Examining selection dynamics and limitations in multi-round protein selection of high diversity libraries

Abstract

Proteins and peptides underpin essential biological functions and technological applications, from targeting disease-relevant interactions to providing broad enzymatic activities. However, engineering molecules with desired properties remains difficult, owing to complex sequence-structure-function relationships and the lack of data on specific systems. Experimental selection strategies, including directed evolution, phage display, and mRNA display, address this challenge by leveraging high diversity libraries and iterative enrichment under defined selection pressures. This allows for the identification of candidates without requiring extensive prior knowledge, and can generate extensive datasets for use in machine learning. While many selection systems exist, comparisons across different selection approaches are hindered by the lack of a unifying analytical framework. Here, we developed a toolset of broadly applicable analyses for assessing selection dynamics in multi-round or multi-condition experiments, ranging from position level analysis of sequence properties to full sequence space mappings through protein language model embeddings. Performing analyses across different systems, we identify desirable traits in selection experiments including enrichment of distinct sequence patterns and correlation between enrichment and final desired functions. Notably, even under weak selection regimes with all sequences <1% frequency, functional sequences (e.g., 70nM IC50 binder to SARS-CoV-2 main protease) are still consistently enriched. We also find repeated selections of the same starting library can help differentiate selection effects of varying conditions (e.g., different delivery of metal ligand) from system noise. These findings, along with the toolset, can be used to guide experimental design, interpretation, and troubleshooting across protein and peptide discovery platforms.