Alignment-free clustering of UMI tagged DNA molecules

Next-Generation Sequencing has led to the availability of massive genomic datasets whose processing raises many challenges, including the handling of sequencing errors. This is especially pertinent in cancer genomics, e.g. for detecting low allele frequency variations from circulating tumor DNA. Barcode tagging of DNA molecules with unique molecular identifiers (UMI) attempts to mitigate sequencing errors; UMI tagged molecules are polymerase chain reaction (PCR) amplified, and the PCR copies of UMI tagged molecules are sequenced independently. However, the PCR and sequencing steps can generate errors in the sequenced reads that can be located in the barcode and/or the DNA sequence. Analyzing UMI tagged sequencing data requires an initial clustering step, with the aim of grouping reads sequenced from PCR duplicates of the same UMI tagged molecule into a single cluster, and the size of the current datasets requires this clustering process to be resource-efficient.

We introduce Calib, a computational tool that clusters paired-end reads from UMI tagged sequencing experiments generated by substitution-error-dominant sequencing platforms such as Illumina. Calib clusters are defined as connected components of a graph whose edges are defined in terms of both barcode similarity and read sequence similarity. The graph is constructed efficiently using locality sensitive hashing and MinHashing techniques. Calib's default clustering parameters are optimized empirically, for different UMI and read lengths, using a simulation module that is packaged with Calib. Compared to other tools, Calib has the best accuracy on simulated data, while maintaining reasonable runtime and memory footprint. On a real dataset, Calib runs with far less resources than alignment-based methods, and its clusters reduce the number of tentative false positive in downstream variation calling.

Calib is implemented in C++ and its simulation module is implemented in Python. Calib is available at https://github.com/vpc-ccg/calib.

Supplementary data are available at Bioinformatics online.