Abstract
In vitro directed evolution techniques, such as mRNA display, enable peptide ligand discovery and optimization. However, physical libraries that rely on a genetic code can only search a small fraction of sequence space due to inherent biases in the genetic code and experimental limitations. To address this challenge, denoising diffusion implicit models (DDIMs) are applied to generate novel peptide ligands against B-cell lymphoma extra-large (Bcl-x(L)), a key cancer target. Starting with high-throughput sequencing data from previous selections, a DDIM is trained to produce novel sequences with high affinity binding. Experimental validation confirms that most generated sequences are functionally equivalent to the original library members for Bcl-x(L) binding and demonstrated comparable binding kinetics and affinity relative to the wildtype and nearest original neighbors. Importantly, this approach generated rare sequences not easily accessible via mutation and directed evolution. These results indicate that DDIMs can complement and expand directed evolution data, efficiently exploring underrepresented regions of sequence space. This approach provides a broadly applicable framework for accelerating ligand discovery and optimizing molecular properties across diverse targets.