Prediction and Structure-Activity Relationship Analysis on Ready Biodegradability of Chemical Using Machine Learning...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Toxicology

Background:

• Persistent industrial contaminants pose significant environmental and public health risks.
• Understanding chemical biodegradability is crucial for risk assessment and sustainable chemical design.

Purpose of the Study:

• To develop accurate predictive models for chemical biodegradability.
• To identify critical molecular descriptors and substructural features influencing biodegradability.
• To guide the design of chemicals with enhanced biodegradability.

Main Methods:

• Collected and characterized a dataset of 1306 not readily biodegradable (NRB) and 622 readily biodegradable (RB) chemicals.
• Utilized CORINA descriptors, MACCS fingerprints, and ECFP_4 fingerprints for chemical characterization.
• Constructed and evaluated 34 classification models, including decision tree (DT), support vector machine (SVM), random forest (RF), and deep neural network (DNN) algorithms, with a focus on a Transformer-CNN model.

Main Results:

• The best Transformer-CNN model achieved 86.29% balanced accuracy and a 0.71 Matthews correlation coefficient on the test set.
• Key molecular properties influencing biodegradability include solubility, atom charges, rotatable bonds, electronegativities, molecular weight, and hydrogen bonding acceptors.
• Aromatic rings and nitrogen/halogen substitutions hinder biodegradation, while ester and carboxyl groups promote it.

Conclusions:

• Accurate prediction of chemical biodegradability is achievable using machine learning models.
• Specific molecular features significantly impact a compound's environmental fate.
• The findings provide valuable insights for designing environmentally benign chemicals.