Predicting aggregate morphology for varying composition and sequences in sequence-defined macromolecules

Dilute sequence-defined macromolecules, through self-assembly, exhibit aggregate morphologies profoundly influenced by sequence variations. While our prior work established a link between fixed composition sequences and their morphologies, introducing variable compositions presents heightened structural diversity and modeling complexity. In this study, we harness our recurrent neural network (RNN) architecture to predict aggregate morphology, accounting for both sequence and composition variations. After constructing a manifold—selecting sequences in proportion to their compositional availability—and training the RNN, we conduct high-throughput screening using molecular dynamics to validate its predictions and identify candidate sequences for self-assembly into specified morphologies using three distinct sampling methods: 1) edge-based sampling along manifold edges; 2) unrestricted composition k-means from respective clusters; and 3) restricted composition k-means guided by compositional constraints. In this screening process, the RNN exhibits its strongest performance with edge-based sampling. While its precision is commendably maintained with unrestricted k-means, it faces more pronounced challenges with restricted k-means, yet still delivers reliable results. These findings accentuate the RNN's versatility in predicting aggregate morphology across diverse sequence-composition landscapes.