Differences between manual and automatic extraction

[jrcodina]

When I run the analysis I get very different results from the manual and the automatic estimation of signatures to extract. How could I look for more detailed information about the factor to consider for the manual method?

I saw in the description that the most common method is to look to the point where the cophenetic starts to go down and also to the point where the RSS is going up. However, when looking at these I usually get a way higher number than with the automatic method which usually recommends only 1 or 2. Do you know if this method is designed for detecting a small number of signatures?

Thank you very much for your help.

[ShixiangWang]

What's the dimension of your matrix? Please try bp_extract_signatures() and see how it works. Maybe you can provide the survey plot so I can put some comments.

Example code:

# e1 <- bp_extract_signatures(
#   t(simulated_catalogs$set1),
#   range = 8:12,
#   n_bootstrap = 5,
#   n_nmf_run = 10
# )
#
# To avoid computation in examples,
# Here just [load](https://rdrr.io/r/base/load.html) the result
# (e1$signature and e1$exposure set to NA to reduce package size)
load([system.file](https://rdrr.io/r/base/system.file.html)("extdata", "e1.RData", package = "sigminer"))


# See the survey for different signature numbers
# The suggested solution is marked as red dot
# with highest integrated score.
p1 <- bp_show_survey(e1)
p1
# You can also exclude plotting and highlighting the score
p2 <- bp_show_survey(e1, add_score = FALSE)
p2

# You can also plot a simplified version
p3 <- bp_show_survey2(e1, highlight = 10)
p3

[jrcodina]

Answer

My matrix is: mt_tally_SNV$nmf_matrix : int [ 1:25 , 1:96 ]

MAF_file <- read_maf(maf = SNV_bone)
mt_tally_SNV <- sig_tally(MAF_file,
                                     ref_genome = "BSgenome.Hsapiens.UCSC.hg19",
                                     useSyn = TRUE)

mt_est <- sig_estimate(mt_tally_SNV$nmf_matrix,
                                  range = 2:15, 
                                  nrun = 50, 
                                  use_random = FALSE, 
                                  cores = 6,
                                  verbose = TRUE)

show_sig_number_survey(mt_est$survey)

With this information, I would choose to extract 10 signatures, however, the auto-extract function tells me to select K = 2.

With your method I also get 10 as the best result. Is bp_extract_signatures() a better option than sig_estimate() ?

Thank you very much for your help.

[ShixiangWang]

In general, bp_extract_signatures() is better, however, in the second figure, the 11 should be the suggested solution.

Beside, for samples only with 25, I also recommend you use sig_fit() to directly obtain signature activity based on COSMIC signatures, then you can visualize the activity with boxplot to check if there are indeed ~11 COSMIC signatures can be found.

[jrcodina]

Why is bp_extract_signatures() better than sig_estimate() ? By following the steps from the manual I haven't seen the bp function.

I have 15 files from 15 different tissues with 25 samples per file, so I would like to extract signatures for each tissue, and then a pan-cancer extraction. Most of the signatures will be in COSMIC but there is a possibility that out of the ~400 samples analyzed we will obtain some de novo signatures.

[ShixiangWang]

https://shixiangwang.github.io/sigminer/reference/bp.html (ref: Degasperi, Andrea, et al. "A practical framework and online tool for mutational signature analyses show intertissue variation and driver dependencies." Nature cancer 1.2 (2020): 249-263.) See this link. BP is better as they use bootstrapped data of raw input instead of raw input matrix (sig_estimate), so the result is more robust to variation.

[jrcodina]

Thank you! Really appreciate it.

[ShixiangWang]

@jrcodina96 you're welcome.