Integrating Bilinear Transduction with Message Passing Neural Networks for Improved ADMET Property Prediction

With the increasing application of machine learning and deep learning in drug discovery comes the significant challenge of addressing censored molecular property datasets. Pharmaceutical assays frequently generate censored data where measurement limitations prevent recording exact values beyond predetermined thresholds. Standard deep learning approaches struggle with this censoring, often producing systematic prediction errors even for in-distribution molecules. Building on the established strengths of Chemprop and the concept of bilinear transduction, we present a method that integrates bilinear transduction into Chemprop's message-passing neural network. This integration allows us to effectively leverage domain-specific structural relationships between molecules, addressing current limitations in molecular property prediction. Our comprehensive evaluation across multiple ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties demonstrate that our method outperforms standard D-MPNN baselines, with improvements exceeding $100\%$ for heavily censored datasets like CYP2C9 inhibition and CYP2D6 inhibition. This practical solution requires no additional experimental measurements while improving ADMET property prediction accuracy, particularly in the challenging high-censoring regimes common in pharmaceutical research.