OpenOmicsBench

OpenOmicsBench is a compact benchmark collection and local toolkit for bioinformatics software testing, method checks and teaching. It contains 12 certified bulk RNA-seq objects, seven deterministic sequence and classification fixtures, a 10x-style single-cell RNA-seq benchmark, a genomic peak-calling benchmark, strict local tools, and a CI-ready evaluation system. Browse the collection before installing.

Version 11 adds aligned bulk RNA-seq data access for the 12 validated biological objects. The CLI can export counts, sample metadata and source details into a new directory. The Python API loads the same checked files as a feature-by-sample integer array with ordered sample records. The installed package still contains the truth sets, so the workspace is not a blinded challenge.

The package runs offline after installation. Sequence files stay on the local computer. The built-in tools cover inspection and lightweight preprocessing; they do not claim to replace aligners, variant callers, taxonomic classifiers or assay-specific statistical workflows.

Install

OpenOmicsBench requires Python 3.11 or newer.

python -m pip install openomicsbench
omicsbench --version
omicsbench demo my-first-run

Open my-first-run/report.html for the example comparisons. The corresponding report.json, reference-results/results.json and evidence.zip are in the same output directory. The demo refuses to replace an existing directory.

List and validate the bundled benchmarks:

omicsbench list
omicsbench validate rnaseq-002
omicsbench validate sequence-004
omicsbench variant compare sequence-005 calls.vcf.gz
omicsbench peptide compare sequence-006 identifications.peptides.tsv
omicsbench validate singlecell-001
omicsbench cell compare singlecell-001 clusters.tsv
omicsbench validate interval-001
omicsbench interval compare interval-001 calls.narrowPeak
omicsbench validate sequence-007
omicsbench taxon compare sequence-007 assignments.kraken

The wheel contains the complete collection. A repository checkout and network connection are not required.

Test your own results

Prepare a workspace for the benchmarks your pipeline supports:

omicsbench suite prepare my-run --id rnaseq-002 --id sequence-007
omicsbench suite verify-workspace my-run

The command copies checked input files into my-run/inputs/ and writes my-run/workspace.json. That JSON lists each case's input roles and hashes, sample design, source citation, rights, limitations, expected result columns and suggested result filename. It does not copy expected answers into the workspace.

Run your pipeline against those inputs, then put its results under my-run/results/. The suggested names are rnaseq-002.tsv and sequence-007.taxonomy.tsv. Check how files will be routed before writing the result index:

omicsbench suite discover my-run
omicsbench suite discover my-run --write
omicsbench suite grade my-run

Discovery accepts all six comparable assay families. It recognizes benchmark IDs in result filenames or parent directories and checks named columns in RNA effect, peptide and cell-cluster tables. If a pipeline uses unrelated filenames or headers, supply an exact path or column mapping:

omicsbench suite discover my-run \
  --result rnaseq-002=analysis/effects.csv \
  --column rnaseq-002:gene_id=Geneid \
  --column rnaseq-002:log2_fold_change=log2FoldChange \
  --write

The command refuses ambiguous matches and writes my-run/results/results.json only when every selected case has a supported file and required headers. --force with --write replaces an existing index after results change. Discovery checks routing, headers and hashes; it does not score results or establish that a file has the right biological meaning. suite grade performs the benchmark comparison. For results outside a prepared workspace, suite index RESULTS_DIRECTORY remains available for explicit mappings.

Grading checks the workspace inputs again, scores only its selected cases, and writes report.json, report.html, report.csv, report.xml and evidence.zip inside my-run/. Missing or incorrect results appear as failed cases. Reports are preserved by default; use suite grade my-run --force only when you intend to replace them after changing results. This workflow tests file handling and declared benchmark outcomes; each object's limitations still apply.

