Hierarchical clustering of DNA k-mer counts in RNA-seq fastq files reveals batch effects

Authors:
  • WKWolfgang Kaisers
    Wolfgang Kaisers
  • HSHolger Schwender
    Holger Schwender
  • HSHeiner Schaal
    Heiner Schaal

Abstract
Batch effects are artificial sources of variation due to experimental design. Batch effect is a widespread phenomenon in high througput data which can be minimized but not always be avoided. Therefore mechanisms for detection of batch effects are needed which requires comparison of multiple samples. Due to large data volumes 1e12 Bytes this can be technical challenging. We describe the application of hierarchical clustering (HC) on DNA k-mer profiles of multiple fastq files which creates tree structures based on similarity of sequence content. Ideally, HC tree adjacency reflects experimental treatment groups but the algorithm may also agglomerate according to sample preparation groups which then indicates the presence of batch effects. This introduces a new perspective in quality control. In order to perform fast analysis on large data sets we implemented a new algorithm. The algorithm and user interface are implemented a new R-package (seqTools). The result is now available on R-Forge for Linux and Windows. Using this implementation, we compared the transcriptome sequence content of 83 human tissue samples on 13 different flowcells. We pairwise compared flowcells where samples from the same tissue type (fibroblasts or jurkat cells) were present and <10% of the Phred scores were 30) does not remove the identified batch effects.