In Silico Analysis of Contaminant Persistence: From QSARs to Machine Learning Models

For over six decades, in silico persistence modeling has evolved from intuitive, data-efficient quantitative structure-activity relationships (QSARs) for families of closely related chemicals to modern machine learning (ML) capable of handling heterogeneous data and broader chemical space. Early QSARs linked a few descriptors to well-defined properties and remain useful, transparent screeners for emerging contaminants, but environmental fate now demands wider coverage and explicit consideration of complex environmental conditions. This review defines end points beyond single numbers─physicochemical properties, partitioning, rate constants and half-lives, time-resolved degradation profiles, identities of transformation products and pathways, and bulk/surface/structural metrics for polymers (including microplastics) and materials. We then describe how to represent reactants (descriptors, fingerprints, molecular graphs, images, and text), capture environmental system features (bulk/surface chemistry, biomarkers, and optical and spectroscopic fingerprints), and improve data availability through curated, "living" data sets and literature-scale curation. Further, we summarize ML concepts and workflows, highlight advances in product and pathway prediction, and emphasize interpretability, uncertainty, and applicability domain. Finally, we chart a practical roadmap─standardized reporting, benchmark data sets, targeted measurements, biological context for biodegradation, and hybrid mechanistic-data-driven modeling─to move persistence assessment from ad hoc studies to coordinated, decision-ready prediction across small molecules, polymers, and materials.