Capability benchmark · Chemistry · Medicine

SMDD-Bench tests agents on multi-turn small-molecule drug design

A benchmark of guaranteed-solvable, long-horizon small-molecule drug-design tasks across protein targets and chemical task types.

Summary

SMDD-Bench introduced a standardized agent benchmark for real-world small-molecule drug design. It covers 502 multi-turn task instances across five design task types and 102 protein targets, with frontier models still solving less than half of the benchmark.

AI role

LLM agents were evaluated on planning, tool use, chemical and biological reasoning, and iterative design under limited oracle calls.

Narrative role

This adds a chemistry and medicine capability measure that is closer to computational drug-discovery work than static science QA, while also showing substantial remaining headroom.

Caveat

The benchmark is a preprint and evaluates computational task completion rather than wet-lab discovery or clinical utility.

Event

Date: 2026-05-20
Timeline role: Notable
Progress relevance: Capability indicator
Validation: Benchmarked

Evidence

Preprint · Benchmark

Observed metrics

task instances: 502; SMDD-Bench contains 502 guaranteed-solvable task instances.
task types: 5; Tasks span pharmacophore identification, interaction-point discovery, scaffold hopping, lead optimization, and fragment assembly.
protein targets: 102; The benchmark covers 102 unique protein targets.
best reported solve rate: 40.2%; The authors report the strongest evaluated model solved 40.2% of tasks.

Source: arXiv ↗