henhao
Sign in
BioinformaticsbioRxiv
Preprint

STAR Suite: Transcriptomics processing in a single binary through AI-assisted development

Ling‐Hong HungiD, Dylan Baker, Bill Flynn, Danwei HuangfuiD, Renhe LuoiD, Paul RobsoniD, Ting ZhouiD, Ka Yee YeungiD

Abstract

Background. STAR is the aligner underlying most transcriptomics processing, and is used inside Cell Ranger. However, the surrounding steps - adapter trimming, sorting, feature-barcode assignment, probe handling, SLAM-seq analysis, quantification, and quality control - run as separate tools scripted around it. Cell Ranger hides this complexity but is proprietary, restricted to 10x products, and barred by its license from redistribution, so it cannot serve as a shared processing layer. While STARsolo supports scRNA-seq, it lacks the feature-barcode detection needed for Perturb-seq, where short tags such as CRISPR guides or lineage barcodes are read alongside the transcriptome. No open-source tool processes Perturb-seq feature barcodes at production scale, and the only open-source processor for 10x Flex - a probe-based assay for fixed and FFPE tissue that multiplexes several samples in one lane - is a recent standalone k-mer tool. Results. STAR Suite compiles these steps into the STAR codebase to create a drop-in executable for bulk RNA-seq, scRNA-seq, Perturb-seq, 10x Flex, and SLAM-seq. It adds no dependencies beyond OpenSSL's libcrypto, and legacy behavior is preserved. On consortium and public benchmarks it is up to 41-fold faster than Cell Ranger 9.0.1 for Flex (20- to 41-fold from binary CBQ input, 17- to 23-fold from FASTQ) and 3.9- to 6.2-fold faster for scRNA-seq and Perturb-seq, timed on one modest server (24 cores, 128 GiB, local SSD) except for the largest dataset, which ran on a cloud instance limited to the same size. The concordance to Cell Ranger is high: gene-level Spearman and Pearson 0.98-1.0, cell-level Pearson on per-cell total counts 0.9999-1.0, feature-UMI Pearson 0.999-1.000, Jaccard agreement of the called-cell sets 0.97-0.995, and 99-100% CRISPR-call agreement. It also needs far less disk: its peak use on the largest Flex dataset is 16.9 GiB, against 1.8 TiB for Cell Ranger. These gains arise from running the steps in one process rather than as separate programs exchanging files, and from four new computational techniques: fast-Hamming, a vectorized exact Hamming-distance search, used for Flex probe matching and optionally for feature barcodes; a disambiguated hash cascade that assigns Flex probes without alignment; a permit-based thread scheduler that interleaves alignment and feature assignment; and variance-based auto-trimming for SLAM-seq conversion calling, which finds the reliable region of each read from how variable the conversion rate is along it. The four modules add 183,606 lines of C/C++ to STAR's 28,228. STAR Suite is the NIH MorPhiC consortium's production processor. The same binary is run by people, by AI agents and by cluster schedulers, and every production run is recorded with its commands, checksums and outputs. It has been built and maintained over eight months and 25 releases under a human-directed, AI-implemented workflow, with its full design and benchmark record public. Conclusions. STAR Suite provides the first production-ready open-source implementation of Perturb-seq feature-barcode processing, and an open end-to-end 10x Flex pipeline matched to Cell Ranger, delivered as a drop-in replacement for the STAR aligner. The design also extends: downstream analysis now done in Python or R packages can be built into the same binary, and further modalities added alongside, as a companion preprint does for ATAC-seq. Source code, workflow recipes, and per-run provenance are released under the MIT license.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
1
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 7 (24 Sept 2026) · CC BY 4.0

Open-source STAR binary for five transcriptomic assays runs 3.9- to 6.2-fold faster than Cell Ranger 9.0.1 for scRNA-seq and Perturb-seq, 17- to 41-fold for Flex, with closely concordant counts.

Around the STAR aligner, trimming, sorting, feature-barcode assignment, probe handling and QC run as separately scripted tools; Cell Ranger integrates them but is proprietary, 10x-only and not redistributable, and the authors report no open tool for production-scale Perturb-seq feature barcodes and only a standalone k-mer tool for Flex. STAR Suite compiles these steps into STAR as a drop-in executable, adding vectorized Hamming search and alignment-free Flex probe assignment.

Key results

  1. ShownSTAR-Flex assigns Flex probes by hash lookup and vectorized Hamming scoring, aligning no reads and loading no genome. It runs in one process with its working state in memory, unlike Cell Ranger's disk-backed chunks, which sharply cuts runtime and peak scratch disk relative to Cell Ranger 9.0.1.

