Evaluation of machine learning algorithms to predict the hydrodynamic radii and transition temperatures of...

Elastin-like polypeptides (ELP) belong to a family of recombinant polymers that shows great promise as biocompatible drug delivery and tissue engineering materials. ELPs aggregate above a characteristic transition temperature (T_t). We have previously shown that the T_t and size of the resulting aggregates can be controlled by changing the ELP's solution environment (polymer concentration, salt concentration, and pH). When coupled to a synthetic polyelectrolyte, polyethyleneimine (PEI), ELP retains its T_t behavior and gains the ability to be crosslinked into defined particle sizes. This paper explores several machine learning models to predict the T_t and hydrodynamic radius (R_h) of ELP and two ELP-PEI polymers in varying solution conditions. An exhaustive design of experiments matrix consisting of 81 conditions of interest with varying salt concentration (0, 0.2, 1 M NaCl), pH (3, 7, 10), polymer concentration (0.1, 0.17, 0.3 mg/mL), and polymer type (ELP, ELP-PEI800, ELP-PEI10K) was investigated. The five models used in this study were multiple linear regression, elastic-net, support vector regression, multi-layer perceptron, and random forest. A multi-layer perceptron model was found to have the highest accuracy, with an R² score of 0.97 for both R_h and T_t. This was followed closely by the random forest model, with an R² of 0.94 for R_h and 0.95 for T_t. Feature importance was determined using the random forest and linear regression models. Both models showed that salt concentration and polymer type were the two most influential factors that determined R_h, while salt concentration was the dominant factor for T_t.