Chemical-Aware Autoencoders for Eco-Selective Adsorption of Wastewater Contaminants

Pharmaceutical contaminants in wastewater represent a persistent environmental challenge with significant ecological and health implications. Among emerging remediation strategies, poly(2-oxazoline)-based polymers have demonstrated high adsorption efficiency for hydrophobic drugs through π-π stacking and electrostatic interactions. However, identifying additional compounds likely to be retained by such materials remains difficult, especially in the absence of comprehensive experimental datasets. In this work, we introduce a data-driven framework to support early-stage screening of contaminants for phenyl-functionalized poly(2-oxazoline) (polyPhOx) polymers. A conditional autoencoder (CAE) is trained on 500,000 druglike molecules from the ZINC database, conditioned on key descriptors. The model learns a latent space in which structurally or functionally similar molecules-based on their adsorption potential-are embedded nearby. We evaluate this latent representation using a curated set of environmental contaminants, including known polyPhOx adsorbates, analogs with varying properties, and unknown-adsorbance compounds. This enables the identification of clusters associated with potential polymer affinity. We also compare against traditional baselines methods (PCA and UMAP), finding that the CAE yields a sharper colocalization of polyPhOx with verified adsorbates and clearer separation from non-binders. The proposed method offers a lowcost, scalable strategy to prioritize compounds for experimental validation, contributing to the development of sustainable, polymer-based water cleaning technologies.