The algorithms described on this page are used to generate the transcript files and decoding metrics in the output bundle:

• transcripts.zarr.zip
• analysis_summary.html: Decoding tab
• metrics_summary.csv
• qc_images.html
• binned_transcripts.zarr.zip

The goal of RCA product image processing is to detect and filter puncta and correct distortion. Performed for every channel and cycle, the 3D image volumes (Z-stacks) obtained for each field of view (FOV) are processed to detect the puncta in 3D space that correspond to labeled RCA products. Images are captured in four color channels and n cycles depending on the assay and panel configuration.

The RNA fluorescence image is scanned for puncta that stand out from the local background. The XYZ coordinates of each punctum are refined by examining local brightness and the signal intensity is determined based on a fitted shape. Atera Instrument-specific calibrations are made to adjust the optical system's point spread function model as a function of FOV position. This calibration enables increased detection and localization of puncta in transcript-dense regions.

Next, the pipeline filters out puncta that are unlikely to be true transcripts (non-punctate or low quality signals). Similar to DAPI images, the puncta location curvature distortion is corrected.

For each FOV, puncta across channels and cycles must be aligned to reduce differences in image offset, rotation, and magnification. This is important for accurate transcript decoding. The localized puncta from each channel and cycle are aligned so puncta corresponding to the same original RNA molecule are aligned tightly in 3D. Nonlinear transformations are fitted such that all the puncta are aligned to the reference morphology image.

Decoding is performed to translate puncta into gene transcripts. The Atera codebooks (WTA panel, Panel A, Panel C) contain a collection of codewords that are assigned to genes in a given gene panel. The panel configurations supported by the Atera Instrument must be used for the pipeline to correctly and unambiguously decode all panel genes. Some panel codebooks have overlaps and cannot be combined.

The pipeline uses the panel_config.json to specify a given gene name to an indexed codeword. Each codeword is defined based on a pattern of fluorescent signal intensities recorded across channels and cycles. Some codewords are reserved for negative controls.

• The Atera WTA codebook requires 140 imaging cycles. 18,000+ genes can be decoded with this codebook.
• The Atera Select codebooks require 20 cycles. 1,000+ genes can be decoded with this codebook type (applies to Panel A, Panel C).

When selecting the codeword assigned to each gene, the Atera panel design process attempts to keep the predicted transcript density in each cycle-channel in each known cell type as uniform as possible, using single cell reference cell type data. In particular, high expression genes that are coexpressed in the same cell type will be assigned codewords in different cycle-channels as much as possible.

The fluorescent signals from all channels and cycles are compared to the codebook using a global (across all FOVs) maximum likelihood approach based on probabilistic modeling. The decoding process itself does not make per cycle-channel calls. Instead, it performs a filtering step to find codewords that are plausible matches for the puncta signal trace. It assigns probabilities to each candidate codeword based on all the intensity information at the spatial location and chooses the most probable. This approach considers attributes such as puncta location, their color and cycle of detections, and signal intensities.

The cell-feature matrix and secondary analyses only include transcripts with a Q-Score ≥ 20. Final Q-Scores are reported in the transcript data output files.

A Phred-style calibrated quality score is assigned to each decoded transcript to signify the confidence in the decoded transcript identity. The quality score is derived from the likelihood of the maximum likelihood codeword (i.e., the codeword that best explains the observed data), compared to the likelihood of other sub-optimal codewords. This yields a raw Q-Score. Codewords are then mapped to targets using the gene panel information.

Final Q-Scores are obtained first by putting the full range of raw Q-Scores into bins. Then, the raw Q-Scores in each bin are calibrated by the proportion of "Negative Control Codewords" in the bin. A final Q-Score is assigned to each raw Q-Score bin, to ensure that Q-Scores in each bin are correctly calibrated. Control codewords are built into the process to ensure that the final Q-Scores are accurately calibrated.

• Negative control codewords are codewords in the codebook that do not have any probes matching that code. They are chosen to meet the same requirements as regular codewords and can be used to assess the specificity of the decoding algorithm.
• Negative control probes are probes that exist in the panels but target non-biological sequences. They can be used to assess the specificity of the assay. They were screened against the human and mouse transcriptomes and showed no predicted binding.
• Genomic control probes are probes that are designed to bind to intergenic genomic DNA but not to any transcript sequence present in the tissue.
• Deprecated codewords are codewords that are deprecated and can be ignored in downstream analysis.
• Unassigned codewords are unused codewords. There is no probe in a particular gene panel that will generate the codeword.

After decoding, the transcript results are combined from all FOVs. Duplicate decoded transcripts in overlapping regions between adjoining FOVs are reconciled into one transcript, based on a nearest neighbor analysis that considers transcript identity. The deduplicated transcripts are assigned global coordinates based on the reference morphology image obtained from the DAPI image processing step (morphology_3d/ch0000_dapi_3d.ome.tif)).