Benchmark finds nine AI language models build 3D molecules from drawings at 25% to 76% success
A 300-molecule benchmark shows multimodal language models can build correct 3D molecular coordinates from drawings at rates from 25.0% to 76.0%, but accuracy drops on larger molecules and metal complexes. The result matters because 3D construction is a routine, error-prone setup step in chemistry simulations, and the benchmark gives researchers a concrete way to judge AI tools before relying on them.
MolDraft, a curated benchmark released as a preprint on arXiv, tests whether AI can turn a 2D chemical drawing plus a short instruction into full atomic coordinates. The set contains 300 drawing-structure pairs drawn from natural products, compounds with hypervalent bonding, and coordination complexes. Each generated structure is checked for molecular identity, stereochemical configuration and coarse ligand geometry, so the task is not just describing a molecule but constructing a usable simulation-ready shape.
Nine multimodal language models were evaluated under the same agent setup, and the share of molecules built correctly, called the resolve rate, ranged from 25.0% to 76.0%. Performance fell as molecules grew, with the steepest decline above about 80 atoms. Only 29 of the 300 molecules were resolved by all nine models. The hardest source category was literature structures, mostly metal complexes, which challenged every system tested.
The result matters because simulations and drug design begin with 3D coordinates, yet chemists often still convert drawings by hand. The benchmark reframes multimodal AI as a geometry engine, not just a text generator: it must respect bonds, stereochemistry and spatial constraints. That puts it close to practical molecular construction, while the 29-of-300 overlap shows current models are uneven and can fail on the same chemistry.
In pharmaceutical and materials research, 3D coordinates are a required setup step for simulations, and errors in atom placement can distort predicted properties or binding behavior. The evidence is limited by a small benchmark and an unreviewed preprint, but it gives a concrete scorecard for AI tools being tested as ways to turn drawings into simulation inputs for working chemists.
For chemists and materials researchers, the benchmark gives a measurable test of whether AI can handle a manual, error-prone step in preparing 3D structures for simulations.
Sources (1)
- [1]Original source(https://arxiv.org/abs/2611.25041)