TxBench-PP: Analyzing AI Agent Performance on Small-Molecule Preclinical Pharmacology

2026-06-17 · Source: Artificial Intelligence · Field: Science & Research — Health & Medical Research, Life Sciences & Biology, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

TherapeuticsBench Preclinical Pharmacology (TxBench-PP) is introduced as a verifiable benchmark for evaluating AI agents in small-molecule preclinical pharmacology, forming the initial segment of a broader TherapeuticsBench initiative. This benchmark features 100 evaluations, categorized by program stage, assay type, and task structure, covering areas like mechanism-of-action, pharmacodynamic reasoning, compound-target engagement, causal target validation, developability, safety, and translational efficacy. Agents interact with realistic workflow snapshots, inspect files in a coding environment, and provide structured answers for deterministic grading. Across 16 model-harness configurations, involving 11 models and 4,800 trajectories, no system consistently achieved reliable preclinical pharmacology decisions. The top performer, Claude Opus 4.8 / Pi, passed 59.3% of endpoint attempts (178/300), while GPT-5.5 / Pi followed at 55.3% (166/300).

Key takeaway

For AI Scientists developing or deploying agents in drug discovery, these findings highlight a significant performance gap in preclinical pharmacology. You should prioritize robust evaluation against realistic, data-driven benchmarks like TxBench-PP before trusting AI agents with critical program decisions. Focus development efforts on improving agents' ability to interpret novel assay data rather than relying on memorized knowledge.

Key insights

AI agents currently struggle to reliably make accurate preclinical pharmacology decisions from real-world assay data.

Principles

• Trusted evaluation of AI agents requires realistic program decisions.
• Evaluation should focus on recovering conclusions from real-world data, not memorized facts.

Method

The TxBench-PP benchmark involves agents inspecting files in a coding environment and returning structured answers for deterministic grading.

In practice

• The benchmark evaluates AI agents on mechanism-of-action tasks.
• It also covers pharmacodynamic reasoning and developability and safety.

Topics

• AI Agents
• Drug Discovery
• Preclinical Pharmacology
• Benchmarking
• Small Molecules
• Assay Data

Best for: AI Scientist, Research Scientist

Related on AIssential

• Benchmarking AI Agents for Addressing Scientific Challenges Across Scales — AI agents excel in structured scientific data analysis but remain unreliable for open-ended discovery and critical validation.
• APCyc: Property-Informed Design of Cyclic Peptides via Automated Cyclization — APCyc automates cyclic peptide design by inferring cyclization topology and optimizing multiple drug-like properties via Bayesian guidance.
• Introducing Perceptic: the AI operating system for drug development — AI can accelerate drug development by connecting multi-modal evidence and decisions across the entire drug lifecycle.
• CP-Agent: Context-Aware Multimodal Reasoning for Cellular Morphological Profiling under Chemical Perturbations — CP-Agent uses an MLLM and context-aware image-metadata embedding to generate interpretable rationales for drug-induced cell changes.
• AWS and Hopkins Engineering announce groundbreaking database for AI/ML antibody design — A new diverse, experimentally validated dataset enables rigorous benchmarking of AI models for antibody developability.
• Two AI-based science assistants succeed with drug-retargeting tasks — AI-powered agentic systems can accelerate scientific discovery by synthesizing vast literature and generating testable hypotheses.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.