Atlas window total corrected to per-strand-length; peptide reproduce block runs verbatim from a clean clone
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The exact off-target atlas of the nucleic-acid medicines — the registry, not a list
Every nucleic-acid substance NCATS GSRS publishes a usable sequence for, screened against every
window of the whole human transcriptome. Complete enumeration, one integer per window per strand,
no sampling and no parameter inside the computation.
THE SCOPE IS THE FIX. The first version of this screen covered 33 medicines from a list written
from memory. That is a defect no harness can catch: a drug nobody recalls is silently absent from a
safety map, the arithmetic correct and the coverage wrong. The registry is enumerated instead --
742 substances of class nucleicAcid, 740 carrying a sequence, and those whose every subunit falls
in the 8-60 nt range give 472 strands across 350 substances. The enumeration provably contains
every one of the original 33. An 85-minute run of the smaller version was retired rather than
published, because correct-and-incomplete is the worse failure for this artifact.
THE TARGET IS NOW MEASURED, NOT DECLARED. A screen that needs to be told the target cannot
enumerate a registry, since most substances carry no target annotation. A strand's target is
therefore read out of the transcriptome -- the set of genes carrying its perfect complement -- and
off-targets are the qualifying windows outside that set. A strand with no perfect complement
anywhere gets no off-target list, because its published sequence and the transcriptome disagree.
Where a declared target exists it is a cross-check, never an input.
Refusals are recorded rather than hidden: 388 substances outside the oligonucleotide length range
(vectors, plasmids, genes, mRNA vaccines), 2 with no sequence, 2 with a residue the program will
not interpret. The aptamers are screened and fail, which is the point -- an instrument returning an
off-target list for a molecule that binds proteins would be wrong.
The screen was sharded by strand and never by transcript, so every shard reads the entire
transcriptome and every strand is scored against every window; splitting the strand list changes no
number. The published seal is the sha256 of the eight shard seals in order.
Cross-check: pelacarsen reaches LPA here, the same result the independent single-drug screen in
Study 26 reached, and the sodium-salt record reaches it separately.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>