Analysis with R

Differential gene expression associated with HPV integration

In this example, we will study the potential effects of HPV integration on the expression of recurrent target genes in CESC and HNSC tumors. This example demonstrates using R to issue BigQuery queries involving multiple tables across multiple data sets. We will also show users how to bring in their own data to use in conjunction with the TCGA data already available as BigQuery tables. In this exercise, we will reproduce some figures from Tang et. al. [1] to visualize altered expression of host genes frequently targeted by HPV.

References: Tang et. al. The landscape of viral expression and host gene fusion and adaptation in human cancer. Nature Communications 4, Article number:2513|doi:10.1038/ncomms3513

Loading libraries

require(bigrquery,quietly = TRUE) || install.packages('bigrquery',verbose = FALSE)
require(httpuv, quietly = TRUE) || install.packages('httpuv',verbose=FALSE)
require(ggplot2,quietly = TRUE) || install.packages('ggplot2',verbose = FALSE)

Your project ID

You will be using your own project ID. At certain points in the code, it will be necessary to complete the code.

main_cloud_project = "isb-cgc"
my_cloud_project   = "your_project_id"
tcga_data_set      = "TCGA_hg38_data_v0"

First query

Now let’s see if things are working.

bigrquery::list_tables(main_cloud_project, tcga_data_set)

In this tutorial, we will be investigating two studies using two existing BigQuery tables. Additionally, we’re going to BYOD “Bring your own data”.

study=c('TCGA-CESC','TCGA-HNSC')

clinical_table = "`isb-cgc.TCGA_bioclin_v0.Clinical`"

Constructing Queries

One of the great things about working in a scripting environment, is that our analysis – the queries – we write, can be constructed programmatically. That makes it easy to apply the same structured queries to many questions. In the next code block is an example of how to do that.

sqlQuery = paste("SELECT case_barcode, project_short_name AS project, hpv_calls, hpv_status ",
                 "FROM ", clinical_table,
                 " WHERE project_short_name in (",paste(shQuote(study),collapse = ','),")",sep="")

sqlQuery

hpv_table = query_exec(sqlQuery,project = my_cloud_project)

dim(hpv_table)

head(hpv_table)

# We can do some quality control ...
# Assert that if hpv_calls is NA, hpv_status is Negative
stopifnot((is.na(hpv_table$hpv_calls) && hpv_table$hpv_status=="Negative") || !is.na(hpv_table$hpv_calls))

# Let's explore the cohort
ggplot(data=hpv_table, aes(x=hpv_status, fill=project)) + geom_bar(stat="count", position=position_dodge())

Uploading Data

The exact location of HPV integration is not available in the TCGA data, so instead we’ll get a list of frequently targeted genes that was published with this paper:

Ka-Wei Tang et. al. The Landscape of viral expression and host gene fusion and adaptation in human cancer. doi:10.1038/ncomms3513

(Supplementary Data 2: Integration analysis results)

We will access the data from our workshop bucket using the command line or from the Google Cloud Console. Using the cloud console, go to https://console.cloud.google.com and find the workshop bucket.

Using the google command line tool:

gsutil cp gs://isb-cgc-workshop/data/Larsson/ncomms3513-s3.tsv .
gsutil cp gs://isb-cgc-workshop/data/Larsson/ncomms3513-s3_Schema.json .

Now the data is in our directory, but we need to transform it into a BQ table. To do that, we need to create a data set in our project. We can do this from within the BigQuery web UI by clicking on the little blue triangle next to your project ID on the left. Or we can do this on the command line using the bq command line tool.

gcloud init

bq help

bq ls

bq mk workspace

bq load --source_format CSV --field_delimiter "\t"  --schema ncomms3513-s3_Schema.json workspace.ncomms3513_s3 ncomms3513-s3.tsv

Integrating with the expression data

Now we can directly query our own data, and start to combine it with other tables. Let’s try it out!

This next query is going to select the genes that were associated with HPV integration in CESC and HNSC tumors.

sqlQuery = "
SELECT
  Overlapping_genes,
  Cancer
FROM
  `isb-cgc-04-0030.workspace.ncomms3513_s3`
WHERE
  Cancer IN ('TCGA-CESC','TCGA-HNSC')
  AND Overlapping_genes <> 'Intergenic'
GROUP BY
  Cancer,
  Overlapping_genes
  "

affected_genes = query_exec(sqlQuery,project = my_cloud_project)

head(affected_genes)

table(affected_genes$Cancer)

Next, with those offen affected genes, we will query gene expression data.

query <- "
SELECT
  project_short_name,
  gene_name,
  AVG(HTSeq__FPKM) as mean_expression
FROM
  `isb-cgc.TCGA_hg38_data_v0.RNAseq_Gene_Expression`
WHERE
  project_short_name IN ('TCGA-CESC','TCGA-HNSC')
  AND gene_name IN (
    SELECT
      Overlapping_genes AS gene_name
    FROM
      `isb-cgc-04-0030.workspace.ncomms3513_s3`
    WHERE
      Cancer IN ('TCGA-CESC','TCGA-HNSC')
      AND Overlapping_genes <> 'Intergenic'
    GROUP BY
      gene_name )
GROUP BY
  project_short_name,
  gene_name
ORDER BY
  mean_expression"

