An ENCODE-only archive bound systematically undercounts chromatin-modifying enzymes, because ENCODE's target catalogue is transcription-factor-shaped - measured 3-for-3 on the H3K9 machinery, with ratios of infinity, 2.8x and 7.3x
also called ENCODE TF bias hypothesis · Q113 archive-bound hypothesis
The big public atlas this project checks first was built mainly to map transcription factors, so enzymes that modify chromatin are thinly represented in it. Every time we checked one of the H3K9 enzymes, other archives had more data than the atlas did - which means an absence declared from the atlas alone is likely to be wrong in one direction.
The same thing, for a scientist
ENCODE's TF ChIP-seq programme was designed around sequence-specific DNA-binding factors and antibody characterisation pipelines built for them. Chromatin-modifying enzymes and heterochromatin readers enter that catalogue only incidentally. If that is right, an archive-bounded absence that consults ENCODE alone is biased toward declaring an absence for exactly the class of protein this graph's H3K9 and resection reasoning is about - and the bias runs one way, so it manufactures absences and never manufactures presences. Measured this cycle on three members of the H3K9 machinery, ENCODE versus ChIP-Atlas released experiment counts: **SUV39H1 0 vs 4 (and 6 occupancy samples in GEO primary)**, **SETDB1 5 vs 14 (2.8x)**, **CBX5 4 vs 29 (7.3x)**. Three of three in the same direction, and the one where the ratio is infinite is the protein the graph's whole H3K9 dispute turns on. This is a hypothesis and not a finding because n=3 on a hand-picked set is not a test, the null has not been written down, and ChIP-Atlas re-processes a superset of ENCODE by construction so some excess is guaranteed. The point of stating it as a hypothesis is that it makes a scoreable prediction over a set nobody has drawn yet.
The observation that prompted it
absence-no-h3k9-writer-or-reader-localised-at-the-kmt2a-locus bounded its archive half with an ENCODE census and recorded SUV39H1: n=0 alongside the remark that every member of the H3K9 machinery this graph names is represented in ENCODE EXCEPT the SUV39H paralogues, and SUV39H1 is the one the dispute is about. That entry read the asymmetry as being about SUV39H1. This hypothesis reads it as being about ENCODE: the paralogue that is missing is missing because ENCODE is not where histone methyltransferase occupancy lives, and the GEO sweep bears that out with six samples in two cell types 12.
Why the direction of the bias is the whole point
A biased instrument that errs in both directions produces noise. This one cannot produce a false presence - an ENCODE experiment that exists, exists. It can only produce a false absence. So every archive-bounded absence in this corpus that consulted ENCODE and stopped is exposed, and none of the corresponding presences is. That asymmetry is what makes it worth a test rather than a caveat.
There is a real competing explanation and it must be stated: ChIP-Atlas re-processes SRA, which includes ENCODE's own submissions plus every individual lab deposit, so ChIP-Atlas >= ENCODE is true almost by construction for any protein. The hypothesis is therefore not ChIP-Atlas has more - that is trivial - but that the excess is larger for chromatin-modifying enzymes than for sequence-specific transcription factors, which is a comparison with a null and is what the falsifier below scores.
| statement | An archive-bounded absence that consults ENCODE alone systematically undercounts chromatin-modifying enzymes and heterochromatin readers relative to sequence-specific transcription factors, because ENCODE's target catalogue and antibody-characterisation pipeline were built around the latter. The bias is one-directional: it can manufacture a false absence and cannot manufacture a false presence. |
|---|---|
| origin | data-pattern |
| prior probability | 0.6 |
| null not yet written | TRUE, and this is the entry's main weakness. Rule 84 requires the base rate before the proportion means anything: what fraction of ALL human proteins with any ChIP-Atlas data have zero ENCODE experiments? Until that number exists, 'three of three' is uninformative, and the falsifiers above are written to supply it rather than to assume it. |
| what it would change if true | every archive-bounded absence in this corpus about a chromatin enzyme or reader would need its ENCODE-only bound re-run against GEO and ChIP-Atlas before it could be quoted. That is at least three committed entries and the class the graph's H3K9 and resection reasoning rests on. |
predictions
[ "Over a set of 30 chromatin-modifying enzymes and 30 sequence-specific transcription factors matched on ChIP-Atlas experiment count, the fraction with ZERO ENCODE released experiments will be higher in the enzyme set.", "Over the same matched sets, the median ratio (ChIP-Atlas experiments)/(ENCODE experiments) will be higher in the enzyme set.", "Proteins whose only human occupancy data are CUT&RUN will be over-represented in the enzyme set, and excluding CUT&RUN-only proteins will shrink but not abolish the difference." ]
experiments
[]
analyses
[ "ana-q113-geo-holds-human-suv39h1-occupancy-data-that-the-encode-only-census-could-not-see" ]
falsifiers
