SMDD-Bench: Can LLMs Solve Real-World Small Molecule Drug Design Tasks?

2026-05-23 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Life Sciences & Biology, Artificial Intelligence & Machine Learning, Research Methodology & Innovation · Depth: Advanced, quick

Summary

SMDD-Bench introduces a new, challenging, multi-turn, long-horizon agentic benchmark designed to standardize the evaluation of LLM agents on real-world small molecule drug design (SMDD) tasks. This benchmark addresses limitations of current evaluation methods, which are often ad hoc, too simple, or restricted to single-turn interactions. SMDD-Bench comprises 502 guaranteed-solvable task instances across 5 task types, including 2D Pharmacophore Identification and Lead Optimization, spanning a wide chemical space and involving 102 unique protein targets. Solving these tasks requires strong chemical and biological reasoning, 3D intuition, specialized tool use, and planning expertise. Initial benchmarking of 7 frontier open and closed-source LLMs revealed that even the most performant model, GPT5.4, solved only 40.2% of the tasks, highlighting significant challenges for autonomous computational drug design. A public leaderboard is hosted at smddbench.com.

Key takeaway

For AI Scientists and Research Scientists developing LLM agents for drug discovery, SMDD-Bench reveals current models, including GPT5.4, solve only 40.2% of real-world tasks. This indicates a critical need for improved chemical and biological reasoning, 3D intuition, and multi-turn planning capabilities. You should focus your development efforts on training agents for complex, long-horizon tasks and integrating specialized tools to achieve truly autonomous computational drug design.

Key insights

SMDD-Bench reveals current LLMs solve only 40.2% of complex drug design tasks, underscoring significant challenges in autonomous scientific discovery.

Principles

• LLM agents show potential for scientific discovery.
• Real-world SMDD needs multi-turn reasoning.
• Standardized benchmarks drive LLM agent progress.

Method

SMDD-Bench evaluates LLM agents using 502 multi-turn, long-horizon tasks across 5 types and 102 protein targets, requiring chemical reasoning, 3D intuition, tool use, and planning.

In practice

• Use SMDD-Bench to evaluate LLM agent performance.
• Contribute to the public SMDD-Bench leaderboard.
• Focus LLM training on multi-turn drug design.

Topics

• Small Molecule Drug Design
• LLM Agents
• Drug Discovery Benchmarking
• Computational Chemistry
• Protein Targets
• GPT5.4 Performance

Best for: AI Scientist, Research Scientist

Related on AIssential

• Probe Before You Edit: Probing-Guided Molecular Optimization for LLM Agents in Structure-Based Drug Design — Probing molecular edit responses guides LLM agents to jointly optimize binding affinity and druggability in drug design.
• TxBench-PP: Analyzing AI Agent Performance on Small-Molecule Preclinical Pharmacology — AI agents currently struggle to reliably make accurate preclinical pharmacology decisions from real-world assay data.
• Mozi: Governed Autonomy for Drug Discovery LLM Agents — Mozi unifies LLM flexibility with scientific rigor via dual-layer governed autonomy for drug discovery.
• Customized Amazon Nova models improve molecular-property prediction in drug discovery — Fine-tuned LLMs can unify complex molecular property prediction, outperforming or matching specialized GNNs.
• Why We're Not Close to AI Drugs Yet | Isomorphic Labs — Drug design comprises multiple distinct ML problems, not a single one, despite AlphaFold 2's protein structure prediction success.
• Why machine learning fails at mass spectrometry for small molecules — ML models for small-molecule mass spectrometry struggle due to data heterogeneity and lack of domain-specific architectural design.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.