IWC-bench

Bioinformatics agent benchmark from peer-reviewed Galaxy workflows with CI-verified ground truth.

Derived from the Intergalactic Workflow Commission (IWC) — peer-reviewed Galaxy workflows with Planemo CI tests. Each task converts a community-validated workflow into a standalone CLI benchmark with real Zenodo test data and ground truth verified by Galaxy's continuous integration.

This benchmark is actively growing — new tasks are added continuously from the IWC catalog.

What makes this benchmark different?

• Peer-reviewed workflows: Every workflow passed human review + automated CI before IWC publication
• CI-verified ground truth: Expected outputs come from Planemo test assertions that run on every Galaxy release
• Real experimental data: All test data hosted on Zenodo with permanent DOIs — no simulated reads
• Dual prompt system: Each task has an open prompt (biological question only) and a guided prompt (with tool names) — measures biological reasoning vs. tool execution ability
• Task-specific rubrics: Auto-generated from Galaxy workflow DAGs with expected steps, artifacts, and scoring weights
• Complex DAGs: Workflows range from 1 to 41 biological steps, depth up to 21, with multi-branch convergence — far beyond simple linear pipelines
• Provenance tracking: Each task includes the original .ga workflow + Planemo test file for full reproducibility

Tasks (75 total)

Difficulty distribution: 9 easy, 29 medium, 27 hard, 10 expert — sorted by difficulty below.

