Summarizing mutation status as a feature matrix
Tutorial: assemble feature matrix
In 2023, the Morin Lab developed a classifier of Follicular Lymphoma (FL) predictive of histological transformation to more aggressive DLBCL. This study was published in Blood (2023). How was the binary feature matrix assembled for that machine learning model?
In this quick tour we will show how GAMBLR.data resources and functionality of GAMBLR.open can help you to generate such matrix.
This tutorial explores how to obtain metadata for the samples, simple somatic mutations in maf format, and auto-magically transform it into a binary matrix of features.
To simplify this tutorial, we will only include a small subset of features (not the whole set as was used in the original paper), but this example will be able to illustrate the functionality and highlight the main steps of the process.
Obtain metadata
In the previous tutorial, we have already explored the function get_gambl_metadata(). We will use it to retreive the metadata:
metadata <- get_gambl_metadata() %>%
filter(
study == "FL_Dreval"
)We can now see that the samples from the FL study are in the cohort FL_Dreval. Let’s explore these samples more:
table(metadata$pathology)
COMFL DLBCL FL
21 209 213 For illustration purposes, let’s take 10 samples from the FL study: 5 FL and 5 DLBCL:
# Only filter for the samples from FL study
metadata <- metadata %>%
filter(pathology %in% c("FL", "DLBCL")) %>%
group_by(pathology) %>%
arrange(sample_id) %>%
slice_tail(n = 5) %>%
ungroup
# How does our metadata looks like now?
str(metadata)tibble [10 × 30] (S3: tbl_df/tbl/data.frame)
$ age_group : chr [1:10] "Other" "Other" "Other" "Other" ...
$ bam_available : logi [1:10] TRUE TRUE TRUE TRUE TRUE TRUE ...
$ cohort : chr [1:10] "DLBCL_ICGC" "DLBCL_ICGC" "DLBCL_ICGC" "DLBCL_ICGC" ...
$ compression : chr [1:10] "cram" "cram" "cram" "cram" ...
$ COO_consensus : chr [1:10] "ABC" "GCB" NA "ABC" ...
$ DHITsig_consensus : chr [1:10] "DHITsigNeg" "DHITsigNeg" "NA" "DHITsigNeg" ...
$ EBV_status_inf : chr [1:10] NA NA NA NA ...
$ ffpe_or_frozen : chr [1:10] "frozen" "frozen" "frozen" "frozen" ...
$ fl_grade : chr [1:10] NA NA NA NA ...
$ genetic_subgroup : chr [1:10] "dFL" "dFL" "dFL" "dFL" ...
$ genome_build : chr [1:10] "hs37d5" "hs37d5" "hs37d5" "hs37d5" ...
$ hiv_status : chr [1:10] NA NA NA NA ...
$ lymphgen : chr [1:10] "Other" "Other" "BN2" "MCD-COMP" ...
$ lymphgen_cnv_noA53 : chr [1:10] "Other" "Other" "BN2" "BN2/MCD" ...
$ lymphgen_no_cnv : chr [1:10] "MCD" "Other" "BN2" "BN2/MCD" ...
$ lymphgen_with_cnv : chr [1:10] "A53" "A53" "BN2" "BN2/MCD" ...
$ lymphgen_wright : chr [1:10] NA NA NA NA ...
$ molecular_BL : chr [1:10] "non-BL" "non-BL" NA "non-BL" ...
$ normal_sample_id : chr [1:10] "SP59410" "SP59446" "SP59450" "SP59454" ...
$ pairing_status : chr [1:10] "matched" "matched" "matched" "matched" ...
$ pathology : chr [1:10] "DLBCL" "DLBCL" "DLBCL" "DLBCL" ...
$ pathology_rank : num [1:10] 19 19 19 19 19 15 15 15 15 15
$ patient_id : chr [1:10] "DO27833" "DO27851" "DO27853" "DO27855" ...
$ reference_PMID : num [1:10] 37084389 37084389 37084389 37084389 37084389 ...
$ sample_id : chr [1:10] "SP59412" "SP59448" "SP59452" "SP59456" ...
