Limitations

Evaluation

• The transcript gene filter runs before the split. Genes are kept when they have ≥ 10 counts in ≥ 20 % of the samples of at least one tissue, computed on all 899 vials. It uses no labels beyond tissue membership and no fitted statistic, but it is a data-dependent step outside the fold. Its effect on the accuracies is unmeasured; it is the one known deviation from the frozen rules.
• Pooled-vial calibration is only approximately valid. The 18 vials of one animal are not exchangeable with vials of other animals, so the pooled conformal rows (n ≈ 396) overstate the effective calibration size; the one-vial-per-animal rows (n = 22) are the conservative version and are the ones reported as the in-distribution guarantee. With 22 points coverage moves in steps of 1/23.
• Small n. 50 animals in transcriptomics (10 test animals per fold), 8 adult BodyMap animals, 3 or 5 target individuals in recalibration. Every accuracy carries a fold sd for that reason, and the certificate is reported as a distribution over splits, not a number.
• The bridge measurement uses a muscle-derived pool. The reference RNA pools run on several plates were made from gastrocnemius (and, within one site, liver and hippocampus), so batch variance is measurable only on the genes those pools express; markers of other tissues read zero on every plate. The all-gene and expressed-in-pool summaries are reported side by side.
• The QC-only baseline is untuned (multinomial logistic regression, C = 1). It is a floor on what QC numbers can do, not a ceiling; the argument it supports only needs the floor.

Transfer

• BodyMap animal ids are assumed. GEO carries no animal identifier; the animal is taken to be the replicate index within organ × age × sex, and every BodyMap split is grouped on that assumption. If replicates were re-indexed across organs, BodyMap recalibration would have leaked animals across calibration and test.
• Super-class scoring. BodyMap “Muscle” counts as correct for either skeletal muscle and “Brain” for any of the three brain regions; GTEx “Muscle – Skeletal” likewise. Artery – Aorta → vena cava is an imperfect mapping, and thymus and uterus have no MoTrPAC tissue at all (they are scored for abstention only).
• Seven panel markers have no BodyMap organ (ovary, adipose, blood, brown fat, small intestine, colon, vena cava): their only external check is the human transfer, where orthology and species confound it.
• GTEx is scored on 1:1 orthologs only (pending genes), and the human panel is re-selected in that space; the recalibration draws include infinite thresholds (pending of the three-donor draws for the 20-gene panel), so the recalibrated GTEx coverage must be read with its set size.
• A recalibrated coverage is observed, not guaranteed. pending
• Within-set scaling depends on what else is uploaded. pending

Design and data

• Sex is confounded with arrival cohort in every group (males and females arrived about a month apart), and only the 8-week animals share arrival cohort and sacrifice dates with the controls; 1-, 2- and 4-week contrasts also compare seasons. Ovary and testes are single-sex tissues, so the held-out-sex shift is also a label shift.
• One data release. Everything here is portal release c1.0 (rn6, the R package MotrpacRatTraining6moData 2.0.0). The rn7 reprocessing (c2.0) shares 21,651 of 32,883 gene ids and changes the training-regulated gene sets; the tissue fingerprint was not rerun on it.
• Consortium outliers remain in the matrices, including the female 1- and 2-week vena cava vials flagged as brown-fat contaminated; they account for most vena cava → brown fat errors.
• Proteomics as distributed is within-tissue only (ratios to per-tissue reference pools), so the core fingerprint is RNA. On the portal's reporter-ion scale the proteome does carry the tissue axis and transfers (see Multiomic), but there too each plex holds one tissue and the missing-value pattern alone identifies it.

Follow-up: the proteome and metabolome (pre-registered; see the Multiomic page): on the portal's reporter-ion scale tissue explains R² pending of the first principal component of the same proteins (pending on the distributed ratios); a 20-protein panel selected there names the tissue of pending of pending human TMT samples (pending donors; 95 % interval pending), with coverage pending under MoTrPAC calibration and pending after recalibrating on five donors at pending of 7 classes per set; and in Jiang 2020's TMT design pending tissue pairs share a run, the counterexample this audit lacked. Within-study numbers there remain context: plex is nested in tissue on that scale too.