Use from Python

from omicsbench import Collection

collection = Collection()
for item in collection.list(assay="bulk_rna_seq", status="validated", kind="real"):
    print(item["id"], item["title"])

benchmark = collection.get("rnaseq-002")
samples = benchmark.file(role="samples", tier="metadata")
print(samples)
print(benchmark.validate()["status"])

data = benchmark.rnaseq("pocket")
print(data.counts.shape, data.samples)

file() returns a declared path only after checking its size and SHA-256 digest. Collection.prepare(), verify_workspace(), discover(), grade() and evaluate() expose the same contracts as the CLI. The API uses the installed collection by default and works offline. Pass a collection path to Collection(root=...) only when using a separate checkout or local collection.

To use the biological RNA-seq counts in R, Python or another tool, export one object:

omicsbench rnaseq export rnaseq-002 my-rna-data
omicsbench rnaseq export rnaseq-002 my-full-data --tier full

Each new directory contains counts.tsv (genes in rows, samples in columns), samples.tsv (one row per sample in matching column order) and export.json with source attribution, limitations, dimensions and SHA-256 hashes. The compact pocket tier is the default; full uses the complete count matrix for the selected samples. Existing directories are preserved. The export checks the source file against its declared digest and refuses a mismatched sample order. These counts are for software tests and method development; consult each object's limitations before interpreting biological effects.

Sequence tools

Run a strict combined QC report on FASTA, FASTQ or gzip-compressed input:

omicsbench seq qc reads.fastq.gz --molecule dna --adapter AGATCGGAAGAGC

Common operations include:

omicsbench seq stats reference.fasta.gz
omicsbench seq validate proteins.fasta --molecule protein
omicsbench seq pair-check sample_R1.fastq.gz sample_R2.fastq.gz
omicsbench seq trim reads.fastq.gz trimmed.fastq.gz --quality 20 --min-length 30
omicsbench seq filter trimmed.fastq.gz clean.fastq.gz --min-length 30 --max-ambiguity 0.05 --min-mean-quality 25
omicsbench seq sample clean.fastq.gz subset.fastq.gz --count 10000 --seed 7
omicsbench seq transform transcripts.fasta proteins.fasta --operation translate --frame 1
omicsbench seq protein-stats proteins.fasta --ph 7.4
omicsbench seq digest proteins.fasta --enzyme trypsin --missed-cleavages 1
omicsbench seq motif reference.fasta ACGTRG --both-strands
omicsbench seq kmers reads.fastq.gz --k 21 --canonical --top 50
omicsbench seq compare sample-a.fasta sample-b.fasta --k 21 --size 5000

The command suite supports:

• multiline FASTA and FASTQ, including .gz input and output;
• DNA, RNA and protein alphabets with IUPAC ambiguity symbols;
• sequence length, N50/L50, composition, GC, ambiguity and exact duplication statistics;
• FASTQ Phred+33 range checks, Q20/Q30 summaries and per-position quality;
• paired-read identity checks, interleaving and deinterleaving;
• fixed, adapter and end-quality trimming;
• filtering by length, GC, ambiguity and mean quality;
• deterministic sampling and deduplication;
• reverse complements, transcription, back-transcription and six translation frames;
• overlapping IUPAC motif searches and six-frame ORF discovery;
• canonical k-mer counts and deterministic bottom-k similarity sketches;
• amino-acid composition, molecular weight, hydropathy, charge, estimated pI, aromaticity and extinction estimates;
• trypsin, Lys-C, Arg-C and chymotrypsin digestion with missed-cleavage and peptide-length controls;
• record extraction with zero-based, half-open slices.

Commands that write files refuse to replace an existing output unless --force is supplied. Outputs are written through a temporary file and moved into place only after the operation succeeds. JSON reports go to standard output so they can be stored or checked in automated workflows.

The FASTA validator follows the nucleotide symbol expectations described by NCBI. FASTQ has no single formal specification; OpenOmicsBench accepts conventional multiline records with printable Phred+33 quality characters and requires the sequence and quality lengths to match. The GA4GH-maintained HTS specifications describe the surrounding SAM, BAM, CRAM and VCF ecosystem.

The full command reference is in docs/sequence-tools.md.

