Explainable random forest predictions of polyester biodegradability using high-throughput biodegradation data

Abstract

The development of new, biodegradable polyesters is becoming increasingly important as legislative and consumer drivers push society towards more sustainable polymers. One key bottleneck in the development of biodegradable polyesters is the slow nature of biodegradability testing, which can often take weeks to months. High-throughput screening assays serve as a rapid tool to determine the biodegradability of materials quickly, without the need for lengthy, expensive testing. When combined with machine learning, the data generated by high-throughput assays can be exploited to predict the properties of other similar materials. Here, we report the development of a high-throughput enzymatic biodegradation assay, which has been used to determine the biodegradability of 48 polyesters. Using data generated from the assay to train a predictive model, we can predict the biodegradability of polyesters using an explainable random forest model with 71% accuracy. Transfer learning and model chaining were investigated as routes to improve the model predictions by exploiting existing literature data. SHAP analysis gives insight into the beneficial structural features of biodegradable polyesters. This understanding can be applied in the development of future biodegradable polyesters.