    From FASTQ about 17-fold faster than Cell Ranger 9.0.1 (JAX SC2300771: 3 min 20 s vs 57 min 6 s; GSE325982: 4 min 17 s vs 74 min 6 s); 20- and 41-fold from CBQ. 320k set on a same-size AWS instance: 17 min 36 s vs 6 h 38 min (22.6-fold), 72 GiB peak memory. Peak disk on 320k: 16.9 GiB vs 1,805 GiB (Cell Ranger) and 218 GiB (cyto). On spinning disk, 12- to 20-fold faster. · 2.5 STAR-Flex: a disambiguated hash cascade for…; 2.6 Scratch storage and Flex on spinning disk; 5.2 Benchmarking metrics

  2. ShownSTAR-Flex calls cells per biological sample, as Cell Ranger does, so its called-cell sets stay close to Cell Ranger's on multi-tag samples. cyto calls each tag separately and diverges there, although counts on shared cells agree closely for both tools.

    Single-tag JAX and GSE325982: pooled Jaccard 0.988 and 0.994; missed 4 of 20,419 and 5 of 38,444 Cell Ranger cells; gene Spearman above 0.9998 per sample. 320k (two tags per sample): pooled Jaccard 0.974 (0.958 to 0.984 per sample) vs 0.883 (0.696 to 0.952) for cyto. On the 314,388 cells all three tools call, mean per-cell Pearson is 0.9991 vs 0.9958. · 2.5 STAR-Flex: a disambiguated hash cascade for…; 5.3 Flex benchmark conditions

  3. ShownOn three 3' scRNA-seq sets, gene counts are closer to Cell Ranger 9.0.1 than upstream STARsolo with the community CellGENI settings. The gains come from poly-G clipping and a multimapper rule inferred from Cell Ranger outputs. Remaining differences are mostly a few extra marginal cells.

    PBMC 10K: Jaccard 0.995, cell Pearson 0.99999, gene Spearman 0.990 vs 0.950 for STAR 2.7.11b with CellGENI; 16 min 18 s vs 63 min for Cell Ranger (3.9-fold). Glioma and organoid: gene Spearman 0.992 and 0.993 vs 0.957 and 0.960. CellGENI over-counts LINC00486 16- to 106-fold and under-counts RPL39, H3-3A and SUMO2 to 8 to 27% of Cell Ranger's counts. · 2.3 scRNA-seq: parity with Cell Ranger 9 for…; 5.2 Benchmarking metrics; Poly-G clipping; Multimapper rescue

  4. ShownSTAR-Perturb handles gene expression, CRISPR guides and a 245,979-barcode LARRY lineage library in one pass. Its guide calls are nearly identical to Cell Ranger's.

    A375: 2 min 26 s, 4.7-fold faster than Cell Ranger 9.0.1; identical CRISPR calls on all 1,076 cells with a guide call in both outputs. MSK 30-KO: 27 min 6 s vs two summed Cell Ranger runs (6.2-fold); 99.0% exact CRISPR-call agreement; per-guide feature-UMI Pearson 0.999. The lineage dominance caller matched production calls on all 26,521 shared cells. · 2.4 STAR-Perturb: open-source feature processing…; 5.10 Feature calling; 5.1 Benchmarking datasets

  5. SuggestedFor bulk RNA-seq, putting trimming, Y-chromosome removal and TranscriptVB quantification into one STAR call is faster than a stepwise pipeline and agrees with Salmon at gene level. STAR-SLAM reproduces GRAND-SLAM's new-to-total RNA ratio (NTR) estimates.

    PPARG (35.1M PE reads): 1.6-fold faster than Trim Galore + STAR + Salmon without Y removal, 4.3-fold with it. TranscriptVB vs Salmon: gene Spearman 0.9992 and Pearson 1.0000; transcript read-count Spearman 0.985. GRAND-SLAM 100K-read fixture: NTR Pearson 0.994 to 0.999 across read-count thresholds; no SLAM-seq timing reported. · 2.2 Bulk RNA-seq: the base production…; 2.7 STAR-SLAM: integrated NTR estimation for…; 5.1 Benchmarking datasets

Methods, data and reproducibility

AI agents implemented a developer's plans, adding 183,606 lines of C/C++ to STAR 2.7.11b that bring trimming, sorting, quantification, feature-barcode assignment, alignment-free Flex probe assignment and SLAM-seq NTR estimation into one binary. Outputs and runtimes were compared with Cell Ranger 9.0.1, cyto, STARsolo (CellGENI) and a stepwise bulk pipeline; SLAM-seq outputs with GRAND-SLAM.

