sig_fit: Fit Signature Exposures with Linear Combination Decomposition

Description

The function performs a signatures decomposition of a given mutational catalogue V with known signatures W by solving the minimization problem min(||W*H - V||) where W and V are known.

Usage

sig_fit(
  catalogue_matrix,
  sig,
  sig_index = NULL,
  sig_db = c("legacy", "SBS", "DBS", "ID", "TSB", "SBS_Nik_lab", "RS_Nik_lab",
    "RS_BRCA560", "RS_USARC", "CNS_USARC", "CNS_TCGA", "CNS_TCGA176", "CNS_PCAWG176",
    "SBS_hg19", "SBS_hg38", "SBS_mm9", "SBS_mm10", "DBS_hg19", "DBS_hg38", "DBS_mm9",
    "DBS_mm10", "SBS_Nik_lab_Organ", "RS_Nik_lab_Organ", "latest_SBS_GRCh37",
    "latest_DBS_GRCh37", "latest_ID_GRCh37", "latest_SBS_GRCh38", "latest_DBS_GRCh38",
    "latest_SBS_mm9", "latest_DBS_mm9", "latest_SBS_mm10", "latest_DBS_mm10",
    "latest_SBS_rn6", "latest_DBS_rn6", "latest_CN_GRCh37", 
    
    "latest_RNA-SBS_GRCh37", "latest_SV_GRCh38"),
  db_type = c("", "human-exome", "human-genome"),
  show_index = TRUE,
  method = c("QP", "NNLS", "SA"),
  auto_reduce = FALSE,
  type = c("absolute", "relative"),
  return_class = c("matrix", "data.table"),
  return_error = FALSE,
  rel_threshold = 0,
  mode = c("SBS", "DBS", "ID", "copynumber"),
  true_catalog = NULL,
  ...
)

Value

The exposure result either in matrix or data.table format. If return_error set TRUE, a list is returned.

Arguments

catalogue_matrix: a numeric matrix V with row representing components and columns representing samples, typically you can get nmf_matrix from sig_tally() and transpose it by t().
sig: a Signature object obtained either from sig_extract or sig_auto_extract, or just a raw signature matrix/data.frame with row representing components (motifs) and column representing signatures.
sig_index: a vector for signature index. "ALL" for all signatures.
sig_db: default 'legacy', it can be 'legacy' (for COSMIC v2 'SBS'), 'SBS', 'DBS', 'ID' and 'TSB' (for COSMIV v3.1 signatures) for small scale mutations. For more specific details, it can also be 'SBS_hg19', 'SBS_hg38', 'SBS_mm9', 'SBS_mm10', 'DBS_hg19', 'DBS_hg38', 'DBS_mm9', 'DBS_mm10' to use COSMIC v3 reference signatures from Alexandrov, Ludmil B., et al. (2020) (reference #1). In addition, it can be one of "SBS_Nik_lab_Organ", "RS_Nik_lab_Organ", "SBS_Nik_lab", "RS_Nik_lab" to refer reference signatures from Degasperi, Andrea, et al. (2020) (reference #2); "RS_BRCA560", "RS_USARC" to reference signatures from BRCA560 and USARC cohorts; "CNS_USARC" (40 categories), "CNS_TCGA" (48 categories) to reference copy number signatures from USARC cohort and TCGA; "CNS_TCGA176" (176 categories) and "CNS_PCAWG176" (176 categories) to reference copy number signatures from PCAWG and TCGA separately. UPDATE, the latest version of reference version can be automatically downloaded and loaded from https://cancer.sanger.ac.uk/signatures/downloads/ when a option with latest_ prefix is specified (e.g. "latest_SBS_GRCh37"). Note: the signature profile for different genome builds are basically same. And specific database (e.g. 'SBS_mm10') contains less signatures than all COSMIC signatures (because some signatures are not detected from Alexandrov, Ludmil B., et al. (2020)). For all available options, check the parameter setting.
db_type: only used when sig_db is enabled. "" for keeping default, "human-exome" for transforming to exome frequency of component, and "human-genome" for transforming to whole genome frequency of component. Currently only works for 'SBS'.
show_index: if TRUE, show valid indices.
method: method to solve the minimazation problem. 'NNLS' for non-negative least square; 'QP' for quadratic programming; 'SA' for simulated annealing.
auto_reduce: if TRUE, try reducing the input reference signatures to increase the cosine similarity of reconstructed profile to observed profile.
type: 'absolute' for signature exposure and 'relative' for signature relative exposure.
return_class: string, 'matrix' or 'data.table'.
return_error: if TRUE, also return sample error (Frobenius norm) and cosine similarity between observed sample profile (asa. spectrum) and reconstructed profile. NOTE: it is better to obtain the error when the type is 'absolute', because the error is affected by relative exposure accuracy.
rel_threshold: numeric vector, a signature with relative exposure lower than (equal is included, i.e. <=) this value will be set to 0 (both absolute exposure and relative exposure). In this case, sum of signature contribution may not equal to 1.
mode: signature type for plotting, now supports 'copynumber', 'SBS', 'DBS', 'ID' and 'RS' (genome rearrangement signature).
true_catalog: used by sig_fit_bootstrap, user never use it.
...: control parameters passing to argument control in GenSA function when use method 'SA'.

