TxBench-PP
A verifiable agentic benchmark for small-molecule preclinical pharmacology: 100 evaluations indexed by program stage, assay type and task structure, testing whether AI agents recover correct conclusions from real-world assay data (mechanism-of-action, pharmacodynamics, target engagement, causal target validation, developability/safety, translational efficacy) rather than recalling memorized literature facts.
Composite
52.0
Experimental validation
Retrospective
Stages
Lead ID / ADMETTarget IDIND-enabling
Modalities
small molecule
Task types
classificationretrievalgeneration
Size
evaluations: 100
model_harness_configs: 16
models: 11
splits: {'train': 0, 'val': 0, 'test': 100}
note: 100 deterministically-graded evaluations; agents inspect files in a coding environment and return structured answers; 16 model-harness configurations spanning 11 models x 4 harnesses (4,800 agent runs reported). First focused slice of a broader TherapeuticsBench effort.
model_harness_configs: 16
models: 11
splits: {'train': 0, 'val': 0, 'test': 100}
note: 100 deterministically-graded evaluations; agents inspect files in a coding environment and return structured answers; 16 model-harness configurations spanning 11 models x 4 harnesses (4,800 agent runs reported). First focused slice of a broader TherapeuticsBench effort.
License
Other — see arXiv 2606.19245 (release terms not yet confirmed)
First release
2026-06-17
Last updated
2026-06-17
Official site
Leaderboard
→ leaderboard
Dataset
→ dataset
Code / GitHub
→ repository
HuggingFace
→ HF
Paper
TxBench-PP: Analyzing AI Agent Performance on Small-Molecule Preclinical Pharmacology · Hannah Le, Ramesh Ramasamy, Alex Urrutia, Mahsa Yazdani, Tim Proctor, Kenny Workman · 2026 · paper · doi:N/A — arXiv preprint 2606.19245 · 0 citations
Flags
none
Experts
—
Groups
—
Hosted by
—
Related benchmarks
Rubric (7-criterion)
rigor
4
coverage
3
maintenance
2
adoption
1
quality
4
accessibility
3
industry_relevance
3
Notes
One of the first agentic drug-discovery benchmarks grounded in realistic program-decision workflows with deterministic grading over real assay data rather than literature recall (rigor 4, quality 4). Broad coverage across MoA/PD/target-engagement/developability but limited to 100 evaluations and small-molecule preclinical only (coverage 3). Brand-new (Jun 2026), no leaderboard or citations yet (maintenance 2, adoption 1); release/license terms not yet fully confirmed (accessibility 3). Industry relevance is real (program-stage framing) but unproven in deployment (3).