File size: 9,406 Bytes
7121295
 
 
 
8f1e292
 
7121295
8f1e292
7121295
16ce2fd
7121295
8f1e292
16ce2fd
7121295
 
 
 
 
 
 
 
 
 
16ce2fd
0fc8227
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e83ebe8
4c06b16
1875280
1203e7c
53ce4d9
a52c376
f26d70f
a6f2f1c
68b0c15
86bc5d8
42758e8
86dc2be
aee1a89
2195b4c
a2e13bd
6c45775
 
4cdf23c
 
a7de712
3905e19
 
5313598
cbe197c
 
 
 
 
 
 
 
 
 
a7c391b
 
 
 
 
 
 
 
 
 
bb88453
087ddb7
f7dcab4
ee6036f
d199590
16ce2fd
7121295
 
8f1e292
01c7a5b
7121295
 
 
 
 
 
 
 
 
 
 
 
 
8f1e292
01c7a5b
7121295
 
 
 
 
 
 
 
 
8f1e292
 
7121295
8f1e292
01c7a5b
 
7121295
 
 
01c7a5b
8f1e292
7121295
8f1e292
7121295
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
---
license: mit
---

# Extended-BioAgentBench

> **A growing benchmark for evaluating LLM agents on complex bioinformatics workflows.**

[![Dataset on HF](https://img.shields.io/badge/Dataset-HuggingFace-yellow)](https://huggingface.co/datasets/lingzhi227/Extended-BioAgentBench)
[![Tasks](https://img.shields.io/badge/Tasks-71-blue)]()
[![Based on](https://img.shields.io/badge/Based%20on-BioAgentBench-green)](https://github.com/bioagent-bench/bioagent-bench)

Building on [BioAgentBench](https://arxiv.org/abs/2601.21800) (10 tasks), this benchmark adds **71 new tasks** that test LLM agents on increasingly complex, multi-tool bioinformatics pipelines. Tasks span 6 domains and range from simple linear workflows to depth-8 diamond DAGs with 16+ CLI tools.

**This benchmark is actively growing** — new tasks are added continuously to cover more domains, increase complexity, and push the boundaries of what AI agents can do in bioinformatics.

## What makes this benchmark different?

- **Diamond DAG complexity**: Tasks require running multiple independent tool branches that converge — not just linear pipelines
- **No pipeline leakage**: Task prompts describe *what* to produce, never *which tools* to use or *how* to structure the pipeline
- **Domain-specific traps**: Tasks include steps where default parameters silently produce wrong results (e.g., Tn5 shift correction in ATAC-seq, Medaka model selection for Nanopore)
- **Real public data**: Every task uses published datasets with ground truth generated by validated reference pipelines

