History / CRISPR Genome Off Target Map

State the verdict on both pages instead of ending in disclaimer Founder, directly: do not overstate and do not understate -- just state that something shows promise and warrants lab follow-up, or state that Affine.Earth does not call it safe. Both pages were failing the second half of that. They ended in pure disclaimer, which is its own kind of dishonesty: a result that clears every control and then says only "this is not a recommendation" has told the reader nothing they can act on. STUDY 26, COMBINATIONS. The call is now stated: these eleven tumour types SHOW PROMISE AND WARRANT LABORATORY FOLLOW-UP, because a result that survives the control which killed the previous stage is worth a bench. And the other half, equally plain: Affine.Earth does not call any of these pairs safe, effective or ready for a patient, and nobody should take anything on the page. The four that do not clear are named as not promising ON THIS EVIDENCE, with the reason given as our own lack of power rather than a statement about those cancers -- so a lab knows not to follow up on those four from this study. CRISPR MAP. What the screen does call: fifteen guides, each with exactly one perfect match and zero sites at a single mismatch, thirteen of fifteen with zero at two. On the question the instrument answers -- how many places in the genome match this guide exactly -- these are clean guides and the count is not an opinion. What it does not call: Affine.Earth does not call any of these therapies safe, because off-target potential is one input to safety among many and this page measures only that one. Where a bench should look: the sites at four mismatches or fewer, published in full with coordinates and strands, for exactly that purpose. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

rg78803 committed Sep 6, 2026

CRISPR genome map: every guide the registry publishes, not the two we could name The first version screened two guides found by searching NCATS GSRS for drug names recalled from memory. That is a completeness defect no harness can catch -- a guide nobody recalls is silently absent from a safety map. All 742 nucleicAcid substances were therefore scanned for the canonical SpCas9 sgRNA scaffold GUUUUAGAGCUAGAAAUAGCAAGU. Fifteen carry it, and in every one the scaffold begins at position 20, which makes the 20-base spacer a measurement rather than a convention. Finding guides by STRUCTURE rather than by NAME vindicated an earlier refusal instead of overturning it. NTLA-2001's own record is the full 4,423-base molecule carrying the Cas9 messenger RNA, with no isolatable spacer, and it was refused. Its guide exists under a separate record -- Nexiguran, UNII 5G537B4BTJ -- and the structural scan found it. Refusing to guess lost nothing. The cut site is now MEASURED, not declared. Most guide records carry no target annotation, so demanding a declared chromosome cannot enumerate a registry -- it would have reported nothing for fourteen of fifteen guides. A guide's measured site is where its zero-mismatch match actually falls in the assembly; off-targets are the sites outside it; and a guide with no zero-mismatch site anywhere gets no off-target list because its published spacer and the assembly disagree. Validated on chromosome 2 before the whole-genome run: only Casgevy, whose BCL11A target is on chr2, found a site there, and all fourteen others correctly found none. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Three defects were caught in the page generator before publication, each by checking output rather than trusting code. The seal regex lacked re.MULTILINE and matched nothing, so the gate refused and published nothing. A prefix-based join then filled the table but gave the two evoncabtagene guides and the two soficabtagene guides IDENTICAL off-target profiles, because the summary table truncates names to 37 characters and both pairs share that prefix -- one guide's numbers under another guide's name. Keying the detail blocks by UNII fixed it, and the published rows now reproduce the per-guide profile computed independently from the raw output.

rg78803 committed Sep 6, 2026

CRISPR map: print the two guide sequences on the page itself The page described where the spacers came from and how they were separated from the sgRNA scaffold, but never printed the sequences. A reader had to open the corpus file to see what was actually screened, which is exactly the kind of thing this program exists to avoid. Both 20-base spacers are now on the page beside their UNII, and both UNIIs are pinned by check_figure so a later edit cannot silently change what the page claims was tested. Found by checking the LIVE page for its own key figures after publishing rather than trusting the push -- the seal, the assembly digest and the site counts were all present, and the guide sequence was the one thing missing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

rg78803 committed Sep 6, 2026

The exact genome-wide off-target map of the approved CRISPR medicines Casgevy (exagamglogene autotemcel) is an approved CRISPR therapy for sickle cell disease and beta-thalassemia; NTLA-2002 is a late-stage in-vivo therapy for hereditary angioedema. Both publish their guide in NCATS GSRS. Neither trial answers the question a patient actually has about a molecule that edits their genome, and that question does not expire when a trial reads out: besides its intended cut site, where else could this guide direct a cut? Screened against every NGG PAM site on both strands of GENCODE GRCh38 primary assembly -- 3,099,750,718 bases over 194 sequences, 304,796,751 candidate sites, one integer each. No sampling, no seed heuristic, no alignment score, no parameter. 1,651 windows containing an N are set aside and counted rather than scored, because absence is neither a match nor a mismatch. Each guide has EXACTLY ONE zero-mismatch site in the whole genome, and it is the intended one. Casgevy: 0mm=1, 1mm=0, 2mm=0, 3mm=6, 4mm=137. NTLA-2002: 0mm=1, 1mm=0, 2mm=0, 3mm=4, 4mm=182. One perfect site and then nothing at all until three mismatches, across three billion bases and both strands. Every site at four mismatches or fewer is named with coordinates and strand. The screen validates itself before reporting: a guide's zero-mismatch site must exist and must fall on the chromosome its published target lies on, and no off-target list is printed for a guide that fails. That check earned its place -- run against chromosome 2 alone, Casgevy passed and NTLA-2002 correctly emitted nothing, its target being on chromosome 4. The whole-genome run was started only after that behaviour was confirmed. The spacer was not assumed to be the first twenty bases. A guide registered in GSRS is the whole molecule, so the canonical sgRNA scaffold was located in each sequence and the spacer taken as everything before it; in both records the scaffold begins at position 20, which is what makes the 20-base reading correct rather than conventional. Two records were refused rather than parsed on a guess and are recorded on the page: NTLA-2001 is registered as the full 4,423-base molecule carrying the Cas9 mRNA, and renizgamglogene has no canonical scaffold at any position. The page is a map, not a verdict on any medicine, and says so: a three-mismatch site is a place the chemistry could direct a cut, and whether it does needs a laboratory. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

rg78803 committed Sep 6, 2026