The problem. The classic RNA-seq route — align every read to the genome with STAR, then count — is accurate but heavy: full alignment is slow and produces information you throw away if all you want is transcript abundance. Do you actually need to know where each read maps, or just which transcript it came from?

The idea. kallisto introduces pseudoalignment: using a k-mer index of the transcriptome, it determines the set of transcripts compatible with each read without computing a base-level alignment. An expectation-maximisation step then resolves reads shared across isoforms into abundance estimates. The result is transcript quantification orders of magnitude faster than alignment, with comparable accuracy.

Why it matters. This is a genuine fork in the road for aml_rnaseq_nf: STAR-then-count versus pseudoalignment (kallisto or Salmon, next). Reading it clarifies the trade — you gain enormous speed and lose read-level positional information, which matters if you also want variants or novel junctions but not if you only want expression. Knowing when each is appropriate is real pipeline-design knowledge.

Verdict. Foundational — it reframed what RNA-seq quantification requires. Salmon (coming up) added bias modelling in a similar spirit. Read it for the core insight: for counting, compatibility is enough; you don’t need the alignment.