## Tasks (71 total)

| # | Task ID | Name |
|---|---------|------|
| 11 | `chipseq-peak-calling` | ChIP-seq Peak Calling: TAL1 Binding Site Comparison |
| 12 | `bacterial-assembly` | Bacterial Genome Assembly and Annotation: MRSA Characte |
| 13 | `mobile-elements` | Bacterial Mobile Genetic Element Characterization: MRSA |
| 14 | `outbreak-investigation` | Foodborne Pathogen Outbreak Investigation via WGS Phylo |
| 15 | `atacseq-accessibility` | ATAC-seq Chromatin Accessibility Profiling |
| 16 | `longread-assembly` | Nanopore Long-read Bacterial Genome Assembly |
| 17 | `hybrid-assembly` | Hybrid Genome Assembly from Illumina and Nanopore Data |
| 18 | `sv-detection` | Bacterial Structural Variant and SNP Detection |
| 19 | `pangenome-evolution` | E. coli Pan-genome and Core Phylogeny |
| 20 | `metagenomic-profiling` | Metagenomic Assembly and Functional Profiling |
| 21 | `phage-characterization` | Bacteriophage Genome Assembly and Functional Characteri |
| 22 | `genome-comparison` | Pairwise Bacterial Genome Comparison |
| 23 | `mapping-qc` | Genome Mapping and Coverage Quality Assessment |
| 24 | `multisample-variants` | Multi-sample Variant Calling and Comparison |
| 25 | `consensus-genome` | Bacterial Consensus Genome Generation |
| 26 | `gene-prediction` | Gene Prediction Method Comparison |
| 27 | `downsampling-analysis` | Read Downsampling and Assembly Quality Titration |
| 28 | `plasmid-typing` | Plasmid Detection and Replicon Typing |
| 29 | `genome-completeness` | Genome Completeness and Quality Assessment |
| 30 | `species-identification` | Multi-reference Bacterial Species Identification |
| 31 | `viral-amplicon` | Viral Amplicon Surveillance Analysis |
| 32 | `bisulfite-methylation` | Bisulfite Sequencing DNA Methylation Analysis |
| 33 | `rnaseq-isoform` | RNA-seq Isoform Assembly and Quantification |
| 34 | `ancient-dna` | Ancient DNA Authentication and Damage Assessment |
| 35 | `mirna-seq` | Small RNA-seq miRNA Discovery and Quantification |
| 36 | `gcms-metabolomics` | GC-MS Metabolomics Profiling: Brown Algae Salinity Adap |
| 37 | `cutandrun` | CUT&RUN Epigenomic Profiling |
| 38 | `scrna-full-pipeline` | Single-Cell RNA-seq Full Pipeline: Multi-Quantifier Ana |
| 39 | `crispr-screen` | CRISPR Screen Analysis: Drug Sensitivity Gene Discovery |
| 40 | `amplicon-microbiome` | 16S Amplicon Microbiome: Community Profiling and Functi |
| 41 | `rna-fusion` | RNA Fusion Detection from RNA-seq |
| 42 | `spatial-transcriptomics` | Spatial Transcriptomics: Visium FFPE Brain Cancer Analy |
| 43 | `taxonomic-profiling` | Multi-classifier Taxonomic Profiling of Metagenomic Rea |
| 44 | `lcms-metabolomics` | LC-MS Untargeted Metabolomics: Urine Feature Discovery |
| 45 | `somatic-variant-calling` | Somatic Variant Calling: Tumor-Normal Paired Analysis |
| 46 | `amr-bgc-screening` | Antimicrobial Resistance and Biosynthetic Gene Cluster  |
| 47 | `variant-trio` | Variant Annotation Trio: Clinical Interpretation of Ash |
| 48 | `clinical-metaproteomics` | Clinical Metaproteomics: Multi-Engine Marine Microbiome |
| 49 | `mhc-immunopeptidomics` | MHC Immunopeptidomics: Peptide Identification and Quant |
| 50 | `riboseq` | Ribosome Profiling Translation Analysis |
| 51 | `neoantigen-prediction` | Neoantigen Prediction: Tumor-Normal Somatic Analysis |
| 52 | `somatic-germline-dual` | Somatic+Germline Dual Analysis: Hereditary Cancer Varia |
| 53 | `hicar-chromatin` | HiCAR Chromatin Interaction: Proximity Ligation and Acc |
| 54 | `radseq-popgen` | RADseq Population Genetics: Stickleback Freshwater-Mari |
| 55 | `mag-recovery` | MAG Recovery: Metagenome-Assembled Genomes from Environ |
| 56 | `viral-phylodynamics` | Viral Phylodynamics (Molecular Clock Analysis) |
| 57 | `edna-metabarcoding` | eDNA Aquatic Metabarcoding Biodiversity Assessment |
| 58 | `metatranscriptomics` | Metatranscriptomics: Active Microbial Community Profili |
| 59 | `nascent-transcription` | Nascent Transcription: GRO-seq Polymerase Activity Prof |
| 60 | `circrna-discovery` | Circular RNA Discovery: C. elegans Wild-type vs fust-1  |
| 61 | `hic-3d-conformation` | Hi-C 3D Genome Conformation Analysis |
| 62 | `genome-scaffolding` | Long-Read Genome Scaffolding of Fragmented Assembly |
| 63 | `longread-rna-isoform` | Long-Read RNA Isoform Discovery from Direct RNA Sequenc |
| 64 | `circrna-detection` | Circular RNA Detection and Quantification |
| 65 | `pharmacogenomics` | CYP2D6 Pharmacogenomic Star Allele Calling |
| 66 | `rna-editing-detection` | RNA Editing Detection: A-to-I Editing from Matched RNA/ |
| 67 | `cnv-detection-wes` | CNV Detection from Whole-Exome Sequencing |
| 68 | `haplotype-phasing` | Haplotype Phasing and Genotype Refinement |
| 69 | `dda-proteomics-simple` | DDA Proteomics: Single-Engine BSA Identification |
| 70 | `immune-repertoire` | Immune Repertoire Analysis (BCR-seq) |
| 71 | `germline-wes-gatk` | Germline WES Variant Calling: Clinical Exome Analysis |
| 72 | `gwas-association` | GWAS Population Association Testing |
| 73 | `dia-proteomics` | Label-Free Proteomics: BSA Standard Identification |
| 74 | `structural-variant-multi` | Structural Variant Detection: Multi-caller Human SV Ana |
| 75 | `dda-lfq-proteomics` | DDA Label-Free Quantitative Proteomics |
| 76 | `clinical-wgs-interpretation` | Clinical WGS Interpretation: Full Clinical Genome Analy |
| 77 | `repeat-element-annotation` | Repeat Element Annotation (Transposable Element Analysi |
| 78 | `msi-detection` | Microsatellite Instability Detection: Multi-Caller Cons |
| 79 | `scatac-seq` | Single-Cell ATAC-seq Chromatin Accessibility Analysis |
| 80 | `multiomics-rna-atac` | Multi-omics Integration (RNA-seq + ATAC-seq) |
| 81 | `methylation-array-epic` | Methylation Array Analysis (Illumina EPIC) |

