Atera Resource Usage Report

Overview

For this report I used a Seurat-based resource-usage benchmark comparing Atera and Xenium spatial transcriptomics datasets processed through the same workflow: zero-count-cell filtering, SCTransform normalization, PCA, nearest-neighbor graph construction, clustering, and UMAP.

Data sources

Six datasets were used in this benchmarking exercise.

Five datasets came from the public 10x Genomics dataset collection: Atera FFPE human breast, Atera FFPE human cervix, Xenium human breast, Xenium human cervix, and Xenium fresh-frozen mouse brain.

The sixth mouse-brain regional dataset came from Margo’s paper about the Isocortex Layer-6 ‘a’ and ‘b’ and was the dataset used in the Xenium demo workflow, represented as z_Region_2_left_margo-obj.rds.

For each of the five datasets obtained from 10x Genomics, two Seurat .rds objects were generated using the custom LoadAtera function derived from Seurat (Sajita lab)‘s LoadXenium function, as each of them contained the modern’.zarr’ file format that LoadXenium does not support. The only significant change is, as follows

• molecule_coordinates = TRUE: Molecule-level spatial coordinates were retained. Default for LoadXenium function.
• molecule_coordinates = FALSE: Molecule-level coordinates were omitted. Default for LoadAtera function.

Files containing false_coords are the the ones with molecule_coordinates set to FALSE, while files without false_coords have the setting set to TRUE.

Filename Examples

File Name	Configuration
Atera_Breast_false_coords-obj.rds	molecule_coordinates = FALSE
Atera_Breast-obj.rds	molecule_coordinates = TRUE

To use the function, first run the setup, and then load the function by running it’s script.

source("python_reticulate_setup.R")
source("LoadAtera.R")

Then you can call the function as shown:

xenium_obj <- LoadAtera(
  data.dir = "path/to/xenium/outs",
  fov = "fov",
  cell.centroids = TRUE,
  molecule.coordinates = FALSE
)

The scripts get_RDS.R and get_RDS_xenium.R were used with rotating raw data directory namess to obtain the requisite Seurat objects, stored as .RDS files. All of the processed files are also archived here

Workflow

All benchmark objects were processed through the same Seurat analysis workflow implemented in the demo. This was done by first loading the requisite functions from mem_test_with_checkpoints.R, then running the Mem_test_loop.R.

Core Steps

1. Load Object: Load the .rds object and update it to the current Seurat object structure.
2. Filter Cells: Remove zero-count cells from the Xenium assay count layer.
3. Normalize Data: Run SCTransform normalization on the Xenium assay.
4. Reduce Dimensions: Run PCA on the SCT assay.
5. Cluster Cells: Build a nearest-neighbor graph, cluster the cells, and run UMAP.
6. Log Metrics: Record runtime, CPU time, object size, and memory-usage summaries in the output workbook.

Note: For the Atera human cervical dataset, the full object exceeded RAM capacity of the workstation during SCTransform. To make benchmarking feasible, a left-half spatial crop of the Atera cervical tissue was created and benchmarked instead.

Output metrics

The Excel workbook records several metrics for each object:

Metric Group	Field Name	Description
Metadata	Datasets	Descriptive dataset label used in the benchmark output.
	Species	Organism represented by the dataset.
	Sample type	Tissue or sample label.
	Cells or nuclei	Indicates whether the object represents cells or nuclei.
	Preservation	Sample preservation method.
	Area (mm²)	Tissue area used in the benchmark.
	Molecule Coordinates	Whether molecule-level spatial coordinates were retained.
Memory	Avg. Memory Usage	Average process memory sampled across checkpoints.
	Min. Memory Usage	Minimum process memory observed across checkpoint samples.
	Max. memory Usage	Peak memory observed during the run.
	Disk Usage	Input .rds file size on disk.
	PeakRAM_total_allocated_GiB	Total RAM allocated by the benchmarked expression.
Dataset Dimensions	RDS_file	Filename of the processed Seurat object.
	Input_cells	Number of cells in the Seurat object.
	Input_features	Number of input features or genes in the Seurat object.
	Clusters	Number of clusters identified after clustering.
Performance	Start_time	Timestamp marking the start of the benchmark run.
	End_time	Timestamp marking the end of the benchmark run.
	Wall_time_min	Total elapsed runtime in minutes.
	CPU_user_sec	CPU time spent executing user-space code.
	CPU_system_sec	CPU time spent in system-level operations.
Normalized Metrics	Peak(GB)/mm^2	Peak memory normalized by tissue area.
	Peak(KB)/Total_features	Peak memory normalized by total feature count (cells × features).

Summary Rows

• Atera: Aggregate Atera values for area-normalized peak-memory comparisons.
• Xenium: Aggregate Xenium values for the same normalized comparisons.
• Ratio: Atera-to-Xenium ratio for each summary metric.

Note: The Results of a run are stored in Resource_usage.xlsx, even though the script creates a .csv. This was done to make the post-processing of Normalized Metrics and Summary Rows easier to Manage. As these are not included in the script or the loop