Programmable Protein Stabilization with Language Model-Derived Peptide Guides

Introduction: Over 20,000 proteins exist in the human proteome, playing a variety of roles including catalysis of biochemical reactions, transport of nutrients and other biomolecules, and communication between cells to different organ systems. Post-translational modifications (PTMs) such as ubiquitination and phosphorylation are drivers of proteome diversity and expand our proteins’ capacities to regulate critical processes contributing to our overall health. As such, regulation of protein expression and PTMs are extremely important. Dysregulation of aberrantly altered proteins – such as improper ubiquitination and subsequent degradation of tumor suppressors or key transcription factors – can induce several disease phenotypes including different types of cancers (i.e., colorectal cancers, Ewing sarcoma, etc.), neurodegenerative diseases (i.e., Alzheimer’s disease and Huntington’s disease), and others. Current methods present significant drawbacks that result in “undruggable” diseases that cannot be addressed as viable therapeutic targets: (1) target proteins may not have existing or known binders, which are both laborious and expensive to engineer using solely protein engineering efforts; (2) even out of existing binders to target proteins, only a restricted selection are motif-specific that would allow for editing of pathogenic proteins; and (3) binding domains of the modifying enzyme itself may not be accessible to target motifs, limiting design and delivery methods. While targeted protein degradation (TPD) modalities have been developed, other proteome-editing modalities have yet to be explored. As such, we sought to engineer an analogous, genetically-encodable modality for programmable targeted protein stabilization (TPS) as a unique opportunity for therapeutic interventions and research tools.

Materials and

Methods: We have previously engineered ubiquibodies (uAbs) by linking computationally designed “guide” peptides to E3 ubiquitin ligase domains for modular, CRISPR-analogous TPD. Not only can we utilize currently existing protein binders such as nanobodies, DARPins, or antibodies, but we can also leverage computationally-derived peptide sequences to redirect engineered E3 ubiquitin ligases for subsequent degradation of a specific target protein. We design peptide binders to any protein using only its amino acid sequence using our protein language model algorithms such as SaLT&PepPr, PepPrCLIP, and PepMLM, trained on millions of natural amino acid sequences, that take an input protein sequence to conditionally generate binding peptides that capture diverse biochemical, structural, and functional features of a protein sequence alone (in other words, not needing protein structure). In this work, we hijack the ubiquitin-proteasome pathway’s counterpart deubiquitinases (DUBs) to prevent unnecessary degradation of key protein targets such as tumor suppressors. Using our various computational algorithms, we designed a protein-based chimera termed “deubiquibodies” (duAbs) that link target-recruiting peptides to the catalytic domain of a potent deubiquitinase. We have also extended our ability to design peptide guides to not just disease-related targets but also PTM effector enzymes. We sought to engineer a highly programmable, heterobifunctional architecture where any protein of interest can be brought in proximity to PTM enzymes. For both architectures, we systemically identified peptide-guided constructs that exhibit robust intracellular deubiquitination in vitro across diverse targets; we further show that stabilization of our targets show downstream functional effects, leading to a decrease in disease phenotypes.

Results, Conclusions, and Discussions: Utilizing a previously published list of catalytically potent DUBs, we first compared different catalytic domain and linker combinations fused with the YFP nanobody to engineer our duAbs. Our preliminary data exhibits significant increases in KCNQ1-YFP with the OTU deubiquitinase OTUB1 in a DUB-dependent manner. We wanted to extend the applicability of our duAbs by leveraging our various generative language models (SaLT&PepPr, PepPrCLIP, and PepMLM) for peptide binder design. We demonstrate we can significantly increase β-catenin expression and show downstream functional effect by increasing Wnt signaling via our TOPFlash-GFP assay. We further demonstrate DUB-dependent stabilization of FOXP3, a disordered transcription factor responsible for T-cell regulation, WEE1, a tumor suppressor kinase downregulated in several cancers, and p53, a hallmark tumor suppressor critical in regulating cell division (Figure 1).

Additionally, we linked targeting peptides to validated, experimentally-derived ligands of the USP11 deubiquitinase called FYLIR and LXLL. We designed SaLT-PepPr-derived peptides to target FOXP3 and incorporated them into our plasmid backbones containing FYLIR and LXLL ligands. Evaluating FOXP3 stabilization via a fluorescence-based assay in human HEK293T cells, we demonstrate that our bifunctional peptide architectures resulted in stabilization of FOXP3 in a DUB-dependent manner (Figure 2).

We seek to achieve CRISPR-like programmability for proteome-editing biologics and provide high-throughput, experimental tools to analyze PTMs in vitro and in vivo – here, we develop a modular, genetically-encoded architecture for targeted protein stabilization by designing guide-like peptides linked to potent DUBs and target-recruiting peptide binders, enabling intracellular stabilization of various target proteins. We leverage our various generative protein language models to derive binding peptides based on sequence only, enabling stabilization of even “undruggable” protein targets lacking previously accessible binding sites and disorderly structural characteristics. Our experimental results show statistically significant stabilization of various disordered proteins with clinical relevancy for modulation of downstream signaling pathways. Future work involves optimizing these architectures further for increased modularity and efficacy of stabilizing targets, as well as characterizing our architecture both computationally and experimentally for translational studies. Collectively, we seek to establish programmable proteome editing architectures, aiming to further our progress in our research and the drug development fields.

Acknowledgements (Optional):