Modeling the Soil Dissipation Kinetics of Small Molecules

Abstract

Understanding the metabolic fate of small molecules is central to risk assessment and the rational design of plant protection products, pharmaceuticals, cosmetics, and food additives. In environmental systems, dissipation is governed by kinetic processes operating across soil, water, and other compartments. For chemicals released into the environment, especially active ingredients of plant protection products, assessing soil persistence is an essential regulatory requirement. The ability to predict soil persistence, ideally in advance of commissioning experiments, is highly valuable for identifying novel active ingredients for plant protection products. Here, we introduce machine learning models that predict the full time course of small-molecule dissipation in soil. These models are explicitly constrained by mechanistic assumptions encoded as ordinary differential equations (ODEs). By design, this physics-constrained parametrization enforces physical plausibility. To the best of our knowledge, our models are the first to predict soil dissipation kinetics from molecular structures using a constrained time course framework. We found that this framework almost always outperformed a conventional QSAR approach of predicting prefitted kinetic parameters. We further investigated the incorporation of soil metadata, i.e., predefined and experimentally measured soil parameters, into the model, with mixed results. Our framework supports screening and design decisions by linking molecular properties and soil metadata to predicted environmental lifetimes and dissipation profiles.