DANDI:000008

an alternative analysis using entropies of Leiden clustering 35 often highlighted the same t-types as variable (Extended Data Fig. 8 ).

Results, Variability in individual t-types, paragraph 3; Extended Data Figure 8g-l

Leiden at resolution 2 gives 15 to 20 clusters per run. Across 35 t-types the Spearman correlation between normalized variance and median entropy is 0.266, one-sided p = 0.061, short of 0.05. With the code's membership of the excluded cells it is 0.313 (p = 0.031); with all cells, as the paper, 0.166 (p = 0.171) and 0.222 (p = 0.097). Four t-types are among the ten most variable by both measures (L5 ET_1, L5 IT_1, Sst Chodl, Vip Mybpc1_3). The paper gives no measure of agreement; the Spearman test is the reader's reading, and the result is near its threshold and moves with three cells.

How the test differed from the paper’s method

• Each recording counted once; results with all cells are shown. leidenalg is a current release; the paper names no version.

were performed on wild-type C57Bl/6 ( n = 27), Viaat -Cre/Ai9 (vesicular inhibitory amino acid transporter, encoded by the Slc32a1 gene, n = 24), Sst -Cre/Ai9 (somatostatin, n = 75), Vip -Cre/Ai9 (vasoactive intestinal polypeptide, n = 46), Pvalb -Cre/Ai9 (parvalbumin, n = 76), Npy -Cre/Ai9 (neuropeptide Y, n = 2), Vipr2 -Cre/Ai9 (vasoactive intestinal peptide receptor 2, n = 7), Scl17a8 -Cre/Ai9 (VGLUT3, vesicular glutamate transporter 3, n = 6), Gnb4 -Cre/Ai9 ( n = 1), and Slc17a8 -iCre/Ai9 ( n = 2) mice

Methods, Animals

Total 266 mice, one genotype each. Nine of ten lines match. Wild type is 26 against 27, and one mouse is Scnn1a-Tg3-Cre/wt; Ai9/wt (Cre label SCNN1), a line the paper does not list; the paper's code counts SCNN1 with the wild type (ttype-assignment.ipynb, not run here). The two Slc17a8-iCre mice are spelled Vglut3-Cre in the genotype and SLC17a8-iCre in the Cre field.

Each neuron was injected with 600-ms-long current pulses starting from −200 pA and up to 1,380 pA with 20-pA increment steps (in some cases stimulation was stopped before reaching 1,380 pA). There were 1.3- or 1.4-s intervals between successive current pulses, depending on the used setup. For most neurons, the stimulation was then repeated multiple times from the beginning.

Methods, Patch-seq recording procedures, paragraph 4

Every step lasts 600 ms (samples 2,500 to 17,500 at 25 kHz) and every series rises in 20 pA steps. 1,183 cells (89.1%) start at -200 pA, 90 at -100 pA and the rest between -320 and -80 pA; 112 cells go above 1,380 pA (87 to 1,480, one to 1,780). 89.0% conform on the steps, between the criterion's 80% and 95%. The departures are mostly one shift of the whole series, as 87 of the 90 cells that start at -100 pA end at 1,480 pA. The intervals between pulses cannot be tested, as no series has a start time, and no file holds a second series (0 cells), so the repeated stimulation is not in the dandiset.

How the test differed from the paper’s method

• The interval part of the criterion could not be applied; conformance is judged on the steps alone.

leaving 1,227 neurons for further analysis (817 inhibitory, 410 excitatory; 369 and 269 with morphological reconstructions, respectively). The resulting data set covered 77 out of the 90 neuronal t-types (Fig. 1a )

Results, Patch-seq of mouse primary motor cortex, paragraph 3; Figure 1a

The files give 1,226 cells with a t-type after the exclusions, 817 inhibitory and 409 excitatory, in 76 t-types. The cell the dandiset lacks (20190211_sample_10) is excitatory, so 1,226 and 409 are accepted by the criterion. It is the only L6 CT Gpr139 cell, so the t-type count is 76, not 77; the paper's table, with that cell, gives 1,227 cells and 77 t-types. Three of the four counts match. Whether the allowance for the missing cell extends to the t-type it alone covers is not in the criterion; the question is in questions/000008.yaml.

