Transcript-to-Gene Summarisation

Aggregating transcript-level estimates to gene level for standard differential expression

Introduction

Pseudo-alignment tools (Salmon, Kallisto) and EM-based quantifiers (RSEM) produce transcript-level (isoform-level) counts. Most differential-expression workflows operate at the gene level, so transcript counts must be aggregated to genes via a transcript-to-gene mapping derived from the annotation. The tximport package handles this aggregation cleanly and produces output ready for DESeq2, edgeR, or limma-voom; the related tximeta extends this with automatic metadata and reference-genome tracking.

Prerequisites

A working understanding of transcript-level vs gene-level expression quantification, RNA-seq annotation files (GTF), and the difference between counts, TPM, and effective length.

Theory

The simplest transcript-to-gene aggregation sums transcript counts within each gene: \(\mathrm{gene\ count} = \sum_{\mathrm{transcripts \in gene}} \mathrm{transcript\ count}\). tximport adds two important refinements:

• Length-scaled TPM (countsFromAbundance = "lengthScaledTPM"): combines transcript abundance and length information to produce gene-level counts that respect both isoform-resolution and gene-level interpretation.
• Effective-length offsets: passes effective-length information to downstream DE tools (DESeq2 with tximeta::tximeta() integration) for accurate normalisation.

Direct transcript-level DE (without aggregation) is supported by tools like swish (fishpond) and DRIMSeq for differential transcript usage; aggregation is the right choice when the question is about gene-level expression.

Assumptions

The transcript-to-gene map matches the annotation used for transcript quantification; sample-file ordering is consistent with metadata.

R Implementation

library(tximport); library(tximeta)

# Example: salmon outputs
# files <- list.files("salmon_out", pattern = "quant.sf", recursive = TRUE, full.names = TRUE)
# tx2gene <- read.table("tx2gene.tsv", header = TRUE)
#
# txi <- tximport(files, type = "salmon", tx2gene = tx2gene,
#                 countsFromAbundance = "lengthScaledTPM")
#
# txi$counts: gene x sample matrix

Output & Results

tximport() returns a list with counts (gene-by-sample matrix), abundance (gene-level TPM), and length (effective lengths used for downstream normalisation). The matrix is ready as input to DESeq2::DESeqDataSetFromTximport() or as a starting point for edgeR / limma workflows.

Interpretation

A reporting sentence: “tximport aggregated 190{,}000 transcript-level Salmon estimates to 22{,}000 gene-level counts using a GENCODE v44 transcript-to-gene map; countsFromAbundance = 'lengthScaledTPM' produced gene counts that DESeq2 then analysed with effective-length offsets passed via the txi$length matrix.” Always document the aggregation method and tx-to-gene source.

Practical Tips

• countsFromAbundance = "lengthScaledTPM" is the standard for DE analysis; it scales raw counts by transcript length while preserving the count interpretation.
• Use tximeta instead of raw tximport when possible — it automatically retrieves the transcript-to-gene map from the Salmon index’s metadata, eliminating annotation-mismatch errors.
• For differential transcript usage (DTU), use DRIMSeq or DEXSeq on the transcript-level matrix; aggregation collapses isoform-level information.
• For differential transcript expression (DTE), use swish (fishpond::swish) on the transcript-level matrix with bootstrap inferential replicates from Salmon.
• Verify that sample order in files matches the rows of your metadata table; off-by-one errors here corrupt the entire downstream analysis silently.
• For very large studies, tximport reads files sequentially; parallelise with BiocParallel for substantial speedups.

R Packages Used

tximport for the canonical aggregation; tximeta for metadata-aware aggregation that handles tx-to-gene mapping automatically; fishpond for swish-based DTE with bootstrap replicates; DRIMSeq and DEXSeq for differential transcript usage at the transcript level.