RAGPPI
The first factual question-answer benchmark for identifying the biological impacts of protein-protein interactions (PPIs) in target identification, designed to evaluate retrieval-augmented generation (RAG) and LLM reasoning. 4,420 QA pairs: a 500-pair expert-annotated gold standard and a 3,720-pair silver standard built via an ensemble auto-evaluation LLM.
Composite
56.5
Experimental validation
N/A — literature-grounded QA benchmark
Stages
Target ID
Modalities
small moleculebiologic (mAb/ADC/bispecific)
Task types
retrievalclassification
Size
qa_pairs: 4,420
gold_standard: 500
silver_standard: 3,720
splits: {'train': 0, 'val': 0, 'test': 4420}
note: 500 expert-annotated gold-standard QA pairs + 3,720 silver-standard pairs curated via ensemble auto-evaluation (fact-abstract F1 + low-similarity fact counts)
gold_standard: 500
silver_standard: 3,720
splits: {'train': 0, 'val': 0, 'test': 4420}
note: 500 expert-annotated gold-standard QA pairs + 3,720 silver-standard pairs curated via ensemble auto-evaluation (fact-abstract F1 + low-similarity fact counts)
License
Other — see arXiv 2505.23823
First release
2025-05-28
Last updated
2026-06-29
Official site
Leaderboard
→ leaderboard
Dataset
Code / GitHub
→ repository
HuggingFace
→ HF
Paper
RAGPPI: RAG Benchmark for Protein-Protein Interactions in Drug Discovery · Youngseung Jeon, Ziwen Li, Thomas Li, JiaSyuan Chang, Morteza Ziyadi, Xiang 'Anthony' Chen · 2026 · paper · doi:N/A — arXiv preprint 2505.23823 · 0 citations
Flags
none
Experts
Groups
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
3
coverage
3
maintenance
3
adoption
1
quality
4
accessibility
3
industry_relevance
3
Notes
Fills a real gap: no prior factual QA benchmark for the biological impact of PPIs in target ID. Expert-driven criteria and a 500-pair gold standard give quality a 4, with a transparent two-tier (gold/silver) construction. Coverage and rigor are moderate because the silver standard relies on LLM auto-evaluation rather than full human annotation. Adoption 1 (preprint, recently revised June 2026). Useful complement to network/knowledge-graph target-ID benchmarks.