WarpSTR: determining tandem repeat lengths using raw nanopore signals

Abstract

Motivation: Short tandem repeats (STRs) are regions of a genome containing many consecutive copies of the same short motif, possibly with small variations. Analysis of STRs has many clinical uses but is limited by technology mainly due to STRs surpassing the used read length. Nanopore sequencing, as one of long-read sequencing technologies, produces very long reads, thus offering more possibilities to study and analyze STRs. Basecalling of nanopore reads is however particularly unreliable in repeating regions, and therefore direct analysis from raw nanopore data is required.

Results: Here, we present WarpSTR, a novel method for characterizing both simple and complex tandem repeats directly from raw nanopore signals using a finite-state automaton and a search algorithm analogous to dynamic time warping. By applying this approach to determine the lengths of 241 STRs, we demonstrate that our approach decreases the mean absolute error of the STR length estimate compared to basecalling and STRique.

Availability and implementation: WarpSTR is freely available at https://github.com/fmfi-compbio/warpstr.

Figure 1.

Overview of the WarpSTR analysis.

Figure 2.

A finite-state automaton modeling a DM1 STR locus (CAG) and its flanking sequences (shown in gray).

Figure 3.

Extended finite-state automaton over the k-mer space.

Figure 4.

A part of the warping path produced by the WarpSTR search algorithm using the state automaton for DM1 locus. States representing k-mers are shown on the left, while the part of the dynamic programming matrix is shown on the right. The warping path, i.e. the path through states with the lowest cost is shown as the sequence of lines with corresponding nucleotides.

Figure 5.

Signal polishing effect on the alignment. Example alignment of the signal to the expected signal from the state automaton before (top) and after polishing (bottom). Before polishing, the normalized signal values are much higher than the expected signal values, and some of these differences decrease after polishing. More importantly, a spurious repeat in the highlighted windows disappears after polishing.

Figure 6.

MAE for WarpSTR and basecalling for individual loci colored by repeating pattern length. 25 loci that have very high MAE in both methods are not shown; these are most likely due to large expansions not captured by VCF callers in the gold standard.

Figure 7.

Clustered predictions of DM2 for NA24385 subject split per repeat unit.