From randomness to recognition: modeling the evolution of DNA sequence information during enrichment for binding.

Motivation: Enrichment of random nucleic acid libraries for binding to a target can result in a specific aptamer sequence, but quantitative models describing the detailed evolution of molecular information during such processes are still lacking. Results: A masked language model (MLM) trained on unlabeled DNA sequences after partial enrichment to multiple targets was used to create an encoder that generates latent space sequence representations encompassing the attributes of binding enriched libraries. Independent replicate enrichments against the same target converged to nearly identical latent representations despite containing no overlapping sequences, demonstrating that the representation captured the functional characteristics of the library specific to the target. The degree of divergence of enriched libraries from the original unenriched library strongly correlated with experimental binding, and classifier performance based on latent embeddings captured target specificity, including single amino acid differences. These results show that latent space models provide a quantitative measure of molecular information evolution during enrichment and can provide evidence of binding outcomes, offering both conceptual insight into the evolution of molecular information and practical strategies for designing more effective initial libraries and enrichment processes. Availability: The datasets and code are available on Zenodo: https://doi.org/10.5281/zenodo.14941815.