A classifier to assess section integrity.

(A) Generation of the transcript mask. The transcript density image was converted to a binary mask using a random forest classifier. (B) Generation of the DAPI mask. The modified DAPI image was converted to a binary mask using a random forest classifier. (C) Generation of the detachment mask. The detachment mask was the difference between DAPI and transcript masks. (D) Generation of the ventricle mask. A binary image summarizing the locations of 11 ventricle boundary genes was summed with the DAPI mask and converted to a binary mask with a random forest classifier. (E) Generation of the damage mask. Two intermediate masks were created via a series of binary operations on the modified DAPI image and DAPI mask, then summed. The ventricle mask was subtracted to remove ventricles. (F) Sequential combination of ventricle, detachment, transcript, and damage masks resulted in the final 5-category image.

Tissue area, detachment and damage.

(A) Accuracy of the pixel classifier, evaluated on a test dataset consisting of a 1 mm2 subregion from each of 621 tissue sections. (B) Probability distributions describing the percentage detachment, ventricle, and damage for each of 621 sections. Insets: false positive distributions calculated for 20 sections without detachment, ventricles, or damage. (C) Percentage of the section identified as on-tissue, detachment, and damage, plot over time (grey), and for the 59 sections in the Yao et al. (2023) Allen Brain Cell Atlas dataset (black).

Transcript density.

(A) Transcript locations for two neighboring sections from the same mouse brain, separated along the A-P axis by 200 µm. (B) Transcript density across A-P locations for a single mouse. 59 sections in the Yao et al. (2023) Allen Brain Cell Atlas dataset. (C) Histograms of transcript density per transcript species per square micrometer for 641 sections. Summed results from 4 gene panels (VA142, VA373, BP0770, VZG147). Black: 59 sections in the Yao et al. (2023) Allen Brain Cell Atlas dataset (using VA142). Arrowheads: transcript densities for the two sections in A (panel VA142). (D) Transcript density over time. Symbols indicate gene panels. (E) Comparison of transcript densities in plane 0 and summed across all planes. Each data point represents one section from the Yao et al. (2023) dataset. Pearson correlation coefficient 0.86, p = 3.1 x 10-18. (F) Mean transcripts per soma vs transcript density. One data point per section, Yao et al. (2023) dataset. Pearson correlation coefficient 0.69, p = 1.07 x 10-9. (G) Distribution of transcripts per soma. Yao et al. (2023) dataset.

Dropped images cause local data loss.

(A) Transcript locations for one transcript species: Gja1. Each point represents one Gja1 transcript. Inset: transcripts around one region of data loss. (B) Output of the data loss detection routine, showing the number of transcript species missing from each FOV. Not all FOVs with dropped genes are missing Gja1. White: off-tissue, as determined by the classifier. (C) Histogram of missing fields of view across 641 mouse sections (grey) and for the 59 sections in the Yao et al. (2023) Allen Brain Cell Atlas dataset (blue). Arrowhead: results for the section in panels A and B. Inset: Number of transcript species missing per affected field of view. (D) Effect of missing transcript species on label transfer. Change in class, subclass, supertype and cluster labels calculated for 10,000 cells from the Yao et al. (2023) Allen Brain Cell Atlas dataset. Median, minimum and maximum % change from 100 trials. Genes to be removed were selected at random. (E) Number of missing fields of view over time and with different acquisition software versions. Each point represents a tissue section.

Uneven detection efficiency across each field of view.

(A) Transcript locations for two coronal sections from the same brain, separated by 100 µm. To the left and below, transcript densities summed along x and y axes. For the section on the left, changes in transcript density occur at anatomical boundaries with little indication of variations in detection efficiency along x or y axes. For the section on the right, superimposed on differences in genes expression are variations in detection efficiency with a periodicity of 200 µm. (B) Periodicity metric, calculated for each z-plane along x and y axes, for the two sections in A. Filled symbols, left example in A. Open symbols, right example in A. Black and grey, metric along x and y axes, respectively. (C) Histogram of minimum periodicity metrics for 641 sections (grey) and for the 59 sections in the Yao et al. (2023) Allen Brain Cell Atlas dataset (black). Arrowheads, the two sections in A. (D) Effect of reduced detection efficiency on label transfer. Change in class, subclass, supertype and cluster labels for 10,000 cells from the Yao et al. (2023) Allen Brain Cell Atlas dataset (VA142 500 probe panel). Median, minimum and maximum % change from 100 trials. (E) Periodicity metric over time. Black: periodicity metric for the 59 sections in the Yao et al. (2023) Allen Brain Cell Atlas dataset.

Uneven detection efficiency along the optical axis.

(A) Transcript locations in three z planes for each of two neighboring sections from the same mouse brain, separated along the A-P axis by 200 µm. Distances are from the coverslip surface. (B) Transcript counts along the z-axis for the section in panel A. (C) Distribution of p6/p0 ratio for 641 sections (grey) and the 59 sections in the Yao et al. (2023) Allen Brain Cell Atlas dataset (black). Arrowheads, sections in A. (D) p6/p0 ratio over time. (E) Mean ± SEM transcript counts along the z-axis for 10, 14 and 20 µm thick sections. 3 sections each at 14 and 20 um, 6 at 10 um.

Imperfections identified by manual inspection.

(A) DAPI image. Data loss results in horizontal stripes in the left half of the image. (B) DAPI in the sub-region in the box in panel A. (C) Transcripts in the corresponding region. (D) Transcripts in a section 6.2 mm posterior to bregma. (D) Transcripts from a neighboring section. (E) Transcript density in cortex and thalamus (boxes in panels D and E) for 6 neighboring sections.

Transcript density across spatial transcriptomics platforms.

(A) Example sections from four commercial platforms. MERSCOPE, mouse liver section. CosMx, human brain. Xenium and Molecular Cartography, mouse brain. Scale bars, 1 mm. Histograms indicate transcript density in (cardinal axes, normalized to peak). (B) Periodicity metric for public datasets (4 sections in A and 3 Vizgen mouse brain datasets). Grey: mean ± stdev periodicity metric for the Yao et al. (2023) Allen Brain Cell Atlas dataset. (C) Transcript count along the z axis for sections in A. (D) Pairwise CV of transcripts across experiments. MERSCOPE, 3 Vizgen mouse brain datasets and Yao et al. (2023) Allen Brain Cell Atlas dataset (open symbol). Xenium, fresh-frozen-mouse-brain-replicates-1-standard dataset from 10x. Grey datapoints (Xenium and CosMx) from Cook et al. (2023).