De Novo Design of Cyclic Peptide Binders Using Generative Models

Many natural product therapeutics are cyclic peptides, which have numerous desirable properties as scaffolds for drug discovery. Most therapeutic cyclic peptides, however, contain non-canonical amino acids (ncAAs), which are not modeled in cyclic peptide design models. Given advances in protein structure prediction, I hypothesize that I can modify an AlphaFold3-based model (AF3) to generate candidate cyclic peptide binder designs against a desired target. With Xuezhi Xie, who implemented an optimized cyclic offset encoding on top of Boltz1 (AF3), we developed CyclicBoltz1 to address this gap by enabling the computational design and structure prediction of cyclic peptides containing ncAAs. In this thesis, I evaluate the performance of this model and implement Markov Chain Monte Carlo (MCMC) optimization for binder design. I benchmark the structure prediction and binder design and show that the model achieves more accurate structure prediction in cyclic peptides with ncAAs and stronger binder designs based on DockQ and RMSD scores.