AIChemy: Probabilistic Virtual Screening for Drug Discovery
Equivariant deep probabilistic graph neural networks that screen compounds with calibrated uncertainty on every prediction.
Challenge
Virtual screening ranks millions of candidate compounds so chemists can spend lab time on the few most likely to bind a target. Classical pipelines return point predictions with no honest measure of confidence, so promising hit lists are riddled with overconfident false positives and the cost of synthesising and assaying dead ends falls on the discovery team.
Solution
AIChemy is our drug-discovery platform: equivariant deep probabilistic graph neural networks built on Borch, our deep universal probabilistic programming language. Modelling molecules as graphs with equivariant architectures respects the symmetries of 3D chemistry, while the Bayesian treatment attaches calibrated uncertainty to every compound. Hit lists arrive with confidence attached, so screening prioritises compounds that are both promising and well understood by the model.
This work was done together with RG Discovery, who bring the medicinal chemistry and assay expertise that grounds the models in real wet-lab data.
The approach is benchmarked against classical pipelines in peer-reviewed work, including our equivariant graph neural networks for drug discovery and the CDK9 inhibitor study presented at DDC Europe 2024.
Business Value
Uncertainty-aware screening focuses expensive wet-lab capacity on the compounds most likely to pay off and flags where the model is guessing rather than knowing. That means fewer wasted synthesis cycles, defensible decisions in a high-stakes domain, and a screening process that improves as it sees more data.
Learn more at aichemy.io.