# running the query.
mean_affected_genes = query_exec(query, project = my_cloud_project)

# we'll create some more meaningful x-axis labels
mean_affected_genes$xlabel <- paste0(mean_affected_genes$project_short_name, "_", mean_affected_genes$gene_name)

# Now we can visualize it.
qplot(data=mean_affected_genes,
      x=factor(x = xlabel, ordered = T, levels = xlabel),
      y=mean_expression,
      col=project_short_name) +
      theme(axis.text.x = element_text(angle = 90, hjust = 1, size=4)) +
      xlab("Project_Gene")

Computing Statistics

Instead, if we want to get the actual gene expression values, we could query for that, and retrieve it as a data.frame.

sqlQuery = "
SELECT
  case_barcode,
  sample_barcode,
  project_short_name,
  gene_name,
  HTSeq__FPKM
FROM
  `isb-cgc.TCGA_hg38_data_v0.RNAseq_Gene_Expression`
WHERE
  project_short_name IN ('CESC','HNSC')
  AND gene_name IN (
  SELECT
    Overlapping_genes as gene_name
  FROM
    `isb-cgc-04-0030.workspace.ncomms3513_s3`
  WHERE
    Cancer IN ('TCGA-CESC','TCGA-HNSC')
    AND Overlapping_genes <> 'Intergenic'
  GROUP BY
    gene_name )
        "

gexp_affected_genes = query_exec(sqlQuery,project = my_cloud_project)

#view results
head(gexp_affected_genes)

# a couple different ways to look at the results
#qplot(data=gexp_affected_genes, x=project_short_name, y=HTSeq__FPKM, col=gene_name, geom="boxplot")
#qplot(data=gexp_affected_genes, x=project_short_name, y=log2(HTSeq__FPKM), col=gene_name, geom="boxplot")
qplot(data=gexp_affected_genes, x=log2(HTSeq__FPKM+1), col=gene_name, geom="density") + facet_wrap(~ project_short_name)

Not all the samples listed in the clinical data have gene expression data, however. Let’s filter the hpv_table to match the samples to those in gexp_affected_genes

require(tidyr,quietly = TRUE) || install.packages('tidyr',verbose = FALSE)
require(dplyr,quietly = TRUE) || install.packages('dplyr',verbose = FALSE)
require(broom,quietly = TRUE) || install.packages('broom',verbose = FALSE)

# let's get rid of 'indeterminate' samples
hpv_table = dplyr::filter(hpv_table, hpv_status != "Indeterminate", case_barcode %in% gexp_affected_genes$case_barcode)

T-tests

Now, we are going to perform t.tests on expression by hpv_status and study.

gxps <- merge(x=gexp_affected_genes, y=hpv_table, by=c("project_short_name","case_barcode"))

# Performing a t-test between hpv+ and hpv- by project_short_name and gene
res0 <- gxps %>%
group_by(project_short_name, gene_name) %>%
do(tidy(t.test(log2(HTSeq__FPKM+1) ~ hpv_status, data=.))) %>%
ungroup() %>%
arrange(desc(statistic))

# These are the top 5 results ...
top5 <- select(top_n(res0, 5, statistic), project_short_name, gene_name)

# Let's subset the data by the top 5 results...
res1 <- merge(x=top5, y=gxps) %>% mutate( Project_Gene = paste0(project_short_name, "_", gene_name))

# now we can plot the results...
ggplot(res1, aes(x=Project_Gene, y=log2(HTSeq__FPKM+1), fill=hpv_status)) + geom_boxplot()

Making BigQueries

Previously, we downloaded data and performed some work on it. But another way to work is to compute as much as possible in the cloud, and use R to visualize summary results.

Please see: https://cloud.google.com/bigquery/query-reference

sqlQuery = "
SELECT
  case_barcode,
  sample_barcode,
  project_short_name,
  gene_name,
  HTSeq__FPKM
FROM
  `isb-cgc.TCGA_hg38_data_v0.RNAseq_Gene_Expression`
WHERE
  project_short_name = 'TCGA-CESC'
  AND case_barcode IN (
  SELECT
    case_barcode
  FROM
    `isb-cgc.TCGA_bioclin_v0.Clinical`
  WHERE
    hpv_status = 'Positive' )
  AND gene_name IN (
  SELECT
    Overlapping_genes AS gene_name
  FROM
    `isb-cgc-04-0030.workspace.ncomms3513_s3`
  WHERE
    Cancer = 'TCGA-CESC'
    AND Overlapping_genes <> 'Intergenic'
  GROUP BY
    gene_name )
"

q1 = query_exec(sqlQuery,project = cloud_project_workshop)

dim(q1)

Now lets make a small change, and get gene expression for subjects that are hpv negative.