[ "SCOREABLE FALSIFIER 1. Draw 30 chromatin-modifying enzymes (SET/SUV/KMT/KDM/HAT/HDAC families) and 30 sequence-specific transcription factors, matched pairwise on ChIP-Atlas hg38 experiment count at threshold 5 (within +/-25%). For each, count ENCODE released experiments under the exact target label plus any tagged label (eGFP-, 3xFLAG-). THE HYPOTHESIS DIES if the fraction with zero ENCODE experiments is not higher in the enzyme set by at least 15 percentage points, or if a two-sided Fisher exact test on the zero/non-zero table gives p > 0.05. State the drawn sets and the seed.", "SCOREABLE FALSIFIER 2. On the same 60 proteins, compute the ratio of ChIP-Atlas to ENCODE experiment counts, treating a zero ENCODE count as the maximum ratio in the set. THE HYPOTHESIS DIES if the median ratio in the enzyme set is not at least 1.5x the median ratio in the transcription-factor set. Report both medians and a Wilcoxon rank-sum p; a p > 0.05 also kills it.", "SCOREABLE FALSIFIER 3, which kills the EXPLANATION rather than the observation. Repeat falsifier 2 after excluding every protein whose only human occupancy data are CUT&RUN or CUT&Tag. IF the difference disappears entirely, the observation is a temporal artefact of a new assay and not a property of ENCODE's target catalogue, and the hypothesis as stated is wrong even though its prediction held.", "IMMEDIATE ONE-PROTEIN CHECK, cashable in minutes and not yet run: pick the five sequence-specific transcription factors with the FEWEST ChIP-Atlas experiments and ask whether any of them has zero ENCODE experiments. IF two or more do, ENCODE's zeros are a coverage-depth phenomenon and not an enzyme-class one, and the hypothesis is badly weakened before the full test is worth running." ]
measured rows 2026 09 22 — the numbers behind the chart
{
"SUV39H1": {
"encode": 0,
"chip_atlas": 4,
"geo_occupancy_samples": 6
},
"SETDB1": {
"encode": 5,
"chip_atlas": 14
},
"CBX5": {
"encode": 4,
"chip_atlas": 29
},
"SUV39H2": {
"encode": 0,
"chip_atlas": 0,
"note": "a genuine zero everywhere - the hypothesis does not predict that every enzyme has hidden data"
}
}- If this is true, the cheap fix - always check GEO too - makes every archive bound slower and none of them conclusive, because a GEO title sweep is itself bounded by what a submitter typed. Replacing a confident false absence with an unbounded one is not obviously progress unless the entry says which it is.
- What IS the base rate? Of all human proteins with at least one ChIP-Atlas experiment, what fraction has zero ENCODE experiments? Nobody in this project has computed it and every claim of this shape depends on it.
- Does the same bias exist in ReMap, which explicitly scopes itself to transcription-factor-like DNA-binding experiments? If ReMap is worse than ENCODE for enzymes, the two secondary archives this project reaches for are biased the same way and GEO primary is the only unbiased route.
- Is the ENCODE gap a target-catalogue decision or an antibody-characterisation bottleneck? The two have different remedies: the first is fixed by asking ENCODE, the second by a reagent.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database
For three members of the H3K9 machinery, ChIP-Atlas holds more released experiments than ENCODE, in the same direction each time, with SUV39H1 at 0 versus 4.
ENCODE portal type=Experiment&target.label=<SYM> versus ChIP-Atlas hg38 target tables: SUV39H1 0 / 4, SETDB1 5 / 14, CBX5 4 / 29 · ENCODE Project and ChIP-Atlas / DBCLS 2026What this evidence cannot show
n=3, hand-picked as the proteins this beat was already auditing, so this is the worst possible sample for a general claim about ENCODE. No null was computed and no transcription-factor comparison set was drawn. The ENCODE counts are released experiments under the exact target label; the ChIP-Atlas counts are SRX columns in the hg38 target table at threshold 5, which is a different unit and inflates the ratio by an unknown amount.
Evidence that argues against this (1)
- ChIP-Atlas re-processes SRA and therefore contains ENCODE's submissions as a subset, so ChIP-Atlas exceeding ENCODE is expected for every protein including transcription factors. Without the TF comparison set this measurement is consistent with no enzyme-specific bias at all, and the honest reading of the three rows on their own is that they show nothing. Oki S et al. 2018
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The one protein with an infinite ratio is the one the graph's H3K9 argument depends on, and its GEO data are in two cell types that no secondary archive fully indexes.
Chaetocin-mediated SUV39H1 inhibition targets stemness and oncogenic networks of diffuse midline gliomas and synergizes with ONC201 (GSE225028, SUV39H1 CUT&RUN, hg38 bigWig) · Xin D, Liao Y, et al. · Neuro-Oncology 2024What this evidence cannot show
ChIP-Atlas indexes the MCF-7 ChIP-seq set but not the CUT&RUN set, so even the secondary archive that did best here saw 4 of the 6 samples. The GEO title sweep is the only route that reached all six and it is bounded by what a submitter typed into a sample title.
Evidence that argues against this (1)
- CUT&RUN is a newer assay and its under-representation in secondary archives is a temporal artefact that will resolve, not a structural property of ENCODE's target catalogue. If the enzyme excess is mostly CUT&RUN, the hypothesis as stated - about ENCODE's TF shape - is the wrong explanation for the right observation. NCBI Gene Expression Omnibus 2024
the SUV39H1 overturn is the observation this hypothesis generalises, and the hypothesis is what makes that observation testable rather than anecdotal
if true, that entry's ENCODE-only archive census is biased toward declaring absence for exactly the protein class it is about
every count in this hypothesis was taken with a same-template positive control, which that entry is the reason for