Single-cell tools

Inspect a 10x-style feature-barcode matrix without loading a dense array:

omicsbench cell qc matrix.mtx barcodes.tsv features.tsv --feature-columns 3

The report includes matrix dimensions, nonzero entries, library totals, detected features per cell, zero-count cells and the mitochondrial count fraction. Plain and gzip-compressed inputs are accepted. Coordinates, counts and companion table dimensions are checked before statistics are returned.

singlecell-001 contains 90 synthetic cells, 60 features, three cell states and two crossed batches. Score a clustering table with cell_id and label columns:

omicsbench cell compare singlecell-001 clusters.tsv

Cluster names may differ from the truth labels. The comparator uses label-invariant ARI and NMI, alongside purity and complete cell coverage. The object tests file handling, QC and recovery of a known synthetic partition. It does not validate biological cell-type annotation, normalization, integration, trajectory inference or batch correction.

Genomic interval tools

Validate a BED or narrowPeak file and summarize its coordinates:

omicsbench interval qc calls.narrowPeak --chrom-sizes chrom.sizes

interval-001 contains 12 deterministic peaks across two synthetic contigs. The fixture covers narrow peaks, one broad peak, a close doublet and positions near chromosome boundaries. Its binned bedGraph signal is small enough for unit tests and teaching while preserving a declared truth set.

omicsbench interval compare interval-001 calls.narrowPeak

Peak matches require at least 50% reciprocal overlap. The report separates interval-level precision, recall and F1 from base-pair precision, recall and F1, so extra calls and over-wide calls remain visible. Parsing follows the UCSC zero-based, half-open BED convention, checks chromosome bounds, and accepts BED3, BED6, narrowPeak and gzip-compressed files. The benchmark measures coordinate and overlap recovery. It does not assess alignment, blacklist filtering, replicate concordance or biological peak validity. See the genomic interval guide for the exact contract.

Mock taxonomy assignment

sequence-007 contains three generated reference sequences and 48 paired reads. Twelve pairs deliberately have no unique reference assignment: six come from a shared segment and six from an absent background sequence. The reference headers and truth labels use project-local codes 1, 2 and 3; they are not NCBI taxonomic identifiers.

Compare read-level assignments written as read_id and taxon_id columns, or pass Kraken 2's standard five-column per-read output directly:

omicsbench taxon compare sequence-007 assignments.tsv
omicsbench taxon compare sequence-007 assignments.kraken

For a small exact-match teaching example, omicsbench taxon classify reference.fasta reads_R1.fastq reads_R2.fastq calls.tsv writes canonical assignments when reference IDs follow mock_taxon_<code>. It assigns 0 when neither mate identifies a unique reference. This narrow tool tests file handling and scoring; it is not a metagenomic classifier for biological samples.

Benchmark suites

Validate the complete installed collection and save reports that can be attached to a CI run:

omicsbench suite validate --json reports/validation.json --markdown reports/validation.md --junit reports/validation.xml

The selection can be narrowed by assay or by repeating --id:

omicsbench suite validate --assay bulk_rna_seq
omicsbench suite validate --id rnaseq-002 --id sequence-005

To test a differential-expression pipeline across several biological objects, place each result in a directory under its benchmark ID. CSV and TSV are accepted, with optional gzip compression:

results/
  rnaseq-002.tsv.gz
  rnaseq-003.csv

Then run:

omicsbench suite compare results --id rnaseq-002 --id rnaseq-003 --junit reports/comparison.xml

For a mixed bulk RNA, single-cell, chromatin, DNA, protein and mock-taxonomy run, add benchmark-named cluster, peak, VCF, peptide and read-assignment output alongside the expression tables:

results/
  method.json
  rnaseq-002.tsv.gz
  singlecell-001.cells.tsv
  interval-001.narrowPeak
  sequence-005.vcf.gz
  sequence-006.peptides.tsv
  sequence-007.taxonomy.tsv

omicsbench suite evaluate results \
  --id rnaseq-002 \
  --id singlecell-001 \
  --id interval-001 \
  --id sequence-005 \
  --id sequence-006 \
  --id sequence-007 \
  --json reports/evaluation.json \
  --html reports/evaluation.html \
  --evidence reports/evaluation-evidence.zip

