GTN-bench

Bioinformatics agent benchmark from Galaxy Training Network peer-reviewed tutorials.

Derived from the Galaxy Training Network (GTN) — 292 peer-reviewed bioinformatics tutorials. Each task converts a community-validated tutorial workflow into a standalone CLI benchmark with real experimental data and Planemo test assertions as ground truth.

Companion to IWC-bench (75 tasks from IWC peer-reviewed workflows with CI-verified ground truth).

What makes this benchmark different?

• Peer-reviewed tutorials: Every workflow comes from a GTN tutorial reviewed by domain-expert maintainers
• Real experimental data: Input data hosted on Zenodo with permanent DOIs or bundled as local test-data
• Planemo test assertions as ground truth: Expected outputs defined in Planemo test format (has_text, has_n_lines, has_line, etc.)
• Dual prompt system: Each task has an open prompt (biological question only) and a guided prompt (with tool names)
• Task-specific rubrics: Auto-generated from Galaxy workflow DAGs with expected steps, artifacts, and scoring weights
• Wide domain coverage: 14 domains from assembly to digital humanities, single-cell to computational chemistry
• Complex workflows: Up to 57 biological steps, DAG depth up to 20, with multi-branch convergence

Ground truth model

GTN-bench uses a two-level verification model:

• Level 1 (structural, all tasks): Workflow parses correctly, test YAML has valid assertions (>=3), input data URLs are accessible. Automated and verified for all 53 tasks.
• Level 2 (execution, sample): Full pipeline execution — download data, run tools, check output assertions match. Done on representative tasks to validate ground truth reliability.

Note: Unlike IWC-bench where every workflow is CI-execution-verified on every PR, GTN workflow tests are format-validated but not automatically executed in GTN's CI. The assertions were generated via planemo workflow_test_init from actual Galaxy server invocations by tutorial authors. See filter.md for full verification analysis.

Tasks (53 total)

Difficulty distribution: 7 easy, 23 medium, 13 hard, 10 expert — all 53 with >=3 ground truth assertions.

