Abstract
To better understand and design proteins, it is crucial to consider the multifunctional landscapes on which all proteins exist. Proteins are often optimized for single functions during design and engineering, without considering the countless other functionalities that may contribute to or interfere with the intended outcome. In this work, we apply deep learning to understand and design the multifunctional host-targeting landscape of the T7 bacteriophage receptor binding protein for enhanced infectivity, pre-defined specificity, and high generality in virulence toward unseen strains. We compare several different model architectures and design approaches and experimentally characterize designed phages optimized for 26 diverse tasks. We demonstrate that with multiobjective machine learning, it is possible to design complex specificities at success rates that can enable low-throughput validation of predicted hits. Our results show that the targeting capabilities of T7 are highly plastic, with opposite specificities often separated by only a few mutations. This level of tunability underscores how models trained on multifunctional data can uncover key principles of phage biology and specificity. The same modeling framework can be applied to guide the multiobjective design of other proteins or mutable biological systems, offering a general strategy for navigating multifunctional landscapes.