Engineering dynamic behaviors in computationally designed proteins

Proteins are the fundamental molecular machines of life, orchestrating virtually every essential biological process. A proteinâ s structural architecture is encoded by its primary amino acid sequence, which in turn determines its biological activity. While computational protein design has made remarkable progress in generating stable, single-state folds, most natural protein functions rely on dynamic conformational changes and tightly regulated interactions. Transitioning from the design of static architectures to the engineering of programmable, dynamic behavior, therefore, represents a central challenge in the field. This thesis integrates physics-based modeling, geometric deep learning, and generative AI to design and experimentally validate functional, controllable, and dynamic protein systems. First, we address the challenge of engineering externally controlled proteinâ protein interactions through the development of humanized, small-molecule-responsive systems. Chemically induced dimerization (CID) is a powerful tool in synthetic biology; however, its clinical translation is often limited by the potential immunogenicity of non-human protein components and the toxicity of commonly used small-molecule inducers. To overcome these limitations, we computationally designed caffeine-inducible heterodimers and humanized their sequences to minimize therapeutic risk. These modules were integrated into cytokine receptors and split transcription factors, enabling precise regulation of therapeutic genes, including chimeric antigen receptors (CARs), in response to non-toxic concentrations of caffeine. Next, we explored the structural plasticity of ligand-mediated recruitment interfaces using Cereblon (CRBN), a key E3 ligase in targeted protein degradation. By combining geometric deep learning (MaSIF-NeoSurf) with high-throughput screening, we identified a pomalidomide-dependent binder. Structural characterization by cryo-electron microscopy revealed that this binder engages CRBN through a previously unrecognized binding mode, distinct from canonical â G-loopâ degrons. Despite its unconventional interface, the complex remains functionally competent to induce proteasomal degradation, demonstrating that de novo design can uncover latent interaction sites inaccessible to traditional motif-driven strategies. Finally, we investigate the de novo design of metamorphic proteins-single sequences capable of accessing two structurally distinct conformations. We developed a generative design workflow based on RFdiffusion and ProteinMPNN to produce candidate sequences intended to interconvert between a natural fold and a synthetic alternative state. Using an exploratory solution NMR spectroscopy approach, we are probing the dynamic conformational landscapes of these metamorphic proteins. In parallel, we are applying conformation-specific trapping strategies to perturb equilibrium populations, to experimentally interrogate and constrain the designed energy landsca