henhaobio
Sign in

bioRxivBioinformaticsPreprint

STARsolo: accurate, fast and versatile mapping/quantification of single-cell and single-nucleus RNA-seq data

Benjamin KaminowiD, Dinar YunusoviD, Alexander DobiniD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

We present STARsolo, a comprehensive turnkey solution for quantifying gene expression in single-cell/nucleus RNA-seq data, built into RNA-seq aligner STAR. Using simulated data that closely resembles realistic scRNA-seq, we demonstrate that STARsolo is highly accurate and significantly outperforms pseudoalignment-to-transcriptome tools. STARsolo can replicate the results of, but is considerably faster than CellRanger, currently the most widely used tool for pre-processing scRNA-seq data. In addition to uniquely mapped reads, STARsolo takes account of multi-gene reads, necessary to detect certain classes of biologically important genes. It has a flexible cell barcode processing scheme, compatible with many established scRNA-seq protocols, and extendable to emerging technologies. STARsolo can quantify transcriptomic features beyond gene expression, which we illustrate by analyzing cell-type-specific alternative splicing in the Tabula Muris project.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (5 May 2021) · CC BY 4.0

Genome-aligning STARsolo matches simulated truth (R≈0.99) and CellRanger's counts almost exactly, running up to 4.4 times faster than CellRanger; pseudoaligners call hundreds of false-positive genes per cell.

Gene/cell count matrices feed every downstream scRNA-seq analysis, yet existing quantifiers either chain generic aligners through slow intermediate files, rely on lightweight alignment to the transcriptome (Alevin, Kallisto-bustools), or are proprietary and tied to 10X Genomics (CellRanger). STARsolo builds whole-genome alignment, read-to-gene assignment, barcode demultiplexing and UMI collapsing into STAR, with configurable barcode geometry and multi-gene read handling.

Key results

  1. ShownPseudoalignment-to-transcriptome tools (Kallisto, Alevin_sketch) call hundreds of spurious genes per cell, which the authors trace to intronic and intergenic reads; whole-genome STARsolo and Alevin with a full genome decoy stay close to truth.

    Simulation from 10X-pbmc-5k without multi-gene reads: gene/cell Spearman R≈0.99 for STARsolo (Alevin_full-decoy equal) vs ≈0.77 Kallisto and 0.8 Alevin_sketch; false-positive genes per cell ~5 vs ~600 and ~700 (≈21% and 24% of expressed genes). Removing non-exonic reads improved the pseudoaligners. On real data they add ~800 (Kallisto) and ~600 (Alevin_sketch) genes per cell absent from CellRanger. · 2.2 Simulations without multi-gene reads; 2.3 Pseudoalignment-to-transcriptome predicts…; 2.5 Mapping/quantification comparison on real…

  2. ShownRun with CellRanger-matching settings, STARsolo reproduces CellRanger's count matrix and its downstream marker genes almost exactly on real 10X PBMC data, whereas the other tools diverge.

    10X-pbmc-5k, 4655 shared cells: of ~17M non-zero gene/cell counts, STARsolo_sparseSA differed from CellRanger by one count in 85 elements (~0.0005%). Wilcoxon marker genes (Benjamini-Hochberg padj<0.01; CellRanger's 9 Leiden clusters; 4,473 cells): near-perfect agreement in marker sets and log2 fold changes for STARsolo, decent for Alevin_full-decoy, poor for Kallisto and Alevin_sketch. · 2.5 Mapping/quantification comparison on real…; 2.6 Differential gene expression is strongly…; 5.3 Real data: clustering and differential gene…

  3. ShownAmong the accurate tools STARsolo is fastest; pseudoaligners are quicker and lighter for mature mRNA, but Kallisto, the one benchmarked for pre-mRNA (single-nucleus) counting, loses that edge, while STARsolo's cost is unchanged.

    10X-pbmc-5k, 5 runs per tool: STARsolo_fullSA ran 4.4-fold faster than CellRanger and 2.8-fold faster than Alevin_full-decoy using 32 GiB (sparseSA: 16 GiB, 1.7 times slower); CellRanger needed 73 GiB at 20 threads; Alevin_sketch was ~2-3 times faster than STARsolo_fullSA, pseudoaligners ≲4 GiB. For pre-mRNA, Kallisto slowed ~5-fold and grew to 68 GiB; STARsolo was unchanged. · 2.7 Benchmarks: trading accuracy for speed; 3 Discussion; 5.4 Benchmarking speed and RAM consumption

  4. SuggestedSimulated multi-gene reads lower accuracy for every tool; enabling STARsolo's multi-gene recovery (EM) trades a few false positives for fewer false negatives, a small net gain the authors argue matters for paralogs.