#	Task ID	Name	Difficulty	Steps	Domain
1	abc_intro_phylo	Phylogenetic Tree Building	easy	3	evolution
2	bacterial-isolate-species-contamination-checking	Bacterial Isolate Species and Contamination Check	easy	3	ecology
3	classification_regression	Classification and Regression	easy	3	statistics
4	debruijn-graph-assembly	De Bruijn Graph Genome Assembly	easy	3	assembly
5	mitochondrion-assembly	Mitochondrial Genome Assembly	easy	3	assembly
6	analysis-md-simulations	Molecular Dynamics Trajectory Analysis	easy	5	comp-chem
7	galaxy-intro-101-everyone	Galaxy Introduction: Data Analysis Basics	easy	5	introductory
8	flexynesis_survival	Flexynesis Survival Analysis	medium	3	statistics
9	encyclopedia	DIA Metaproteomics with EncyclopeDIA	medium	4	proteomics
10	ewas-suite	Epigenome-Wide Association Study (Methylation BeadChip)	medium	4	epigenetics
11	small_ncrna_clustering	Small Non-coding RNA Clustering	medium	4	transcriptomics
12	assembly-decontamination	Genome Assembly Decontamination	medium	5	assembly
13	introduction_to_dh	Introduction to Text Analysis for Digital Humanities	medium	5	digital-humanities
14	marine_omics_bgc	Marine Biosynthetic Gene Cluster Annotation	medium	5	ecology
15	quality-contamination-control	Quality and Contamination Control	medium	5	sequence-analysis
16	scrna-data-ingest	Single-Cell RNA-seq Data Ingestion and Format Conversion	medium	5	scRNAseq
17	scrna-scater-qc	Single-Cell RNA-seq Quality Control (scater)	medium	5	scRNAseq
18	braker3	Genome Annotation with Gene Prediction	medium	6	annotation
19	metatranscriptomics-short	Metatranscriptomics Short Analysis	medium	6	microbiome
20	mrsa-illumina	MRSA Genome Assembly from Illumina MiSeq	medium	6	assembly
21	unicycler-assembly	Hybrid Bacterial Genome Assembly (Illumina + Nanopore)	medium	6	assembly
22	flexynesis_classification	Flexynesis Classification Analysis	medium	7	statistics
23	med-chem-data	Medicinal Chemistry Data Import and Analysis	medium	7	comp-chem
24	pseudobulk-analysis	Pseudobulk Differential Expression Analysis	medium	7	scRNAseq
25	scrna-case_alevin-combine-datasets	Single-Cell RNA-seq: Combining Multiple Datasets	medium	7	scRNAseq
26	mapping	Sequence Read Mapping	medium	8	sequence-analysis
27	scrna-raceid	Single-Cell RNA-seq RaceID Analysis	medium	8	scRNAseq
28	comparison-braker-helixer-annotation	Genome Annotation Tool Comparison	medium	9	annotation
29	metaquantome-data-creation	Metaquantome Data Preparation	medium	9	proteomics
30	taxonomic-profiling	Taxonomic Profiling	medium	9	microbiome
31	flexynesis_unsupervised	Flexynesis Unsupervised Analysis	hard	8	statistics
32	metagenomics-assembly	Metagenomic Sequencing Assembly	hard	9	assembly
33	qcxms-predictions	In-silico Mass Spectra Prediction (QM-based)	hard	9	metabolomics
34	scrna-case_alevin	Single-Cell RNA-seq: Count Matrix Generation	hard	11	scRNAseq
35	galaxy-intro-ngs-data-managment	NGS Data Logistics and Management	hard	12	introductory
36	hybrid_denovo_assembly	Hybrid De Novo Assembly (Nanopore + Illumina)	hard	12	assembly
37	ref-based	Reference-Based RNA-Seq Differential Expression	hard	12	transcriptomics
38	tb-variant-analysis	TB Variant Analysis	hard	12	variant-calling
39	mrsa-nanopore	MRSA Genome Assembly from Nanopore MinION	hard	13	assembly
40	tnseq	Essential Gene Detection with Transposon Insertion Sequencing	hard	16	annotation
41	gtn-hic-drosophila	Hi-C 3D Genome Organization Analysis (Drosophila)	hard	18	epigenetics
42	clustering_machinelearning	Machine Learning Clustering	hard	20	statistics
43	mirna-target-finder	Whole Transcriptome Analysis and miRNA Target Prediction	hard	25	transcriptomics
44	Ecoregionalization_tutorial	Marine Ecoregionalization Analysis	expert	18	ecology
45	metaplasmidome_query	Metaplasmidome Plasmid Identification and Annotation	expert	23	microbiome
46	ERGA-post-assembly-QC	Post-Assembly Quality Control (ERGA Pipeline)	expert	24	assembly
47	somatic-variant-discovery	Somatic Copy Number Variant Detection	expert	24	variant-calling
48	metatranscriptomics	Metatranscriptomics Analysis from Microbiome RNA-seq	expert	27	microbiome
49	scatac-standard-processing-snapatac2	Single-Cell ATAC-seq Processing	expert	29	scRNAseq
50	formation_of_super-structures_on_xi	ChIP-seq: Super-Structure Formation on X Chromosome	expert	35	epigenetics
51	differential-isoform-expression	Differential Isoform Expression Analysis	expert	38	transcriptomics
52	scrna-case_basic-pipeline	Single-Cell RNA-seq Basic Analysis Pipeline	expert	41	scRNAseq
53	tal1-binding-site-identification	TAL1 Transcription Factor Binding Site Identification	expert	57	epigenetics

