Building Comprehensive Toxicity Data Libraries of Short-Chain Length PHA-Based Materials for the Development of...

Area of Science:

• Biomaterials Science
• Toxicology
• Computational Chemistry

Background:

• Polyhydroxyalkanoates (PHAs) offer biodegradable alternatives to plastics, with potential in medical applications.
• Assessing PHA toxicity is challenging due to complex, costly, and limited data.
• Machine learning (ML) offers a promising approach to predict toxicity and reduce in vivo testing.

Purpose of the Study:

• To build standardized cytotoxicity and ecotoxicity data libraries for PHAs.
• To develop and validate ML models predicting PHA toxicological outcomes based on composition.
• To identify key descriptors influencing PHA toxicity predictions.

Main Methods:

• Construction of comprehensive, standardized PHA cytotoxicity and ecotoxicity datasets.
• Development and evaluation of ML models, including Extra Trees Classifier and Gradient Boosting Classifier.
• Application of SHAP analysis to determine influential toxicity descriptors.

Main Results:

• Cytotoxicity model achieved MCC of 0.678 and balanced accuracy of 0.901; key predictors were additive type, exposure conditions, and particle morphology.
• Ecotoxicity model achieved MCC of 0.639 and balanced accuracy of 0.818; organism and exposure descriptors were dominant.
• Polymer composition had minimal impact on predictions, aligning with known biocompatibility of PHB and PHBV.

Conclusions:

• ML models show potential for predicting PHA cytotoxicity and ecotoxicity.
• Additive type, exposure conditions, and particle morphology are critical for cytotoxicity.
• Ecotoxicity predictions are driven by organism and exposure factors, with polymer composition being less influential.