Measuring sequencing accuracy
Every base a long-read sequencer reports carries a quality score, and each of those scores comes from the instrument's own basecalling model. We measure error per read directly, against plasmid standards whose sequence is known before they are measured.
Measurement and standards
Accuracy is treated here as a property of one molecule, measured against material of known sequence and fed back into the model that calls it.
Single-Molecule Accuracy sequencing (SMA-seq)
A per-run accuracy figure describes a library, not any molecule in it. SMA-seq scores error read by read against plasmid standards of known sequence, which makes ground truth a physical molecule instead of a consensus drawn from the reads under examination.
Each physical standard is assembled by Golden Gate cloning, and its sequence is confirmed by Sanger sequencing, restriction digest, and capillary electrophoresis before it is used. The current panel tiles a stretch of the CYP2D6 region. Barcoded adapters keep each molecule's intended identity readable after the panel is pooled for nanopore sequencing, so the quality score the basecaller reports for a read can be compared with the error measured on that read. The constructs are grown in bacteria, which lets one panel be prepared with or without DNA methylation and carry its own controls. No standard is perfectly pure, and a standard's purity limits what it can measure.
Basecaller fine-tuning
A basecaller is a neural network trained on the manufacturer's data, and its accuracy on a given library depends on how closely that training set resembles it. We fine-tune these models on the lab's own standards, using the per-read error SMA-seq measures as ground truth labels and the raw current signal from the same reads as input.
One-pot library preparation
Conventional long-read preparation moves DNA through separate shearing, end repair, and ligation steps, each of which loses material and blurs the correspondence between an input molecule and a read. The lab builds libraries in one reaction that defines both ends of every molecule, and that correspondence is what a per-read score depends on.
Read-end integrity and end-reason filtering
A nanopore records one continuous current trace, and the instrument software divides it into reads and records why each one stopped. That annotation is absent from the files most analyses consume, and it answers a question a quality score cannot: whether the molecule was sequenced to completion.
Scarless Type-IIS assembly and DNA standards
Ground truth has to be established independently of the instrument being measured, and no vendor sells a standard shaped like a particular CYP2D6 star allele. The lab assembles its own plasmid standards by scarless Type-IIS cloning and qualifies each construct by sequence before it is used to score anything.
Targeting and enrichment
A single locus reaches the sequencer by one of two routes, enrichment before the run or selection during it.
Cas9-targeted sequencing and nuclease activity
Cas9 enrichment makes long-read sequencing of a single pharmacogene practical at clinical scale, and the libraries it produces do not behave as a clean blunt cut would predict. The lab treats the geometry of the cut as an experimental variable, since a staggered end and a blunt end are not interchangeable inputs to a ligation.
Nanopore adaptive sampling for pharmacogene panels
Adaptive sampling moves enrichment onto the instrument: a molecule is basecalled and mapped while it is still translocating, and rejected if it falls outside the panel. The panel becomes a configuration file instead of a chemistry, and the open question is whether the depth this reaches supports a reproducible diplotype.
Interpretation and infrastructure
Per-read accuracy has to reach a clinical or genomic conclusion, and the provenance that lets it be checked travels with it.
Pharmacogenomics and CYP2D6 diplotyping
CYP2D6 governs how patients metabolize a large share of prescribed drugs, and it sits beside the near-identical pseudogene CYP2D7 in a region short reads cannot resolve. A diplotype call inherits the error model of the reads beneath it, which is why this work is done against the same standards the accuracy program uses.
Long-read telomere profiling in liposarcoma
Standard telomere assays report an average over a population of cells, which hides the differences between chromosome arms and between alleles. A read that passes through the telomere into unique subtelomeric sequence can be assigned to one arm, and the lab is preparing a study that will apply that per-arm measurement to liposarcoma.
Pipelines, provenance, and reproducible analysis
Most available tooling accounts for accuracy at the level of a run, and a pipeline that reports one mean quality cannot be asked what happened to a single molecule. The lab versions its registries, references, and computing environments as part of the method: a figure can then name the molecule and the code behind it.
Epigenetic assays
These methods are being extended to epigenetic assays, among them chromatin conformation, ChIP-seq, and RNA-seq. The work is early and has produced no published results yet.
