The leukocyte artefact runs the opposite way to the one I pre-registered: it makes a haematopoietic antigen look dangerous by loading spleen and lymph node, and the fix is to read which tissue carries the signal rather than how big the ratio is
also called CD19 scores 12.0 on spleen B cells · a normal-tissue ratio mixes parenchyma with tissue cellularity · my S3(b) had the polarity backwards
Whole-tissue gene measurements include the blood cells sitting inside each organ. For a blood-cell target that makes organs full of lymphocytes look like dangerous off-targets, when really they are just more of the cells you meant to treat.
The same thing, for a scientist
🔵 **CONFIRMED AND BOUNDED 2026-09-24 (cycle 137).** The re-run this entry asked for has been done. **The correction works** - the CD19 control collapses from **12.0 to 0.63**, a 95% fall, exactly as predicted here. **And its reach is narrow**: 9 of 29 genes change, all of them the lymphoid-topped ones, and **not one of the published worst cases moves**. The most consequential shifts are **PSIP1/LEDGF 2.07 -> 1.62**, which sits inside the menin-LEDGF-KMT2A interface, and **ACTB 1.21 -> 0.94** - the reference point itself was inflated by lymph node, so the band for 'unremarkable' sits lower than the published table implies. --- This corpus already records `concept-a-normal-tissue-expression-value-includes-the-leukocytes-inside-the-tissue-and-it-killed-cd19`. Running the normal-tissue filter on surface antigens put a number on it and **corrected the direction I had assumed.** 🔴 **I pre-registered the artefact backwards.** My condition S3(b) read: *"no verdict if CD19 also scores CLEAN, because then the metric is reading leukocyte content."* That expects contaminating leukocytes to make a haematopoietic antigen look **safe**. **CD19 scored 12.0** - one of the worst ratios in the run - because its top tissue is **spleen at 70.8 nTPM**, and spleen and lymph node are full of B cells. **The leukocytes inflate the ratio, they do not deflate it.** My escape condition could never have fired, because it was pointed at an outcome the mechanism does not produce. **The correct statement:** in a whole-tissue RNA panel, the ratio *highest normal tissue / bone marrow* conflates two different things - **genuine parenchymal expression**, which is a real off-target liability, and **tissue cellular composition**, which for a haematopoietic antigen just means the organ contains more of the lineage you are already treating. 🔵 **So the discriminator is not the magnitude, it is the identity of the tissue.** Two rows from the same run make the point: | antigen | ratio | top tissue | reading | |---|---:|---|---| | CD19 | 12.0 | **spleen** | lymphoid organ - the signal IS B cells; the ratio is an artefact | | PROM1 / CD133 | 5.27 | **kidney** | not a lymphoid organ - the signal is parenchymal; the ratio means what it says | **CD19's larger number is the less alarming one.** A ranking by ratio alone would have put CD19 above PROM1 and been exactly wrong about which target carries parenchymal risk.
The rule, stated so the next filter implements it
A normal-tissue ratio is not interpretable without its tissue label. Any table of these ratios must carry the top tissue in the same row, and any automated ranking must either:
- Exclude the lymphoid and haematopoietic organs (spleen, lymph node, thymus, bone marrow, and arguably tonsil and appendix) from the numerator when the target is a haematopoietic antigen - reporting both the all-tissue and the parenchyma-only maximum; or
- Deconvolve, using an immune-cell signature to estimate what fraction of a tissue's signal is infiltrating leukocytes.
Option 1 is a two-line change and should be the default. Option 2 is better and needs a reference this project does not currently hold.
And the denominator fails separately, which is a different fault in the same statistic: with bone marrow at 0.0 the ratio is undefined, as it was for both CSPG4 and the ALB control. A near-zero denominator does not make a safety number, and the raw pair of values must be reported instead of a ratio.
| preregistration error | S3(b) expected contaminating leukocytes to make a haematopoietic antigen score CLEAN; they make it score BADLY |
|---|---|
| cd19 ratio | 12 |
| cd19 top tissue | spleen |
| cd19 top ntpm | 70.8 |
| prom1 ratio | 5.27 |
| prom1 top tissue | kidney |
| prom1 top ntpm | 19.5 |
| rule | report the top tissue beside every ratio; exclude lymphoid organs from the numerator for haematopoietic antigens, or deconvolve |
| separate fault | a bone-marrow denominator at or near 0.0 makes the ratio undefined (CSPG4, ALB); report the raw pair instead |
measurements
[
{
"quantity": "highest normal tissue / bone marrow ratio and the tissue carrying it",
"value": {
"CD19": 12,
"PROM1": 5.27
},
"unit": "ratio of nTPM",
"subject": "CD19 (top tissue spleen, a lymphoid organ) versus PROM1 (top tissue kidney, parenchyma)",
"universe": "one 16-tissue HPA consensus panel; the ordering by ratio inverts the ordering by parenchymal risk"
}
]rerun cycle 137
{
"n_genes": 29,
"n_changed": 9,
"n_dropping_20pct_or_more": 5,
"CD19": [
12,
0.63
],
"PSIP1": [
2.07,
1.62
],
"ACTB": [
1.21,
0.94
],
"worst_cases_unchanged": [
"LAMP5",
"MEIS1",
"MEF2C",
"MLLT1",
"PROM1",
"MEN1"
],
"regression_vs_published": "13 of 13 within tolerance"
}- ANSWERED in cycle 137. The re-run excluded spleen, lymph node and thymus and reproduced 13 of 13 published ratios first. Nine of 29 genes have a lymphoid top tissue and exactly those nine change; five lose 20% or more - CD19 12.00->0.63 (-95%), KDM6A -29%, KDM4C -23%, ACTB 1.21->0.94 (-22%) and PSIP1/LEDGF 2.07->1.62 (-22%). Crucially NONE of the worst cases moves: LAMP5, MEIS1, MEF2C, MLLT1, PROM1 and MEN1 all have parenchymal top tissues and are identical under both rules. See ana-excluding-lymphoid-organs-moves-nine-of-twenty-nine-safety-ratios-and-none-of-the-worst-cases.
- Is there a usable immune-deconvolution reference for HPA consensus tissues? Option 2 is the better fix and the corpus does not currently hold the reference it needs.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.
On one 16-tissue HPA consensus panel, CD19 gives a highest-normal-tissue-over-bone-marrow ratio of 12.0 with its maximum in spleen at 70.8 nTPM, while PROM1 gives 5.27 with its maximum in kidney at 19.5 nTPM; the larger ratio belongs to the antigen whose off-tissue signal is lymphocytes rather than parenchyma.
What this evidence cannot show
Run 2026-09-24 (cycle 136). CD19's Ensembl id was taken from HPA's own exact-symbol row and is flagged HPA-sourced rather than independently verified; the URL above is constructed from that id and should be re-resolved before being quoted elsewhere. The claim that spleen's CD19 signal is B cells rather than parenchyma is inference from known tissue composition, not a deconvolution performed here.
Evidence that argues against this (1)
- The inflation and the deflation stories are not mutually exclusive. Leukocytes inside a solid organ inflate the numerator for a haematopoietic antigen, but leukocytes inside the BONE MARROW denominator also inflate that, which pushes the ratio DOWN for the same antigen. Both act at once and this run did not separate them, so the direction stated here is the dominant effect for CD19 specifically and not a general theorem.
the CD19 control taught the metric its polarity and made the PROM1 row interpretable
ranking by ratio alone mixes parenchymal liability with tissue cellular composition
a near-zero denominator is the ratio metric failing at its other end
runs the exclusion that concept proposed; the CD19 control falls 95%, confirming it