DANDI:000691

Reproducing Scherrer et al. (2023) from dandiset 000691 7

Notebook reproduce_scherrer_2023.ipynb

• ◐Partially reproducedExtended Data Fig. 2b; Methods, Coupling optics, paragraph 3data in GitHub FeeLab/Feescope-Paper-Code at commit d596646b16a505907203828d8026feb7736ca69c, sf3/feather (images of a thin layer of fluorescent dye; axicons, led and reference folders)

  For the Featherscope, the axicon correction optics give a flat signal profile across the field of view, where total vignetting (optical vignetting with Gaussian illumination) makes the signal fall toward the edges.

  Total vignetting (yellow) illustrates the decrease in fluorescent signal caused by the combination of optical vignetting and a Gaussian-shaped excitation light profile. The curve with correction optics (blue) illustrates the flat signal intensity profile that can be achieved with an axicon pair that increases the excitation intensity at the edges of the field of view.

  Extended Data Fig. 2b; Methods, Coupling optics, paragraph 3

  At 0.4 mm the corrected profile is 0.742, total vignetting 0.342 and optical vignetting 0.628; the corrected profile is higher, but falls below 0.8 (minimum 0.742 over 0 to 0.4 mm). The minimum is 0.785 out to 0.38 mm and 0.698 out to 0.42 mm; with the centre moved by 100 um it is 0.626 to 0.867, and 0.766 about the image centre.

  How the test differed from the paper’s method

  • The radius of 0.4 mm and the level of 0.8 that define "flat" were chosen here; the legend gives no number.
• ◐Partially reproducedExtended Data Fig. 2cdata in GitHub FeeLab/Feescope-Paper-Code at commit d596646b16a505907203828d8026feb7736ca69c, sf3/widefield

  For the Kiloscope, the axicon correction optics give a flat signal profile across the field of view, where total vignetting makes the signal fall toward the edges.

  c, same as in (b) but for the Kiloscope.

  Extended Data Fig. 2c

  At 1.5 mm the corrected profile is 0.797, total vignetting 0.253 and optical vignetting 0.389. The corrected profile is higher, and its minimum over 0 to 1.5 mm, 0.797, is 0.003 below the criterion's 0.8; at 1.125 mm it is 0.922. The minimum is 0.846 out to 1.425 mm and 0.742 out to 1.575 mm; with the centre moved by 100 um it is 0.765 to 0.831 (above 0.8 for one of four moves), and 0.695 about the image centre.

  How the test differed from the paper’s method

  • The radius of 1.5 mm and the level of 0.8 were chosen here; the result sits at the level.
• ✓ReproducedMain text, page 546, paragraph 3; Methods, Data analysis, paragraph 1

  The Kiloscope records thousands of neurons at once in a freely behaving mouse; in the mouse with the best expression EXTRACT gave 2,948 footprints.

  The second design, the Kiloscope, provides cellular resolution imaging across a 4.8 × 3.6 mm FOV in a device weighing only 1.4 g, allowing us to simultaneously record from thousands of neurons in a freely behaving mouse ... For the mouse with the best calcium indicator expression, this extraction yielded 2,948 footprints

  Main text, page 546, paragraph 3; Methods, Data analysis, paragraph 1

  All 2948 footprints have a trace above zero at some time; no two have traces correlated above 0.77, so near-duplicate merging leaves 2948. The mouse visits all four arms and runs 48.5 loop lengths during the 1081 s of imaging that overlap the tracked video. Tracking has 2 steps above 40 px between video frames (one glitch at 1064.8 s) and 55 frames with a blob under 50 px; leaving them out changes the arm of 0.05% of imaging frames and the octant of 0.06%.

  How the test differed from the paper’s method

  • The count is of EXTRACT footprints before the paper's curation (2382 kept); the dandiset does not mark the curated set.
• ✓ReproducedMain text, page 549, paragraph 1; Extended Data Fig. 8; Methods, Data analysis, paragraph 1

  The correlated activity in each NMF component is driven by footprints spread across the field of view.

  Note that the correlated activity visible in each component is driven by footprints from across the FOV (Extended Data Fig. 8).

  Main text, page 549, paragraph 1; Extended Data Fig. 8; Methods, Data analysis, paragraph 1

  For all 8 components the event has 196 to 320 active footprints, more than half of the component's members, with a spread of 0.945 to 1.055 times that of all footprints (0.98 mm) and a 90th percentile of pairwise distance of 2.05 to 2.29 mm. Active thresholds of 0.3 and 0.7 also give 8 of 8. The best of 4 NMF replicates was kept (errors 97.55 to 97.58). Two events (80.4 and 122.9 s) fall in the opening 153 s the behaviour analysis leaves out; with event times from 153 s on, 8 of 8 pass (spread ratios 0.94 to 1.05). The test is weak. In 100 random 50 s windows, the median fraction of members active is 0.399 to 0.519, and the event's fraction (0.568 to 0.851) is above the 95th percentile of random windows for 6 of 8 components. Random sets of footprints of the same size have spread ratios of 0.951 to 1.042 (5th to 95th percentiles), so any set this large passes the criterion's half. The event's spread ratio lies within its own component's random range for 3 of 8 components, below it for 2 (0.945 and 0.956, against 5th percentiles of 0.962) and above it for 3 (1.042, 1.055 and 1.050, against 95th percentiles of 1.037, 1.042 and 1.036).