  • 10x PBMC 10K (3' v3) scRNA-seq; compared with Cell Ranger 9.0.1 and STAR 2.7.11b CellGENI baselines · n = 638.9 million read pairs
  • Public scRNA-seq: glioma cell line (3' v3) and GEM-X 3' v4 organoid · n = 117.0 and 301.6 million read pairs · GSE261042 (SRR28247645); GSE316010 (SRR36749062)
  • 10x A375 CRISPR 5' GEX Perturb-seq dataset (11 features) · n = 1,083 cells with a Cell Ranger guide call
  • MSK 30-KO Perturb-seq on ES cells: gene expression, 30 CRISPR guides, 245,979 LARRY lineage barcodes (MorPhiC, embargoed) · human (inferred from MorPhiC's human-gene knockout scope) · n = 33,226 cells
  • JAX SC2300771 10x Flex dataset (MorPhiC, embargoed) · human (processed with GRCh38-2024-A, probe set v1.1.0) · n = 4 biological tags, 8 lanes, 2.011 billion read pairs
  • Public Flex sets: GSE325982 pool (four single-tag samples) and 10x 320k scFFPE GEM-X (16 tags in 8 samples) · human (processed with GRCh38-2024-A probe set) · n = 1.121 and 7.303 billion read pairs · GSE325982
  • PPARG paired-end bulk RNA-seq benchmark in KOLF2.2J iPSC · human (KOLF2.2J iPSC line) · n = 35.1M PE reads
  • GRAND-SLAM SLAM-seq fixture with reference NTR values · human · n = 100K reads
Methods & tools
STAR 2.7.11b / STARsolo, Cell Ranger 9.0.1, cyto 0.4.7, Cutadapt 5.1, Trim Galore, Salmon, GRAND-SLAM (GEDI), samtools, EmptyDrops, rapidgzip 0.16.0, BINSEQ (CBQ), Model Context Protocol (MCP) server
Code
https://github.com/morphic-bio/STAR-suitehttps://github.com/morphic-bio/STAR-suite-paperhttps://github.com/morphic-bio/STAR-suite-provenancehttps://github.com/morphic-bio/STAR-suite-recipeshttps://github.com/morphic-bio/morphic-provenancehttps://doi.org/10.5281/zenodo.21632664
Data & accessions
https://github.com/erhard-lab/gedi/wiki/GRAND-SLAMhttps://www.10xgenomics.com/datasets/1k-CRISPR-5p-gemxhttps://morphic.bio/data/
Software & environment
STAR Suite v1.9.5 (on STAR 2.7.11b), MIT license. Releases ship Linux x86-64 binaries, x86-64/arm64 Docker images and Debian packages. Runtime needs zlib and OpenSSL libcrypto; source builds need g++ (C++17), GNU make and zlib.
Assay conditions & compute
Intel i9-13900KF server (24 cores, 32 threads, 128 GiB, local NVMe) for all timings except the 320k set. The 320k set ran on AWS m6id.16xlarge (32 vCPUs, 128 GiB, two NVMe in RAID0). STAR-Flex peak memory 18 to 72 GiB; Cell Ranger's 320k run needs about 3 TB scratch.
Reproducibility gaps
MSK 30-KO and JAX SC2300771 data are embargoed until MorPhiC release, obtainable from the authors on request meanwhile under consortium policy. CBQ input is experimental. Release binaries use a software popcount unless rebuilt with -mpopcnt. HTML QC reports load plotting code from a CDN.
Limitations (6)
  • Cell Ranger components were reimplemented clean-room, so some rules (e.g. the scRNA-seq multimapper policy) are inferred from outputs. Monte Carlo sampling in EmptyDrops and thread-dependent read order also prevent exact parity. · stated by the authors
  • STAR-Flex scores the unanchored probe half by Hamming distance, so it rejects reads with an indel there that Cell Ranger accepts. The half-probe route drops the earlier cache's genomic check, which the authors say may be biologically more accurate. · stated by the authors
  • The MSK 30-KO and JAX SC2300771 benchmarks are under consortium embargo, CBQ input is experimental, and there is no speed comparison for SLAM-seq. · stated by the authors
  • For most of development the AI agents could not reliably run the implement-and-test loop unsupervised, so human review was needed at most stages. · stated by the authors
  • The developers designed and ran every benchmark, including the Cell Ranger reference runs for PBMC 10K, the glioma and organoid sets, and 320k. No independent evaluation is reported. · noted in this summary
  • No replicate timings or run-to-run variance are given in the main text. The bulk speedup rests on one 35.1M-read sample, and SLAM-seq parity on a 100K-read fixture. · noted in this summary

Who should read:Read this if you run STAR, STARsolo or Cell Ranger on 10x scRNA-seq, Perturb-seq or Flex data and need an open, redistributable processor, or want a case study of maintaining a large scientific codebase with AI agents.

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
THISKaminow '21Hung '26Dobin '12Ballouz '18Edwards '22Xu '25
60% semantic matchBioinformatics
Benjamin Kaminow, Dinar Yunusov, Alexander Dobin·bioRxiv·May 5, 2021· 433 citations