Signatures

In YAPSA, the patterns of nucleotide exchanges in their triplet context constituting the mutational signatures are stored in data frames which can be loaded as follows (more information provided by the help function, cf.  (1. Usage of YAPSA) for an overview of how these data frames can be re-created from data available online):

data(sigs)
data(sigs_pcawg)

The first command, data(sigs), loads eight dataframes into the workspace. Among these, four contain the patterns of nucleotide exchanges in their triplet context, i.e., the mutational signatures themselves, and the remaining four dataframes contain additional information on the signatures, including a naming and numbering convention, a description of the asserted underlying mutational process and a choice of generic colour coding. In particular, these eight signature dataframes are:

• Two signature and two additonal information dataframes containing SNV Alexandrov COSMIC signatures (available at https://cancer.sanger.ac.uk/cosmic/signatures_v2, downloaded on January 15th, 2016):
  • AlexCosmicArtif_sig_df and AlexCosmicArtif_sigInd_df,
  • AlexCosmicValid_sig_df and AlexCosmicValid_sigInd_df
• Two signature and two additonal information dataframes containing SNV Alexadrov Initial signatures (the signatures presented in the initial publication (Alexandrov et al. 2013)):
  • AlexInitialArtif_sig_df and AlexInitialArtif_sigInd_df,
  • AlexInitialValid_sig_df and AlexInitialValid_sigInd_df

Of note, all data provided in YAPSA follows a consistent naming convention:

• Names containing Valid or Real refer to those subsets of signatures which have been identified by NMF in the respective discovery analyses and validated using orthogonal sequencing technologies.
• Names containing Artif refer to extended sets of signatures including also those signatures which have later been ascribed to be artifact signatures (reflecting, among others, sequencing errors).

The second command in the above code block, data(sigs_pcawg), loads six additional dataframes into the workspace, three dataframes containing mutational signatures themselves and three dataframes containing additional information on the signatures. In particular, these six dataframes are:

• Two signature and two additional information dataframes containing SNV Alexandrov PCAWG signatures (available at https://cancer.sanger.ac.uk/cosmic/signatures_v3, downloaded on August 16th, 2018):
  • PCAWG_SP_SBS_sigs_Artif_df and PCAWG_SP_SBS_sigInd_Artif_df,
  • PCAWG_SP_SBS_sigs_Real_df and PCAWG_SP_SBS_sigInd_Real_df
• One signature and one additional information dataframe containing Indel Alexandrov PCAWG signatures (also available at https://cancer.sanger.ac.uk/cosmic/signatures_v3, downloaded on August 16th, 2018):
  • PCAWG_SP_ID_sigs_df and PCAWG_SP_ID_sigInd_df

As of december 2019 and to our knowledge, it is a unique feature of YAPSA to include PCAWG mutational signatures. Of note, a separate vignette describes the usage of and analysis with Indel signatures.

Optimal cutoffs

For each of the signature dataframes, corresponding dataframes storing numerical values of the signature-specific cutoffs are available. The cutoff dataframes can be loaded to the workspace analogously to the signatures themselves:

data(cutoffs)
data(cutoffs_pcawg)

The command data(cutoffs) loads cutoff dataframes corresponding to the four Alexandrov COSMIC and Alexandrov initial sets of signatures. In the following, an additional naming convention is introduced:

• Absolute (abs) signature-specific cutoffs were computed using an ROC analysis performed on absolute exposures as provided by the initial NMF analysis
• Relative (rel) signature-specific cutoffs were computed using on ROC analysis performed on relative, i.e., normalized exposures as provided by the initial NMF analysis.

This double training has proven to be both useful and necessary in order to account for the heterogeneity of the underlying training data: on one hand, a training with absolute exposures weighs all mutations equally, however a whole genome sequenced (WGS) sample gets a much higher weight due to the high number of mutations detected in it than a whole exome sequenced (WES sample). On the other hand, a training with relative or normalized signature exposures weighs all samples equally, but at the cost of having different weights for mutations originating from samples of different mutation count. Based on these considerations, cutoffs derived from the training with absolute exposures should be used for the supervised analysis of WGS data, whereas cutoffs trained on relative exposures are better suited for the analysis of WES data. The two different sets of cutoffs thus account for the different maximal feature occurrence between WGS and WES data. The correspondence between signatures and cutoffs is as follows:

• For AlexCosmicValid_sig_df, the corresponding cutoff dataframes are cutoffCOSMICValid_abs_df and cutoffCOSMICValid_rel_df.
  
• For AlexCosmicArtif_sig_df, the corresponding cutoff dataframes are cutoffCOSMICArtif_abs_df and cutoffCOSMICArtif_rel_df.

With data(cutoffs_pcawg) the signature-specific cutoff dataframes for the PCAWG signatures are loaded. In this case, only one dataframe with cutoffs per signature dataframe is available. Signature-specific cutoffs are valid for WGS and WES data analysis as differences in the feature occurrence is taken into account during the training procedure.

