In 2026, two teams mapped the genes that mini-tumours grown in dishes
(“organoids”) cannot live without — a goldmine of candidate drug targets. But organoids are fed a
defined, serum-free diet that blood never resembles. We re-ran the numbers every way the public
data allow, and found that about 85% of the 456 differential dependencies in our reanalysis are consistent
with the food, not the 3D shape.
Every point of light behind this text is one real gene’s measured effect.
Scroll — and watch what the medium explains. Hover a point to meet a gene.
Live data, not a picture: 17,525 points = 17,525 genes from DepMap 26Q3 Chronos screens. Colour = audit verdict.
Chapter 01 — the confound
Shape changed. So did supper.
An organoid differs from a classic flat culture in two ways at once: its
3D shape — and its diet. Flat cultures bathe in serum, rich in lipids and cholesterol. Organoids
get a precise, serum-free recipe. So when a gene suddenly looks essential in organoids, is that
the shape talking… or starvation?
Luckily, the public data contain an observational control: other
models grown flat, but fed organoid-type medium. Food changed, shape didn’t. That single
comparison — observational, with caveats detailed below — splits the story in two.
Left: grown flat. Right: grown as a 3D dome. The question is which difference matters.
Chapter 02 — the audit
85%
follow the food.
Now every gene sits at its real coordinates: organoid effect (→) versus
flat-grown, same-medium effect (↑). Genes near the dashed diagonal show a similar-sized effect
without any dome. 387 of 456 differences meet our medium-consistency criteria — an
observational consistency label, not causal proof — cholesterol factories like SQLE first
among them.
Δ organoid vs conventional 2D →
Δ flat-grown, organoid medium →
Chapter 03 — what survives
Almost nothing needs the dome.
Fade out everything consistent with the menu. What remains: one nominal
shape candidate (LSM7 — not significant genome-wide) and a short ring of honestly
unresolved genes — the credible follow-ups, including Sanger headline names
SCAP, DDX3X and ITGB1. Click one to interrogate it.
Every gene’s organoid effect (Δ, organoid minus conventional 2D; negative = more
essential in organoids), what happens in the same medium grown flat, and our verdict — with uncertainty.
Click any gene for the full picture, a copyable citation, and its source record.
17,525
genes audited, each with a verdict and uncertainty
85%
of 456 differential dependencies are consistent with medium effects
bootstrap slope: flat cells in the same medium reproduce the organoid effect [95% CI 0.82–0.99]
Popular checks:
· shareable links like #gene=SQLE · press / to search from anywhere
Gene
Verdict
Δ organoid
Δ flat, same medium
Δ shape only
Reproduced flat
Bootstrap stability
The whole result in one picture
Each dot is one of the 456 differences. Horizontal: how much more/less essential a
gene is in organoids. Vertical: the same measurement in flat-grown cells fed organoid medium. Dots on
the dashed line (A = D) show a similar-sized effect without any dome — consistent with a medium
effect under our criteria. Hover or click a dot.
Red: medium-consistent · amber: unresolved · blue diamond: the one format candidate (LSM7).
Dashed line = A = D (equal-sized effect without domes).
05
Findings
1 · Most of it is medium-consistent — gains and losses alike
Of 456 genes whose essentiality differs between organoids and conventional cultures,
86% of gains and 84% of losses meet our medium-consistency criteria (same-direction effect, at least half as large, in flat-grown cells fed organoid medium)
(75–82% under conservative robust statistics). The strongest class is cholesterol/sterol manufacture
(SQLE, MSMO1), followed by fatty-acid and heme synthesis — exactly the pathways a serum-free,
lipid-poor diet would stress. Gene-expression data show the mirror image (cholesterol genes switched
on in organoid medium, the classic starvation response), and in independent screens using
blood-like medium, the flagship gene SQLE moves the opposite way.
Figure 1. Left: flat-grown, same-medium effect vs organoid effect — points hug the diagonal.
Right: the shape-only contrast (dome vs flat within organoid medium) — a shapeless cloud near zero.
2 · Auditing the published target lists
Against the Broad catalogue’s own published hits: 47% are medium-consistent overall, and 87% of those
that replicate under lineage adjustment. Against Sanger’s “organoid-specific core fitness”
headline-18: 8 of 16 replicate here, 5 are medium-consistent — four of them the SREBP/sterol axis
(MVD, IDI1, MBTPS1, MBTPS2) — and 3 remain unresolved (SCAP, DDX3X, ITGB1), making them the more
credible follow-up candidates. Sanger’s own calls are prevalent in their assay (median depletion
in 70–80% of organoids); the issue is attribution, not call quality.
