Topology-directed optimization of block copolymer architecture for self-assembly into spherical vesicles using machine learning.

Amphiphilic block copolymers are a promising class of macromolecules used for creating new functional containers for delivering active agents into the cells. The diversity of these macromolecules' architectures allows for the selection of the necessary system parameters to generate the desired morphologies of these containers. Where the standard grid-search approach proves costly, optimization algorithms can directly search for the desired structure using a data-aware approach. The challenge here lies in the numerical characterization of such vesicles with a single objective function, that is, finding a functional capable of indicating the degree of proximity of the aggregate to the desired class of structures, for example, vesicles. This paper describes a novel pipeline for the topology-directed search for molecular parameters that enable the spontaneous formation of spherical vesicles in a system of amphiphilic comb-coil copolymers with variable architectural and solvent parameters within simulations. A key feature of the presented method is the use of a probabilistic classification model, trained on ideal structures, as a functional describing the "vesicularity" of the aggregate obtained in the point-based simulation. The classification of structures is based on topological data analysis, and, as shown, topological characteristics are sufficient to accurately distinguish typical structures self-assembling in solutions of amphiphilic macromolecules. Optimization of this probabilistic functional using a classic Bayesian optimization algorithm allows finding parameters consistent with spherical vesicle formation in a reasonable number of iterations. By design, the usage of this methodology is not restricted by the considered class of polymers and can be applied to other simulated macromolecular systems.