Comparative ngram analysis of whole-genome sequences

A current barrier for successful rational drug design is the lack of understanding of the structure space provided by the proteins in a cell that is determined by their sequence space. The protein sequences capable of folding to functional three-dimensional shapes of the proteins are clearly different for different organisms, since sequences obtained from human proteins often fail to form correct three-dimensional structures in bacterial organisms. In analogy to the question "What kind of things do people say?" we therefore need to ask the question "What kind of amino acid sequences occur in the proteins of an organism?" An understanding of the sequence space occupied by proteins in different organisms would have important applications for "translation" of proteins from the language of one organism into that of another and design of drugs that target sequences that might be unique or preferred by pathogenic organisms over those in human hosts. 

Here we describe the development of a biological language modeling toolkit (BLMT) for genome-wide statistical amino acid n-gram analysis and comparison across organisms (freely accessible at www.cs.cmu.edu/~blmt). Its functions were applied to 44 different bacterial, archaeal and the human genome. Amino acid n-gram distribution was found to be characteristic of organisms, as evidenced by (1) the ability of simple Markovian unigram models to distinguish organisms, (2) the marked variation in n-gram distributions across organisms above random variation, and (3) identification of organism-specific phrases in protein sequences that are greater than an order of magnitude standard deviations away from the mean. These lines of evidence suggest that different organisms utilize different "vocabularies" and "phrases", an observation that may provide novel approaches to drug development by specifically targeting these phrases. The results suggest that further detailed analysis of n-gram statistics of protein sequences from whole genomes will likely - in analogy to word n-gram analysis - result in powerful models for prediction, topic classification and information extraction of biological sequences.