Hauser7733/VenusX_Frag_Evo_MF50_MCQ4

VenusX Fragment MCQ — Evo (MF50)

4-choice multiple-choice question reformulation of the VenusX Fragment-level Evo sub-task from the paper VenusX: Unlocking Fine-Grained Functional Understanding of Proteins (ICLR 2026).

This fork adapts the original N-way classification task into a 4-choice MCQ so that zero-shot LLMs can be evaluated fairly. The original task required the model to output one of 739 InterPro class labels—infeasible for a CausalLM that cannot know which IPR IDs are in the benchmark's label subspace.

Schema

Each sample is one multiple-choice question with exactly one correct answer.

Field	Type	Description
uid	str	Original VenusX sample UID
seq_fragment	str	The protein fragment amino acid sequence
annotation	str	"Evo" (sub-task name)
interpro_label	int	Original VenusX integer label (preserved for compatibility)
correct_ipr	str	The correct InterPro accession (e.g. IPR000169)
correct_letter	str	"A" / "B" / "C" / "D" — the letter whose option matches correct_ipr
option_{a,b,c,d}_ipr	str	InterPro accession for each option
option_{a,b,c,d}_desc	str	Human-readable description from InterPro entry.list
distractor_source	str	How the 3 distractors were picked: hierarchy, mixed, or pool (see below)

How the MCQ is constructed

For each sample, we start with one golden InterPro ID (the original label) and pick 3 distractors via a 2-tier fallback:

1. Tier 1 — InterPro hierarchy siblings: If the golden IPR is in the InterPro hierarchy tree, take true siblings (same parent, not the golden itself, not any ancestor/descendant of the golden). We use up to 3.
2. Tier 2 — Same sub-task label pool (random): Fill the remaining slots by random sampling from the full label pool of the same sub-task, excluding the golden, all its ancestors/descendants, and any Tier 1 picks.

All 4 options (1 golden + 3 distractors) are then shuffled to randomize letter positions (A/B/C/D), using a deterministic per-sample seed derived from {annotation}:{split}:{uid}:{golden} so every dataset build is bit-identical and reviewers can independently verify each MCQ.

The distractor_source column records which strategy was used:

• hierarchy — all 3 distractors are InterPro siblings
• mixed — some are siblings, some are random pool samples
• pool — all 3 are random pool samples (most common for Active_site / Binding_site / Conserved_site types, which have no InterPro hierarchy per EBI convention)

Why descriptions are included

The original free-text task expected the LLM to directly output an IPR ID like IPR019757. This is unfair because the LLM has no way to know which specific IPRs are in the benchmark's small label subspace. Our MCQ format exposes the 4 candidate options with human-readable names (e.g. IPR019757 — Peptidase S26A, signal peptidase I, lysine active site) so that the LLM can use its biological knowledge to match the fragment's features against the candidate functional descriptions.

Example

Fragment: IHCIAGLGRTP

A) IPR033694 — Pyroglutamyl peptidase I, Cys active site
B) IPR023411 — Ribonuclease A, active site
C) IPR016130 — Protein-tyrosine phosphatase, active site     ← correct
D) IPR000169 — Cysteine peptidase, cysteine active site

[ANSWER]C[/ANSWER]

Build Reproducibility

This dataset is fully reproducible from the included build scripts and reference files:

scripts/
  parse_interpro.py     # Parses InterPro flat files into a queryable cache
  build_mcq.py          # Builds MCQ samples with 2-tier distractor fallback
reference/
  entry.list            # InterPro entries dump (downloaded 2026-04-09)
  ParentChildTreeFile.txt  # InterPro hierarchy tree (downloaded 2026-04-09)
  label_pool.json       # Union label pool across all 5 sub-tasks

To rebuild:

# 1. Refresh InterPro flat files (optional — pinned versions included)
curl -O https://ftp.ebi.ac.uk/pub/databases/interpro/current_release/entry.list
curl -O https://ftp.ebi.ac.uk/pub/databases/interpro/current_release/ParentChildTreeFile.txt

# 2. Parse into cache
python scripts/parse_interpro.py

# 3. Build MCQ (loads original VenusX Fragment datasets from AI4Protein/*)
python scripts/build_mcq.py

Known Limitations

1. Distractors are automatically generated, not peer-reviewed. Unlike MMLU or MedMCQA whose distractors are human-written by exam boards, our distractors come from random sampling / hierarchy traversal. Some may be too easy (e.g. a completely unrelated domain for a specific active site), inflating accuracy.

2. InterPro hierarchy coverage is low for Active_site, Binding_site, Conserved_site, and Repeat entry types — EBI does not arrange these into hierarchies. As a result, Tier 1 (sibling-based) distractors only apply to a minority of samples (see distractor_source column for each sample's strategy).

3. Random baseline is 25% (not 1/N of the original label space). Accuracy numbers from this MCQ benchmark should be interpreted against this 25% baseline, not the paper's full-label-space accuracy.

4. interpro_label field is preserved for traceability but not used in MCQ scoring. MCQ scoring compares pred_letter to correct_letter.

5. Not comparable to VenusX paper Table 4 numbers. The paper reports ESM2 probe accuracy on the full label space; we report LLM accuracy on a 4-way MCQ. Same metric name (ACC), different semantics.

Citation

If you use this MCQ-reformulated dataset, please cite both the original VenusX paper and this fork:

@inproceedings{venusx2026,
  title={VenusX: Unlocking Fine-Grained Functional Understanding of Proteins},
  author={Tan, Yang and others},
  booktitle={ICLR},
  year={2026},
  url={https://arxiv.org/abs/2505.11812}
}

References

The MCQ reformulation methodology draws from the following literature:

• Hendrycks et al., Measuring Massive Multitask Language Understanding (MMLU), ICLR 2021
• Pal et al., MedMCQA: A Large-scale Multi-Subject Multi-Choice Dataset for Medical domain Question Answering, CHIL 2022. https://arxiv.org/abs/2203.14371
• El-Sanyoury et al., Automatic distractor generation in multiple-choice questions: a systematic literature review, PeerJ Computer Science 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11623049/
• Susanti, Iida & Tokunaga, Automatic Generation of English Vocabulary Tests (WordNet-based distractor), CSEDU 2015
• Gene Ontology sibling negatives: Frontiers in Genetics 2020, BMC Bioinformatics 2009

License

Derived from VenusX (AI4Protein). InterPro data is licensed CC-BY-4.0 from EMBL-EBI. This fork is released under CC-BY-4.0.

Contact

Built by hauser7733 as part of the SiEval evaluation framework. Questions or issues: open an issue at the SiEval repo.