How the test differed from the paper’s method

• The five exclusions come from the paper's table (link-metadata-table); the files alone give 1,231 cells.

Within-type distributions of soma depths (Fig. 1b ) agreed well with previous data 4 and with the layer-specific nomenclature of excitatory t-types, confirming the validity of our t-type assignment.

Results, Patch-seq of mouse primary motor cortex, paragraph 3; Figure 1b

18 of 19 excitatory t-types with at least three depths (95%) have a median in a layer their name allows. L5 IT_4 (four cells, four mice, depths 0.753 to 0.835) has its median at 0.757, in L6. The paper's µm depths give the same (median 0.767).

How the test differed from the paper’s method

• The normalized depth is the ratio of the files' 4x image measurements of soma depth and cortical thickness; the paper used µm measurements, which are not in the dandiset (median absolute difference 0.005, 14 cells differ by more than 0.05). The µm reading from the paper's table is shown.

All 1,328 cells of the dandiset match a row of the paper's table by cell id; the table's one other cell, 20190211_sample_10 (L6 CT Gpr139), has no file. For all 1,328 cells the records agree on mouse, slice, genotype, sex, Cre label, targeted and inferred layer, exon, intron and intergenic reads, batch, genes detected, t-type confidence, the ALM/VISp mapping and age, and on t-type and family with the files' PT read as the paper's ET (1,279 agree in the files' spelling). The files' cortical_thickness_4x equals the table's soma depth in 4x units and cortical_thickness_um its cortical thickness in 4x units for all 1,328 cells; the µm columns are not in the files. The five cells with a t-type that the table excludes are all in the dandiset and are not marked. Numbers are compared exactly, as floats without rounding.

Every file matches a published row; 1,320 of 1,328 cells (99.4%) agree to a relative 1e-4 on all 28 features other than the rheobase. Three cells (20190128_sample_3, 20180215_sample_1, 20190222_sample_5) differ in 22 to 25 features, so their files hold other traces than those the published features were computed from; four differ in 5 or 6 features of the highest-firing trace (Fano factors, CVs, spike frequency adaptation, wildness, or the spike count), and one in the time constant only. The rheobase differs for 41 cells (-40 to 20 pA); it is fitted by RANSAC with no seed in the paper's code, and for 4 of the first 8 such cells one of seeds 0 to 9 gives the published value.

How the test differed from the paper’s method

• The rheobase is computed with the global numpy seed set to 0 before each cell; the paper's code sets no seed.

In total, we performed whole-cell recordings from more than 2,000 cells, of which 1,329 cells (from 266 mice) passed initial quality control.

Results, Patch-seq of mouse primary motor cortex, paragraph 1; Methods, Animals

The dandiset has files for 1,328 distinct cells from 266 mice, which matches reading (b) and the mouse count. Three pairs of files hold identical voltage matrices, so the files hold 1,325 distinct recordings, but the six cells of these pairs are distinct cells (different read counts, genes detected and depths, four of the six on different slices). The criterion counts distinct recorded cells, so the duplicated recordings are a fault of the deposit, recorded as a finding, not missing cells. Changed from partially_reproduced after verification round 1, which counted the cells this way.

Experiments on adult male and female mice ( n = 266; median age 75 days, interquartile range 64–100, full range 35–245 days, Extended Data Fig. 2a )

Methods, Animals; Extended Data Figure 2a

One age per mouse (no mouse has two). Over 266 mice the median is 75 days, the quartiles 64 and 100 days, the range 35 to 245 days, under the linear, lower, higher and nearest quantile rules alike.

