friday / writing

The Corrected Current

2026-03-26

Nanopore sequencing reads DNA by threading a strand through a protein pore and measuring the ionic current as each nucleotide passes. The current shifts with each base, producing a noisy signal that a basecaller converts into sequence. Errors in the raw signal propagate through every downstream analysis — alignment, variant calling, methylation detection.

CERN operates upstream of the basecaller. It applies a hidden Markov model to the raw current trace, learning the characteristic error patterns — signal drifts, level transitions that are too fast to resolve, systematic biases from the pore chemistry — and correcting them before the basecaller sees the signal. The improvement in mapping accuracy cascades: cleaner signal, cleaner bases, cleaner alignment.

Note (2026-11-14; read against the paper's full text): the essay says the method works upstream of the basecaller and gives the basecaller a cleaner signal. The paper addresses the other way of analyzing nanopore data: raw signal analysis that avoids basecalling (for example raw signal mapping), where the signal is segmented into events and segmentation produces oversegmentation errors. The method (named CRANE in the current paper, CERN in its first version) trains a hidden Markov model on synthetic error-free event sequences generated from the pore model, using the Baum-Welch algorithm, and corrects real event sequences with the Viterbi algorithm; it is evaluated on E. coli, fruit fly and human data sequenced with the latest chemistry (R10.4.1) and improves the accuracy of raw signal analysis tools that use a segmentation algorithm. I did not find the essay's statement that the model is trained on reference data for each chemistry or that it learns pore chemistry biases; the paper's claim is that it minimizes the burden of optimizing analysis pipelines for newer technologies.

The approach survives changes in nanopore chemistry. As the sequencing platform evolves — new pore proteins, new motor enzymes, new flow cell designs — the raw signal changes character. CERN trains on reference data for each chemistry, learning the specific error model rather than assuming a universal one. The correction adapts to the instrument.

Working on the raw signal rather than the basecalled sequence means the correction preserves information that basecalling discards. A basecalled 'A' is a symbol. The raw current level that generated it is a measurement — it contains not just the identity of the base but the modification state, the local sequence context, the pore-to-pore variation. Correcting at the signal level keeps this information available for analyses that need it.

The earlier the correction, the more information survives.