-corrected spike raster aligns visually with the spline-corrected raster.
We found that the motion traces estimated using the spline-based AP method and DREDge-LFP overlapped strongly and that the LFP-corrected spike raster visually aligned with the spline-corrected raster
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 1; Fig. 2bHow it will be judged
Not tested here. The paper reports no number for the overlap.
- fig2-lfp-vs-manual
DREDge's LFP-based motion estimate is highly correlated with manual LFP-based tracking in three human participants (Pt01 r = 0.98, Pt02 r = 0.99, Pt03 r = 0.85).
DREDge’s estimate was highly correlated with manual LFP-based tracking (Pt01, r = 0.98; Pt02, r = 0.99; Pt03, r = 0.85; Pearson’s r , P < 0.000001 for all three instances
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 1; Extended Data Fig. 2b - fig2e-isi-violations
In human recordings, the inter-spike interval violation proportion of sorted clusters decreases as the temporal resolution of LFP-based motion tracking used for interpolation increases from 1 Hz to 250 Hz.
the clusters’ inter-spike interval violation proportions decreased as the temporal resolution of LFP-based motion tracking increased from 1 Hz to 250 Hz
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 3; Fig. 2e - fig2-sorted-units
After LFP-based motion correction, the number of sorted units in the human recording falls from more than 500 to around 50, unregistered clusters being oversplit.
the number of sorted units decreased from more than 500 without correction to around 50
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 3; Fig. 2d and 2hHow it will be judged
The count before correction is a bound ("more than 500"); it holds when the unregistered sort gives more than 500 units, and the corrected count is within the expected value's tolerance.
- fig3b-evoked-potentials
After motion correction and Zapline-plus, the dmPFC LFP shows depth-specific evoked potentials that differ between black and white squares.
the preprocessed dmPFC LFP showed depth-specific evoked potential responses to the visual stimuli that differentiated between the black and white squares ( n = 1; P < 0.01, Wilcoxon rank-sum test per time point, false discovery rate-corrected for multiple comparisons)
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 4; Fig. 3b - fig3e-trial-variance
Before correction, motion produces significantly higher voltage variance across visual trials (channel average at 0.25 s after image onset) than after correction.
significantly higher motion-induced voltage across trials with visual presentations (averaged across channels at 0.25 s after image onset; n = 1; P < 0.000001, Kruskal–Wallis test;
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 4; Fig. 3e - fig3f-baseline-variance
In baseline data without stimuli, the correction pipeline decreases voltage variance over time and channels in each of six participants.
this correction pipeline decreased voltage variance over time and channels ( n = 6; P < 0.000001, pairwise Wilcoxon rank-sum test per participant;
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 4; Fig. 3f - edf8-burst-suppression
Burst suppression detections in the common median LFP are highly correlated before and after correction (Pt01 r = 0.93, Pt03 r = 0.95).
with high correlation in burst detections before and after correction in the common median voltage (Pt01, r = 0.93, P < 0.00001; Pt03, r = 0.95, P < 0.00001)
Results, DREDge rescues spike sorting and LFP features in human brain activity, paragraph 5; Extended Data Fig. 8 - fig4-nhp-insertion
During deep NP1-NHP insertions in macaques, DREDge tracks motion across centimeters where Kilosort's template-based tracking fails, and spike positions resolve into well-isolated clusters after registration.
While single-unit clusters were completely obscured before correction, spike positions resolved into well-isolated clusters after registration.
Results, Tracking drift during deep probe insertions in nonhuman primates, paragraphs 1 and 2; Fig. 4a and 4b; Supplementary Fig. 6 - fig5b-ibl
In 47 IBL recordings, DREDge improves the mean template correlation over no registration (mean difference 0.04) and over Kilosort (0.01), and scores higher than Kilosort in every recording.
