PlastiGuard: ML/Molecular Descriptor based Biodegradability Analyzer of Persistent Chemical Carcinogens in Plastics

This study uses machine learning models and molecular descriptor data to evaluate the biodegradability of harmful plastic carcinogens and identify safer structural alternatives. The research shows that neural networks perform the best at predicting biodegradability, and it highlights several naturally produced substitutes that could reduce the environmental and health risks linked to persistent toxins in plastics.

Plastics pose a significant threat to human health by introducing carcinogenic chemicals through inhalation and digestion. Carcinogens such as Bisphenol A (BPA), Diethyl Phthalate, Per- and Polyfluoroalkyl Substances (PFAS), and Polyvinyl Chloride (PVC) disrupt reproduction, growth, and metabolism by mimicking hormonal activity. Their persistence in the environment increases human exposure, particularly when non-biodegradable. To evaluate biodegradability, four machine learning models—logistic regression, k-nearest neighbors, random forest, and neural network—were trained on molecular descriptor data generated using RDKit. The dataset was refined through Principal Component Analysis (PCA) to assess its impact on performance. Results indicate that the neural network achieved the highest accuracy with RDKit descriptor data (88.1%) and PCA-transformed data (78.2%). Across all models, the RDKit dataset outperformed PCA-applied data, with accuracy improvements of +9.9% for the neural network, +14.3% for logistic regression, +9.7% for random forest, and +6.9% for k-nearest neighbors. Notably, all models classified BPA, Perflurorooctanic Acid (PFOA), and Perfluorooctanesulfonic Acid (PFOS) as non-biodegradable, identifying them as the most persistent and hazardous carcinogens in plastics. To identify safer alternatives, cosine similarity was employed to pinpoint structurally similar compounds, which were subsequently assessed for biodegradability. This approach revealed that viable naturally produced substitutes include 3-Propionylchromone, N-Acetyl-D-glucosamine, and Butyl Paraben. This study highlights the potential of machine learning in identifying biodegradable alternatives to persistent carcinogens in plastics. By reducing the environmental and health risks associated with plastic-related toxins, these findings offer a crucial step toward more sustainable and responsible plastic manufacturing practices.

Explore More!

Research Paper