thebench.bio #1 Ecosystem for the Bio-Tech Industry
All RNA-seq questions

Question 12 of 20 Intermediate Concept

How does lightweight mapping differ from genome alignment for RNA-seq quantification?

Short answer

Tools such as Salmon and kallisto work out which transcripts each read is compatible with, without writing a base-by-base genome alignment, then estimate transcript abundances statistically. They are fast and handle isoform ambiguity well, but they only see transcripts in their index.

Why

A genome aligner places each read at a position in the genome, including across splice junctions, and writes a large alignment file. Counting then happens as a separate step. Lightweight methods skip most of that. kallisto's pseudoalignment and Salmon's mapping find the set of transcripts a fragment could have come from.

A statistical model then shares out reads that fit several isoforms. Salmon adds sample-specific models for sequence, position and fragment GC content bias, and its authors report it was the first transcriptome-wide quantifier to correct for fragment GC bias. The speed is real: the Salmon paper reports about 600 million paired-end reads quantified in roughly 23 minutes on 30 threads.

Since version 1.0.0, Salmon's default is selective alignment, which scores candidate mappings, and its documentation recommends a decoy-aware index so reads from unannotated genomic regions that resemble a transcript are not forced onto it. The trade-off is scope. A transcriptome index cannot report novel splice junctions, intronic reads, structural variants or variant calls.

When those matter, a spliced genome alignment is still the right tool.

A strong answer covers

  • Compatibility with transcripts replaces full alignment to the genome
  • A statistical model distributes reads shared between isoforms
  • Bias models such as fragment GC correction improve accuracy
  • Decoy-aware indexes reduce spurious assignment from unannotated loci
  • Genome alignment is still needed for novel junctions and variants

Common mistakes

  • Believing lightweight mapping is less accurate simply because it is faster
  • Using a transcriptome-only index and expecting to discover new transcripts
  • Feeding Salmon TPM directly into a count-based differential expression tool