    Gene/cell Spearman R vs truth fell from ≈0.99 to 0.95 for STARsolo and Alevin_full-decoy, and further for Kallisto (to 0.69) and Alevin_sketch (to 0.61). Without recovery, STARsolo and Alevin_full-decoy had a ~5% false negative rate; recovery reduced it at some cost in false positives, improving overall accuracy slightly. · 2.4 Simulation with multi-gene reads; 5.1.7 Multi-gene reads

  5. SuggestedSTARsolo's splice-junction counts let droplet scRNA-seq, despite 3′-end bias, resolve cell-type-specific junction usage: hundreds of junctions per Tabula Muris tissue switch usage between cell types.

    Tabula Muris 10X data (~100 thousand cells, 20 mouse organs/tissues): ~20-40% of reads spliced; pseudobulk junction usage per cluster with 100 bootstrap resamples, Wilcoxon test and Benjamini-Hochberg correction (padj < 0.05) gives hundreds of significant junctions per tissue, e.g. an exon-skipping event between Bladder and Bladder-Urothelial cells. · 2.8 Beyond gene expression: cell-type specific…; 5.5 Alternative splicing analysis; Figure legends

Methods, data and reproducibility

STARsolo adds barcode demultiplexing, UMI deduplication, multi-gene read assignment (uniform, rescue or EM) and CellRanger-style cell calling to STAR's whole-genome alignment. It was tested on reads simulated from real 10X PBMC data (true loci from BWA-MEM alignments, 0.5% mismatches) and on the real data against CellRanger, Alevin and Kallisto.

  • 10X Genomics peripheral blood mononuclear cell library, Chromium v3 (10X-pbmc-5k); real-data comparisons and speed/RAM benchmarks · human (benchmarks report human genome index memory) · n = 384 million reads, ~5000 cells; 4655 cells shared by all tools, 4,473 after QC
  • Simulated reads derived from 10X-pbmc-5k (BWA-MEM true loci, 0.5% substitution rate), with and without multi-gene reads · human (derived from 10X-pbmc-5k) · n = top 5000 cells by true UMI count, intersected with each tool's cells
  • Tabula Muris 10X scRNA-seq libraries, reprocessed for splice-junction usage · mouse (Mus musculus) · n = ~100 thousand cells from 20 organs and tissues
  • GRCm38 genome assembly with Gencode M20 annotations (STAR index for the splicing analysis) · mouse
Methods & tools
STAR, STARsolo, CellRanger, Alevin, Alevin-fry, Kallisto-bustools, BWA-MEM, Scanpy 1.6.0, Leiden clustering
Software & environment
Scanpy 1.6.0; per-tool parameters in Supp. Table 1; STAR distributed under the MIT license on GitHub with its full release history.
Assay conditions & compute
Two Intel Xeon Gold 6140 CPUs (36 cores), 384 GiB DDR4 RAM, eight 12TB SATA drives in RAID0; 5 runs per tool, wall time and peak RSS via /usr/bin/time -v. Peak RAM for human: 32 GiB (fullSA), 16 GiB (sparseSA).
Reproducibility gaps
No accession or URL printed for 10X-pbmc-5k (deferred to Supp. Table 1) or Tabula Muris (no source given); versions of STAR and the compared tools are not stated in the main text.
Limitations (6)
  • The simulation keeps only reads whose barcodes exactly match the passlist, so barcode and UMI error correction are not evaluated. · stated by the authors
  • Real-data accuracy is judged against CellRanger, which is a reference point rather than ground truth. · stated by the authors
  • Genome alignment costs memory: 32 GiB (full) or 16 GiB (sparse suffix array) for human vs ≲4 GiB for pseudoaligners, and quantification keeps 16 bytes per read in RAM instead of processing barcodes in sorted batches. · stated by the authors
  • Low sequencing depth and 3′-end bias mean droplet data capture only a small share of cell-type-specific splice junctions. · stated by the authors
  • Every benchmark uses a single 10X PBMC library and the authors' own simulation, whose truth comes from BWA-MEM alignments to genome plus transcriptome; none of the other barcode protocols STARsolo supports is benchmarked. · noted in this summary
  • Numbers in Results 2.4 conflict with 2.2/2.3: Kallisto's no-multi-gene R is ≈0.86 there vs ≈0.77 earlier, and ~600 to ~800 false-positive genes are attributed to Alevin_full-decoy, earlier reported at ~5 per cell; check the figures. · noted in this summary

Read this if you quantify droplet scRNA-seq or snRNA-seq and are choosing between STARsolo, CellRanger, Alevin and Kallisto, or need CellRanger-equivalent counts from an open, protocol-flexible pipeline.

Citations
433

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBrüning '22Zentgraf '25Hung '26Li '25You '23Sullivan '24
72% semantic matchBioinformatics
Ralf Schulze Brüning, Lukas Stefan Tombor, Marcel H. Schulz+2 more·GigaScience·Jan 1, 2022· 43 citations