$ seq_type : chr [1:10] "genome" "genome" "genome" "genome" ...
$ sex : chr [1:10] "M" "F" "M" "F" ...
$ study : chr [1:10] "FL_Dreval" "FL_Dreval" "FL_Dreval" "FL_Dreval" ...
$ time_point : chr [1:10] "A" "A" "A" "A" ...
$ Tumor_Sample_Barcode: chr [1:10] "SP59412" "SP59448" "SP59452" "SP59456" ...Our subsetting worked and we can now proceed with matrix assembly.
Generate feature matrix
We will now use the metadata from the previous step to assemble the binary feature matrix.
Return the simple somatic mutations
First, lets return the data frame with simple somatic mutations in maf format and store it in a variable. Technically this step is not strictly necessary, as each function used below will be able to retreive it for you when the maf data is not provided, but we will advocate for good practice here and have the maf data stored in a designated variable. As we will be utilizing both coding and non-coding mutations, we will take advantage of returning the variants per sample (without necessarily restricting to coding-only variants).
# Obtain simple somatic mutations
maf <- get_ssm_by_samples(
these_samples_metadata = metadata
)Wow, that was super easy and blazing fast! Did it even work?? Let’s confirm:
[1] 10
# Are all these samples the ones we are interested in and requested with metadata?
sort(unique(maf$Tumor_Sample_Barcode)) == sort(metadata$Tumor_Sample_Barcode) [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
# What are the mutations in the maf? Are they just coding?
table(maf$Variant_Classification)
3'Flank 3'UTR 5'Flank
29 20 127
5'UTR Frame_Shift_Del Frame_Shift_Ins
61 7 1
IGR In_Frame_Del Intron
46 2 1502
Missense_Mutation Nonsense_Mutation RNA
78 9 15
Silent Splice_Region Splice_Site
29 6 8
Translation_Start_Site
1 We can see from the above outputs that we got somatic mutations for all requested samples and the maf contains both coding and non-coding variants. We can now proceed to the next steps.
Coding mutations as features
The feature matrix in the FL study contained coding mutation status at selected genes denoted as 1 when the mutation was present and 0 when there was no mutation. In addition to that, mutation hotspots at some genes were also taken into account. To complicate things more, annotation of hotspots was performed differently depending on the specific gene and mutational effect. Specifically, the CREBBP missemse mutations were considered hotspots when they occured in the KAT domain, but not outside of it. The mutations in FOXO1 were considered hotspots when the AA change was at M1.
GAMBLR.data provides a one-stop shop to achieve this level of details out-of-the-box in a simple and convenient way, and is directly used by the get_coding_ssm_status function. This function has a logical arguments include_hotspots to separate regular mutations from the ones occurring at hotspots, and review_hotspots, which will handle the specific cases we described above in an automated way. Both of these arguments are TRUE by default, so you do not need to toggle them separately, but in this example we will specify them explicitly just to illustrate that this will happen during the function call. To keep the example clean and concise, we will also annotate the mutation status only for a few of selected genes.
# Specify genes to annotate
our_genes <- c(
"CREBBP", "MYD88", "RRAGC",
"PIM1", "BCL2", "BCL6"
)
# Generate binary matrix
coding_matrix <- get_coding_ssm_status(
gene_symbols = our_genes,
these_samples_metadata = metadata,
maf_data = maf,
include_hotspots = TRUE,
review_hotspots = TRUE
)
coding_matrix sample_id PIM1 CREBBP RRAGC MYD88 BCL6 BCL2 CREBBPHOTSPOT MYD88HOTSPOT
1 SP59412 1 1 0 0 0 0 0 0
2 SP59448 0 0 0 0 0 0 0 0
3 SP59452 1 0 0 1 1 0 0 0
4 SP59456 0 0 0 0 0 0 0 0
5 SP59460 0 0 0 0 1 1 0 1
6 SP59420 0 1 0 0 0 0 0 0
7 SP59424 0 0 0 0 0 0 1 0
8 SP59432 0 0 1 0 0 0 0 0
9 SP59436 0 1 0 0 0 0 0 0
10 SP59464 0 0 1 0 0 0 0 0We can see that in this example there is only one sample with hotspot mutation in CREBBP, SP59424 (annotated as 1 in the column CREBBPHOTSPOT). Let’s sanity check this annotation for illustration purposes:
crebbp_hotspot_mutation <- maf %>%
filter(
Tumor_Sample_Barcode == "SP59424",
Hugo_Symbol == "CREBBP"
) %>%
select(Chromosome, Start_Position, Variant_Classification)
crebbp_hotspot_mutationgenomic_data Object
Genome Build: grch37
Showing first 10 rows:
Chromosome Start_Position Variant_Classification
1 16 3788650 Missense_MutationThis mutation falls within KAT domain and is a missense variant, so it is indeed correct to be annotated as hotspot. How can we check it does fall within KAT domain? This is easy with GAMBLR.data!
