Designing symmetrical multi-component proteins using a hybrid generative AI approach

Proteins, the fundamental building blocks of biological function, orchestrate complex cellular processes by assembling into intricate structures through meticulous interactions. The design of specific protein-protein interfaces to create customized protein assemblies holds immense potential for various biotechnological applications. To address the current limitations in designing heteromeric interactions for multi-component assemblies, we developed a hybrid generative AI design approach. This method combines deep learning and automated reasoning, explicitly considering both positive and negative interaction states to favor heteromeric desired over undesired interactions. The approach leverages Effie, a deep-learned pairwise decomposable scoring function, and an advanced reasoning tool extended for multicriteria optimization of this function. Here, we tested the ability of this hybrid AI method to redesign homomeric interfaces of bacterial microcompartment components (BMC-H) into heteromeric assemblies. We benchmarked its performance against ProteinMPNN, a sequence design autoregressive model. Our in silico assessment, complemented by experimental validation, highlights the outperformance of the hybrid AI generative design approach, and its potential to unlock the engineering of complex multi-component self-assembling protein entities.