In NCBI-Hackathons/SuMu:

Here we are demonstrating the feasibility of analyzing genomic data using Stan. The first use case is to analyze somatic mutations for association with survival, after adjusting for key clinical variables with known prognostic status.

knitr::opts_chunk$set(fig.width=8, fig.height=6, fig.path='Figs/',
                      echo=TRUE, warning=FALSE, message=FALSE)

# library(SuMu)
devtools::load_all('.')
library(dplyr)
library(survminer)
library(survival)
library(ggplot2)
library(scales)
library(tidyr)
library(rstanarm)
options(mc.cores = 4)

Data Exploration

Clinical Data

First, download the clinical data. Here we are using the TCGA skin cutaneous melanoma (SKCM) cohort.

clin_df <- SuMu::get_tcga_clinical(cohort = "SKCM")

## format some clinical data variables

clin_df2 <- clin_df %>% 
  dplyr::mutate(stage_part1 = gsub(pathologic_stage,
                                   pattern = '(Stage [0I]+).*',
                                   replacement = '\\1'),
                diagnosis_year_group = cut(year_of_initial_pathologic_diagnosis,
                                           breaks = c(1975, 1990, 1995, 2000,
                                                      2005, 2010, 2015, 2020),
                                           include.lowest = TRUE),
                os_10y = ifelse(OS_IND == 1 & OS <= 10*365.25, 1, 0),
                sample = sampleID
                )

Review clinical data

For this analysis we will consider the survival time in r print(unique(clin_df['OS_UNIT'])) since initial pathologic diagnosis.

fit <- survfit(Surv(OS, OS_IND) ~ 1,
                  data = clin_df2)
survminer::ggsurvplot(fit) + 
  ggtitle('Survival since diagnosis in full cohort')

Plotting by stage, although the time of 'stage' determination may be confounded if not collected at time of initial diagnosis.

fit <- survfit(Surv(OS, OS_IND) ~ pathologic_stage,
                  data = clin_df2)
survminer::ggsurvplot(fit, legend = "right")

There also seem to be differences by tumor type.

fit <- survfit(Surv(OS, OS_IND) ~ sample_type,
                  data = clin_df2)
survminer::ggsurvplot(fit, legend = "right")

(Aside: I wonder how similar tumor type is to sample type? For example, we could have a metastatic patient where the sample was obtained from the primary tumor. We will want to adjust our genetic data analysis for the sample type but may want to estimate prognosis according to the tumor type?)

A variable like year_of_initial_pathologic_diagnosis is guaranteed to be unconfounded since we can safely assume it was collected at the time of diagnosis.

fit <- survfit(Surv(OS, OS_IND) ~ diagnosis_year_group,
                  data = clin_df2)
survminer::ggsurvplot(fit, legend = 'right')

This makes it pretty clear that we have a strong "survival" bias to our data. This would suggest that, among people whose diagnosis was made in the 90s, only those who survived long enough to be enrolled were included in the study.

Let's look at a histogram of years of initial diagnosis.

ggplot(clin_df2, aes(x = year_of_initial_pathologic_diagnosis,
                     fill = diagnosis_year_group)) + 
  geom_histogram() + 
  theme_minimal()

Let's look at the time since initial diagnosis (presumably, the time from enrollment to diagnosis).

Finally, we can visualize a more comprehesive set of clinical variables.

fit <- survival::coxph(Surv(OS, OS_IND) ~
                         age_at_initial_pathologic_diagnosis + 
                         sample_type +
                         breslow_depth_value + initial_weight + 
                         strata(year_of_initial_pathologic_diagnosis),
               data = clin_df2)
print(fit)

Somatic Mutations Data

We can download the somatic mutations to supplement the phenotypes.

mut_df <- SuMu::get_tcga_somatic_mutations(cohort = "SKCM") %>%
  dplyr::mutate(gene_aa = paste0(gene, ".", Amino_Acid_Change),
                gene_effect = paste0(gene, ".", effect)
                )

Check the most frequent mutations.

mut_df_missense = mut_df %>% dplyr::filter(effect == "Missense_Mutation")
mut_df_missense$gene_aa = paste0(mut_df_missense$gene, ":", mut_df_missense$Amino_Acid_Change)
mut_df_missense %>% select(gene_aa) %>% table %>% sort %>% rev %>% as.data.frame %>% head(10)

Filter to top genes

top_genes <- mut_df %>%
  dplyr::group_by(gene) %>%
  dplyr::mutate(gene_count = n()) %>%
  dplyr::ungroup() %>%
  dplyr::distinct(gene, .keep_all = TRUE) %>%
  dplyr::top_n(gene_count, n = 10) %>%
  dplyr::select(gene)

mut_df_topgenes <- mut_df %>%
  dplyr::semi_join(top_genes)