We obtained soma depth information for 1,284 neurons out of 1,329 (45 neurons were mistakenly recorded using a solution with insufficient biocytin concentration, and we could measure soma depths for only 2 of those; for 2 other neurons the measurements could not be carried out because the slices were lost).

Methods, Cortical thickness normalization and layer assignment

1,283 cells have a soma depth, in the field cortical_thickness_4x (4x image units), which the data check shows to be a depth (its ratio to cortical_thickness_um lies between 0.031 and 0.977 and gives the files' inferred layer for 1,264 of 1,283 cells) and the link shows to be the table's 4x soma depth. The field soma_depth_um is present in 43 files and empty in all. The cell missing from the dandiset has a depth, so 1,283 is the count the criterion accepts (the table has 1,284).

How the test differed from the paper’s method

• The depth is read from a field whose name says cortical thickness; the identification rests on the values and on the link to the paper's table.

Hyperpolarizing currents were used to compute the input resistance (274.80 MOhm) and membrane time constant tau (21.95 ms). ... It was used to compute AP threshold (−40.18 mV), AP amplitude (81.17 mV), AP width (0.80 ms), AHP (−12.60 mV), ADP (0 mV), UDR (1.62) and latency of the first spike (69.28 ms). d , Regression line gives the rheobase estimate (20.44 pA).

Extended Data Figure 4a-f legend

All 23 values within tolerance, each equal to the paper's own feature table. The AP threshold is -43.18 mV as in that table, against -40.18 mV in the legend, inside the 10% tolerance.

n = 1,320 cells used to construct the embedding, 1,219 cells with t-type labels shown.

Figure 3a legend; Methods, t-SNE visualization of the morpho-electric phenotypes

Counting each recording once, 1,317 cells have all 17 features and 1,216 of them a t-type and no exclusion; with all cells, as the paper, exactly 1,320 and 1,219. Without the paper's exclusions, which the files do not mark, the second count is 1,221 (1,224 with all cells).

How the test differed from the paper’s method

• The three cells whose recording duplicates another cell's are left out (each recording counted once); the counts with all cells are given.
• The five excluded cells are taken from the paper's table (link-metadata-table).

These groups had distinct electrophysiological properties: for example, as expected, Pvalb neurons were characterized by high firing rates while Sst neurons had high values of the hyperpolarization sag and rebound

Results, Distinct phenotypes of major families, paragraph 2; Figure 2 (right biplots)

Reading (b), five groups, holds in full; Pvalb median max APs 71 (others at most 27), Sst median sag ratio 1.154 and rebound 5.70 mV, the highest. Under reading (a), seven families, ET has the highest median sag ratio (1.252), so two of three hold. Same with all cells and with per-mouse medians. 1,216 cells from 264 mice. Mann-Whitney p (information only) below 1e-37 for the three comparisons.

How the test differed from the paper’s method

• Each recording counted once; the result with all cells is shown and is the same.

We quantified this separation using a confusion matrix for k -nearest neighbours (kNN) classification of cells into families: it was mostly diagonal, with only the ET and IT subclasses strongly overlapping (Fig. 3d ).

Results, Distinct phenotypes of major families, paragraph 3; Figure 3a, d

Both handlings of the excluded cells give ET to IT 49.0% (IT to ET 1.2%) and every other off-diagonal entry below 25%, the largest Lamp5 to Sst 14.4%. Diagonal from 49.0% (ET) and 65.6% (Lamp5) to 97.6% (Pvalb). Mean diagonal 0.808 against 0.157 with permuted labels. The 24 ET cells classified as IT come from 16 mice. All cells give ET to IT 49.0% and largest other entry 14.3%.

How the test differed from the paper’s method

• Each recording counted once; the result with all cells is shown and is the same.

The first four panels correspond to features showing the largest IT/ET differences at room temperature, according to the two-sided Wilcoxon-Mann–Whitney test statistic (and omitting several features that are very correlated with the shown ones ... upstroke-to-downstroke ratio, sag time, and sag area).

