Sequence-Based Computational Design of High-Affinity Amphiphilic Copolymers for Protein Targeting: A Machine Learning and Coarse-Grained Simulation Approach

Synthetic polymers have been developed as abiotic affinity reagents for biomarker proteins due to their capacity for high-affinity and specific binding to target proteins. However, current structure-based methods for designing polymers that target specific proteins are often inefficient and unsuitable, primarily because of the structural flexibility of most polymers and the lack of experimental data. In this work, we introduce a novel sequence-based computational paradigm for the rational design of high-affinity amphiphilic copolymers for proteins. The approach involves efficient on-lattice coarse-grained simulations to explore the sequence–affinity relationship for short copolymers, followed by a machine learning (ML) model that predicts properties for longer and more complex copolymers based on the previous simulation data. As a demonstration of this method, we constructed a binding data set comprising over 34,000 random copolymers, specifically assessing their binding affinity toward the epithelial cell adhesion molecule (EpCAM) protein. Direct simulations estimated the binding affinities for approximately 3500 copolymers, while the remaining copolymers were accurately predicted by the ML model, achieving an overall mean absolute error of less than 2 kBT. Notably, our method identified a distinctive pattern in copolymer sequences, emphasizing hydrophobic segments on margins favored by the protein, which was corroborated by all-atom molecular dynamics simulations, thus validating our approach. Collectively, we believe that this numerical framework provides a valuable toolkit for the design of polymers as affinity reagents.