Skip to contents

This page goes from a folder of RSEM files to values ready for plots, using only functions of rsemverse 0.3.0. The folder is assumed to hold one *.genes.results file per sample (and, for the last section, the *.isoforms.results files), with a sample table that names the groups:

rsem/                     metadata.csv
  CTRL_1.genes.results      sample_id,ref,group
  CTRL_2.genes.results      CTRL_1,CTRL_1.genes.results,control
  DRUG_1.genes.results      CTRL_2,CTRL_2.genes.results,control
  DRUG_2.genes.results      DRUG_1,DRUG_1.genes.results,drug
                            DRUG_2,DRUG_2.genes.results,drug

1. Read

library(rsemverse)
tpm <- rsem_read("rsem", metadata = NULL, species = "human", expr = "tpm")

metadata = NULL finds the files and prints which file became which sample. With a sample table, rsem_read("rsem", "metadata.csv", ...) reads the files it lists instead.

2. Check the effective lengths before tximport or DESeq2

len <- rsem_read("rsem", metadata = NULL, species = "human", expr = "length")
attr(len, "zero_length")           # genes with length 0, per sample
lp <- rsem_length_profile(len)
subset(lp, zero_fraction > 0 & zero_fraction < 1)

Genes with length 0 in some samples are the ones that stop DESeqDataSetFromTximport(). Removing them, or importing the transcript-level files, avoids changing their length offset.

3. Choose a detection threshold

rsem_detection_curve(tpm, "tpm", thresholds = c(0.5, 1, 2, 5), min_samples = c(1, 0.5))
filtered <- rsem_filter(tpm, "tpm", min_value = 1, min_samples = 0.5)

Each row of the curve is the number of genes rsem_filter() keeps with those settings.

4. Remove rRNA and see the gene-type composition

gtf <- "gencode.v29.annotation.gtf.gz"   # the annotation the RSEM reference was built from
rsem_biotype_profile(tpm, gtf, types = c("protein_coding", "lncRNA", "rRNA", "Mt_rRNA"))
filtered <- rsem_subset(filtered, remove = c("rrna", "mito_rrna"), annotation = gtf)

5. Z-scores against the controls

ref <- rsem_fit_reference(filtered, "control", metadata = "metadata.csv", group = "group")
z <- rsem_apply_reference(filtered, ref, zero_sd = "drop")

ref keeps the control means and standard deviations on the log2(TPM + 1) scale. A later batch on the same scale is scored against the same controls with rsem_apply_reference(new_tpm, ref).

6. Plot or share

long <- rsem_long(z, metadata = "metadata.csv")
ggplot2::ggplot(subset(long, gene == "CDKN1A"), ggplot2::aes(group, value)) +
  ggplot2::geom_point() + ggplot2::labs(y = "z-score against controls")

rsem_write(tpm, "tpm.tsv.gz", annotation = gtf)   # gene, gene IDs, gene type, samples

7. Transcripts

iso <- rsem_read("rsem", metadata = NULL, species = "human", expr = "tpm", level = "isoform")
main <- rsem_select_isoforms(iso, min_tpm = 1, min_fraction = 0.1, rule = "all")

The selected values are the original estimates; to quantify against the selection, build a new reference and rerun RSEM.