Extended Data Figure 6a-e legend; the code, physiological-temperature.ipynb, cells 22 and 23

Layer as the claims file's opening comment defines it, the files' normalized depth cut at the Nissl boundaries (the targeted layer without a depth). 111 L5 IT cells (61 mice) against 46 L5 ET cells (26 mice). After leaving out UDR, sag time and sag area, the four smallest p are sag ratio (7.9e-14), input resistance (5.9e-12), AP width (6.4e-11) and rebound (1.8e-05); the fifth is the AHP (8.7e-05). With the files' inferred_layer, or the same rule on the paper's µm depths (45 L5 ET cells each), the same four come first, rebound 2.9e-05 against AHP 7.7e-05. Same with all cells.

How the test differed from the paper’s method

• The normalized depth is the ratio of the files' 4x image measurements; the paper's code uses µm depths, which are not in the dandiset. The µm reading from the paper's table is shown and gives the same.
• Cells with an undefined feature are dropped for that feature, the reading the claims file names; here no L5 IT or ET cell lacks a feature among the top ranks.

We were also able to identify a t-type, L4/5 IT_1, that was located on the boundary between L2/3 and L5 and probably corresponds to the quasi-L4 neurons described previously in motor cortex 31 .

Results, Patch-seq of mouse primary motor cortex, paragraph 4; Figure 1b

22 L4/5 IT_1 cells from 14 mice; median depth 0.279, inside 0.235 to 0.40 (reading a); interquartile range 0.259 to 0.306, which contains 0.29 (reading b). With the paper's µm depths, 0.277 and 0.252 to 0.303. The correspondence to quasi-L4 neurons is not tested.

How the test differed from the paper’s method

• Normalized depth from the files' 4x fields, as in fig1b-depth-layers.

The separability between inhibitory families was weaker than with electrophysiological features (Fig. 3d ). The between-family separability was strongest when we had combined electrophysiological and morphological features into a joint representation (Fig. 3c, d ), showing that these sets of properties are not redundant.

Results, Distinct phenotypes of major families, paragraph 4; Figure 3b-d

Inhibitory mean diagonal 0.737 with morphology against 0.851 with electrophysiology; seven-family mean highest for the combined representation, 0.872 (16 PCs) or 0.874 (17 PCs), against 0.808 and 0.753 (627 cells with both). Both handlings of the excluded cells, both block sizes and all cells give the same orderings; with all cells, as the paper, 628 cells have both, the criterion's count, with inhibitory means 0.737 against 0.847 and the combined seven-family mean 0.869 or 0.872. The morphology covers 635 cells (367 inhibitory, 268 excitatory); the paper's 636 includes the cell missing from the dandiset.

How the test differed from the paper’s method

• Each recording counted once; the result with all cells is the same.

Most t-types could be unambiguously placed into the correct family (Fig. 5a ), but some t-types were in between two families. For example, many Sst Pvalb Calb2 neurons were classified as belonging to the Pvalb subclass on the basis of electrophysiology. ... Vip Mybpc1 neurons often had Sst -like firing.

Results, Variability in individual t-types, paragraph 1; Figure 5a and legend

28 of 35 t-types with at least 10 cells have at most 25% classified into another family (i). Sst Pvalb Calb2 as Pvalb 38.5% (ii); Lamp5 Egln3_1 as Vip or Sst 90% of 10 cells (iii); Vip Mybpc1_3 as Sst 37.5% (iv, reading a), Vip Mybpc1_2 23.5%, pooled 25.4% (reading b). The code's handling (36 t-types, 29 at most 25% wrong) and all cells give the same.

