Shining a light on dark sequencing : characterising errors in Ion Torrent PGM data

The Ion Torrent Personal Genome Machine (PGM) is a new sequencing platform that substantially differs from other sequencing technologies by measuring pH rather than light to detect polymerisation events. Using re-sequencing datasets, we comprehensively characterise the biases and errors introduced by the PGM at both the base and flow level, across a combination of factors, including chip density, sequencing kit, template species and machine. We found two distinct insertion/deletion (indel) error types that accounted for the majority of errors introduced by the PGM. The main error source was inaccurate flow-calls, which introduced indels at a raw rate of 2.84% (1.38% after quality clipping) using the OneTouch 200 bp kit. Inaccurate flow-calls typically resulted in over-called short-homopolymers and under-called long-homopolymers. Flow-call accuracy decreased with consecutive flow cycles, but we also found significant periodic fluctuations in the flow error-rate, corresponding to specific positions within the flow-cycle pattern. Another less common PGM error, high frequency indel (HFI) errors, are indels that occur at very high frequency in the reads relative to a given base position in the reference genome, but in the majority of instances were not replicated consistently across separate runs. HFI errors occur approximately once every thousand bases in the reference, and correspond to 0.06% of bases in reads. Currently, the PGM does not achieve the accuracy of competing light-based technologies. However, flow-call inaccuracy is systematic and the statistical models of flow-values developed here will enable PGM-specific bioinformatics approaches to be developed, which will account for these errors. HFI errors may prove more challenging to address, especially for polymorphism and amplicon applications, but may be overcome by sequencing the same DNA template across multiple chips