Skip to main content
- Reference-based accuracy assessment uses one algorithm both to produce a result and to judge it, so a systematic bias can propagate without ever being visible. A physical standard breaks that circle for the standard itself. Whether the resulting error model transfers to an arbitrary patient sample is untested.
- Error is not uniform along a read or across sequence context. How much of the per-position structure is a property of the chemistry, and how much of the basecalling model, is not yet separated.
- How much of the terminal signal loss is recoverable by software and how much is physically absent because the strand has already left the pore. Those need different remedies, and one is a chemistry problem.
- Whether end-reason filtering changes downstream calls enough to justify itself in routine pipelines, as opposed to being demonstrably correct in principle.
- How far a targeted method generalizes. Long reads are not a uniform upgrade; they are a targeted intervention for the structurally complex minority of cases, and the boundary of that minority is not sharp.
- Whether a diplotype call carrying an explicit confidence changes clinical decisions in practice, which is a question about clinical workflow rather than about sequencing.
- Whether cut geometry can be designed for rather than merely measured, and what that would cost in guide choice.
- How far a one-pot approach scales in multiplexing before the reaction stops being a single well-behaved system.
- A construct that assembled correctly is not the same as a construct that behaves as a standard. How much of the residual error measured against one of these plasmids belongs to the sequencer and how much to the assembly has not been disentangled.
- The platform is application-agnostic by design, which means nothing in it checks whether a requested target is a sensible standard for the measurement it will be used in. That judgment currently sits with the person requesting the construct.
- Rejecting a molecule mid-pore is a decision made on partial evidence, and the reads it discards are not recorded the way a completed read is. How much of the apparent on-target yield depends on where the rejection threshold sits has not been separated from the enrichment itself.
- The comparison against Cas9 enrichment is currently a comparison of depth. Whether the two enrichments differ in which alleles they can resolve, rather than only in how deeply they cover them, remains to be shown.
- Arm assignment depends on how much subtelomere a read retains, and the requirement is in the kilobase range rather than the hundreds of bases. Whether that is reachable on clinical liposarcoma material, as opposed to on reference assemblies, is unknown.
- Separating telomerase-positive from alternative-lengthening tumors is the point of the atlas. Whether per-arm length variance separates those classes in liposarcoma specifically, rather than in cell lines, is not yet established.
- Whether TERRA overexpression in tumors that lengthen their telomeres without telomerase follows from a definable state of TAR1, the subtelomeric repeat that carries the TERRA promoters, in its sequence, copy number or DNA methylation. Short reads cannot resolve TAR1, and a 2026 to 2027 collaboration with the Cancer Science Institute of Singapore is designed to test it in sarcomas.
- Provenance is only as good as its least-recorded stage. Which stages of the current stack can genuinely reproduce a figure from its inputs, and which only record that they ran, has not been audited end to end.
