Filtering of Clinical NGS Data to Improve Low Allele Frequency Variant Calling

Massive parallel sequencing (NGS) is useful in detecting and later classifying somatic driver mutations in cancer tumours. False-positive variants occur in the NGS workflow and they may be mistaken for low frequency somatic cancer mutations in a patient sample. This pushes the need for decreasing the noise rate in the NGS workflow since it may improve the detection of rare allele frequency variants, in particular cancer mutations. In this project, the aim was to reduce the level of false-positive variants in an NGS workflow. The scope was limited to looking at substitution errors and their neighbouring nucleotides. Alongside this, it was also a way to understand how different types of substitution errors are distributed in the data, if their frequencies are affected by neighbouring nucleotides and how data processing may affect these substitution rates. A bioinformatic pipeline was set up where a commercially available genomic DNA sample with known variants was subjected to different trimming and filtering settings. The goal was to reduce the substitution error rate as much as possible, without removing any true variants from the data. The optimised settings were trimming the sequencing reads with 5 bp from the tail and filtering sequencing reads that contained 5 or more substitutions. Three additional samples, whereof two were clinical and the third commercial, were tested with these settings. The results showed that in all samples, C:G>T:A substitutions were of a higher frequency compared to the rest of the substitution types. For all samples, A:T>C:G substitutions, where the neighbouring nucleotide was a C or a G on each side, had a higher frequency compared to A:T>C:G substitutions with other neighbouring nucleotides on both sides. Those substitution types were especially targeted by the trimming. For the two commercial samples, substitutions that resulted in the nucleotide combinations >XAA or >XTT were of a higher frequency compared to the same substitution types that did not result in those nucleotide combinations. Filtering reads with 5 or more substitutions particularly targeted these substitution types. Consequently, filtering had a greater effect on the commercial samples, compared to the clinical samples. Overall, trimming and filtering helped reduce transversions more than the transitions, increasing the transition/transversion ratio after processing the data. The results suggest that trimming and filtering can be a useful method to computationally reduce the transversion errors introduced in an NGS workflow, but transition errors to a lesser extent, in particular A:T>G:C transitions. To confirm these findings, more samples should be tested using this methodology. To better understand the effect of trimming and filtering on variant calling, the scope could in the future be expanded to also look at small insertions and deletions