nf-core-bench

Bioinformatics agent benchmark from nf-core peer-reviewed Nextflow pipelines.

Derived from nf-core — 147 peer-reviewed Nextflow bioinformatics pipelines. Each task converts a CI-tested pipeline into a standalone benchmark with nf-test snapshot assertions as ground truth.

Companion to IWC-bench (75 Galaxy tasks) and GTN-bench (53 Galaxy tasks).

What makes this benchmark different?

• CI-verified execution: Every nf-core pipeline is tested in GitHub Actions CI across Docker, Conda, and Singularity profiles
• MD5 snapshot ground truth: nf-test snapshots contain file trees + MD5 checksums of stable output files
• Peer-reviewed: Release PRs require 2 reviewer approvals (at least 1 core team), 30+ automated lint checks
• Standardized invocation: Every pipeline runs via nextflow run nf-core/<pipeline> -profile test,docker
• Real sub-sampled test data: Hosted on GitHub (nf-core/test-datasets), publicly accessible
• Wide domain coverage: Transcriptomics, variant calling, metagenomics, epigenetics, proteomics, immunogenomics, virology, and more

Tasks (65 total)

Difficulty distribution: 22 easy, 21 medium, 11 hard, 11 expert — sorted by difficulty below.

#	Pipeline	Difficulty	Domain	MD5 Assertions	Total Assertions
1	cageseq	easy	sequence-analysis	0	0
2	callingcards	easy	other	8	8
3	chipseq	easy	epigenetics	31	31
4	circdna	easy	other	10	10
5	createtaxdb	easy	metagenomics	35	36
6	denovotranscript	easy	transcriptomics	0	0
7	drugresponseeval	easy	other	7	8
8	fastqrepair	easy	other	3	3
9	fastquorum	easy	other	15	15
10	fetchngs	easy	metagenomics	3	5
11	funcscan	easy	metagenomics	83	84
12	mag	easy	metagenomics	383	385
13	metaboigniter	easy	metagenomics	0	0
14	metapep	easy	metagenomics	116	116
15	mhcquant	easy	proteomics	0	2
16	multiplesequencealign	easy	sequence-analysis	0	0
17	pacvar	easy	variant-calling	1	3
18	pangenome	easy	pangenomics	0	0
19	phaseimpute	easy	other	47	48
20	phyloplace	easy	other	25	25
21	proteinfold	easy	proteomics	87	89
22	readsimulator	easy	simulation	0	0
23	reportho	easy	other	8	10
24	ribomsqc	easy	other	6	7
25	scnanoseq	easy	scRNAseq	0	0
26	viralmetagenome	easy	variant-calling	324	326
27	airrflow	medium	other	250	252
28	ampliseq	medium	metagenomics	159	160
29	bacass	medium	assembly	211	213
30	bamtofastq	medium	other	29	31
31	detaxizer	medium	other	19	20
32	differentialabundance	medium	other	24	26
33	eager	medium	ancient-dna	0	1
34	genomeassembler	medium	assembly	18	20
35	hic	medium	3d-genome	20	21
36	hlatyping	medium	immunogenomics	673	675
37	isoseq	medium	transcriptomics	38	39
38	metatdenovo	medium	metagenomics	11	13
39	methylong	medium	epigenetics	16	18
40	molkart	medium	other	52	54
41	nanostring	medium	gene-expression	17	19
42	pairgenomealign	medium	sequence-analysis	22	23
43	pathogensurveillance	medium	other	0	1
44	rnafusion	medium	transcriptomics	108	110
45	rnaseq	medium	transcriptomics	948	950
46	scrnaseq	medium	transcriptomics	385	387
47	viralintegration	medium	virology	8	9
48	coproid	hard	other	0	1
49	cutandrun	hard	epigenetics	78	79
50	demultiplex	hard	data-processing	671	673
51	methylseq	hard	epigenetics	753	754
52	nascent	hard	other	300	302
53	oncoanalyser	hard	transcriptomics	0	2
54	proteinannotator	hard	proteomics	47	49
55	proteinfamilies	hard	proteomics	8	10
56	rnavar	hard	transcriptomics	128	130
57	sarek	hard	variant-calling	1376	1378
58	taxprofiler	hard	metagenomics	19	20
59	atacseq	expert	epigenetics	194	196
60	bactmap	expert	other	129	131
61	epitopeprediction	expert	other	32	34
62	magmap	expert	metagenomics	76	78
63	rangeland	expert	metagenomics	0	2
64	raredisease	expert	variant-calling	0	2
65	riboseq	expert	other	256	258
66	seqinspector	expert	other	633	635
67	smrnaseq	expert	transcriptomics	578	580
68	variantbenchmarking	expert	variant-calling	221	223
69	viralrecon	expert	variant-calling	610	611

Quick start

# Clone
git clone https://github.com/lingzhi227/nf-core-bench.git
cd nf-core-bench

# List tasks
python3 src/dataset.py list-tasks -v

# Show task details
python3 src/dataset.py show-task --task rnaseq

# Run a pipeline test (requires Nextflow + Docker)
nextflow run nf-core/rnaseq -profile test,docker --outdir results

Task format

tasks/{pipeline}/
  prompt_open.md              # Biological question (no tool names)
  prompt_guided.md            # Pipeline invocation command
  rubric.json                 # Ground truth: MD5 checksums, file tree, scoring
  test_config.txt             # Original conf/test.config content
  nfcore/
    test.config               # Nextflow test configuration
    nextflow_schema.json      # Pipeline parameter schema
    tests/*.nf.test           # nf-test definitions
    tests/*.nf.test.snap      # Snapshot assertions (ground truth)
    tests/.nftignore          # Files excluded from MD5 checks

Ground truth model

nf-core uses nf-test snapshots as ground truth:

• MD5 checksums: Exact byte-level verification of stable output files
• File tree: Complete list of expected output paths
• Task count: Expected number of successful Nextflow tasks
• Software versions: Pinned tool versions for reproducibility

Files excluded from MD5 checking (via .nftignore): BAMs, HTML reports, logs, PNGs — anything non-deterministic.

Evaluation

40% pipeline completion (did it produce the expected file tree?)
40% result correctness (do MD5 checksums match snapshot?)
20% biological validity (are results biologically meaningful?)

Citation

@article{ewels2020nf,
  title={The nf-core framework for community-curated bioinformatics pipelines},
  author={Ewels, Philip A and Peltzer, Alexander and Fillinger, Sven and others},
  journal={Nature Biotechnology},
  volume={38},
  pages={276--278},
  year={2020}
}