Figure 3. Composition of the published hit lists under our flags. Grey = does not reproduce
as a hit under lineage adjustment (attribution unestablished — not evidence of real biology).
3 · The weakening that isn’t (mostly) biology
Ribosome and spliceosome genes look less essential in organoids. With curated gene sets, only
the ribosomal set shows a clear whole-complex shift (+0.17 vs +0.03 baseline), present in both labs’
assays after global corrections; Sanger’s organoid screens are globally shallower (essential
depth 0.63×). Slower growth → fewer cell divisions → long-lived protein machines don’t have time
to wear out is the leading explanation — but it’s a hypothesis: cumulative doublings are
recorded for only 1 of 1,241 Broad screens, so nobody can settle it from this data release.
Figure 2. Left: curated-set median shifts by contrast (Broad). Right: Sanger essential
depth and depth-scaled ribosomal/spliceosomal ratios — spliceosome weakening vanishes after scaling,
ribosomal persists.
What would settle it
Add LDL cholesterol back to organoid medium. If the sterol dependencies are starvation,
SQLE-inhibitor and statin sensitivity should collapse (≥5× potency shift) while genuine candidates
like ITGB1 don’t budge. Pre-agreed falsification: if sensitivity survives LDL repletion in
≥3 of 4 models, our interpretation fails. The full protocol is in the
wet-lab proposal (Downloads) — a ~2-week experiment,
no new screen needed.
06
How this was made
An honest changelog — including the mistakes. This audit was produced by an
AI-assisted research programme under human direction, then attacked by an independent AI reviewer
five times before release. This website got the same treatment.
2026 · The catalogues
Two Nature papers map organoid weak spots
Broad (Neiswender et al.) and Sanger (Herranz-Ors et al.) publish genome-wide CRISPR dependency
maps for patient-derived organoids, nominating organoid-specific drug targets. Broad’s own
1,000-gene side experiment already hints some hits are medium-driven — but nobody counts,
genome-wide, across both catalogues.
The design
Four comparisons that separate food from shape
Public data already contain the controls: organoid medium grown flat (food without shape),
domes vs flat within the same food family (shape without food change), and an unrelated
serum-free brain-cancer diet. Add corrections for the two different CRISPR toolkits used
(measured from 22 models screened with both), cancer-type mix, and screen strength — and hold
out entire patients to check the answer on unseen data.
Round 1
First pass: 85% food, 0% shape
The headline landed immediately, including a dramatic “spliceosome collapse” (+0.9).
Confidence was high. Too high.
Self-correction
The spliceosome that wasn’t
Swapping a small hand-picked gene panel for the full curated sets shrank the spliceosome effect
from +0.90 to +0.02. The panel had been cherry-picked by history. The ribosome effect survived;
the spliceosome story was demoted in print, in the abstract, everywhere.
Triangulation
Three independent witnesses
Gene-expression data show the cholesterol-starvation mirror image (+0.36 vs −0.38).
Sanger’s independent assay shows the same cholesterol gain (percentile 0.099 vs 0.394).
Blood-like medium screens from 2021 move SQLE the opposite way. Different labs, different
methods, same conclusion.
Adversarial review
Five rounds against a skeptical reviewer
An independent AI reviewer (Codex, highest reasoning) audited code and claims, recomputing from
source. It found real faults every round — contaminated calibration, an impossible statistical
test for shape effects, a bootstrap that estimated a different quantity than claimed, a NaN bug
corrupting one gene’s stability score — each verified, fixed, and re-reviewed, until it
scored the work 95/100. This website was then reviewed the same way.
Novelty audit
“Hasn’t somebody done this?” — 230+ web searches later
A separate uncapped deep-research run hunted for prior art. Verdict: partially true — the
general ideas have real precedents (2019–2026, now cited below), one web report claiming similar
scope turned out to carry a fabricated arXiv ID, and a commentary nine days earlier asked the same
question without numbers. The catalogue-wide per-gene audit itself appears unreported. The title
was downgraded accordingly: not “attribution” — an audit.
