We downloaded a previously generated simulated scATAC-seq dataset for human bone marrow cells, generated with 250–5,000 average counts per cell and a noise rate of 0.2 (ref. 52). Data were downloaded from GitHub (https://github.com/pinellolab/scATAC-benchmarking). For each simulated dataset, we ran each dimension reduction method as described above for the PBMC dataset, except that we used the first five dimensions for each method rather than 20. For SCALE, we used the full ten-dimension latent space.