#	Task ID	Name	Difficulty	Steps	Domain
1	avg-bigwig	Average BigWig Between Replicates	easy	1	other
2	fastq-to-matrix-10x	Single-Cell RNA-seq Preprocessing: 10X CellPlex Multiplexed	easy	1	scRNAseq
3	velocyto	RNA Velocity Analysis: Velocyto for 10X Data	easy	1	scRNAseq
4	parallel-accession-download	Parallel SRA Accession Download	easy	2	data
5	mag-genome-annotation-parallel	MAG Genome Annotation Parallel	easy	2	microbiome
6	short-read-qc	Short-Read Quality Control and Trimming	easy	2	QC
7	brew3r	BREW3R 3-prime UTR Extension from RNA-seq	easy	3	transcriptomics
8	vgp0-mitogenome	Vgp0 Mitogenome	easy	3	VGP assembly
9	sra-manifest-fastqs	SRA Manifest to Concatenated FASTQs	easy	4	other
10	baredsc	Single-Cell Bayesian Gene Expression Density Estimation	medium	2	scRNAseq
11	cgmlst-bacterial-genome	core genome MLST Bacterial Typing	medium	3	bacterial
12	hyphy	Phylogenetic Selection Analysis	medium	4	comp-genomics
13	assembly-with-flye	Genome Assembly with Flye	medium	4	assembly
14	qcxms-sdf	EI Mass Spectra Prediction from Molecular Structure	medium	4	metabolomics
15	lncrna-annotation	Lncrna Annotation	medium	4	annotation
16	bacterial-genome-assembly	Bacterial Genome Assembly from Short Reads	medium	5	assembly
17	raw-reads-qc	Raw Read Quality and Contamination Control	medium	5	other
18	goseq	Gene Ontology and KEGG Pathway Enrichment Analysis	medium	5	transcriptomics
19	vgp-plot-nx-size	VGP Nx and Size Plots for Assemblies	medium	6	VGP assembly
20	protein-ligand-param	Protein-Ligand Complex Parameterization for MD	medium	6	other
21	rnaseq-pe	RNA-Seq Paired-End Analysis and Quantification	medium	6	transcriptomics
22	rnaseq-sr	RNA-Seq Single-End Analysis and Quantification	medium	6	transcriptomics
23	annotation-braker3	Annotation Braker3	medium	6	annotation
24	nuclei-segmentation	Fluorescence Nuclei Segmentation and Counting	medium	7	imaging
25	gcms-metams	GC-MS Metabolomics Data Processing	medium	7	metabolomics
26	pseudobulk-decoupler-edger	Single-Cell Pseudobulk Differential Expression Analysis	medium	7	other
27	openms-metaprosip	MetaProSIP Stable Isotope Probing Proteomics	medium	7	proteomics
28	rnaseq-de	RNA-Seq Differential Expression with Visualization	medium	7	transcriptomics
29	annotation-helixer	Annotation Helixer	medium	7	annotation
30	vgp1-kmer-profiling	VGP1 K-mer Profiling and Read Statistics	medium	8	VGP assembly
31	bacterial-genome-annotation	Bacterial Genome Annotation	medium	8	bacterial
32	polish-long-reads	Assembly Polishing with Long Reads	medium	8	other
33	vgp7-scaffolding-bionano	Vgp7 Scaffolding Bionano	medium	8	VGP assembly
34	fragment-docking	Fragment-Based Virtual Screening with Docking and Scoring	medium	9	comp-chem
35	sars-cov2-consensus	SARS-CoV-2 Consensus Construction from Variants	medium	9	SARS-CoV-2
36	sars-cov2-se-wgs	SARS-CoV-2 SE Illumina WGS Variant Calling	medium	9	SARS-CoV-2
37	vgp2-kmer-trio	VGP K-mer Profiling HiFi Trio	medium	10	VGP assembly
38	lcms-preprocessing	LC-MS Metabolomics Preprocessing	medium	12	metabolomics
39	chipseq-pe	ChIP-seq Paired-End Analysis	hard	5	epigenetics
40	chipseq-sr	ChIP-seq Single-End Analysis	hard	5	epigenetics
41	bacterial-qc-post-assembly	Post-Assembly Quality Control and Contamination Check	hard	7	bacterial
42	cutandrun	CUT&RUN Protein-DNA Interaction Mapping	hard	7	epigenetics
43	mags-taxonomy-annotation	MAGs Taxonomy Annotation and Classification	hard	7	microbiome
44	amr-gene-detection	AMR Gene Detection	hard	7	other
45	annotation-maker	Genome Annotation with Maker	hard	8	annotation
46	mfassignr	Molecular Formula Assignment from FT-MS Data	hard	9	metabolomics
47	vgp9-decontamination	Vgp9 Decontamination	hard	9	VGP assembly
48	haploid-vc	Haploid Variant Calling WGS PE	hard	11	other
49	ploidy-vc	Ploidy-aware Genotype + Variant Calling WGS PE	hard	11	other
50	sars-cov2-pe-wgs	SARS-CoV-2 PE Illumina WGS Variant Calling	hard	11	SARS-CoV-2
51	dada2	DADA2 16S Amplicon Analysis Pipeline	hard	12	amplicon
52	gromacs-mmgbsa	GROMACS MMGBSA Free Energy Calculation	hard	12	comp-chem
53	generic-vc	Generic WGS PE Variant Calling	hard	12	other
54	metagenomic-raw-reads-amr	Metagenomics Taxonomic and AMR Gene Analysis from Raw Reads	hard	13	microbiome
55	sars-cov2-ivar	SARS-CoV-2 iVar Amplicon Pipeline with Lineage Classification	hard	13	SARS-CoV-2
56	gromacs-dctmd	GROMACS dcTMD Free Energy Simulations	hard	14	comp-chem
57	consensus-peaks	Consensus Peaks from ATAC-seq Replicates	hard	14	epigenetics
58	viral-generic	Generic Non-segmented Viral Variant Calling	hard	14	other
59	variation-reporting	Generic Variation Analysis Reporting	hard	14	SARS-CoV-2
60	vgp6b-purge-dups-haplotype	Vgp6B Purge Dups Haplotype	hard	14	VGP assembly
61	sars-cov2-ont-artic	SARS-CoV-2 ONT ARTIC Variant Calling	hard	16	SARS-CoV-2
62	sars-cov2-reporting	SARS-CoV-2 Variation Analysis Reporting	hard	16	SARS-CoV-2
63	vgp3-hifi-assembly	VGP3 HiFi-only Genome Assembly	hard	20	VGP assembly
64	vgp5-hifi-trio-assembly	VGP5 HiFi+Trio Phased Genome Assembly	hard	20	VGP assembly
65	influenza	Influenza A Isolate Subtyping and Consensus	hard	20	virology
66	atacseq	ATAC-seq Chromatin Accessibility	expert	16	epigenetics
67	sars-cov2-pe-artic	SARS-CoV-2 ARTIC PE Illumina Variant Calling	expert	20	SARS-CoV-2
68	vgp4-hifi-hic-assembly	VGP4 HiFi+HiC Phased Genome Assembly	expert	23	VGP assembly
69	vgp6-purge-dups	VGP6 Purge Duplicate Contigs	expert	23	VGP assembly
70	metagenomic-genes-catalogue	Metagenomic Genes Catalogue Analysis	expert	24	microbiome
71	hic-contact-map	Hic Contact Map	expert	26	VGP assembly
72	mags-building	Metagenome-Assembled Genomes (MAGs) Generation	expert	30	microbiome
73	vgp8-scaffolding-hic	Vgp8 Scaffolding Hic	expert	32	VGP assembly
74	binning-evaluation	MAGs Binning Evaluation with CAMI AMBER	expert	35	microbiome
75	scanpy-clustering	Single-Cell RNA-seq Analysis: Scanpy Preprocessing and Clustering	expert	41	scRNAseq