DREDge consistently improved over both no registration and KS in n = 47 recordings (metric mean differences 0.04 and 0.01 respectively; two-sided paired t -test P = 2.1 × 10 −9 and 2.9 × 10 −18 with 46 d.f.;
Results, Estimating motion in acute mouse recordings, paragraph 2; Fig. 5b - fig5c-np-ultra-metric
In 12 NP Ultra recordings, DREDge improves the template correlation metric by 0.1 over no correction and 0.07 over Kilosort, while Kilosort does not improve over no correction.
mean differences 0.1 and 0.07 relative to uncorrected and KS; two-sided paired t -test P =0.0004 and P = 0.015), whereas KS did not improve over no correction (two-sided paired t -test P = 0.27)
Results, Estimating motion in acute mouse recordings, paragraph 3; Fig. 5c - fig5d-np-ultra-jumps
In the NP Ultra recordings, DREDge produces fewer physically implausible jumps (time bins with drift above 10 um/s) than Kilosort, 392 fewer per recording on average.
DREDge also produced fewer physically implausible jump artifacts than KS on these datasets (392 fewer per recording on average; paired t -test P = 7.1 × 10 −5 with 11 d.f.;
Results, Estimating motion in acute mouse recordings, paragraph 3; Fig. 5d - supp13-simulated-short-probe
In six simulated recordings on a shortened NP2-like layout, DREDge improves over Kilosort.
We replicated these improvements in a diverse set of six simulated extracellular recordings created with a shortened NP2-like layout
Results, Estimating motion in acute mouse recordings, paragraph 4; Supplementary Fig. 13 - supp14-probe-density
In simulation, spike-based motion estimation stays reliable as channels are subsampled until the interchannel distance reaches about 50 um.
Motion estimation was consistently reliable until the probe density became too low to localize spikes (interchannel distance ≈ 50 μm)
Results, Estimating motion in acute mouse recordings, paragraph 4; Supplementary Fig. 14 - fig6-chronic-first-sessions
In ten chronic datasets, restricted to the first four sessions, DREDge improves the template correlation metric over Kilosort by 2.6% on average (s.e. 0.47%).
corresponding to a quantitative improvement in the correlation metric when restricted to the first four sessions of the ten datasets (mean improvement 2.6%, s.e. 0.47% over n = 40 sessions; two-sided paired t -test P < 1.95 × 10 −6 with 39 d.f.)
Results, Estimating motion in chronic mouse recordings, paragraph 3; Fig. 6eHow it will be judged
The tolerance on the mean improvement is the paper's stated s.e. (0.47%).
- fig6e-chronic-jumps
In the chronic datasets DREDge always reduces jumps relative to Kilosort (by 0.5 to 22.2 per minute, eight on average) and has no jumps in 74% of sessions against 30% for Kilosort.
DREDge always reduced jumps, with differences in mean jumps per minute ranging from 0.5 to 22.2, with an average of eight more jumps per minute in the KS motion trace. DREDge had no jumps in 74% of sessions versus 30% for KS.
Results, Estimating motion in chronic mouse recordings, paragraph 3; Fig. 6eHow it will be judged
The paper does not say over what unit the differences in mean jumps per minute are taken; the readings are (a) per dataset, the mean over its sessions (ten values), and (b) per session. The reported range 0.5 to 22.2 holds when every difference lies within it widened by 10% of each bound (0.45 to 24.42).
- edf1-lfp-runtime
Online nonrigid LFP motion estimation at 250 Hz runs at roughly one quarter of real time.
Online nonrigid motion estimation at 250Hz temporal resolution ran at roughly one quarter of real time (right).
Extended Data Fig. 1b legend - fig2c-lfp-tracking-clusters
In a human NP1-S recording with millimeters of drift, LFP-based DREDge tracks features of both the LFP and the spike raster, and offsetting spike positions by the tracked motion reveals isolated clusters with well-stereotyped waveforms.
Offsetting spike positions to compensate for the tracked motion revealed isolated clusters of spike activity ( Fig. 2c(ii , iii )).
Results, DREDge rescues spike sorting and LFP features in human brain activity; Fig. 2c - edf4a-nonrigid-human
In human recordings, nonrigid LFP-based correction gives more stable spiking activity than rigid correction, visibly and in a metric of spike activity stability.
it led to visible improvements in the stability of spiking activity over rigid correction, as well as improvements in a metric of spike activity stability
Results, DREDge rescues spike sorting and LFP features in human brain activity; Extended Data Fig. 4a and 4c - supp2-rat-spike-tracking
In ketamine-xylazine-anesthetized rats, LFP-based tracking gives an artifactually oscillatory motion estimate, while spike-based tracking is stable.
were dominated by global slow-wave activity, leading to an artifactually oscillatory motion estimate; however, spike-based tracking was stable in this case.
