tailfindr: Alignment-free poly(A) length measurement for Oxford Nanopore RNA and DNA sequencing

Maximilian Krause; Adnan M. Niazi; Kornel Labun; Yamila N. Torres Cleuren; Florian S. Müller; Eivind Valen

Oxford Nanopore Technologies recently presented an updated basecalling strategy with Flip-flop models instead of standard models used for neural network nucleotide decoding [35] . In this approach, the raw current level data is decoded by averaging 2 sample points instead of 5, resulting in higher-resolution basecalling. Additionally, we could not observe moves of 2, which in most cases represent missed nucleotide translocations. When we calculate the readspecific nucleotide translocation rate from reads basecalled with standard and Flip-flop models, we routinely observe lower average values for Flip-flop model basecalling. The calculated geometric mean of moves is often below 8. The geometric mean calculated from standard model basecalling routinely results in translocation rates between 8 and 9, which agrees with average translocation rates communicated by ONT (unpublished communications). Compared to standard model basecalling it is possible that Flip-flop model basecalling over-segments the raw data, resulting in too low average nucleotide translocation rates when calculated by the geometric mean. This is exemplified by the observation that the most frequently observed nucleotide translocation rate is 2 samples, the lowest possible value considering the resolution of 2 sample points. Instead of calculating the average rate by geometric mean, we used the arithmetic mean after discarding the 5% highest outliers as an approximation of the right normaliser, based on comparisons to results obtained by standard model basecalling on the same PCR-DNA molecules.