From Randomness to Recognition: Modeling the Evolution of DNA Sequence Information During SELEX

S ystematic E volution of L igands through E xponential E nrichment (SELEX) was used as a model system to explore the evolution of DNA sequence information and function during enrichment of molecular recognition to a series of related target molecules. Using a Natural Language Processing (NLP) based approach, a model was trained on an unlabeled mixture of oligonucleotide sequences sampled from both unenriched libraries and from six libraries that had been enriched for binding to either peptide or protein targets. This general model (pre-trained model) was then used to generate latent space representations of sequences that contained embedded binding information. Unsupervised clustering was used to create two dimensional maps that allowed comparison of the overlap between the latent space representations of sequences from different unenriched and enriched libraries. Replicate, independent enrichments to the same targets starting from completely unique random libraries gave rise to essentially indistinguishable latent space representations. However, similar representations between unenriched and enriched sequences, or between enriched sequences from different targets, resulted in distinct clustering patterns. The extent of overlap between the patterns from unenriched and enriched libraries was correlated with overall target binding by the enriched library. Further, the pre-trained model was fine-tuned to classify which target library each sequence belonged to, and the accuracy of classification between unenriched and enriched sequences was also correlated with the overall binding of the enriched library to its target. Finally, it was possible to distinguish libraries enriched for binding to two different targets differing by as few as 1 in 30 amino acids. Thus, using NLP-based approaches, it is possible to relate the DNA sequences present in an enriched library with the ability of the library to specifically bind its intended target, providing a tool for both guiding the enrichment process and more deeply understanding the structure-binding relationships that evolve.