Nanopore tRNA reads skip basecalling: tyCooNN classifies all 48 E. coli tRNAs at 98.2% accuracy from raw current signals

Nanopore direct-RNA reads of tRNAs can be assigned to their tRNA species directly from ionic current signals, without basecalling and without sequence alignment. The prior route basecalled each read and mapped it back to reference sequences, but dense tRNA modifications corrupt basecalling, so mapping either loosened stringency and lost accuracy or tightened it and lost usable reads. On 1000 Guppy-basecalled reads per tRNA, the ntc44 classifier averaged 99.1% classification accuracy while retaining 82.7% of reads; Minimap2 averaged 75.6% accuracy with 72.5% retention, BWA-MEM reached 97.8% accuracy but retained only 50.7%, and Parasail retained 94.4% at 71.2% accuracy.

Sequencing an equimolar mixture of 44 purified tRNAs on both platforms yields a per-species read-count bias parameter, and correcting with it brings read fractions close to true molecular proportions. Read counts cannot be converted directly into molecular copy numbers because each tRNA species carries its own read bias on each platform, and that bias was previously hard to estimate accurately. In the equimolar mixture each tRNA should account for about 2.27% of reads, but the observed distribution ran from 0.5% to 5.9%, with bias up to 5.48-fold between platforms (tRNAPro2), reproduced across three independent experiments; after correction, nanopore and Illumina abundance estimates for E. coli total tRNA rose from R2=0.23 to R2=0.86.

The corrected pipeline returns absolute abundances for all 48 E. coli tRNA species within one day and supports comparisons across culture conditions and modification-deficient strains. Earlier tRNA sequencing mostly reported relative changes between samples and could not directly compare copy numbers across tRNA species. In stationary-phase E. coli, tRNAGly3 was the most abundant at 7.08% of the total tRNA pool and tRNASec the least at 0.07%, roughly a 101-fold difference; in a thiI deletion strain, Northern blotting showed s4U8-bearing tRNACys fell to about one-fifth of the wild-type level.

Binary classifiers built on tyCooNN judge whether a given tRNA modification is present read by read and estimate its frequency. Mapping-based detection that relies on basecalling errors varies widely in sensitivity by modification type, and modifications such as s4U are barely detectable. The binary classifier for Ψ55 in tRNAGly3 averaged 96% validation accuracy, its predictions on samples mixed 1:3, 2:2 and 3:1 matched LC-MS ratios, and its false positive and false negative rates were 3.5% and 4.8%, versus 19.0% and 23.8% for the mapping-based method on the same samples.

The demonstrated scope is E. coli tRNA: 44 purified species support training and the equimolar calibration, classification accuracy across 48 species is 98.2%, and absolute abundance and differential analysis are carried out in laboratory-cultured E. coli and modification-enzyme knockout strains. The authors propose extending the method to other organisms including humans, but no such experiment is reported.

The authors note that trained models do not transfer directly between RNA002 and RNA004 chemistry and must be retrained for a new chemistry; ntc48 is slightly less accurate than ntc44 because of ambiguity among eight isodecoders; and against earlier methods, the nanopore abundances correlate moderately with 2D-PAGE (R2=0.56) and DORQ-seq (R2=0.52) but only weakly with microarrays (R2=0.14), so abundance estimates differ appreciably between methods.