Pipeline outputs do not need to be renamed or moved into one flat directory. Index their existing relative paths and native headers:

omicsbench suite index results \
  --result rnaseq-002=rna/deseq2.csv \
  --result singlecell-001=single-cell/clusters.csv \
  --result interval-001=chromatin/peaks.narrowPeak \
  --result sequence-005=dna/final.vcf.gz \
  --result sequence-006=protein/identifications.tsv \
  --result sequence-007=taxonomy/assignments.kraken \
  --column rnaseq-002:gene_id=Geneid \
  --column rnaseq-002:log2_fold_change=log2FoldChange \
  --column singlecell-001:cell_id=Barcode \
  --column singlecell-001:label=Cluster \
  --column sequence-006:protein_id=Protein \
  --column sequence-006:peptide=Sequence \
  --column sequence-006:q_value=QValue

This writes results/results.json. Evaluation detects it automatically, verifies every recorded hash and applies the declared column mappings. A changed file, unsafe path, duplicate benchmark, unknown field or incompatible file type stops the run before scoring. Existing benchmark-named files continue to work when no manifest is present.

method.json is optional. When supplied, it records the exact method version, command, container or source revision, parameters, runtime, peak memory and thread count. Missing selected outputs fail as benchmark cases.

Compare several tools or parameter sets in one matrix:

omicsbench suite matrix \
  --method deseq2=results/deseq2 \
  --method edger=results/edger \
  --json reports/matrix.json \
  --csv reports/matrix.csv \
  --html reports/matrix.html

Use evaluate-matrix when each method directory contains a mixture of RNA tables, VCF calls and peptide tables:

omicsbench suite evaluate-matrix \
  --method current=results/current \
  --method candidate=results/candidate \
  --id rnaseq-002 \
  --id sequence-005 \
  --id sequence-006 \
  --html reports/multi-assay.html

The leaderboard uses pass counts and keeps assay metrics separate. It does not average unlike measurements into a single score.

Each evaluation records SHA-256 hashes for the submitted results and benchmark manifests, then builds a run_fingerprint from those hashes, the effective input contract and the validated method receipt. Local directory names and manifest formatting do not affect this identity.

Apply a checked-in policy to an evaluation before accepting a run:

omicsbench suite gate reports/evaluation.json policy.json \
  --json reports/gate.json \
  --junit reports/gate.xml

A policy can require named benchmarks, overall and per-assay pass rates, a method receipt, runtime and memory ceilings, and metric floors scoped to an assay or comparator. Missing required evidence fails the gate.

Track several accepted or candidate runs in chronological order:

omicsbench suite trend \
  --run v4.0=reports/v4.0.json \
  --run candidate=reports/candidate.json \
  --json reports/trend.json \
  --html reports/trend.html

Trend reports preserve each source fingerprint, method receipt, resource measurements, pass rate and per-benchmark metric history.

Use a previously accepted report as a regression baseline:

omicsbench suite regress accepted.json candidate.json --absolute-tolerance 0.01 --junit reports/regression.xml

Omit --id from suite compare to require results for all 12 comparable RNA-seq objects. Omit it from suite evaluate to require every benchmark with a supported comparator. A completed suite exits with status 0 when every check passes, status 2 when a benchmark, policy or regression gate fails, and status 1 for invalid input. Existing report files are protected unless --force is supplied. The suite reference describes the file convention and report fields.

Verify an evaluation evidence crate without extracting it:

omicsbench evidence verify reports/evaluation-evidence.zip

Portable bundles

Create a deterministic ZIP containing selected benchmarks, their manifests, declared files and the project licence material:

omicsbench bundle create rna-and-dna.zip --id rnaseq-002 --id sequence-005
omicsbench bundle verify rna-and-dna.zip

Bundle verification checks the complete file inventory, byte counts, SHA-256 digests and embedded dataset manifests without extracting the archive. Recreating the same selection with the same package version produces the same archive bytes. See docs/portable-bundles.md.

