Fig 1.
In situ sequencing involves image stitching and alignment, dot detection, genotyping and phenotyping of cells.
A) Barcodes introduced into cells are read out with 12 cycles of sequencing-by-synthesis. The amplicon colonies that correspond to two barcodes in different cells are highlighted in each cycle, and the sequences shown below. A 10 micron scale bar is included in the bottom right image. B) Stitching and alignment must be performed on both sequencing and phenotyping images. Then, amplicon colonies (dots) are detected and cells are segmented. Reads are generated from amplicon colonies and assigned to cells, and various features are calculated for each cell to represent phenotypes. Two example features, area and intensity, are shown.
Fig 2.
ConStitch performs stitching and alignment jointly on multi-cycle image sets.
A) Imaging procedure for in situ sequencing. Each sequencing cycle contains four channels corresponding to each nucleotide, and a fiducial channel to help with alignment. Positional errors are introduced into the collected images in two ways: intra-cycle, due to inaccuracies of microscope stage motor and sensors, and inter-cycle, due to removal and replacement of the well plate from the microscope. B) Procedure for stitching and aligning multi-cycle image sets. Overlapping pairs of images are aligned to each other, regions with no features are excluded from the alignment process, then the system of equations containing all pairwise alignments is solved to find global positions for each image. C) Performance of different solving methods with high levels of noise: Mean absolute error solver (default), mean squared error solver, spanning tree solver, and an individual cycle solver where each cycle is stitched independently with the MAE solver, then all stitched cycles are aligned with a single translation. Performance is tested on image set 4.1. To better show performance differences in suboptimal conditions, Gaussian noise of sigma = 100 was added to the pixel intensities of the image. D) Ratio of mean dot intensity to mean pixel intensity (top) and percent of reads miscalled (bottom) when sequencing images are shifted by a fractional pixel offset using spline interpolation. E) Percent of reads miscalled when alignment error is introduced into a random cycle.
Table 1.
Image sets used to test the performance of STARCall compared to MIST, ASHLAR and the Feldman et al pipeline.
Fig 3.
ConStitch provides comparable stitching performance to MIST and superior alignment performance to ASHLAR.
A-B) Stitching accuracy of MIST and ConStitch using the MIST evaluation framework, measuring differences in colony area (A) and colony position (B) compared to ground truth images. C) Alignment error of ASHLAR and ConStitch on ASHLAR evaluation dataset derived by breaking stitched and aligned images into 200 pixel by 200 pixel blocks, the alignment of which was evaluated between the two cycles with phase correlation. Histograms show magnitudes of alignment vectors for blocks in the resulting optical flow field. D-E) Stitching performance of ConStitch and ASHLAR on all imaging rounds of the MIST image set when aligned together. Area error and positional error are reported across the five imaging rounds. F) Percent area of well with less than 1 pixel of alignment error. Alignment error was measured as in (C), applied to each pair of cycles, taking the maximum error at each location. G) Histogram of alignment error on image set 1.1, calculated as in (C).
Fig 4.
STARCall filters out background and corrects for differences in intensity when calling reads.
A) Key steps for read calling by STARCall, including background filtering, amplicon colony detection, base calling and read assembly. B) Sequencing images before and after the Laplacian of Gaussian filter from Feldman et al. and the Gaussian filter and normalization from STARCall. Amplicon colonies in each cell appear as small bright dots in the channel corresponding to their nucleobase. Violet = G, Blue = T, Green = A, Red = C. A 10 micron scale bar is included in the bottom right image. C) 50th (left) and 99.9th (right) percentile pixel intensity of sequencing images in VIS-seq image set 1.1. The 50th percentile approximates background and the 99.9th percentile approximates the intensity of signal generated by amplicon colonies. D) The percent of cells with barcodes found in the previously determined lookup table for each experiment when reads were called by STARCall (green) or the Feldman et al. pipeline (blue). E) Test set performance of STARCall (green) and the Feldman et al. pipeline (blue). Five hundred cells were selected and manually annotated with their sequence F) The read count per cell of STARCall (left) and the Feldman et al. pipeline (right) on image set 2.1. G) The nucleotide frequency at each cycle from reads called by STARCall (left) or the Feldman et al. [4] pipeline (right), with the expected frequency in the barcode library shown by dashed lines.
Fig 5.
STARCall resource requirements.
A) Performance metrics for a full run of STARCall on image set 1.1. Shown is the number of jobs running colored by their section of the pipeline (top), the percent of CPU usage measured (middle), and the resident set size (RSS) memory usage in gigabytes (bottom). B) Performance metrics for ConStitch, the package that performs stitching and alignment in STARCall, compared to ASHLAR. The percent of CPU usage (top) and RSS memory usage (bottom) over time are shown. C) The same performance metrics as in (B) shown for just the read calling portion of STARCall, compared to the Feldman et al. [4] pipeline.