Results, DREDge rescues spike sorting and LFP features in human brain activity; Supplementary Fig. 2 - fig2fg-unit-spatial-spread
After 250 Hz LFP-based motion-corrected interpolation, the spatial distribution of voltage in each sorted human unit is more concentrated, with higher amplitudes.
the spatial distribution of voltages in each unit was more concentrated with higher amplitudes with this correction
Results, DREDge rescues spike sorting and LFP features in human brain activity; Fig. 2f and 2g; Extended Data Fig. 5b - edf6-128-channel
In three human awake intraoperative recordings with a 128-channel probe of 30 um vertical spacing, DREDge's tracking stabilizes the spike rasters.
In three such recordings, DREDge’s tracking yielded stabilized spike rasters after motion correction of spike locations
Results, DREDge rescues spike sorting and LFP features in human brain activity; Extended Data Fig. 6 - fig3gh-iids
Epileptiform interictal discharges are detectable in the median signal over channels both before and after correction, and are not obscured by the preprocessing.
was large enough to be detected in the median signal over channels before and after correction ( Fig. 3h ) and was not obscured by this preprocessing
Results, DREDge rescues spike sorting and LFP features in human brain activity; Fig. 3g and 3h; Supplementary Fig. 4 - fig4f-insertion-anatomy
During the deep macaque insertion, averages of waveform features as a function of motion-corrected spike depth vary in a way that roughly aligns with the expected anatomical boundaries along the trajectory.
Visualizing averages of these features as a function of motion-corrected spike depth revealed variations that roughly aligned with expected anatomical boundaries along the insertion trajectory
Results, Tracking drift during deep probe insertions in nonhuman primates; Fig. 4d to 4f - fig5a-imposed-zigzag
In NP1 and NP2 recordings with imposed zigzag motion of about 50 um, DREDge recapitulates the performance of Kilosort.
In these recordings, DREDge recapitulated the performance of KS
Results, Estimating motion in acute mouse recordings; Fig. 5a(i, ii) - edf4d-nonrigid-ibl
In nine further IBL recordings across brain regions, DREDge's nonrigid correction gives higher template correlation scores than its rigid correction.
it led to visual improvements ( Extended Data Fig. 4b ) and higher correlation scores when compared to rigid correction
Results, Estimating motion in acute mouse recordings; Extended Data Fig. 4b and 4d - supp10-11-simulated-drift
In MEArec simulations mimicking the IBL recordings, DREDge and Kilosort both track the simulated motion accurately, with DREDge better under discontinuous bump drift.
DREDge and KS accurately tracked simulated motion ( Supplementary Figs. 10 and 11 ), with DREDge showing improvements
Results, Estimating motion in acute mouse recordings; Supplementary Figs. 10 and 11 - supp12-subsetted-channels
When the NP1 and NP2 recordings of Fig. 5a are subsetted to the 282 um span of the Ultra probe, DREDge improves over Kilosort as in the NP Ultra data.
we spatially subsetted the channels of the NP1 and NP2 recordings of Fig. 5a to fit inside the 282-μm span of the Ultra probe and found similar improvements of DREDge over KS
Results, Estimating motion in acute mouse recordings; Supplementary Fig. 12 - fig6cd-chronic-clusters
In a chronic NP2 recording (AL032, shank 1), clusters isolated on DREDge-corrected positions have well-stereotyped waveforms across sessions, while the Kilosort estimate gives them drifting, jumpy trajectories, evidence that DREDge tracks the probe more accurately.
This provided evidence that DREDge was tracking the probe trajectory more accurately than KS.
Results, Estimating motion in chronic mouse recordings; Fig. 6c and 6d; Supplementary Fig. 16
Pipeline
- Paper
- done (1 of 1)
- Claims
- done
- Claims check
- done (ready)
- Attempt
- todo
- Questions
- none
- Verified
- none