Design of Peptide-Based Protein Degraders via Contrastive Deep Learning

Abstract

Most disease-associated proteins are considered “undruggable” by standard small molecule-based pharmaceutical approaches. Ubiquibodies are a promising experimental therapeutic intervention that can effectively treat so-called undruggable targets by hijacking the natural ubiquitin-proteasome pathway. In this work, we develop a sequenced-based machine learning model to predict protein-peptide binding for the purpose of designing ubiquibodies. By leveraging known experimental binding proteins as scaffolds, we create a streamlined inference pipeline, termed Cut&CLIP, that efficiently selects ubiquibody peptide candidates for downstream screening. Unlike existing computational methods, this approach is applicable to the vast majority of human proteins, including disordered and unstable targets. Finally, we test ubiquibodies generated by the Cut&CLIP pipeline experimentally and demonstrate robust intracellular degradation in human cells of therapeutic targets for pancreatic cancer, COVID-19, Ewing sarcoma, and fatty liver disease, motivating further development of the technology for clinical translation.