An AceProp submission to the PXR Blind Challenge
Published on July 9, 2026
Acellera entered the OpenADMET PXR Blind Challenge using AceProp, the molecular property trainer in PlayMolecule AI. Training used the data supplied by the challenge rather than a proprietary dataset.
Result source pending. The external scorer and final results link have not yet been recorded in our public evidence set. Quantitative placement and comparative claims remain off this page until that source is attached.
The challenge
PXR, the pregnane X receptor, is a xenobiotic sensor relevant to drug interactions. The challenge asked participants to predict PXR activation potency directly from chemical structure. Model development and final blind evaluation used separate analogue series.
The AceProp workflow
AceProp trained a supervised ensemble from challenge provided measurements. The ensemble combined a neural network regressor on pretrained molecular embeddings with a gradient boosted model on molecular descriptors and embeddings. Automated search selected hyperparameters under cross validation, followed by multi seed ensembling and calibration.
An AI agent sequenced featurisation, training, tuning and inference. The scientific team retained responsibility for the dataset, evaluation design and interpretation.
What we learned
Additional labelled compounds did not automatically improve transfer to the second analogue series. The useful lesson was methodological: data relevance and evaluation design matter more than dataset size alone.
AceProp is available through PlayMolecule AI. Teams with a live molecular property dataset can discuss one Active Program.