STAR vs HISAT2: RNA-Seq Aligner Comparison

Every bulk RNA-Seq pipeline needs a spliced aligner, and for the last decade the choice has boiled down to two: STAR (Dobin et al., 2013) and HISAT2 (Kim et al., 2019). Both are actively maintained, both are used in flagship publications, and both handle short-read paired-end data well. So which do you pick?

This article is a practical decision guide, not a re-benchmark. We assume you have short paired-end Illumina reads and a decent reference. We’ll compare on the axes that actually matter in a real project: accuracy, memory, speed, splice-junction discovery, downstream compatibility, and reproducibility.

Short answer

• Use STAR if you have 32 GB+ RAM per job, care about novel splice-junction discovery, or are building a fusion / splice-QTL pipeline.
• Use HISAT2 if you’re on shared HPC with per-user RAM caps, aligning bacterial / non-model species with tight compute, or running a huge cohort where indexing time dominates.
• Either is fine for standard human/mouse bulk RNA-Seq → featureCounts → DESeq2. Pick and stick.

1. Algorithm at a glance

Aspect	STAR	HISAT2
Core algorithm	Suffix array + seed extension	Hierarchical FM-index (global + local)
Splice-aware	Yes (two-pass optional)	Yes (single-pass)
Multi-threaded index	Yes	Yes
Multi-mapper resolution	Custom MAPQ + primary flag	Similar
Output	SAM/BAM, wig, tab-separated splice junctions	SAM/BAM, splice sites file

STAR uses uncompressed suffix arrays for fast seed lookup — that’s where the RAM cost comes from. HISAT2 uses a graph-based FM-index with a two-level hierarchy that lets it fit human in about 8 GB.

2. Building the index

STAR:

STAR --runMode genomeGenerate \
  --genomeDir ref/star_index \
  --genomeFastaFiles ref/GRCh38.primary_assembly.genome.fa \
  --sjdbGTFfile ref/gencode.v46.annotation.gtf \
  --sjdbOverhang 100 \
  --runThreadN 16

Index size: ~30 GB on disk, ~30 GB RAM at build.

HISAT2:

hisat2_extract_splice_sites.py ref/gencode.v46.annotation.gtf > ref/splice_sites.txt
hisat2_extract_exons.py       ref/gencode.v46.annotation.gtf > ref/exons.txt
hisat2-build -p 16 \
  --ss ref/splice_sites.txt \
  --exon ref/exons.txt \
  ref/GRCh38.primary_assembly.genome.fa \
  ref/hisat2_index/grch38

Index size: ~8 GB on disk, ~200 GB RAM at build (yes, really — build once on a fat node).

3. Alignment

STAR (recommended flags for typical bulk RNA-Seq):

STAR --runMode alignReads \
  --genomeDir ref/star_index \
  --sjdbGTFfile ref/gencode.v46.annotation.gtf \
  --readFilesIn sample_R1.fastq.gz sample_R2.fastq.gz \
  --readFilesCommand zcat \
  --outSAMtype BAM SortedByCoordinate \
  --twopassMode Basic \
  --outSAMstrandField intronMotif \
  --outFileNamePrefix bam/sample_ \
  --runThreadN 16

HISAT2:

hisat2 -x ref/hisat2_index/grch38 \
  -1 sample_R1.fastq.gz -2 sample_R2.fastq.gz \
  --dta --rna-strandness RF \
  -p 16 -S sample.sam
samtools sort -@ 8 -o bam/sample.bam sample.sam
samtools index bam/sample.bam

The --dta flag emits XS attributes for downstream StringTie transcript assembly. Add --rna-strandness RF for standard TruSeq stranded libraries; drop it for unstranded.

4. What benchmarks say

Multiple community benchmarks — Baruzzo et al. (2017), the SEQC2 consortium, and follow-up work — agree on the same broad picture:

• STAR has a slight edge in splice-junction sensitivity and novel-junction discovery, especially with two-pass mode.
• HISAT2 has a slight edge in speed per CPU-hour and a dramatic edge in RAM footprint.
• Gene-level counts from featureCounts on the two aligners’ BAMs correlate above r = 0.99. Downstream DE calls with DESeq2 or edgeR are essentially identical.
• Both handle circRNA back-splice junctions poorly compared to specialized tools like CIRCexplorer2 or CIRI2.

If your goal is a differential expression table, the choice of aligner does not meaningfully change the biological conclusions of your study.

5. Downstream compatibility

Both produce standard sorted BAMs, so featureCounts, RSeQC, Salmon-in-alignment-mode, and Qualimap all work identically.

Where compatibility matters:

• StringTie works with either, but expects --dta from HISAT2 or --outSAMstrandField intronMotif from STAR.
• rMATS and MISO for alternative splicing prefer STAR’s splice-junction files.
• Arriba and STAR-Fusion for gene-fusion detection require STAR output specifically.

6. Reproducibility gotchas

• Multi-mappers. STAR’s default caps at 10 secondary alignments per read; HISAT2 defaults to primary-only. If you count multi-mappers with featureCounts (-M --fraction), the two aligners will diverge.
• Soft-clipping. STAR clips more aggressively in two-pass mode. For splice-QTL calling, this improves junction accuracy but reduces read length uniformity.
• Version drift. Both projects push meaningful bug fixes yearly. Pin the version in your conda environment (star=2.7.11b, hisat2=2.2.1).

7. Decision matrix

Scenario	Recommendation
Standard human/mouse bulk RNA-Seq, plenty of RAM	STAR
Shared HPC with strict per-job RAM caps	HISAT2
Fusion detection, splice-QTL, novel junctions	STAR (two-pass)
Non-model organism, small budget	HISAT2
Very large cohort (500+ samples), throughput-limited	HISAT2 (or Salmon pseudo-alignment)
Publishing in a journal that requires the “standard” pipeline	STAR
You need transcript-level TPMs, not gene counts	Use Salmon or Kallisto pseudo-alignment instead

8. When neither is the right tool

If your primary output is transcript-level abundance or you’re dealing with 1000+ samples, skip alignment entirely and use pseudo-alignment (Salmon or Kallisto). It’s 10-50× faster and gives you TPMs directly. See our Kallisto vs Salmon comparison.

For long-read RNA-Seq (PacBio Iso-Seq, Nanopore direct-RNA), forget both — use minimap2 with -ax splice or -ax splice:hq.

Bottom line

The STAR vs HISAT2 debate is much older than it is important. Both are excellent, both are well-maintained, both give you the same DE genes on standard bulk RNA-Seq. Optimize for your compute constraints and downstream tools, pin the version in your workflow, and stop worrying.

Related reading: our end-to-end RNA-Seq guide uses STAR by default; swap in HISAT2 with the commands above and the rest of the pipeline is identical.