We found that many t-types had total variance substantially above the variances of the k -means clusters (Fig. 5b ) ... some of them, such as Pvalb Vipr2_2 (chandelier cells), appeared morpho-electrically homogeneous. By contrast, Vip Mybpc1_2 was marked as having high electrophysiological variability

Results, Variability in individual t-types, paragraphs 2 and 3; Figure 5b

With n_init=10 and the Methods' membership, the k-means band is 0.173 to 0.468; 15 of 35 t-types (43%) are above it, many under both readings; Pvalb Vipr2_2 (0.449) is not above it and Vip Mybpc1_2 (0.692) is. With all cells, 17 of 35. The result depends on the k-means initialization; with n_init='auto' only 8 of 35 (23%) are above the band, and with the code's membership of the excluded cells n_init=10 gives 8 of 36 and 'auto' 16 of 36.

How the test differed from the paper’s method

• n_init=10 is set explicitly. The criterion says scikit-learn defaults as the code; the code ran under Python 3.7.9 (notebook metadata), for which every scikit-learn release has the default n_init=10, while the installed scikit-learn defaults to 'auto'. 'auto' is shown as a variant.
• Each recording counted once; results with all cells are shown.

The resulting cDNA yield was not correlated with the hold time (Spearman correlation −0.01).

Methods, Patch-seq recording procedures, last paragraph

Spearman correlation between hold time and log cDNA yield over the 1,319 dandiset cells with both values, -0.012 (p = 0.65); over all 1,320 cells of the table, also -0.012. Neither value is in the dandiset's files; the data check lists every DandiIcephysMetadata and subject field in all files and every entry under general of one file, and none is a hold time or a yield. The dandiset contributes only the cells.

Pooling the pairs across all families, we found that these two distance measures were correlated, with r = 0.60 (Fig. 4c , n = 200 pairs; Extended Data Fig. 7 ).

Results, Continuous phenotypic variation, paragraph 3; Figure 4c and legend

Under reading (a), the code's, 50 t-types and 200 within-family pairs, r = 0.597 (p = 9.9e-21). Reading (b) gives 52 t-types, 222 pairs and r = 0.563, so the paper's counts are reading (a)'s. All cells give 0.597 and 0.562.

The correlation was also observed within multiple individual families and for many individual electrophysiological features (Extended Data Fig. 7 ).

Results, Continuous phenotypic variation, paragraph 3; Extended Data Figure 7c-n

All six families have at least three t-types and a positive correlation (Pvalb 0.747, Sst 0.494, Vip 0.403, Lamp5 0.781, IT 0.328, CT 0.967), so reading (b) holds. 16 of 17 features correlate positively in at least half of the families. Same with all cells.

Transcriptomic distances between t-types were strongly correlated with the average soma depth differences ( r = 0.70; Fig. 4d ).

Results, Continuous phenotypic variation, paragraph 4; Figure 4d legend

Eight IT t-types with at least five cells with a depth (249 cells); r = 0.690 (p = 4.9e-05) with the files' 4x depths; 0.698 with the paper's µm depths.

How the test differed from the paper’s method

• Normalized depth from the files' 4x fields; the paper's µm depths are shown as a variant.

Here we also found that morpho-electric properties varied continuously across the transcriptomic landscape, with neighbouring t-types consistently showing similar morphologies and similar rebound values (Fig. 4b ).

Results, Continuous phenotypic variation, paragraph 2; Figure 4b; Extended Data Figure 6i

Layer as the claims file defines it, from the files' normalized depth. 11 L5 Sst t-types with at least five cells (143 cells, 63 mice), 55 pairs; r = 0.339, p = 0.0113 with all 17 features. Rebound alone r = 0.275, p = 0.042. With the files' inferred_layer or the paper's µm depths (64 mice), r = 0.347, p = 0.0094. The morphological part is not tested.

How the test differed from the paper’s method

• The normalized depth that gives the layer is the ratio of the files' 4x image measurements; the paper's µm depths are shown as a variant.

Similarly, Sst Pvalb Calb2 cells had high variability in terms of the maximum firing rate, but high-firing cells were mostly grouped in one part of the transcriptomic landscape (Fig. 5d ).

