Aim

Pharmaceutical residues in wastewater are difficult to study one compound at a time. The experimental work behind this paper considers phenyl-functionalized poly(2-oxazoline), or polyPhOx, a polymer that can retain some hydrophobic pharmaceutical compounds. The open question is which other contaminants may show similar affinity and therefore deserve laboratory testing.

We investigated whether representation learning could support this first screening stage. The goal was not to predict a final adsorption value from a large labelled dataset, because such a dataset was not available. Instead, we asked whether a chemically informed latent space could place known adsorbates close to compounds with related properties and separate them from known non-binders.

The paper was developed collaboratively across machine-learning and chemistry expertise. My contribution was part of the computational work: preparing the representation-learning approach, analysing the latent space, and comparing it with simpler dimensionality-reduction baselines.

Method

The conditional autoencoder was trained on 500,000 drug-like molecules from the ZINC database. This broad training set gives the model examples of chemical structures beyond the small group of environmental contaminants used in the final analysis. Chemical descriptors are supplied as conditioning information, encouraging the latent representation to organize molecules using both structural and physicochemical information.

After training, we projected a curated set containing verified polyPhOx adsorbates, related compounds, non-binders, and contaminants with unknown adsorption behaviour. This turns the latent space into an exploratory map: compounds located near verified adsorbates become candidates for closer experimental attention.

We compared the autoencoder representation with PCA and UMAP. These baselines are useful because they show whether the neural model adds a clearer organization than standard linear or manifold-based projections.

Findings

On the curated contaminant set, the conditional autoencoder produced sharper co-localization of verified adsorbates and clearer separation from non-binders than the PCA and UMAP views reported in the paper. The useful outcome is a ranking and visualization aid, not a replacement for adsorption measurements: it helps narrow a large chemical space to a smaller set of compounds that can be tested in the laboratory.

This project was also a useful lesson in interdisciplinary modelling. A latent representation is only meaningful when its neighbourhoods can be interpreted against chemical knowledge and experimental evidence. Working with domain specialists shaped both the descriptors used by the model and the way we evaluated the resulting clusters.

The paper was presented at IEEE MetroGREENST 2025 and appears on pages 256-261 of the proceedings.

Resources

• Paper DOI
• University of Siena research record