Exploring a large cancer cell line RNA-sequencing dataset with k-mers

Abstract

Analyzing the immense diversity of RNA isoforms in large RNA-seq repositories requires laborious data processing using specialized tools. Indexing techniques based on k-mers have previously been effective at searching for RNA sequences across thousands of RNA-seq libraries but falling short of enabling direct RNA quantification. We show here that RNAs queried in the form of k-mer sets can be quantified in seconds, with a precision akin to that of conventional RNA quantification methods. We showcase several applications by exploring an index of the Cancer Cell Line Encyclopedia (CCLE) collection consisting of 1019 RNA-seq samples. Non-reference RNA sequences such as RNAs harboring driver mutations and fusions, splicing isoforms or RNAs derived from repetitive elements, can be retrieved with high accuracy. Moreover, we show that k-mer indexing offers a powerful means to reveal variant RNAs induced by specific gene alterations, for instance in splicing factors. A web server allows public queries in CCLE and other indexes: https://transipedia.fr. Code is provided to allow users to set up their own server from any RNA-seq dataset.

Constrained enumeration ofk-mers from a collection of references with metadata

“…However, recent advances in the indexing of large-scale biological sequence datasets have brought solutions leading to different large scale queryable indexes. Examples include the entirety of assembled bacterial genomes [5], or creating a searchable, quantitative index for RNA-seq data in cancer research [1]. Various methodologies, all of them based on k -mers, have been introduced to this extent.…”

“…Examples include the entirety of assembled bacterial genomes [5], or creating a searchable, quantitative index for RNA-seq data in cancer research [1]. Various methodologies, all of them based on k-mers, have been introduced to this extent.…”

“…• (|B * | 2 + |C * | 2 ) ; where η = max max B∈B |B|, max C∈C |C| would be the size of the longest tag in either B or C. It appears to be impossible to beat this worst-case complexity, unless the Strong Exponential Time Hypothesis [6] is false -see for instance this discussion on Theoretical Computer Science Stack Exchange [2].However, in practice it can be useful to partition sets intelligently in order to limit the number of inclusion tests to be performed. With Algorithm 1 we propose such a method, running in O(η • |P| 2 ) -as per the previous discussion -provided that checking whether f (n) is defined for some n can be done in O(1). :6Constrained enumeration of k-mers from a…”