Quick start

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

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

# Show task details
python src/dataset.py show-task --task mrsa-nanopore

# View ground truth assertions
python src/dataset.py show-assertions --task mrsa-nanopore

# Download input data for a task
python src/dataset.py download --task mrsa-nanopore

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

Task format

tasks/{task-id}/
  prompt_open.md          # Biological question only (no tool names)
  prompt_guided.md        # Pipeline steps with tool names
  rubric.json             # Evaluation rubric (steps, assertions, scoring)
  environment.yml         # Conda dependencies
  Dockerfile              # Reproducible container
  workflow/
    {name}.ga             # Original GTN peer-reviewed workflow (provenance)
    {name}-test*.yml      # Planemo test assertions (ground truth)
    test-data/            # Local test data files (if any)

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 Unicycler to assemble, Prokka to annotate"). 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_line, etc.)
• metadata: Difficulty (easy/medium/hard/expert), domain, DAG metrics, tool count
• scoring_weights: 40% pipeline completion + 40% result correctness + 20% biological validity

Ground truth

Derived from Planemo test assertions — the same format Galaxy uses to validate workflows:

• workflow/{name}-test*.yml — Planemo test specification with input data URLs and output assertions
• rubric.json — structured assertions mapped to pipeline steps

Input data

Input data comes from two sources:

• Zenodo (most tasks): Permanent DOI-backed URLs referenced in the test YAML
• Local test-data (some tasks): Small files bundled in workflow/test-data/

Use python src/dataset.py download --task <id> to fetch all input data for a task.

Domain coverage

Domain	Tasks	Examples
Assembly	9	De Bruijn graph, hybrid Nanopore+Illumina, MRSA, mitochondrial, metagenomics, ERGA QC
Single-cell	8	scRNA-seq (Alevin, RaceID, scater, basic pipeline), scATAC-seq, pseudobulk, data ingestion
Statistics	5	Classification/regression, flexynesis (survival, classification, unsupervised), ML clustering
Epigenetics	4	EWAS methylation, Hi-C 3D genome, ChIP-seq super-structures, TAL1 binding
Transcriptomics	4	RNA-seq DE, miRNA targets, differential isoform expression, small ncRNA clustering
Microbiome	4	Metaplasmidome, metatranscriptomics (full + short), taxonomic profiling
Ecology	3	Ecoregionalization, marine BGC, bacterial isolate checking
Annotation	3	Braker3, tool comparison, TnSeq essential genes
Comp-chem	2	MD trajectory analysis, medicinal chemistry
Introductory	2	Galaxy 101, NGS data management
Proteomics	2	DIA metaproteomics, metaquantome
Sequence analysis	2	Quality/contamination control, read mapping
Variant-calling	2	Somatic CNV detection, TB variant analysis
Evolution	1	Phylogenetic tree building
Digital humanities	1	Text analysis introduction
Metabolomics	1	QCxMS mass spectra prediction

Evaluation

Tasks are evaluated by checking:

1. Pipeline completion (40%): Did the agent execute all necessary analysis steps?
2. Result correctness (40%): Do output values match the Planemo test assertions?
3. Biological validity (20%): Are the results biologically meaningful?

Assertions are checked against the Planemo test YAML:

outputs:
  assembly_info:
    asserts:
      has_text:
        text: "contig_1"
      has_n_lines:
        n: 2

For evaluation with an LLM agent harness, see IWC-bench which provides scripts/run_agent.py (3-tool agent harness) and scripts/judge_agent.py (agentic LLM judge) — both compatible with GTN-bench's identical format.

Source

Filtered from 292 GTN tutorials:

• 147 had workflow + test files
• 141 had parseable output assertions
• 53 passed verification (52 Level 1 + 1 Level 2 verified pilot)

Full filtering analysis: filter.md

Citation

Galaxy Training Network tutorials are CC-BY-4.0 licensed.

@article{hiltemann2023galaxy,
  title={Galaxy Training: A powerful framework for teaching!},
  author={Hiltemann, Saskia and Rasche, Helena and Serrano-Solano, B{\'e}r{\'e}nice and others},
  journal={PLoS Computational Biology},
  volume={19},
  number={1},
  pages={e1010752},
  year={2023}
}

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}
}