Evaluating AceProp on public Polaris benchmarks
Published on July 29, 2026
After the PXR Blind Challenge, we wanted to test whether one training workflow could be applied across a broader set of molecular property endpoints. We used AceProp, the molecular property trainer in PlayMolecule AI, across 20 public ADMET datasets available through Polaris.
Leaderboard sources pending. Direct links to the submitted results have not yet been recorded in our public evidence set. Comparative placement claims remain off this page until each result can be checked at its source.
One training recipe
For each dataset, the same AceProp workflow handled featurisation, training, tuning and evaluation. The objective was not to write a separate pipeline for every endpoint, but to see where a consistent procedure worked and where the endpoint required different scientific judgement.
Why public evaluation matters
A public benchmark makes the dataset, split and metric visible. That is more useful than a capability claim because another scientist can inspect the evaluation context and decide whether it resembles their own data.
Teams with a live molecular property dataset can discuss one Active Program.