| --- |
| pretty_name: Agent Eval Effector Hunt |
| tags: |
| - biology |
| - genomics |
| - ai-agents |
| - agent-evaluation |
| - scientific-discovery |
| task_categories: |
| - question-answering |
| - text-generation |
| language: |
| - en |
| --- |
| |
| # Agent Eval: Effector Hunt |
|
|
| While AI scientist agents like Claude Science and Google's AI co-scientist highlight the potential of autonomous research, compact and reproducible datasets for evaluating these agents on real scientific workflows remain scarce. |
|
|
| **Agent Eval: Effector Hunt** is a genomics benchmark package designed around a real scientific discovery workflow from the *Science* paper [Chen et al. 2017](https://doi.org/10.1126/science.aao4810). It asks an AI agent, a computational biologist, or a hybrid human-agent workflow |
| to analyze anonymized paired-end sequencing reads against an anonymized fungal |
| reference genome and recover a biologically meaningful effector signal. |
|
|
| ## Task |
|
|
| Given two anonymized paired-end readsets and an anonymized reference genome with |
| annotations, identify the key effector-region difference between the samples. |
|
|
| An agent should be able to: |
|
|
| 1. Inspect the provided reference genome and annotation files. |
| 2. Align or otherwise compare the anonymized reads to the reference. |
| 3. Detect the major sample-specific genomic signal. |
| 4. Prioritize candidate effector genes or regions. |
| 5. Produce a concise evidence-backed report explaining the finding. |
|
|
| The dataset card includes an evaluator-facing rubric below so that users can |
| score agent outputs without requesting a separate answer by email. For blind |
| agent evaluation, do not include the dataset card or rubric in the agent prompt. |
|
|
| ## Dataset Contents |
|
|
| ```text |
| data/raw/fastq_anonymized/ |
| sample_1_R1.fastq.gz |
| sample_1_R2.fastq.gz |
| sample_2_R1.fastq.gz |
| sample_2_R2.fastq.gz |
| |
| reference/reference_anonymized/ |
| reference_genome.fna.gz |
| reference_annotations.gff.gz |
| reference_annotations.gtf.gz |
| reference_cds.fna.gz |
| reference_protein.faa.gz |
| checksums.md5 |
| |
| data_manifest.md |
| ``` |
|
|
| The sample names, read names, contig identifiers, gene identifiers, transcript |
| identifiers, and protein identifiers have been replaced with neutral names. |
| Read sequences, quality strings, genome sequences, protein sequences, annotation |
| coordinates, strands, feature types, and phases are preserved. |
|
|
| ## Suggested Evaluation Setup |
|
|
| For a clean agent evaluation, give the agent access to this dataset and a |
| standard command-line bioinformatics environment, but do not provide the source |
| paper, original sample names, original reference identifiers, or the private |
| answer key. |
|
|
| Recommended evaluation criteria: |
|
|
| - Correctly identifies the major genomic difference between `sample_1` and |
| `sample_2` |
| - Localizes the relevant region in the anonymized reference |
| - Connects the region to plausible effector biology using the provided |
| annotations |
| - Provides reproducible commands or analysis steps |
| - Separates evidence from speculation |
| - Reports uncertainty and checks alternative explanations |
|
|
| If internet access is enabled during evaluation, benchmark integrity may be |
| weaker because an agent could search for external provenance clues instead of |
| solving the task from the anonymized data. |
|
|
| ## Rubric |
|
|
| This rubric is intended for evaluators, not for the agent being tested. A strong |
| submission should recover the anonymized target locus and explain the |
| sample-specific evidence for it. |
|
|
| ### Expected Finding |
|
|
| The key finding is a loss-of-heterozygosity signal in `sample_2` relative to |
| `sample_1` that affects an effector-region candidate on: |
|
|
| - Target contig: `ref_contig_000132` |
| - Target interval: approximately `2553060-2553458` |
| - Strand: `+` |
| - Target gene: `gene_027287` |
| - Target transcript: `transcript_028378` |
| - Target protein: `protein_027771` |
| - Public alias: `locus_X` |
|
|
| The best answers should identify this region as the central candidate, describe |
| the sample-specific haplotype/heterozygosity pattern, and avoid treating the |
| signal as a simple loss of read depth unless their analysis supports that claim. |
|
|
| ## Intended Use |
|
|
| This dataset is intended for research and evaluation of AI agents in scientific |
| workflows. It is especially useful for testing whether an agent can: |
|
|
| - plan a multi-step genomics analysis |
| - choose appropriate command-line tools |
| - recover from failed or uninformative analyses |
| - synthesize sequence-level evidence into a scientific claim |
| - write a report that a scientist can audit |
|
|
| It is not intended as a clinical, diagnostic, agricultural decision-making, or |
| production pathogen-surveillance dataset. |
|
|
| ## Public-Safe Anonymization |
|
|
| The uploaded package is designed to be public-facing. Private mappings from |
| neutral identifiers back to source identifiers are not included. Original raw |
| downloads, source metadata, paper-derived notes, and answer-key files should be |
| kept outside the Hugging Face dataset repository. |
|
|
|
|
| ## Citation And Provenance |
|
|
| This benchmark package is derived from publicly available genomics data |
| associated with a published plant-pathogen study in *Science*. The public upload |
| intentionally uses anonymized names so that the dataset can function as an |
| evaluation task rather than a paper-reading exercise. |
|
|
| If you use this dataset in a paper, report, or benchmark suite, cite this |
| dataset card and describe whether agents were allowed to use internet search, |
| external biological databases, or only the files provided here. |
|
|
