"Zero called peaks" is a statement about a reproducibility threshold, not about whether the protein was there - and in this graph the difference was worth a whole cycle's conclusion

**The sibling of `concept-a-genome-wide-null-flatters-any-interval-that-belongs-to-an-enriched-class`.** That concept is about the null you compare a signal to. This one is about the filter that decides whether a signal is reported at all. **IDR (Irreproducible Discovery Rate) does not ask "is there signal here".** It takes each replicate's ranked peak list and asks whether a candidate's *rank* is consistent between them, then reports the estimated fraction of calls at that level that are irreproducible. A region with strong, real, reproducible-in-both-replicates enrichment can fail it because the two replicates disagree about where it sits in their rankings. ENCODE's `conservative` and `optimal IDR thresholded peaks` files - the ones an analyst naturally downloads - are already past that filter at IDR < 0.05. The pre-filter evidence is in the separate `IDR ranked peaks` and `peaks and background as input for IDR` files, which most analyses never open. **The worked case in this graph** [E1]. Cycle 63 read the thresholded files for 13 ENCODE BRCA1/CBX3/CBX1 experiments at seven *KMT2A* intervals, found **zero peaks in 13 of 13**, and concluded that no occupancy exists above background. Cycle 64 opened the pre-threshold files for the same experiments: ``` ENCSR000BRT CBX3 K562 ENCFF951BQB optimal IDR 26,789 peaks nothing at the element ENCFF271ZAZ IDR ranked 133,088 peaks signalValue 22.39 rep1 14.01 rep2 8.49 globalIDR 0.227 -> REJECTED ENCSR000EDB BRCA1 HeLa-S3 ENCFF668JZI conservative 9,452 peaks nothing at the element ENCFF457RFT IDR ranked 62,322 peaks signalValue 15.66 rep1 9.21 rep2 7.13 globalIDR 0.0588 -> REJECTED by 0.009 ``` **Both peaks are present in both biological replicates.** Neither is an artefact of a single library. One is rejected at IDR 0.227, which is a fair rejection; the other at 0.0588, nine thousandths from the cut-off. **Why this is a concept and not a complaint about IDR.** IDR is doing its job. A genome-wide peak list needs a reproducibility filter and 0.05 is a defensible line. **The error is entirely on the reading side**: an analyst asking a *locus-specific* question - "is this protein at this 147 bp element" - inherits a threshold that was set for *genome-wide list quality*, where the cost of a false positive among 27,000 calls is very different from the cost of a false negative at the one interval the whole project is about. **A threshold tuned for a list is being applied to a single query.** **Three rules follow, and they generalise past ChIP-seq.** 1. **For a locus-specific question, never stop at the thresholded file.** Read the pre-filter file too, and report the filter statistic - here the global IDR - rather than reporting absence. `IDR 0.0588` is a far more useful sentence than `0 peaks`. 2. **State which file a zero came from.** "Zero conservative IDR thresholded peaks in ENCFF668JZI" is checkable; "no BRCA1 peak at KMT2A" is not, and the two were treated as interchangeable for a full cycle. 3. **The general form.** Any pipeline whose last step is a multiple-testing or reproducibility filter converts "below a list-wide threshold" into "absent", and downstream readers cannot tell the two apart because the filtered object is the only one published. **The same shape covers an FDR-filtered differential-expression table, an IDR peak set, a called-variant VCF and a curated interaction database** - in each case the unfiltered layer exists and is rarely opened. **And the honest counterweight.** Turning the threshold down finds things everywhere. The same reads that yield the two rejected peaks above also yield 133,088 and 62,322 pre-filter candidates, and in a 3-billion-base genome most of those are noise. **The defence against confirmation bias is not to avoid the pre-filter file, it is to bring a class-matched null with you when you open it** - which is the other concept, and the two are only useful together.