sqlQuery = "
SELECT
  case_barcode,
  sample_barcode,
  project_short_name,
  gene_name,
  HTSeq__FPKM
FROM
  `isb-cgc:TCGA_hg38_data_v0.RNAseq_Gene_Expression`
WHERE
  project_short_name = 'TCGA-CESC'
  AND case_barcode IN (
  SELECT
    case_barcode
  FROM
    `isb-cgc.TCGA_bioclin_v0.Clinical`
  WHERE
    hpv_status = 'Negative' )
  AND gene_name IN (
  SELECT
    Overlapping_genes AS gene_name
  FROM
    `isb-cgc-04-0030.workspace.ncomms3513_s3`
  WHERE
    Cancer = 'TCGA-CESC'
    AND Overlapping_genes <> 'Intergenic'
  GROUP BY
    gene_name )
"

q2 <- query_exec(sqlQuery,project = cloud_project_workshop)

dim(q2)

Now we merge the previous two queries, and compute T statistics using BigQuery built in functions, SQRT, MEAN, STDDEV, POW, COUNT, and LOG2.

Please see: https://cloud.google.com/bigquery/query-reference

q <- "
        WITH
          ## we use the 'WITH' keyword in order to create a few
          ## preliminary tables that we will use later on in the
          ## query -- this helps in writing 'modular' SQL that
          ## is easier to read
          --
          ## start by getting the list of genes of interest from
          ## the paper's table (will result in a list of 106 genes)
          geneList AS (
          SELECT
            Overlapping_genes AS gene_name
          FROM
            `isb-cgc-04-0030.workspace.ncomms3513_s3`
          WHERE
            Overlapping_genes <> 'Intergenic'
          GROUP BY
            gene_name ),
          ## next, get the identifiers (barcodes) for all HPV+ cases
          ## (will result in a list of 383 cases)
          posCaseList AS (
          SELECT
            case_barcode
          FROM
            `isb-cgc.TCGA_bioclin_v0.Clinical`
          WHERE
            hpv_status = 'Positive' ),
          ## now do the same for all HPV- cases
          ## (will result in a list of 664 cases)
          negCaseList AS (
          SELECT
            case_barcode
          FROM
            `isb-cgc.TCGA_bioclin_v0.Clinical`
          WHERE
            hpv_status = 'Negative' )
          ##
          ## Now we being our main SELECT statement, which actually
          ## wraps a pair of SELECTs that are JOINed together
          ##
        SELECT
          p.gene_name AS gene,
          p.project_short_name,
          p.x AS x,
          p.sx2 AS sx2,
          p.nx AS nx,
          o.y AS y,
          o.sy2 AS sy2,
          o.ny AS ny,
          (p.x-o.y) / SQRT((p.sx2/p.nx) + (o.sy2/o.ny)) AS T
        FROM (
            # first the gene expression summaries for hpv+ tumors
          SELECT
            project_short_name,
            gene_name,
            AVG(LOG(HTSeq__FPKM+1,2)) AS y,
            POW(STDDEV(LOG(HTSeq__FPKM+1,2)),2) AS sy2,
            COUNT(case_barcode) AS ny
          FROM
            `isb-cgc.TCGA_hg38_data_v0.RNAseq_Gene_Expression`
          WHERE
            project_short_name = 'TCGA-CESC'
            AND case_barcode IN (
            SELECT
              case_barcode
            FROM
              posCaseList )
            AND gene_name IN (
            SELECT
              gene_name
            FROM
              geneList )
          GROUP BY
            project_short_name,
            gene_name
          HAVING
            ny>0
            AND sy2>0) AS o
        JOIN (
            # Then we get the gene expression summaries from hpv-
          SELECT
            project_short_name,
            gene_name,
            AVG(LOG(HTSeq__FPKM+1,2)) AS x,
            POW(STDDEV(LOG(HTSeq__FPKM+1,2)),2) AS sx2,
            COUNT(case_barcode) AS nx
          FROM
            `isb-cgc.TCGA_hg38_data_v0.RNAseq_Gene_Expression`
          WHERE
            project_short_name = 'TCGA-CESC'
            AND case_barcode IN (
            SELECT
              case_barcode
            FROM
              negCaseList )
            AND gene_name IN (
            SELECT
              gene_name
            FROM
              geneList )
          GROUP BY
            project_short_name,
            gene_name
          HAVING
            nx>0
            AND sx2>0) AS p
        ON
          p.gene_name = o.gene_name
          AND p.project_short_name = o.project_short_name
        GROUP BY
          gene,
          project_short_name,
          x,
          sx2,
          nx,
          y,
          sy2,
          ny,
          T
        ORDER BY
          T DESC
 "

 t_test_result <- query_exec(q, project = cloud_project_workshop)

 head(t_test_result)


# and we can see the same results in the previously done work.
 res0

Extras

Transform gexp_affected_genes_df into a gexp-by-samples feature matrix

gexp_fm = tidyr::spread(gexp_affected_genes,gene_name,HTSeq__FPKM)

gexp_fm[1:5,1:5]

Have feedback or corrections? You can file an issue here or email us at feedback@isb-cgc.org.