Some TADs are offset/don't align visually

[Naveen-Ahuja]

Hi,

Spectral works really well but in some regions the TADs are offset or don't align well with what one would expect when visually inspecting them (see attached image). I wanted to know if you have any advice on fixing this or any particular parameter I should play with to fix this.

Here is how I called the TADs:

for (chr in chromosomes) {
out_path <- paste0("ctrl_p14_", chr, ".bedpe")
ctrl_p14<- strawr::straw("NONE", "/Users", chr, chr, "BP", 50000)

ctrl_p14_tad = SpectralTAD(ctrl_p14, chr = chr, resolution = 50000, levels = 3, qual_filter = FALSE, z_clust = FALSE, out_format= "juicebox", out_path=out_path)
}

Thank you

[mdozmorov]

Hi @Naveen-Ahuja, thanks for reporting. This may happen due to data sparsity. I suggest two solutions:

1. Set levels = 1. First-level TADs are typically the most robust.
2. Set qual_filter = TRUE, z_clust = TRUE. This put additional checks on the quality of TADs. In fact, we now use these settings by default. We need to change the defaults on Bioconductor.

If you would like to make a more thorough test, here is our code checking the effect of various combinations of parameters. As said, qual_filter = TRUE, z_clust = TRUE should work best.

# SpectralTAD settings
qual_filter = TRUE; z_clust = TRUE # Silhouette score filtering
# qual_filter = FALSE; z_clust = TRUE # Z-score filtering
# qual_filter = FALSE; z_clust = FALSE # Mixed filtering
max_tad_size <- 2000000; window_size <- max_tad_size / resolution
gap_threshold <- 0.8
# Save directory
save_dir_tads <- file.path(save_dir, paste0(s, "_", normalization, "_", resolution, "_", ifelse(qual_filter, "qualT", "qualF"), "_", ifelse(z_clust, "zT", "zF"), "_", window_size, "_", gap_threshold))
if (!dir.exists(save_dir_tads)) {
  dir.create(save_dir_tads, recursive = TRUE)
}

for (chr in chromosomes) {
  # Input file
  fileNameIn <- file.path(data_dir, sample, paste0(s, "_", chr, "_", normalization, "_", resolution, ".txt.gz"))
  mtx <- fread(fileNameIn)
  # Output file
  fileNameOut <- file.path(save_dir_tads, paste0(chr, ".bedpe"))
  # SpectralTAD run depending on parameters
  SpectralTAD(mtx, chr = chr, levels = 1, qual_filter = qual_filter, z_clust = z_clust, window_size = window_size, resolution = resolution, gap_threshold = gap_threshold, grange = FALSE, out_format = "bedpe", out_path = fileNameOut)
}

[mdozmorov]

@Naveen-Ahuja, can you make a minimal reproducible example? You can share data subset privately. Your results do look confusing, we never encountered them and I would dig deeper.

[gorliver]

Hi @Naveen-Ahuja, thanks for reporting. This may happen due to data sparsity. I suggest two solutions:

1. Set levels = 1. First-level TADs are typically the most robust.
2. Set qual_filter = TRUE, z_clust = TRUE. This put additional checks on the quality of TADs. In fact, we now use these settings by default. We need to change the defaults on Bioconductor.

If you would like to make a more thorough test, here is our code checking the effect of various combinations of parameters. As said, qual_filter = TRUE, z_clust = TRUE should work best.

# SpectralTAD settings
qual_filter = TRUE; z_clust = TRUE # Silhouette score filtering
# qual_filter = FALSE; z_clust = TRUE # Z-score filtering
# qual_filter = FALSE; z_clust = FALSE # Mixed filtering
max_tad_size <- 2000000; window_size <- max_tad_size / resolution
gap_threshold <- 0.8
# Save directory
save_dir_tads <- file.path(save_dir, paste0(s, "_", normalization, "_", resolution, "_", ifelse(qual_filter, "qualT", "qualF"), "_", ifelse(z_clust, "zT", "zF"), "_", window_size, "_", gap_threshold))
if (!dir.exists(save_dir_tads)) {
  dir.create(save_dir_tads, recursive = TRUE)
}

for (chr in chromosomes) {
  # Input file
  fileNameIn <- file.path(data_dir, sample, paste0(s, "_", chr, "_", normalization, "_", resolution, ".txt.gz"))
  mtx <- fread(fileNameIn)
  # Output file
  fileNameOut <- file.path(save_dir_tads, paste0(chr, ".bedpe"))
  # SpectralTAD run depending on parameters
  SpectralTAD(mtx, chr = chr, levels = 1, qual_filter = qual_filter, z_clust = z_clust, window_size = window_size, resolution = resolution, gap_threshold = gap_threshold, grange = FALSE, out_format = "bedpe", out_path = fileNameOut)
}

Hi @mdozmorov, Thanks for the code. Would you please explain the parameter "normalization"? From what I understand, this parameter is for matrix generation, not for spectralTAD?

[mdozmorov]

This code is inspirational, shouldn't be used verbatim. Yes, "normalization" refers to our data preprocessing, it is not related to SpectralTAD. Most important are qual_filter = TRUE; z_clust = TRUE; max_tad_size <- 2000000; window_size <- max_tad_size / resolution; gap_threshold <- 0.8

[Naveen-Ahuja]

@mdozmorov sorry for the delay, I had to get permission to share the data. Where can I privately share with you my hic matrix?