Assay profiles

The assay profiles explain what the local tools can check and where a dedicated workflow becomes necessary:

omicsbench assay list
omicsbench assay info whole-genome
omicsbench assay plan whole-genome sample_R1.fastq.gz sample_R2.fastq.gz --adapter AGATCGGAAGAGC

Profiles are available for whole-genome, exome, targeted-panel, bulk RNA-seq, single-cell RNA-seq, ATAC-seq, ChIP-seq and related assays, amplicon sequencing, shotgun metagenomics, long-read DNA and RNA, genome or transcriptome references, RNA FASTA and protein FASTA. A plan contains executable validation and preprocessing commands plus a clear boundary for the downstream tools that still need a reference, database or assay-specific model.

Bundled sequence benchmarks

These seven objects are project-authored synthetic fixtures. Each has an exact file inventory, checksums, format and molecule declarations, expected summary metrics and a deterministic rebuild workflow. sequence-005 verifies that every truth-set reference allele matches the bundled reference and that every alternate allele is supported by the paired reads. sequence-006 verifies exact protein-peptide pairs from a declared tryptic digest. sequence-007 checks read-level labels against exact mock-reference matches. The objects test software behavior and do not represent a biological cohort or sequencing instrument.

The version 3 VCF comparator for sequence-005 matches contig, one-based position, REF and ALT exactly. It reports precision, recall and F1 for A/C/G/T substitutions and verifies REF against the packaged synthetic reference. Genotypes, indels, complex representations and confident-region stratification are outside this fixture. Use a haplotype-aware benchmarking engine such as hap.py or vcfeval with an appropriate truth set and confident regions for real germline benchmarking; the GA4GH benchmarking project documents that broader problem.

The version 4 peptide comparator accepts protein_id and peptide columns, with an optional q_value column. It scores unmodified pairs after the declared q-value filter and reports peptide precision, recall and F1, protein recall and covered truth residues. It does not parse spectra or claim to replace a full identification standard. The HUPO-PSI mzIdentML specification is the appropriate interchange model for complete mass-spectrometry identification results.

Bulk RNA-seq benchmarks

Version 4 retains the complete version 1 biological collection unchanged.

Each biological object includes the source counts for its selected samples, a smaller pocket matrix, ordered sample metadata, the declared design, source attribution, rights evidence, reference details, transformations and quantitative validation. All 12 pockets pass their predeclared DESeq2 fidelity thresholds.

Compare another tool's gene-level effects with a bundled full-source DESeq2 reference:

omicsbench compare rnaseq-002 results.csv

The input must contain unique gene_id and finite log2_fold_change columns. The report includes rank correlation, top-50 overlap, sign agreement, coverage and identifier differences. Status 0 means the declared checks passed, status 2 means a completed comparison missed a scientific threshold and status 1 means the input was invalid.

Collection records

The JSON catalog contains every benchmark family. catalog/collection.tsv lists the biological objects, catalog/sequences.tsv lists the sequence fixtures, catalog/single-cell.tsv lists the single-cell objects, and catalog/intervals.tsv lists genomic interval benchmarks.

Expression Atlas and BioStudies supplied the biological count matrices. Their bundled material is recorded as CC BY 4.0 with provider attribution. The sequence fixtures and software are project-authored and distributed under Apache-2.0. Project-written descriptive metadata is CC BY 4.0.

The objects are research and software-test material. They are not clinical references and do not replace the complete source studies.

Development checks

python -m unittest discover -s tests -v
python workflows/build_sequence_fixtures.py
python workflows/build_singlecell_fixture.py
python workflows/build_interval_fixture.py
python scripts/build_catalog.py
python scripts/audit_collection.py
python scripts/check_repository.py
python scripts/certify_release.py --preflight

The repository checker validates every manifest, file inventory, byte count, SHA-256 digest and expected result. It also rejects undeclared files, schema drift, unsupported metadata and prohibited attribution traces.

Citation

Zenodo releases list Vivaan Patni as the author. GitHub development and commits use the account vxxqv. Cite the exact archived release used in an analysis. The concept DOI 10.5281/zenodo.22551734 resolves to the latest archived version.