This pipeline enables exploring the interactions between two types of multi-omic features in relation to cancer patient outcomes. The pipeline runs an iterative procedure consisting of subsampling, feature preprocessing and elastic net regularization of the multivariate models to identify the most frequently selected interactions across the iterations. These candidate interactions are a minimal subset of input interactions with robust explanatory information. The pipeline also provides a series of ANOVA P values which can be used for further highlighting the interactions whose explanatory information is complementary their individual components as well as the baseline variables.
Dependencies:
glmnetsurvival
Clone the repository (git clone https://github.com/reimandlab/PACIFIC.git), then run the following command in R: install.packages("path/to/PACIFIC", repos = NULL, type = "source").
In this example, PACIFIC is applied to a dataset of 463 TCGA-LUSC primary tumors annotated with overall survival outcomes, clinical variables, mutational status of cancer driver genes, and relative abundance of immune cell types in tumors (see IGX paper for data collection details). Here, we define the first feature set (feat1) as the mutational status of driver genes, the second feature set (feat2) as the immune cell levels, and the baseline variables as patient age, sex, and tumor stage. For simplicity, we use a subset of available features for the two feature sets in this example.
library(PACIFIC)
####
# Run an example using the dataset included in the PACIFIC package.
####
fname_data <- system.file("extdata", "example_dataset.rds", package = "PACIFIC")
data <- readRDS(fname_data)
####
# Define the baseline and the two sets of features.
####
baseline <- c("age", "sex", "stage")
feat1 <- c("TP53", "KMT2D", "CDKN2A", "KRAS")
feat2 <- c("B_cells_memory", "Plasma_cells",
"Macrophages_M1", "Macrophages_M2",
"Monocytes", "NK_cells_activated",
"T_cells_CD8","T_cells_regulatory_Tregs")
####
# View a few rows of the data with the selected columns:
####
data[1:5 , c("time", "status", baseline, feat1, feat2)]
# time status age sex stage TP53 KMT2D CDKN2A KRAS
#1: 371 1 67 MALE Stage_I mutated not_mutated not_mutated not_mutated
#2: 136 1 72 MALE Stage_I mutated mutated mutated not_mutated
#3: 2304 1 77 FEMALE Stage_I mutated not_mutated not_mutated not_mutated
#4: 3650 0 74 MALE Stage_I mutated not_mutated not_mutated not_mutated
#5: 146 1 81 MALE Stage_II mutated mutated not_mutated not_mutated
# B_cells_memory Plasma_cells Macrophages_M1 Macrophages_M2 Monocytes
#1: 0.00000000 0.26404279 0.09956779 0.1293732 0.038421812
#2: 0.01379474 0.10389209 0.03997567 0.2588654 0.028434340
#3: 0.00000000 0.01028084 0.05054219 0.3876288 0.003981384
#4: 0.00000000 0.04098218 0.09988443 0.4618061 0.000000000
#5: 0.00000000 0.07996814 0.10241677 0.1191702 0.021436310
# NK_cells_activated T_cells_CD8 T_cells_regulatory_Tregs
#1: 0.000000000 0.10903708 0.003766431
#2: 0.000000000 0.17287965 0.046420556
#3: 0.001416327 0.07765524 0.000000000
#4: 0.003790870 0.03094862 0.000000000
#5: 0.036007745 0.17967422 0.040558222
####
# Apply PACIFIC to the data with the defined features (using default settings):
####
set.seed(1) # for reproducibility of this demo
results <- PACIFIC(data = data,
baseline = baseline,
feat1 = feat1,
feat2 = feat2,
num_iterations = 10,
output_dir = 'out')
# ----------------------------------------------------
# PACIFIC step 1:
# preprocessing ... 0.06038189 secs
# iteration 1/10 : 11 features selected by Elastic net. 0.227638 secs
# iteration 2/10 : 8 features selected by Elastic net. 0.2311988 secs
# iteration 3/10 : 1 features selected by Elastic net. 0.2655289 secs
# iteration 4/10 : 10 features selected by Elastic net. 0.2355261 secs
# iteration 5/10 : 13 features selected by Elastic net. 0.2278142 secs
# iteration 6/10 : 16 features selected by Elastic net. 0.2753899 secs
# iteration 7/10 : 14 features selected by Elastic net. 0.2690251 secs
# iteration 8/10 : 13 features selected by Elastic net. 0.231288 secs
# iteration 9/10 : 4 features selected by Elastic net. 0.2364419 secs
# iteration 10/10 : 12 features selected by Elastic net. 0.2437558 secs
# elapsed time: 2.506445 secs
# ----------------------------------------------------
# PACIFIC step 2:
# elapsed time: 0.4059999 secs The returned results is saved in ".RDS" format in output_dir. View the top interactions (frequently selected in the iterations) in the $top_interactions facet of results:
results$top_interactions
# intr feat1_level feat2_level EN_score
#1: KMT2D*Monocytes mutated higher_than_median 100
#2: CDKN2A*T_cells_CD8 mutated lower_than_or_equal_to_median 80
#3: TP53*Monocytes mutated lower_than_or_equal_to_median 60
# intr_P intr_P_C1 intr_P_C2 intr_P_C3 feat1_P feat2_P
#1: 0.0004387153 0.0002226526 0.001773958 0.001535063 0.3606967 0.07902884
#2: 0.0659434636 0.1744545811 0.096514456 0.295044751 0.2137792 0.36907611
#3: 0.0103065474 0.0337056835 0.047693554 0.121954591 0.1398141 0.07902884
# intr_logHR intr_logHR_lower95 intr_logHR_upper95 feat1_logHR
#1: 0.8232534 0.40706933 1.23943739 0.1575879
#2: 0.4981860 -0.00184557 0.99821767 0.2442130
#3: -0.3784980 -0.67170399 -0.08529202 -0.2483703
# feat1_logHR_lower95 feat1_logHR_upper95 feat2_logHR feat2_logHR_lower95
#1: -0.1754631 0.49063898 0.2555551 -0.03018218
#2: -0.1311699 0.61959601 0.1291237 -0.15292406
#3: -0.5713498 0.07460914 -0.2555551 -0.54129234
# feat2_logHR_upper95
#1: 0.54129234
#2: 0.41117153
#3: 0.03018218The KM plots for the top interactions are stored in a named list in the $km_plot_list facet of results. View the KM plots for the "KMT2D*Monocytes" interaction:
plot(results$km_plot_list[['KMT2D*Monocytes']])
- We recommend highlighting only the interactions for which all of the following ANOVA tests are significant after FDR correction (available in
results$top_interactions):intr_P: interaction vs. baselineintr_P_C1: interaction vs. baseline + feat1intr_P_C2: interaction vs. baseline + feat2intr_P_C3: interaction vs. baseline + feat1 + feat2
- By default, each feature in
feat1andfeat2is treated according to the following rules:- If it is categorical, it must be
factorwith two levels. In this case, the second level is considered as the "active" state. In other words, only the second level may contribute to any potential interaction. - If it is continuous (
numeric), it is converted into high or low levels (within each iteration of feature selection) based on> medianor<= median, respectively. In this case, both high and low levels may contribute to any potential interaction.
- If it is categorical, it must be
- Usually more than 1000 iterations are needed for stable results (depending on the complexity of the input data).