## Quick start

```bash
# Clone and install
git clone https://github.com/lingzhi227/Extended-BioAgentBench.git
cd Extended-BioAgentBench
pip install click requests

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

# Download a specific task (data + reference + ground truth)
python src/dataset.py download --task outbreak-investigation --reference --results

# Download everything
python src/dataset.py download --all --reference --results
```

## Task format

Each task follows the BioAgentBench format:

```
tasks/<task_id>/
  Dockerfile          # Reproduce ground truth with this
  environment.yml     # Conda environment specification
  run_script.sh       # Reference pipeline (ground truth generator)
```

Data, reference, and results are downloaded via `dataset.py` from [HuggingFace](https://huggingface.co/datasets/lingzhi227/Extended-BioAgentBench):

```
tasks/<task_id>/
  data/               # Input data (FASTQ, FASTA, etc.)
  reference/          # Reference genomes (if needed)
  results/            # Ground truth output
```

## Evaluation

Each task prompt provides the expected output format (CSV columns + example values). Evaluation uses **GPT-5.1 as LLM-as-Judge**, scoring:

- `steps_completed` / `steps_to_completion` — how many pipeline stages the agent executed
- `completion_rate` — fraction of the pipeline completed
- `results_match` — full_match / partial_match / no_match against ground truth

## Contributing new tasks

This benchmark is designed to grow. To add a new task:

1. Pick a bioinformatics domain not yet covered
2. Find small public data (< 1 GB, runtime < 4h on 8 CPUs)
3. Write `run_script.sh` to generate ground truth
4. Write a task prompt that says *what* to produce, not *how*
5. Verify the prompt doesn't leak tool names or pipeline structure
6. Submit a PR

## Citation

Based on BioAgentBench:
```bibtex
@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}
}
```