  How the test differed from the paper’s method

  • All 2948 footprints in place of the 2382 curated ones.
  • Event times are each component's largest temporal weight; the code sets them by hand.
  • Distances use the code's 5.94 um per pixel, which the dandiset does not state; the outcome uses ratios and does not depend on it.
• ✓ReproducedMain text, page 549, paragraph 1; Extended Data Fig. 9a and 9b

  Position within the maze can be decoded with octant resolution from the activity, above decoders trained on shuffled data.

  We used the same approach to decode even more fine-grained position information (Extended Data Fig. 9, P < 0.001, Monte Carlo sampling as in Fig. 2j).

  Main text, page 549, paragraph 1; Extended Data Fig. 9a and 9b

  What the paper reports

  • Monte Carlo sampling of shuffled datasets, p < 0.001 (n = 1,000 shuffled datasets (as in Fig. 2j))

  Median class accuracy 0.313 over 50 decoders (0.211 to 0.416), chance 0.125, null 95th percentile 0.144; all 50 exceed it, and each octant's mean accuracy (0.236 to 0.499) is above chance. On six splits fit by both solvers, SGD stops at objectives of 3.89 to 6.18 and the exact solver at 0.071 to 0.094 (at its limit of 10,000 iterations in all six); the exact solver gives 0.220 to 0.292 for three real decoders and 0.092 to 0.145 for three with shifted labels, so the claim also holds with it.

  How the test differed from the paper’s method

  • Position tracked here from the video (the paper's tracking output is not in the dandiset), in the unrotated frame, with the loop's sides at the 2nd and 98th percentiles of position.
  • All 2948 footprints in place of the 2382 curated ones.
  • scikit-learn's SGD classifier (hinge loss, ridge alpha equal to the code's lambda, 10 passes) in place of MATLAB's fitcecoc with templateLinear, whose default for more than 100 predictors is 10 passes of dual SGD; 50 decoders in place of 200.
• ✓ReproducedExtended Data Fig. 9c; Methods, Data analysis, paragraph 5

  Decoders built from the 32 footprints closest to a point decode octant position significantly in parts of the field of view, with accuracy that varies across it.

  Not quoted: The Extended Data Fig. 9c legend and the Methods give the method only. The main text states that accuracy differs across the field of view for the quadrant and direction map (Fig. 2l) and that octant position was decoded with the same approach, but states no result for the octant map. That it is significant in parts of the field of view and varies across it is read from the legend (non-significant pixels plotted in black) and the main text.

  With 32 footprints within 270 um (45 px), 150 grid points qualify and 18 (12%) are significant at Benjamini-Hochberg 5%, with scores of 0.112 to 0.187 (range 0.075, where the criterion asks for 0.05). Scores over all points 0.068 to 0.187, median 0.125, against null means of 0.065 to 0.123. The outcome depends on the solver. With 10 passes of SGD, as in round 1, on the same points and seeds, 1 point is significant and the criterion is not met; at four points SGD stops at objectives of 19.8 to 68.9 where the exact solver reaches 1.976 to 2.000, close to the 2.0 of predicting every class absent, so the scores rest on weak classifiers. liblinear reached its limit of 10,000 iterations in 2268 of 3000 fits. Null standard deviations per point 0.018 to 0.046 (5th to 95th percentile) from 10 shifted decoders; a point alone needs z above 3.40, and the smallest z among the significant points is 2.53. A Welch t-test of the 10 real against the 10 shifted decoders marks 80 points significant (range 0.079), and 5 with SGD (range 0.052). With the code's 64 footprints within 50 px, on a 60 px grid, 22 of 34 points are significant (scores 0.141 to 0.211, range 0.070), and 32 with the t-test. The exact fits stop at an iteration limit, so the counts can differ slightly between machines.

  How the test differed from the paper’s method

  • The legend says 32 footprints within 270 um; the code uses 64 within 50 pixels. The legend is followed for the outcome and the code's choice is shown as a variant.
  • The paper does not describe its significance test. Here, 10 decoders and 10 shifted-label decoders per point and a one-sided z-test, chosen before the analysis.
  • Local decoders are fit by liblinear (LinearSVC) on the same objective, as MATLAB's default for 100 or fewer predictors is BFGS run to convergence; liblinear also penalizes the intercept, weakly. Round 1 used 10 passes of SGD here, which is shown as a variant.
  • A 30 pixel grid where the code uses 5 pixels; all 2948 footprints in place of 2382; position tracked from the video, as for octant decoding.
  • The code's variant is run on a 60 pixel grid where the criterion says 40; on the 40 pixel grid the notebook took 32 minutes on four cores, over the 30 minute limit.
• ✓ReproducedMain text, page 549, paragraph 1; Extended Data Fig. 10

  Some footprints have activity localized to subregions of the maze.

  An examination of individual neural firing rate maps yielded footprints whose activity was localized to subregions of the maze (Extended Data Fig. 10).

  Main text, page 549, paragraph 1; Extended Data Fig. 10

  100 of 2948 footprints (3.4%) have an entropy over linear position below all 100 shifted values, against 29 expected (binomial p = 4.5e-25); 448 against 175 at a looser threshold. For the 100, the entropy is 0.20 to 0.90 bits below the shifted mean (5th to 95th percentile), out of a maximum of 6.64 bits. The binomial test treats footprints as independent, but the same 100 shifts serve every footprint; treating each shifted copy in turn as the observed data gives counts with a median of 28, a 95th percentile of 54 and a maximum of 62, and none of the 100 copies reaches 100.

  How the test differed from the paper’s method

  • The paper selected its examples by eye and shows two-dimensional maps; here a test over all footprints on linear position, with the entropy measure of place_cells.m.
  • All 2948 footprints in place of 2382; position tracked from the video, as for octant decoding.