Quick start

# Clone and install
git clone https://github.com/lingzhi227/IWC-bench.git
cd IWC-bench
pip install click requests

# List tasks
python src/dataset.py list-tasks

# Download input data for a specific task
python src/dataset.py download --task atacseq

# Download all input data
python src/dataset.py download --all

# View ground truth (Planemo assertions)
python src/dataset.py show-assertions --task atacseq

Task format

tasks/{task-id}/
  prompt_open.md          # Biological question only (no tool names)
  prompt_guided.md        # Pipeline steps with tool names
  rubric.json             # Task-specific evaluation rubric (auto-generated)
  environment.yml         # Conda dependencies
  Dockerfile              # Reproducible container
  workflow/
    {name}.ga             # Original IWC peer-reviewed workflow (provenance)
    {name}-tests.yml      # Planemo test assertions (ground truth)

Dual prompt system

• prompt_open.md: Describes the biological question only. The agent must figure out which tools to use, what pipeline to build, and what parameters to set. Tests biological reasoning + autonomous planning.
• prompt_guided.md: Provides the pipeline structure with tool names (e.g., "use Bowtie2 to align, MACS2 to call peaks"). Tests whether the agent can execute a known pipeline correctly.

The delta between open and guided scores reveals where agents struggle — biological reasoning vs. tool execution.

Evaluation rubrics

Each rubric.json contains:

• expected_steps: Pipeline steps with primary tools and valid alternatives
• expected_artifacts: Output files with Planemo assertions (has_text, has_n_lines, has_size)
• metadata: Difficulty (easy/medium/hard/expert), domain, DAG depth, convergence points
• scoring_weights: 40% pipeline completion + 40% result correctness + 20% biological validity

Ground truth

Derived from Planemo test assertions — the same CI checks Galaxy uses to validate every workflow:

• workflow/{name}-tests.yml — full Planemo test specification
• rubric.json — structured assertions mapped to pipeline steps

What "CI-verified ground truth" means

Every workflow in IWC-bench has ground truth that was verified by automated execution, not just written by a human. Specifically:

1. Planemo CI pipeline: The IWC repository runs a GitHub Actions CI pipeline on every pull request. This pipeline:

   • Spins up a real Galaxy server instance
   • Feeds the test inputs (from Zenodo URLs) into the workflow
   • Executes the entire workflow end-to-end on the Galaxy server
   • Checks every output assertion (has_text, has_n_lines, has_size, has_text_matching, etc.) against the actual workflow output
   • Blocks the PR from merging if any assertion fails

2. What this guarantees: Every -tests.yml file in this benchmark has been proven correct by execution — the assertions match real workflow output, not just what an author believed the output should be. This eliminates:

   • Hand-written assertions with typos or incorrect expected values
   • Assertions that were correct for an older tool version but are now stale
   • Assertions that were never actually tested

3. Ongoing validation: IWC workflows are re-tested on Galaxy release candidates, so assertions stay current with tool updates. If a Galaxy upgrade breaks a workflow, the test fails and the assertion is updated before release.

This is the key differentiator from other benchmarks where ground truth is manually curated — CI-verified means machine-checked correctness.

Source

75 tasks converted from the IWC repository, which hosts peer-reviewed Galaxy workflows with Planemo CI tests. All tasks have Zenodo-hosted test data and CI-verified assertions.

Evaluation

Tasks are evaluated by checking:

• Pipeline completion: Did the agent execute all necessary analysis steps?
• Result correctness: Do output values match the Planemo-verified ground truth?
• Biological validity: Are the results biologically meaningful?

Citation

Based on IWC workflows:

@misc{iwc2024,
  title={Intergalactic Workflow Commission},
  author={Galaxy Community},
  url={https://github.com/galaxyproject/iwc},
  year={2024}
}

Based on BioAgentBench evaluation framework:

@article{patino2025bioagentbench,
  title={BioAgentBench: A Benchmark for Evaluating LLM Agents in Bioinformatics},
  author={Patino, Luis and others},
  journal={arXiv preprint arXiv:2601.21800},
  year={2025}
}