GLM model to all genes

Prepare mutation data for analysis

clin_df2_nonmiss <- clin_df2 %>% 
  dplyr::mutate(
    revised_breslow_depth = ifelse(is.na(breslow_depth_value),
                                   0, breslow_depth_value)) %>%
  tidyr::drop_na(os_10y,
                 age_at_initial_pathologic_diagnosis,
                 initial_weight, 
                 revised_breslow_depth,
                 sample_type,
                 diagnosis_year_group)

mutation_matrix <- SuMu:::prep_biomarker_data(
  biomarker_data = mut_df_topgenes,
  data = clin_df2_nonmiss,
  biomarker_formula = 1 ~ gene_effect,
  .fun = sum,
  id = 'sample'
)

glm_df2 <- mutation_matrix %>%
  dplyr::left_join(clin_df2_nonmiss %>% 
                     dplyr::select(sample, os_10y),
                   by = 'sample')

Fit stan-glm model to these genetic data

# construct input formula
gene_names2 <- mutation_matrix %>% 
  head(1) %>% 
  dplyr::select(-sample) %>% 
  names()

gene_subformula <- stringr::str_c('`',
                 stringr::str_c(gene_names2,
                                collapse = '` + `'),
                 '`')

my_formula2 <- stringr::str_c('os_10y ~ ', gene_subformula)

# call to `stan_glm`
glmfit2 <- rstanarm::stan_glm(
  data = glm_df2,
  formula = my_formula2,
  sparse = TRUE,
  family = binomial(),
  chains = 4,
  prior = rstanarm::hs_plus()
)

GLM(er) model including clinical data only

rescale <- function(x) {
  (x - mean(x, na.rm=T))/(2*sd(x, na.rm=T))
}

clin_df3 <- clin_df2 %>%
  dplyr::mutate(
    rescale_age_at_initial_pathologic_diagnosis =
       rescale(age_at_initial_pathologic_diagnosis), 
    rescale_initial_weight = 
      rescale(initial_weight),
    rescale_breslow_depth_value = rescale(breslow_depth_value)
  ) 


glmfit_clin <- rstanarm::stan_glmer(
  os_10y ~
    rescale_age_at_initial_pathologic_diagnosis + 
    sample_type +
    rescale_breslow_depth_value + 
    rescale_initial_weight + 
    (
      rescale_age_at_initial_pathologic_diagnosis + 
        sample_type +
        rescale_breslow_depth_value + 
        rescale_initial_weight
      | diagnosis_year_group
      ),
  data = clin_df3,
  init_r = 1,
  family = binomial()
)

print(glmfit_clin)

GLM(er) model with clinical + genetic data

glm_df3 <- clin_df3 %>%
  dplyr::inner_join(mutation_matrix,
                   by = 'sample')

# construct input formula
clinical_formula <- os_10y ~
    rescale_age_at_initial_pathologic_diagnosis + 
    sample_type +
    rescale_breslow_depth_value + 
    rescale_initial_weight + 
    `__BIOMARKERS__` +
    (
      rescale_age_at_initial_pathologic_diagnosis + 
        sample_type +
        rescale_breslow_depth_value + 
        rescale_initial_weight + 
        `__BIOMARKERS__`
      | diagnosis_year_group
      )

gene_subformula <- stringr::str_c('`',
                 stringr::str_c(gene_names2,
                                collapse = '` + `'),
                 '`')

my_formula3 <- stringr::str_c(
  as.character(clinical_formula)[2],
  as.character(clinical_formula)[3],
  sep = as.character(clinical_formula)[1])
my_formula3 <- as.formula(gsub(my_formula3, 
                               pattern = '`__BIOMARKERS__`',
                               replacement = gene_subformula))

update(clinical_formula,
  stringr::str_c('~ . ',
                 gene_subformula,
                 stringr::str_c('(', gene_subformula, '| diagnosis_year_group)'),
                 sep = '+')
  )


# call to `stan_glm`
glmfit_clingen <- rstanarm::stan_glmer(
  data = glm_df3,
  formula = my_formula3,
  sparse = TRUE,
  family = binomial(),
  chains = 4,
  prior = rstanarm::hs_plus()
)

summarize output

Visualize

summary_table=feature_table(glmfit_clingen)
feature_graph(glmfit_clingen)
view_feature(mutation_matrix,clin_df2,gsub("`","",rownames(summary_table)[1]))
view_feature(mutation_matrix,clin_df2,gsub("`","",rownames(summary_table)[nrow(summary_table)]))

AUC

auc(clin_df3 %>% 
      dplyr::select(os_10y,
                    rescale_age_at_initial_pathologic_diagnosis,
                    sample_type,
                    rescale_breslow_depth_value, 
                    rescale_initial_weight,
                    diagnosis_year_group) %>%
      tidyr::drop_na(),
    "os_10y", glmfit_clin, h_gram = T, roc_plot = T)

Future directions

Our function syntax will look like the following.

fit <- fit_glm(
  data = clin_df,
  formula = os_10y ~ rescale_.. + `__BIO__`,
  biomarker_data = mut_df,
  biomarker_formula = 1 ~ (1|gene_aa) + (1|gene) + (1|effect),
  id = 'sample'
)

It currently supports the following

fit <- fit_glm(
  data = clin_df,
  formula = os_10y ~ age + weight + tumor_type,
  biomarker_data = mut_df,
  biomarker_formula = DNA_VAF ~ gene_aa,
  id = 'sample'
)

NCBI-Hackathons/SuMu documentation built on May 31, 2019, 8:02 a.m.