From protein science, it is well understood that ordered folding and 3D structure mainly arises from balanced and noncovalent polar and nonpolar interactions, such as hydrogen bonding. Similarly, it is understood that single-chain polymer nanoparticles (SCNPs) will also compact and become more rigid with greater hydrophobicity and intrachain hydrogen bonding. Here, we couple automated combinatorial polymer chemistry, high throughput small-angle X-ray scattering (SAXS), high throughput dynamic light scattering (DLS), and machine learning to design libraries of SCNPs with protein-like features. Taking this approach, we are able to identify copolymers that exhibited compactness and flexibility similar to that of bovine serum albumin (BSA). As artificial neural network models do not directly allow for structure-function relationship evaluation due to the black box nature of these models, we perform a SHapely Additive exPlanations (SHAP) analysis to reveal these relationships. Interestingly, this helps identify monomer sets that collaborate to produce compacted and ordered characteristics. Continued training of this machine learning model will allow us to design these materials without taking a high throughput screening approach.

Design of bio(mimetic) polymers using automation driven by active machine learning

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.