Alignment vs Pseudo-alignment

Once your reads pass QC, the next decision is how to map them to a reference. Two families dominate RNA-Seq: spliced alignment to a genome, and pseudo-alignment to a transcriptome. They answer different questions, so the right choice depends on what you plan to do downstream.

Spliced genome alignment

HISAT2 and STAR are spliced aligners: they place each read at genomic coordinates and split reads that span an intron across the two exons they belong to. This resolves splice junctions and gives you base-level positions you can pile up, view in IGV, or hand to a variant caller. Both work from the genome FASTA, and supplying a GTF annotation lets known junctions guide the alignment for better accuracy.

# Build a splice-aware index from genome FASTA + GTF
STAR --runMode genomeGenerate \
     --genomeDir star_index \
     --genomeFastaFiles GRCh38.primary_assembly.genome.fa \
     --sjdbGTFfile gencode.v44.annotation.gtf \
     --sjdbOverhang 99 \
     --runThreadN 8

# Align paired-end reads, emit a coordinate-sorted BAM
STAR --genomeDir star_index \
     --readFilesIn sample_R1.fastq.gz sample_R2.fastq.gz \
     --readFilesCommand zcat \
     --outSAMtype BAM SortedByCoordinate \
     --runThreadN 8 \
     --outFileNamePrefix sample_

# sample_Log.final.out (excerpt)
                        Number of input reads |	28451003
                    Average input read length |	200
                                UNIQUE READS:
                 Uniquely mapped reads number |	25783461
                      Uniquely mapped reads % |	90.62%
           % of reads mapped to multiple loci |	6.11%
                % of reads unmapped: too short |	2.84%

Transcriptome pseudo-alignment

Salmon and kallisto skip base-level alignment entirely. They match the k-mers in each read against a precomputed index of transcript sequences, assign the read to the set of transcripts it is compatible with, then resolve multimappers with an expectation-maximization model. kallisto calls this pseudoalignment, and Salmon adds a lightweight selective-alignment scoring step. Neither needs a GTF or a spliced alignment to the genome: the core input is a transcriptome FASTA. Salmon's recommended decoy-aware index does fold in the genome too, but only as decoy sequence that soaks up reads of genomic origin so they are not forced onto a transcript.

# Build a decoy-aware index: transcripts + genome as decoy
grep "^>" GRCh38.primary_assembly.genome.fa | cut -d " " -f 1 | sed 's/>//g' > decoys.txt
cat gencode.v44.transcripts.fa GRCh38.primary_assembly.genome.fa > gentrome.fa
salmon index -t gentrome.fa -d decoys.txt -i salmon_index -k 31 -p 8

# Quantify against the transcriptome (-l A auto-detects library type)
salmon quant -i salmon_index -l A \
     -1 sample_R1.fastq.gz -2 sample_R2.fastq.gz \
     --gcBias -p 8 -o sample_quant

[2024-01-18 11:42:07.331] [jointLog] [info] Automatically detected most likely library type as ISR
[2024-01-18 11:47:52.008] [jointLog] [info] Mapping rate = 93.4127%

$ head -2 sample_quant/quant.sf
Name	Length	EffectiveLength	TPM	NumReads
ENST00000456328.2	1657	1483.421	2.87451	38.142

Accuracy, speed, disk

Spliced alignment is slower and heavier: a human STAR index needs roughly 30 GB of RAM to build and each sorted BAM runs to several gigabytes, but you gain genomic coordinates for variant calling, novel-isoform and junction discovery, and coverage QC. Pseudo-alignment finishes in minutes on a laptop, writes a tiny quant.sf, and gives accurate transcript-level counts, but it cannot detect anything missing from the transcriptome and is only as complete as your annotation. Reach for STAR or HISAT2 when you need base-level positions or work with a poorly annotated organism, and reach for Salmon or kallisto when the goal is transcript- or gene-level quantification for differential expression that you will import with tximport into DESeq2 or edgeR.