Predicting Drug Interactions to Unassociated Biomedical Implants Using Machine Learning Techniques and Model Polymers
Kerner, J. J.; von Recum, H.
Show abstract
Affinity based drug delivery mechanisms increase efficacy and minimalize off target effects when compared to non-specific methods due to the localization of drugs within target areas. While this is beneficial for targeted delivery, introduction of foreign polymeric medical devices into the body provide a potential area of localization due to high affinity between administered drugs and polymers. Previous attempts at creating models to predict affinity between small molecule drugs and polymers require a specific model be trained for each individual polymer failing to incorporate input features of both the polymer (host) and small molecule drug (guest). Within, we propose a universal model built using a neural network and quantitative structure activity relationships to predict the binding energy between guest and host molecules using input features. The trained model returned a correlation value, R2, of 0.9806 and 0.9958 between predicted and experimental binding affinity for the training and validation sets, respectively. This correlates to a mean absolute error of 0.951 kJ/mol and 0.771 kJ/mol for the training and validation sets, respectively. While limited to the current polymers used to train the model, the dataset can be expanded, and models retrained for further applications.
Matching journals
The top 9 journals account for 50% of the predicted probability mass.