The fundamental principle that selection acts on a gene's function often assumes implicitly that this function is fixed and intrinsic. However, empirical evidence from pangenomics, synthetic biology, and GWAS consistently demonstrates that organismal function is highly context-dependent, varying across genomic backgrounds and cellular states, even for core genes. Drawing a conceptual parallel with modern large language models (LLMs), I propose that genomes, like LLMs, do not encode fixed functions but rather probability distributions over functional and phenotypic outcomes. This framework draws a conceptual analogy between epistasis and transformer-style "attention mechanisms," suggesting that genomic context weights the influence of distant genetic elements. I also introduce the concept of genomic perplexity - an information-theoretic measure of the statistical unexpectedness and incompatibility of a genetic element within its host context. I demonstrate how perplexity serves as a quantifiable metric for the well-known fitness cost associated with inter-species gene flow (e.g. horizontal gene transfer (HGT) and introgression), where a new gene represents a high-perplexity token. This perspective formalizes longstanding observations of genomic fit and provides a testable framework for predicting the integration potential of accessory genes and directing future research in synthetic biology and evolutionary modelling.
James O. McInerney (Wed,) studied this question.