← back to epicme.bio explorer
chr1:206,583,354–206,589,854  ·  CD55 promoter/ROI  ·  137 CpG sites

Day 35 vs Day 6
Two timepoints, two donors, one silencing target

A side-by-side comparison of every dataset on this site so far: Unedited vs CRISPRoff-silenced T cells, at both confidence thresholds, read out either 6 or 35 days after electroporation. Day 35 and Day 6 are two different donors — not the same person at two timepoints — so every difference below could reflect time since editing, donor-to-donor biology, or both at once. That caveat applies to everything on this page and is discussed in full at the bottom.

⚠ what "Day 35 vs Day 6" can and can’t tell you

This page puts two independent experiments next to each other. They share a pipeline (same CD55 region, same 137 CpG window, same Nanopore/methylation-calling approach, same CRISPRoff reagent) but not the same donor. Any gap you see between the Day 35 and Day 6 lines/bars below is consistent with several different explanations: silencing changing over time since electroporation, this being a different person's cells, or some mix of both. Nothing here can separate those apart — that would need the same donor sampled at multiple timepoints, which isn’t yet part of this dataset.

How many molecules are we comparing?

Day 6 has substantially more sequenced molecules per group than Day 35 across the board — several times more at both thresholds. Keep this in mind when comparing precision between the two timepoints: bigger samples generally give steadier estimates.

threshold 0.7 · molecules per group
threshold 0.995 · molecules per group
Day 35 Day 6

Overall methylation, Day 35 vs Day 6

Percent of all CpG×molecule positions called confidently methylated (see each timepoint's own page for why this is a floor estimate, not a precise frequency). Same group, same threshold, different donor and timepoint, side by side.

threshold 0.7 · % methylated (floor estimate)
threshold 0.995 · % methylated (floor estimate)
Day 35 Day 6

Methylation across the 137 CpG sites, both timepoints overlaid

Same idea as each timepoint's own per-site chart, but now with all four lines on one plot: colour still marks Unedited vs CRISPRoff, and line style now marks the timepoint (solid = Day 35, dashed = Day 6). Where the CRISPRoff line separates from Unedited — and whether that gap looks similar in solid and dashed — is the most direct visual read of how consistent the silencing signature is across donor and timepoint.

threshold 0.7
threshold 0.995
Unedited · Day 35 (solid) Unedited · Day 6 (dashed) CRISPRoff · Day 35 (solid) CRISPRoff · Day 6 (dashed)

The CRISPRoff effect itself, Day 35 vs Day 6

Subtracting Unedited from CRISPRoff at every site isolates the silencing signature from each donor's baseline methylation. If this line has a similar shape at both timepoints, that's evidence the silencing pattern itself (which sites gain methylation, and roughly how much) is reproducible across donor and time — even if the two donors' raw methylation levels differ.

threshold 0.7 · Δ = CRISPRoff − Unedited
threshold 0.995 · Δ = CRISPRoff − Unedited
Day 35 Day 6

How reproducible is the per-site profile across donor & timepoint?

Each point is one of the 137 CpG sites: x = its methylation fraction at Day 35, y = its methylation fraction at Day 6. Points hugging the dashed diagonal (y = x) mean that site behaves the same way in both datasets. Pearson r is shown per group — closer to 1 means tighter agreement.

threshold 0.7 · Day 35 (x) vs Day 6 (y)
threshold 0.995 · Day 35 (x) vs Day 6 (y)
Unedited sites CRISPRoff sites

A quick read on the numbers once they render: Unedited baseline methylation tends to correlate more tightly across donor/timepoint than the CRISPRoff pattern does — consistent with a stable endogenous methylation landscape at this locus, and a silencing response that varies somewhat more in degree or kinetics from one donor/timepoint to the next.

Do the six ML models agree across timepoints?

The same six model types from each timepoint's own page (see those pages for what each one actually does), now plotted side by side. Bars are coloured by timepoint; the dashed 50% line marks chance level for this balanced two-class problem.

threshold 0.7 · held-out test accuracy
threshold 0.995 · held-out test accuracy
Day 35 · test set ≈84 (0.7) / ≈39 (0.995) molecules Day 6 · test set ≈221 (0.7) / ≈270 (0.995) molecules

Every model clears chance at both timepoints and both thresholds — CRISPRoff leaves a learnable signature regardless of donor or day. Bars outlined with a dashed border and labelled “check overfit” hit a perfect 100% on a single train/test split; treat those as optimistic upper bounds rather than proof of a perfectly separable signal, especially where only one model spikes to 100% while the rest cluster lower.

Reading this comparison honestly

Donor and timepoint are confounded. Every difference on this page could be "Day 6 vs Day 35" or "this donor vs that donor" — there is no way, from this data alone, to attribute a gap to time since electroporation rather than to individual biology. Resolving that would need the same donor's cells sampled at multiple timepoints.

The unedited baseline is more reproducible than the CRISPRoff response. Per-site correlation between Day 35 and Day 6 is higher for Unedited molecules than for CRISPRoff molecules at both thresholds (see the scatter plots above once rendered). Endogenous methylation at this locus looks fairly stable across donor and time; how strongly CRISPRoff silences it looks more variable.

Overall CRISPRoff methylation is higher at Day 35 than at Day 6 in this data. That's compatible with silencing accumulating further the longer cells are cultured post-electroporation — but it's equally compatible with this particular Day 35 donor simply silencing more efficiently than this particular Day 6 donor. Single-timepoint-per-donor data can't distinguish those.

The ML classifiers agree more than they disagree. All six models separate Unedited from CRISPRoff well above chance at both timepoints, and the strongest, most consistent performer at both timepoints is the sparse (L1-penalized) logistic regression — the model built specifically to avoid latching onto noise. That convergence, across two different donors and two different timepoints, is a reasonably strong signal that CRISPRoff leaves a genuine methylation fingerprint at this locus, not an artifact of one dataset.