Details

The method 'NNLS' solves the minimization problem with nonnegative least-squares constraints. The method 'QP' and 'SA' are modified from SignatureEstimation package. See references for details. Of note, when fitting exposures for copy number signatures, only components of feature CN is used.

References

Daniel Huebschmann, Zuguang Gu and Matthias Schlesner (2019). YAPSA: Yet Another Package for Signature Analysis. R package version 1.12.0.

Huang X, Wojtowicz D, Przytycka TM. Detecting presence of mutational signatures in cancer with confidence. Bioinformatics. 2018;34(2):330–337. doi:10.1093/bioinformatics/btx604

Kim, Jaegil, et al. "Somatic ERCC2 mutations are associated with a distinct genomic signature in urothelial tumors." Nature genetics 48.6 (2016): 600.

Examples

Run this code

# \donttest{

# For mutational signatures ----------------
# SBS is used for illustration, similar
# operations can be applied to DBS, INDEL, CN, RS, etc.

# Load simulated data
data("simulated_catalogs")
data = simulated_catalogs$set1
data[1:5, 1:5]

# Fitting with all COSMIC v2 reference signatures
sig_fit(data, sig_index = "ALL")
# Check ?sig_fit for sig_db options
# e.g., use the COSMIC SBS v3
sig_fit(data, sig_index = "ALL", sig_db = "SBS")

# Fitting with specified signatures
# opt 1. use selected reference signatures
sig_fit(data, sig_index = c(1, 5, 9, 2, 13), sig_db = "SBS")
# opt 2. use user specified signatures
ref = get_sig_db()$db
ref[1:5, 1:5]
ref = ref[, 1:10]
# The `sig` used here can be result object from `sig_extract`
# or any reference matrix with similar structure (96-motif)
v1 = sig_fit(data, sig = ref)
v1

# If possible, auto-reduce the reference signatures
# for better fitting data from a sample
v2 = sig_fit(data, sig = ref, auto_reduce = TRUE)
v2

all.equal(v1, v2)

# Some samples reported signatures dropped
# but its original activity values are 0s,
# so the data remain same (0 -> 0)
all.equal(v1[, 2], v2[, 2])

# For COSMIC_10, 6.67638 -> 0
v1[, 4]; v2[, 4]
all.equal(v1[, 4], v2[, 4])

# For general purpose -----------------------

W <- matrix(c(1, 2, 3, 4, 5, 6), ncol = 2)
colnames(W) <- c("sig1", "sig2")
W <- apply(W, 2, function(x) x / sum(x))

H <- matrix(c(2, 5, 3, 6, 1, 9, 1, 2), ncol = 4)
colnames(H) <- paste0("samp", 1:4)

V <- W %*% H
V

if (requireNamespace("quadprog", quietly = TRUE)) {
  H_infer <- sig_fit(V, W, method = "QP")
  H_infer
  H

  H_dt <- sig_fit(V, W, method = "QP", auto_reduce = TRUE, return_class = "data.table")
  H_dt

  ## Show results
  show_sig_fit(H_infer)
  show_sig_fit(H_dt)

  ## Get clusters/groups
  H_dt_rel <- sig_fit(V, W, return_class = "data.table", type = "relative")
  z <- get_groups(H_dt_rel, method = "k-means")
  show_groups(z)
}

# if (requireNamespace("GenSA", quietly = TRUE)) {
#   H_infer <- sig_fit(V, W, method = "SA")
#   H_infer
#   H
#
#   H_dt <- sig_fit(V, W, method = "SA", return_class = "data.table")
#   H_dt
#
#   ## Modify arguments to method
#   sig_fit(V, W, method = "SA", maxit = 10, temperature = 100)
#
#   ## Show results
#   show_sig_fit(H_infer)
#   show_sig_fit(H_dt)
# }
# }

Run the code above in your browser using DataLab