DRUID: DRUg Indication Discoverer

DRUID, or DRUg Indication Discoverer, is an algorithm that identifies drug profiles that revert or mimic a condition of interest. For example, DRUID can be used to repurpose compounds for novel indications given a gene expression profile, it can be used to prioritize compounds for a compendium of disease states, and it can be incorporated into computational chemistry pipelines for the identification and characterization of drug properties for novel design.

Installation

Install from GitHub using devtools as:

devtools::install_github("midas-wyss/druid")

Running

The easiest way to run DRUID is to use its wrapper concoct as:

library(DRUID)
res <- concoct(dge_matrix, tfidf_matrix, crossproduct_vector, number_random, effect_direction, fold_thr, pvalue_thr, entrez_ids)

where dge_matrix is a 2-column matrix for the query gene expression signature (column 1: fold-changes; column 2: p-values); tfidf_matrix is the calculated corrected TF-IDF matrix (see ctfidf function); crossproduct_vector is the crossproduct of the TF-IDF matrix (see crossprod_matrix function); number_random is the number of random simulations to be run to assess significance of scores (defaults to 1,000 - see random_probability function); effect_direction is the desired effect of the drug ("neg" for a reversal of the query phenotype, "pos" for a mimicking of the phenotype. Defaults to "neg"); fold_thr is a threshold for the fold changes (defaults to log2 = 0); pvalue_thr is the threshold for expression change significance (defaults to 0.05); and entrez_ids are the EntrezIDs for the genes in the query signature.

DRUID comes pre-packaged with a TF-IDF matrix and cross-product vector that were derived from the Connectivity Map data (as available in the Harmonizome).

Example

As an example, I will use the CMAP TF-IDF to generate a query vector and run DRUID on it.

gset <- unique(gsub(" down", "", gsub(" up", "", sample(colnames(DRUID::cmap_druid$tfidf), 100))))

query_matrix <- matrix(1, ncol = 2, nrow = length(gset))
query_matrix[, 2] <- 0
query_matrix[sample(x = seq(1, length(gset)), size = 0.25 * length(gset)), 1] <- -1

With the generated query matrix, we can now run DRUID on it using the concoct wrapper:

example_druid <- concoct(dge_matrix = query_matrix, tfidf_matrix = DRUID::cmap_druid$tfidf, tfidf_crossproduct = DRUID::cmap_druid$cpm, num_random = 10000, druid_direction = "neg", fold_thr = 0, pvalue_thr = 0.05, entrez = gset)

The output of DRUID is a tibble data frame with all the scores for all the drugs. Specifically, the columns in this data frame are:

• number_matches: number of genes in query signature found in drug signature;
• cosine_similarity: similarity of query signature to drug signature;
• probability_random: probability of cosine similarity being better than random;
• druid_score: calculated DRUID score, taking into account the cosine similarity and the random probability.

We can expand this data frame using magrittr and tibble together with the information on the drugs as:

example_druid <- example_druid %>% 
  tibble::add_column(., drug_name = DRUID::cmap_druid$drugs$name, .before = 1) %>%
  tibble::add_column(., concentration = DRUID::cmap_druid$drugs$concentration, .before = 2) %>%
  tibble::add_column(., cell_line = DRUID::cmap_druid$drugs$cell_line, .before = 3)

We can now use ggplot2 to visualize the results:

example_druid %>% dplyr::filter(., cosine_similarity == 0) %>% ggplot() + geom_point(aes(x = drug_name, y = druid_score, color = cell_line), alpha = 0.5) + facet_grid(. ~ cell_line, scales = "free") + theme_bw() + theme(axis.text.x = element_blank())

Using other sources

DRUID comes pre-packaged with the TF-IDF for the Connectivity Map data, but it's simple to generate a TF-IDF to meet your needs. For that, the ctfidf function will be used, which uses the tidytext package by Julia Silge. Inputs to this function are a data matrix where the columns are the words (eg, genes and their direction) and the rows are the documents (eg, drugs).

A drug profile will generate a different response on the transcriptome, with genes being differentially expressed. As such, we can generate a one-hot encoded vector in which all possibilities of change are represented for each gene in a given condition. These are the vectors that will be present in the data matrix that serves as input to ctfidf, where each "word" represents a gene (as an Entrez ID) and the direction of change that the "document" (drug) caused.

(NOTE: a one-hot encoding function was not included in DRUID. Will be included in future releases.)

With a one-hot encoded matrix, we can generate a TF-IDF matrix as:

ex_tfidf <- ctfidf(data_matrix)

and the corresponding cross-product vector as:

ex_cp <- crossprod_matrix(ex_tfidf)