Results, Variability in individual t-types, paragraph 3; Figure 5d

26 Sst Pvalb Calb2 cells (19 mice), max APs 15 to 87, none at the median (32.5), so 13 above it. Reading (a), max APs rise towards the Pvalb centroid (rho = 0.363, one-sided p = 0.034). Reading (b), the 13 high-firing cells (mean pairwise distance 7.05, against 8.49 for random subsets) under the criterion as the answer of 2026-10-21 fixes it. With 10,400,600 subsets of that size, more than 1,000,000, p is estimated from 100,000 random subsets, p = 0.0485 with a 95% Clopper-Pearson interval of 0.0472 to 0.0499, so (b) holds and both readings hold. What the answer changed, before it the criterion's 1,000 permutations gave p = 0.052 with this seed and 0.045 to 0.057 with five others, and the claim was blocked as methods_underspecified because the readings disagreed; the answer replaced them with 100,000 subsets and the bound, under which the claim is reproduced. The margin is small, and the seed still matters. The exact p-value over all 10,400,600 subsets, computed as a variant, is 0.0497, below 0.05, but the 100,000-subset estimate with five other seeds gives 0.0496 to 0.0514, and each interval reaches 0.05 or above, so (b) would not hold under the criterion with those seeds. The section's seed (55) was fixed before the answer. A follow-up question on this is in questions/000008.yaml.

How the test differed from the paper’s method

• The seed of the 100,000 random subsets is a standard choice (55, the section's seed before the answer); it can decide reading (b) here, so five other seeds and the exact p-value are shown.

The selected model used 25 genes with a 5-dimensional latent space and achieved a cross-validated R 2 of 0.38. ... The cross-validated correlations between the first three pairs of projections were 0.90, 0.74, and 0.67, respectively.

Results, Distinct phenotypes of major families, paragraph 1; Figure 2; Methods, Reduced-rank regression

Main reading (four log-transformed features), 1,216 cells from 264 mice. At lambda 0.5, 24 genes, cross-validated R2 0.370 and correlations 0.895, 0.725, 0.648; at lambda 0.45, 29 genes, 0.382, 0.898, 0.744, 0.669. All within tolerance at both lambdas; also under the code's three-feature transform and with all 1,219 cells. The folds are those of the paper's elastic_rrr_cv (seed 42, one permutation, ten contiguous test blocks of 121 cells, 6 cells never tested).

How the test differed from the paper’s method

• The elastic net inside the paper's sparseRRR.py is fitted with scikit-learn (ElasticNet, MultiTaskElasticNet), which minimizes the same objective, because glmnet_python has no distribution for Python 3.13; the paper reports 25 genes at lambda 0.5, the substitute selects 24.
• Each recording counted once; the result with all cells is shown.

A reduced-rank regression model restricted to using only ion channel genes (Extended Data Fig. 5 ) did not perform much worse than the full model (cross-validated R 2 = 0.33 and correlations 0.86, 0.71, and 0.56, respectively, with regularization set to select 25 genes).

Results, Distinct phenotypes of major families, paragraph 2; Extended Data Figure 5e, f

307 ion channel genes detected in at least 10 cells, lambda 0.303, 26 genes selected; R2 0.341 and correlations 0.865, 0.706, 0.565, all within tolerance; the code's three-feature transform gives 0.335, 0.862, 0.707, 0.565.

How the test differed from the paper’s method

• The scikit-learn substitute for glmnet, as in fig2-rrr; 26 genes are selected where the paper reports 25.

For example, electrophysiological properties of the t-types within the Vip subclass varied continuously across the transcriptomic landscape; the membrane time constant, for instance, had its largest values close to the Sncg subclass and gradually decreased towards Vip Gpc3 (Fig. 4a ).

