Sequence analysis paradigm shift reveals unsuspected semantic properties of species proteomes

Alignment-based methods dominate molecular biology. However, by primarily allowing one-to-one comparisons, these methods are focused on a gene-centered viewpoint and lack the broad context essential to analyze how complex biological systems function and evolve. In actuality, a gene is part of genome where more than one sequence contributes to the functional network and evolutionary trajectory of the cell. The need for conservation of established interactions, is arguably more important to the evolutionary success of species than conservation of individual function. To test whether such contextual information exists, a distributional semantics method - Latent Semantic Analysis (LSA), was applied to thousands of species proteomes. Using natural language processing, Latent Taxonomic Signatures (LTSs) were identified that outperformed existing alignment-based BLAST methods when random protein sequences were being mapped to annotated taxonomy according to GenBank. LTSs are a novel proteome distributed feature, suggesting the existence of evolutionary constraints imposed on individual proteins by their proteome context. Even orphan proteins are exhibiting LTSs, which makes their uniqueness linked to a specific taxonomic level questionable. Unlike more simple bias, LTSs represent a self-similarity pattern, where random sets of species proteins show the same statistical properties of a complete proteome at many scales. Natural language processing and machine learning provide insights not easily discernable using alignment based methods suggestive there is more to species related differences than just translational optimization.