QGramProjector: Q-Gram Projection for Indexing Highly-Similar Strings

Abstract

Q-gram (or n-gram, k-mer) models are used in many research areas, e.g. in computational linguistics for statistical natural language processing, in computer science for approximate string searching, and in computational biology for sequence analysis and data compression. For a collection of N strings, one usually creates a separate positional q-gram index structure for each string, or at least an index structure which needs roughly N times of storage compared to a single string index structure. For highly-similar strings, redundancies can be identified, which do not need to be stored repeatedly; for instance two human genomes have more than 99 percent similarity.

In this work, we propose QGramProjector, a new way of indexing many highly-similar strings. In order to remove the redundancies caused by similarities, our proposal is to 1) create all q-grams for a fixed reference, 2) referentially compress all strings in the collection with respect to the reference, and then 3) project all q-grams from the reference to the compressed strings.

Experiments show that a complete index can be relatively small compared to the collection of highly-similar strings. For a collection of 1092 human genomes (raw data size is 3 TB), a 16-gram index structure, which can be used for instance as a basis for multi-genome read alignment, only needs 100.5 GB (compression ratio of 31:1). We think that our work is an important step towards analysis of large sets of highly-similar genomes on commodity hardware.