Mapping Antibiotic Photocatalytic Transformation and Resistance Risks with a DFT-Informed Machine Learning Workflow

The photocatalytic degradation of antibiotics is effective but may yield transformation products (TPs) that sustain or amplify ecological risks, including antibiotic resistance gene (ARG) induction. This study developed a predictive framework that couples photocatalytic experiments, high-resolution mass spectrometry, density functional theory (DFT) calculations and machine learning (ML) to assess risks of TPs. Using tetracycline as a model compound, we constructed a reaction network over 120 steps and 9 533 reactions, and trained an ML model to rapidly predict Gibbs free energy changes with DFT accuracy. Automatic transition-state searches were integrated to evaluate kinetic accessibility within the network. The generalizability of this approach was validated with pathways of five different antibiotics involving 545 reactions. Furthermore, a multi-dimensional scoring system was developed that integrates diversity, ecotoxicity, biodegradability, and feasibility (DEBF) to prioritize pathways by both reactivity and sustainability. Several hydroxylated, aminated, and amide-ketone TPs were identified as high-risk species with enhanced ARG-binding potential. By bridging molecular energetics with ecological outcomes, this work offers a generalizable, mechanism-anchored, and risk-aware approach for analyzing photocatalytic transformations and deriving design principles for pollutant degradation that balance efficiency with ecological safety.