Results, Continuous phenotypic variation, paragraph 1; Figure 4a

The two readings disagree, and the criterion then blocks the claim. Reading (a), 153 Vip cells from 51 mice, Spearman rho between tau and position from the Sncg to the Vip Gpc3 centroid = -0.582, one-sided p = 1.6e-15. Reading (b), nine Vip t-types, rho between median tau and transcriptomic distance to Sncg = -0.383, one-sided p = 0.154 (-0.367, p = 0.166 with the mean of the Sncg t-type means). The direction agrees in both; the t-type reading has nine points.

How the test differed from the paper’s method

• The Sncg subclass mean of reading (b) is taken over the Sncg reference cells; the mean of the Sncg t-type means is shown and gives the same.

This variability was not random: overlaying the rebound values on the t-SNE embedding (Fig. 5c ) showed that cells with low rebound were located close to the boundary with the low-rebound Vip Sncg type.

Results, Variability in individual t-types, paragraph 3; Figure 5c

The two readings disagree. 34 Vip Mybpc1_2 cells (22 mice), rebound 0.3 to 19.8 mV, median 11.31 mV against 2.26 mV for Vip Sncg. Reading (a), rebound rises with the distance to the nearest Vip Sncg reference cell (rho = 0.356, one-sided p = 0.019). Reading (b), rebound does not fall with the correlation to the Vip Sncg mean (rho = 0.067, one-sided p = 0.648; 0.066 and 0.644 with the log taken after the mean). Scaling the Patch-seq counts to a million per cell first gives rho = 0.068, p = 0.648, and raw counts per million rho = 0.016, p = 0.536.

How the test differed from the paper’s method

• The transform of reading (b) is not given by the claims file; the Methods' log2(x + 1) of length-normalized Patch-seq counts (the code's rnaseqTools.map_to_tsne with normalizeNew=True) against the mean of log2(UMI + 1) of the reference is the main reading; log2 of the mean UMI, and two library-size normalizations of the Patch-seq counts, are variants, and all give the same.

The mRNA of these 1,329 cells was sequenced, yielding on average 1.3 million exonic and 0.7 million intronic reads (medians ... 6.0 ± 0.6 and 5.6 ± 0.8, respectively) and 9,100 ± 3,500 (mean ± s.d.) detected genes per cell

Results, Patch-seq of mouse primary motor cortex, paragraph 1; Extended Data Figure 2

Over the 1,329 cells of the paper's count tables, median exonic count 1,281,492 and intronic 687,965 (log10 means 5.99 and 5.61), and 9,097 detected genes on average (s.d. 3,533). The read fields of the paper's table, which the dandiset's files repeat, give 1,190,890 and 688,140, also within 10%.

cells with the highest correlation below 0.4 (78 cells); cells that would be assigned to non-neural t-types, presumably owing to RNA contamination 51 (14 cells; see also Extended Data Fig. 2j–n ); cells with the highest correlation less than 0.02 above the maximal correlation in one of the other two large transcriptomic groups (5 cells). The remaining 1,232 cells passed quality control

Methods, t-Type assignment, paragraphs 1 and 3

The code's first-round mapping to the Tasic et al. (2018) reference gives 78, 14, 0 (Meis2 or CR) and 5 failing cells, 1,232 passing; every cell's pass or fail agrees with the paper's table, and the 1,232 include the 5 excluded cells, leaving 1,227.

We found that Patch-seq expression profiles were most similar to the single-nucleus Smart-seq2 data (Extended Data Fig. 2g, h ). At the same time, there was good agreement between t-type assignments based on Smart-seq2 and those based on 10x reference data (Extended Data Fig. 2i )

Results, Patch-seq of mouse primary motor cortex, paragraph 2; Extended Data Figure 2g-i

Over 1,232 cells, the median maximal correlation is highest for the single-nucleus Smart-seq2 reference (0.703; then single-cell 10x v2 0.623 and single-cell Smart-seq2 0.621), and 75.7% of cells correlate best with it. The single-cell Smart-seq2 and 10x v2 references assign the same t-type for 78.6% of cells.

How the test differed from the paper’s method

• The assignment is the direct one, without the bootstrap over genes that the paper uses for confidence.