Now
Over to the wet lab
Everything computational that can be squeezed from public data has been — twice-checked. What
remains is a two-week experiment: add the cholesterol back. The proposal, with falsification
criteria agreed in advance, is in the downloads.
07
Methods
In plain language
For every gene we ask four questions of the public data: (1) Does its importance change in
organoids vs conventional cultures? (2) Does the same change happen in flat-grown cells fed organoid
medium? (3) Is there anything left over when comparing blobs to flat on the same food? (4) Does an
unrelated serum-free diet do something similar? If (2) reproduces most of (1) and (3) is small, we
call the gene medium-consistent. Mirror-image logic gives format candidate; everything
else is honestly unresolved. These are consistency labels from observational data — not causal
proof. Proof needs the add-back experiment.
Technical details
DepMap 26Q3 Chronos gene effects (PassesQC screens, Avana + Humagne-CD), development patients only
for all calibration, filtering, and imputation. Per-gene library offset from 22 development FBS
models screened with both libraries (all Avana screens are FBS). Per-gene OLS on model-mean Chronos
with lineage dummies; Benjamini–Hochberg q per contrast; HC3 robust SEs shipped and used for
sensitivity. Hits: D q<0.01 and |Δ|≥0.25 (dynamic-range-normalized scale primary; raw
concordant). Medium-consistent: same-sign adherent ratio ≥0.5, nominal p<0.05, 90% delta-method
CI excluding 0, format not dominant under mirror criteria; strict-FDR variant released as
flag_fdr. Joint bootstrap (B=150) shares one model resample between calibration and
analysis; identity replicate reproduces point flags exactly. Held-out evaluation is fully disjoint
(25 reserved-patient domes vs 78 reserved-patient FBS controls). Common-support sensitivity
restricts to the six lineages with adherent comparators (38 domes): slope 0.89. Humagne-only
sensitivity filters at screen level (7 FBS controls, r=0.78). Sanger: supplementary tables 4/5/6 and
cell-line EssMatrix; within-model percentiles are primary across differently processed matrices.
Expression: Omics TPM logp1 mapped by ModelConditionID. HPLM: Rossiter et al. Table S1 scaled gene
scores. Full code: see the release bundle in Downloads.
Limitations, stated plainly
Adherent comparators: 9 models, 6 of 9 lineages, mixed (not identical) formulations — per-gene
verdicts carry real uncertainty (use the bootstrap frequencies). Model derivation/selection
differences are not excluded by lineage adjustment. Library handling rests on 22 dual-screened
models. Cumulative doublings are recorded for 1 of 1,241 Broad screens, so the weakening mechanism
is unresolved. The 2026 catalogues themselves were not re-screened; all inference is from their
public data. What this site claims is an audit with quantified consistency — nothing more.
08
Sources & citing
Every number on this page traces to a public dataset below. Gene pages link
straight to the source record in the DepMap Portal.
Dataset
What we used it for
Citation / DOI
DepMap 26Q3
Screen gene effects (Chronos), screen maps, model conditions, expression
Medium-associated dependencies & corrections predating this audit
Lagziel 2019 doi ·
Pacini 2021 doi ·
Dempster 2019 doi ·
Liang 2025 doi ·
Zaidi & Sawyers 2026 doi
Cite this audit
It’s the Medium: a genome-wide medium-consistency audit of the 2026 organoid
dependency catalogues. Unpublished draft technical note, 2026 (not yet posted
as a preprint; the draft PDF contains author placeholders).
https://www.itsthemedium.com
Data: medium_attribution_v2 tables (v2r2), release bundle v1.
BibTeX
@misc{itsthemedium2026,
title = {Most reproduced "organoid-specific" cancer dependencies are
consistent with culture-medium effects: a genome-wide
medium-consistency audit of the 2026 organoid dependency catalogues},
year = {2026},
note = {Unpublished draft technical note},
url = {https://www.itsthemedium.com}
}
Per-gene citations: open any gene above and use its “Copy gene citation”
button — it includes the exact table version and values so others can reproduce your number.
09
Downloads
Manuscript (draft PDF, 7 pp) — the full technical note with figures; unpublished draft with author placeholders, not yet posted as a preprint312 KB
Release bundle v1 (ZIP) — full analysis code (Python), adversarial review trail, supplementary tables and manuscript source11.5 MB
