Can We Replace Reads by Numeric Signatures? Lyndon Fingerprints as Representations of Sequencing Reads for Machine Learning

Representations of biological sequences facilitating sequence comparison are crucial in several bioinformatics tasks. Recently, the Lyndon factorization has been proved to preserve common factors in overlapping reads, thus leading to the idea of using factorizations of sequences to define measures of similarity between reads. In this paper we propose as a signature of sequencing reads the notion of fingerprint, i.e., the sequence of lengths of consecutive factors in Lyndon-based factorizations of the reads. Surprisingly, fingerprints of reads are effective in preserving sequence similarities while providing a compact representation of the read, and so, k-mers extracted from a fingerprint, called k-fingers, can be used to capture sequence similarity between reads. We first provide a probabilistic framework to estimate the behaviour of fingerprints. Then we experimentally evaluate the effectiveness of this representation for machine learning algorithms for classifying biological sequences. In particular, we considered the problem of assigning RNA-Seq reads to the most likely gene from which they were generated. Our results show that fingerprints can provide an effective machine learning interpretable representation, successfully preserving sequence similarity.