In Acare/hacksig: A Tidy Framework to Hack Gene Expression Signatures

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
options(tibble.print_min = 5L, tibble.print_max = 5L)

Hacksig is a collection of cancer transcriptomics gene signatures and it provides a simple and tidy interface to compute single sample enrichment scores.

This document will show you how to getting started with hacksig, but first, we must load the following packages:

library(hacksig)

# to plot and transform data
library(dplyr)
library(ggplot2)
library(purrr)
library(tibble)
library(tidyr)

# to get the MSigDB gene signatures
library(msigdbr)

# to parallelize computations
library(future)

theme_set(theme_bw())

Available signatures from the literature

In order to get a complete list of the implemented signatures, you can use get_sig_info(). It returns a tibble with very useful information:

the signature_id;
a string of keywords associated to a signature (separated by the "pipe" | symbol);
the publication_doi linking to the original publication;
a brief description.

get_sig_info()

If you want to get the list of gene symbols for one or more of the implemented signatures, then use get_sig_genes() with valid keywords:

get_sig_genes("ifng")

Check your signatures

The first thing you should do before computing scores for a signature is to check how many of its genes are present in your data. To accomplish this, we can use check_sig() on a normalized gene expression matrix (either microarray or RNA-seq normalized data), which must be formatted as an object of class matrix or data.frame with gene symbols as row names and sample IDs as column names.

For this tutorial, we will use test_expr (an R object included in hacksig) as an example gene expression matrix with 20 simulated samples.

By default, check_sig() will compute statistics for every signature implemented in hacksig.

check_sig(test_expr)

You can filter for specific signatures by entering keywords in the signatures argument (partial matching and regular expressions will work too):

check_sig(test_expr, signatures = c("metab", "cinsarc"))

We can also check for signatures not implemented in hacksig, that is custom signatures. For example, we can use the msigdbr package to download the Hallmark gene set collection as a tibble and transform it into a list:

hallmark_list <- msigdbr(species = "Homo sapiens", collection = "H") %>%
    distinct(gs_name, gene_symbol) %>%
    nest(genes = c(gene_symbol)) %>%
    mutate(genes = map(genes, compose(as_vector, unname))) %>%
    deframe()

check_sig(test_expr, hallmark_list)

Missing genes for the HALLMARK_NOTCH_SIGNALING gene set are:

check_sig(test_expr, hallmark_list) %>% 
    filter(signature_id == "HALLMARK_NOTCH_SIGNALING") %>% 
    pull(missing_genes)

Compute single sample scores

hack_sig

The main function of the package, hack_sig(), permits to obtain single sample scores from gene signatures. By default, it will compute scores for all the signatures implemented in the package with the original publication method.

hack_sig(test_expr)

You can also filter for specific signatures (e.g. the immune and stromal ESTIMATE signatures) and choose a particular single sample method:

hack_sig(test_expr, signatures = "estimate", method = "zscore")

Valid choices for single sample methods are:

"zscore", for the combined z-score;
"ssgsea", for the single sample GSEA;
"singscore", for the singscore method.

Run ?hack_sig to see additional parameter specifications for these methods.

As in check_sig(), the argument signatures can also be a list of gene signatures. For example, we can compute normalized single sample GSEA scores for the Hallmark gene sets:

hack_sig(test_expr, hallmark_list, 
         method = "ssgsea", sample_norm = "separate", alpha = 0.5)

There are three methods for which hack_sig() cannot be used to compute gene signature scores with the original method. These are: CINSARC, ESTIMATE and the Immunophenoscore.

hack_cinsarc

For the CINSARC classification, you must provide a vector with distant metastasis status:

set.seed(123)
rand_dm <- sample(c(0, 1), size = ncol(test_expr), replace = TRUE)
hack_cinsarc(test_expr, rand_dm)

hack_estimate

Immune, stromal, ESTIMATE and tumor purity scores from the ESTIMATE method can be obtained with:

hack_estimate(test_expr)

hack_immunophenoscore

Finally, the raw immunophenoscore and its discrete (0-10 normalized) counterpart can be obtained with:

hack_immunophenoscore(test_expr)

You can also obtain all biomarker scores with:

hack_immunophenoscore(test_expr, extract = "all")

Stratify your samples

If you want to categorize your samples into two or more signature classes based on a score cutoff, you can use stratify_sig() after hack_sig():

test_expr %>% 
    hack_sig("estimate", method = "singscore", direction = "up") %>% 
    stratify_sig()

By default, stratify_sig() will stratify samples either with the original publication method (if any) or by the median score (otherwise). stratify_sig() will work only with signatures implemented in hacksig.

Speed-up computation time

Our rank-based single sample method implementations (i.e. single sample GSEA and singscore) are slower than their counterparts implemented in GSVA and singscore. Hence, to speed-up computation time you can use the future package:

plan(multisession)
hack_sig(test_expr, hallmark_list, method = "ssgsea")

Use case

Let's say we want to compute single sample scores for the KEGG gene set collection and then correlate these scores with the tumor purity given by the ESTIMATE method.

First, we get the KEGG list and use check_sig() to keep only those gene sets whose genes are more than 2/3 present in our gene expression matrix.

kegg_list <- msigdbr(species = "Homo sapiens", subcollection = "CP:KEGG_MEDICUS") %>%
    distinct(gs_name, gene_symbol) %>%
    nest(genes = c(gene_symbol)) %>%
    mutate(genes = map(genes, compose(as_vector, unname))) %>%
    deframe()

kegg_ok <- check_sig(test_expr, kegg_list) %>% 
    filter(frac_present > 0.66) %>% 
    pull(signature_id)

Then, we apply both the combined z-score and the ssGSEA method for the resulting list of 10 KEGG gene sets using purrr::map_dfr():

kegg_scores <- map_dfr(list(zscore = "zscore", ssgsea = "ssgsea"), 
                       ~ hack_sig(test_expr,
                                  kegg_list[kegg_ok],
                                  method = .x,
                                  sample_norm = "separate"),
                       .id = "method")

We can transform the kegg_scores tibble in long format using tidyr::pivot_longer():

kegg_scores <- kegg_scores %>% 
    pivot_longer(starts_with("KEGG"), 
                 names_to = "kegg_id", values_to = "kegg_score")

Finally, after computing the tumor purity scores, we can merge the two data sets and plot the results:

purity_scores <- hack_estimate(test_expr) %>% select(sample_id, purity_score)

kegg_scores %>% 
    left_join(purity_scores, by = "sample_id") %>% 
    ggplot(aes(x = kegg_id, y = kegg_score)) +
    geom_boxplot(outlier.alpha = 0) +
    geom_jitter(aes(color = purity_score), alpha = 0.8, width = 0.1) +
    facet_wrap(facets = vars(method), scales = "free_x") +
    coord_flip() +
    scale_color_viridis_c() +
    labs(x = NULL, y = "enrichment score", color = "Tumor purity") +
    theme(legend.position = "top")