Aim
Pharmaceutical residues in wastewater are costly to screen experimentally one compound at a time. This study focuses on phenyl-functionalized poly(2-oxazoline) (polyPhOx), a polymer capable of selectively retaining hydrophobic pharmaceutical pollutants, and addresses a core early-stage screening problem: identifying which untested contaminants exhibit comparable affinity and warrant laboratory evaluation.
Because large-scale labelled adsorption datasets for polyPhOx are unavailable, the problem is formulated through unsupervised and conditional representation learning rather than direct supervised regression. The objective is to construct a chemically informed latent space that co-localizes verified polyPhOx adsorbates with structurally and physicochemically analogous compounds while separating confirmed non-binders.
Method
A conditional autoencoder was pretrained on 500,000 drug-like molecules sampled from the ZINC database, exposing the encoder-decoder network to a broad distribution of molecular scaffolds beyond the limited set of assayed environmental pollutants. Physicochemical descriptors are injected as conditioning vectors, guiding the bottleneck layer to organize molecules according to both topological structure and adsorption-relevant properties.
Following pretraining, a curated evaluation set—comprising verified polyPhOx adsorbates, structurally related analogues, confirmed non-binders, and uncharacterized wastewater contaminants—is projected into the learned latent space. Compounds clustering within the neighbourhood of verified binders are prioritized as candidates for wet-lab testing.
To evaluate whether the neural bottleneck captures non-trivial chemical structure beyond standard projections, the learned embeddings are benchmarked directly against Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP).
Findings
Across the curated contaminant panel, the conditional autoencoder produces tighter co-localization of verified polyPhOx adsorbates and clearer separation from non-binders than either PCA or UMAP baselines. Rather than replacing wet-lab adsorption assays, the resulting latent map serves as a computational triage tool that narrows candidate selection across a large chemical space.
Developed across machine learning and organic chemistry research groups, the work was presented at IEEE MetroGREENST 2025 (pp. 256–261).