# GAMBLR.data has the regions to be considered as hotspots
hotspot_regions_grch37 chrom start end
CREBBP 16 3785000 3791000
EZH2 7 148508764 148506238
NOTCH1 9 139391455 139391455
NOTCH2 1 120459150 120459150
CD79B_trunc 17 62007172 62007172
CD79B_NONtrunc 17 62006800 62006800
# Now check that CREBBP mutations
between(
crebbp_hotspot_mutation$Start_Position,
hotspot_regions_grch37["CREBBP", "start"],
hotspot_regions_grch37["CREBBP", "end"]
)[1] TRUEWe have now generated matrix of coding mutations and SSM hotspots in a binary format and are ready to proceed to the next step.
aSHM mutations as features
The FL study also annotated the non-coding mutations at selected aSHM targets as features of binary matrix. Again, all the necessarily means to do it in a simple step are provided by GAMBLR.data. The regions targeted by aSHM are available in the GAMBLR.data. Since they are always complemented with new regions once they are identified, the latest and most comprehensive version is always available by referring to {projection}_ashm_regions. However, this use case is a perfect example to demonstarate how to operate on versioned iterations of the aSHM target list, as the list has been updated since the time the study was published and at the time of publication version 0.2 was used. We can refer to it directly by the version number:
regions_bed <- somatic_hypermutation_locations_GRCh37_v0.2
head(regions_bed) chr_name hg19_start hg19_end gene region regulatory_comment
1 chr2 96808901 96811913 DUSP2 intron-1 enhancer
2 chr17 56407732 56410140 TSPOAP1 intergenic enhancer
3 chr11 128339774 128345731 ETS1 introns enhancer
4 chr11 128388492 128394163 ETS1 TSS-2 active_promoter
5 chr6 31548325 31550717 LTB intron-1 enhancer
6 chr3 32020518 32024930 OSBPL10 TSS-1 active_promoterWe will perform some simple modifications to it to make our experience better and only select few regions for illustrative purposes:
our_regions <- c(
"BCL6-TSS",
"BCL7A-TSS",
"RHOH-TSS",
"ZFP36L1-TSS"
)
regions_bed <- create_bed_data(
regions_bed,
fix_names = "concat",
concat_cols = c("gene","region"),
sep = "-",
genome_build = "grch37"
)
regions_bed <- regions_bed %>%
filter(name %in% our_regions)
regions_bedgenomic_data Object
Genome Build: grch37
Showing first 10 rows:
chrom start end name region regulatory_comment
1 3 187458526 187464632 BCL6-TSS TSS <NA>
2 12 122456912 122464036 BCL7A-TSS TSS poised_promoter
3 14 69257848 69259739 ZFP36L1-TSS TSS active_promoter
4 4 40193105 40204231 RHOH-TSS TSS active_promoterNow we can see whether or not there are any mutations within these regions:
ashm_matrix <- cool_overlaps(
maf,