In the provided dataframes with numerical values for optimal signature-specific cutoffs, the columns correspond to the different signatures of the chosen signature set, whereas the rows encode different values of cost_factor (cf. introduction). For an actual analysis of mutational signatures, only one optimal cutoff value per signature is required, i.e., the user has to choose one row from the chosen cutoff dataframe. As already explained above, optimal values for cost_factor have also been determined. These are listed in the table below. Unless a user has a specific wish to use a different cost_factor, it is generally recommended to use these default parameter choices.

Thus, if a user wants to do an analysis with the Alexandrov COSMIC signatures, discarding the artifact signatures, on WGS data, he/she would use the following code snippet:

data(cutoffs)
current_cutoff_vector <- cutoffCosmicValid_abs_df[6, ]
current_cutoff_vector

##   AC1        AC2        AC3        AC4 AC5         AC6        AC7      AC8
## 6   0 0.01045942 0.08194056 0.01753969   0 0.001548535 0.04013304 0.242755
##         AC9       AC10       AC11      AC12        AC13      AC14       AC15
## 6 0.1151714 0.01008376 0.09924884 0.2106201 0.007876626 0.1443059 0.03796027
##        AC16        AC17      AC18      AC19      AC20      AC21       AC22
## 6 0.3674349 0.002647962 0.3325386 0.1156454 0.1235028 0.1640255 0.03102216
##         AC23       AC24       AC25       AC26        AC27       AC28       AC29
## 6 0.03338659 0.03240176 0.01611908 0.09335221 0.009320062 0.05616434 0.05936213
##         AC30
## 6 0.05915355

Comparison: analysis without any cutoffs

For the purpose of comparison, we first create a zero cutoff vector:

current_cutoff_vector <- rep(0, dim(AlexCosmicValid_sig_df)[2])

Next, we compute the exposures to the mutational signatures using LCD analysis and this zero cutoff vector:

lymphoma_COSMIC_zero_listsList <-
  LCD_complex_cutoff_combined(
    in_mutation_catalogue_df = lymphomaNature2013_mutCat_df,
    in_cutoff_vector = current_cutoff_vector, 
    in_signatures_df = current_sig_df, 
    in_sig_ind_df = current_sigInd_df)

The function returns an object with the results per cohort, per PID and the consensus between both analyses.

Annotation of a subgroup to each PID allows to group the samples per subgroup upon visualization:

data(lymphoma_PID)
colnames(lymphoma_PID_df) <- "SUBGROUP"
lymphoma_PID_df$PID <- rownames(lymphoma_PID_df)
COSMIC_subgroups_df <- 
  make_subgroups_df(lymphoma_PID_df,
                    lymphoma_COSMIC_zero_listsList$cohort$exposures)

We select the cohort-wide analysis for visualization and make use of the custom plotting function exposures_barplot() in order to display signature exposures as stacked barplots. Note the occurrence of multiple signatures with partially very low exposures.

result_cohort <- lymphoma_COSMIC_zero_listsList$cohort
exposures_barplot(
  in_exposures_df = result_cohort$exposures,
  in_signatures_ind_df = result_cohort$out_sig_ind_df,
  in_subgroups_df = COSMIC_subgroups_df)        

Absoute exposures of the COSMIC signatures in the lymphoma mutational catalogs, signature-specific cutoffs with a cost factor of 6 used for the LCD

Making use of optimal signature-specific cutoffs

The zero-cutoff vector will now be replaced by the signature-specific cutoff vector. After the initial LCD analysis only signatures that contribute higher then their signature specific cutoff value will be considered, leading to a reduction of the overall number of detected signatures.

Based on @ref(tab:Table1) we know that for the validated COSMIC Signatures the optimal cutoffs are the once computed with a cost factor of six.

current_cutoff_df <- cutoffCosmicValid_abs_df
current_cost_factor <- 6
current_cutoff_vector <- current_cutoff_df[current_cost_factor,]

We again compute signature exposures, but this time using the vector of optimal signature-specific cutoffs:

lymphoma_COSMIC_listsList <-
  LCD_complex_cutoff_combined(
      in_mutation_catalogue_df = lymphomaNature2013_mutCat_df,
      in_cutoff_vector = current_cutoff_vector, 
      in_signatures_df = current_sig_df, 
      in_sig_ind_df = current_sigInd_df)

And again proceeding to a visualization (of note, the sample subgrouping information can be used as in the previous subsection):

result_cohort <- lymphoma_COSMIC_listsList$cohort
exposures_barplot(
  in_exposures_df = result_cohort$exposures,
  in_signatures_ind_df = result_cohort$out_sig_ind_df,
  in_subgroups_df = COSMIC_subgroups_df)               

Absolute exposures of the COSMIC signatures in the lymphoma mutational catalogs, signature-specific cutoffs with a cost factor of 6 used for the LCD

We note that the number of identified signatures is smaller. Those signatures which had been identified in the analysis without any cutoffs but not in the analysis with optimal cutoffs can be considered false positive calls.