k-mer based analysis for cancer transcriptomics at nucleotide resolution

Overview & Synopsis

The transcriptome integrates RNA variations produced by two main processes: DNA alteration (e.g., mutations, gene-fusions, etc.) and post-transcriptional modifications (e.g., alternative splicing, etc.). It is an ideal research object for genotype-phenotype association. Next-generation sequencing (NGS) techniques allow a measurement of transcriptome at single-nucleotide resolution, both rapidly and at a relatively low cost. Conventional transcriptome analyses based on gene/transcript quantification do not make use of the full precision of this NGS data. Instead, they inherit the perspective of microarray measurements that consider gene or transcript as the elementary features for statistical analysis or machine learning. In this thesis, we discuss and develop a novel perspective of transcriptome analysis based on k-mers (substrings with fixed length k, typically k = 31) signals. Rather than quantifying predefined genes/transcripts, we count short and arbitrary k-mers and use them directly as elementary features. This allows representing gene expression events at the single-nucleotide resolution, and thereby driving insights into local events occuring at sub-transcript level. Also, this approach prevents the informative signals from cancelling each other at gene/transcript quantification stage. This thesis presents different aspects of this endeavor: (i) The KaMRaT software (chapter 3), developed during the thesis, supports various methods for reducing k-mer dimensionality and improving their specificity. This includes: a ranking-selecting module by evaluating association between sample counts and conditions; an extension module merging overlapped k-mers; a filter module considering their expression level; and a masking module to extract k-mers with a given sequence list. Results show that KaMRaT is an effective and efficient software, with highly flexible and easy-to-use characteristics. (ii) A comparison between classifiers obtained using k-mer or conventional gene features (chapter 4) shows that k-mer-feature-based classifier performs as well as classical gene-feature-based one, in the prostate prognostic problem, with the former further supporting finding novel unannotated events. (iii) The analysis of the replicability of informative k-mer signals in an inter-cohort research (chapter 5) shows that k-mer signals are replicable across independent data sets, and the k-mer-based inter-cohort research allows finding stable informative signals. (iv) The Kmerator software (chapter 6) allows utilization of reference-free k-mer signals as a proxy to reference-based gene expression measures. (v) Application of the REINDEER and Kmerator software (chapter 7) allows for arbitrary sequence indexing and abundance query across RNA-seq samples.