regions_bed,
columns2 = c("chrom", "start", "end")
) %>%
group_by(Tumor_Sample_Barcode, name) %>%
summarize(n = n()) %>%
pivot_wider(
id_cols = Tumor_Sample_Barcode,
names_from = name,
values_from = n,
values_fill = 0
) %>%
column_to_rownames("Tumor_Sample_Barcode")
ashm_matrix BCL6-TSS ZFP36L1-TSS BCL7A-TSS RHOH-TSS
SP59412 4 1 0 0
SP59420 1 0 1 0
SP59424 0 0 0 1
SP59436 1 0 0 0
SP59448 6 4 14 3
SP59452 18 2 12 28
SP59456 2 6 0 9
SP59460 1 1 1 2
SP59464 1 0 0 0We now calculated the number of mutations in each region for each sample. The original study used pathology-specific values per region to convert these counts to binary, but here we will use an arbitrary cutoff of 5 to binarize these counts. We will also convert rownames to column so it will be easier for us to unite all matrices into single one at the later steps:
ashm_matrix[ashm_matrix <= 5] = 0
ashm_matrix[ashm_matrix > 5] = 1
ashm_matrix <- ashm_matrix %>%
rownames_to_column("sample_id")
ashm_matrix sample_id BCL6-TSS ZFP36L1-TSS BCL7A-TSS RHOH-TSS
1 SP59412 0 0 0 0
2 SP59420 0 0 0 0
3 SP59424 0 0 0 0
4 SP59436 0 0 0 0
5 SP59448 1 0 1 0
6 SP59452 1 0 1 1
7 SP59456 0 1 0 1
8 SP59460 0 0 0 0
9 SP59464 0 0 0 0We have now generated matrix of non-coding mutations in a binary format and are ready to unite all matrices together.
Unite into full feature matrix
We can now combine both coding and non-coding mutations into single matrix:
feature_matrix <- left_join(
coding_matrix,
ashm_matrix
) %>%
mutate(
across(
where(is.numeric), ~ replace_na(.x, 0)
)
)
feature_matrix sample_id PIM1 CREBBP RRAGC MYD88 BCL6 BCL2 CREBBPHOTSPOT MYD88HOTSPOT
1 SP59412 1 1 0 0 0 0 0 0
2 SP59448 0 0 0 0 0 0 0 0
3 SP59452 1 0 0 1 1 0 0 0
4 SP59456 0 0 0 0 0 0 0 0
5 SP59460 0 0 0 0 1 1 0 1
6 SP59420 0 1 0 0 0 0 0 0
7 SP59424 0 0 0 0 0 0 1 0
8 SP59432 0 0 1 0 0 0 0 0
9 SP59436 0 1 0 0 0 0 0 0
10 SP59464 0 0 1 0 0 0 0 0
BCL6-TSS ZFP36L1-TSS BCL7A-TSS RHOH-TSS
1 0 0 0 0
2 1 0 1 0
3 1 0 1 1
4 0 1 0 1
5 0 0 0 0
6 0 0 0 0
7 0 0 0 0
8 0 0 0 0
9 0 0 0 0
10 0 0 0 0That’s it!
Happy GAMBLing!
/$$$$$$ /$$$$$$ /$$ /$$ /$$$$$$$ /$$ .:::::::
/$$__ $$ /$$__ $$ | $$$ /$$$ | $$__ $$ | $$ .:: .::
| $$ \__/ | $$ \ $$ | $$$$ /$$$$ | $$ \ $$ | $$ .:: .::
| $$ /$$$$ | $$$$$$$$ | $$ $$/$$ $$ | $$$$$$$ | $$ <- .: .::
| $$|_ $$ | $$__ $$ | $$ $$$| $$ | $$__ $$ | $$ .:: .::
| $$ \ $$ | $$ | $$ | $$\ $ | $$ | $$ \ $$ | $$ .:: .::
| $$$$$$/ | $$ | $$ | $$ \/ | $$ | $$$$$$$/ | $$$$$$$$ .:: .::
\______/ |__/ |__/ |__/ |__/ |_______/ |________/
~GENOMIC~~~~~~~~~~~~~OF~~~~~~~~~~~~~~~~~B-CELL~~~~~~~~~~~~~~~~~~IN~~~~~~
~~~~~~~~~~~~ANALYSIS~~~~~~MATURE~~~~~~~~~~~~~~~~~~~LYMPHOMAS~~~~~~~~~~R~