Chronoatin

An in-silico experimental environment for examining single-cell chromatin over time

It exists to answer two questions that are hard to answer at the bench, and one that cannot be answered at the bench at all:

1. H1 — does chromatin conformation vary systematically by tissue in a way that predicts the order in which cancer drivers are acquired?
2. H2 — in the absence of cancer, how does chromatin change with age? Specifically: do cells drift apart faster than they drift?
3. H3 / E5 — how much of a true epigenetic aging effect survives realistic single-cell measurement, and how many donors does detecting it actually need?

The third is the one that justifies the project. See DESIGN.md for the full scientific design, the sim-to-real limits, and the benchmark ladder.

Install and run

pip install -r requirements.txt
python experiments/run_all.py

Individual experiments:

python experiments/h1_driver_order.py      # H1 + power sweep over tissue distinctness
python experiments/h2_aging_drift.py       # H2 location-scale + AAC + pseudo-replication
python experiments/h3_power_attenuation.py # donor-count and depth requirements

Everything runs on synthetic data with no downloads. Real-data adapters are described in DESIGN.md §8 and match schemas already in use in Methylation Aging.

What it does

genome/      1 kb windows keyed like selected_percent_methylation.csv, CTCF sites
engine/      epigenome state -> conformation, advanced by cell division
assay/       virtual scHi-C and RRBS -- sparsity, doublets, dropout, batch
cohort/      donors with random effects and a dial-able age-batch confound
inference/   compartment/insulation callers, location-scale age models, driver-order
benchmarks/  the validation ladder, Tier 0-5
experiments/ H1, H2, E5 as reproducible scripts

Nothing analyses ground truth except the attenuation benchmark. Every measurement passes through assay/ first, because the gap between the two is the result.

Aging enters the model in exactly three places — heterochromatin maintenance decay, lamina tethering decay, and a constant per-division epimutation rate — and nowhere else. That constraint is what makes it falsifiable rather than merely flexible.

Findings from the synthetic core

These are statements about estimators, not about chromatin. That distinction is the whole methodological point (DESIGN.md §6).

• A constant per-division epimutation rate is sufficient to produce rising cell-to-cell heterogeneity with age (SD 0.039 → 0.106 over 1600 divisions) with almost no shift in the mean. No age-specific programme is required to generate the variance signal.
• Compartment amplitude erodes while the compartment pattern is preserved (strength 0.58 → 0.31; r with tissue prior stays ≈ 0.92). Aging blurs A/B rather than scrambling it.
• H1 is seed-unstable at 25 driver–tissue pairs. τ_full = 0.485 ± 0.128 vs τ_tissue-swap = 0.305 ± 0.097; beats label permutation in 7/8 seeds but beats the tissue-swap null in only 5/8. The test only has reliable power when tissues are genuinely distinct (r < ~0.3), and real tissues share far more than that. H1 as stated needs a larger panel or locus-level resolution.
• Methylation-based aging signals survive realistic assay noise; conformation-based ones do not. Assay Attenuation Coefficients: meth_mean 0.07, meth_sd 0.53, versus comp_strength 1.23 — above 1, meaning the measurement reverses the sign of the true effect at scHi-C depth.
• Pseudo-replication inflates p-values by ~10⁵. Methylation entropy vs age: p = 0.23 at the donor level, p = 2.4 × 10⁻⁶ treating cells as independent.
• Donor-count requirement: ~16. Detecting the age effect on meth_sd at 10 cells/donor: 0/4 runs at 6 donors, 2/4 at 10, 4/4 at 16, 4/4 at 24. The slope estimate stays flat (~0.00065) across all four — the estimator is unbiased, only the standard error moves, which is the internal check that this is a power curve and not a bias curve.
• P(s) slope is the most depth-sensitive statistic there is. The Tier 4 discriminator separates two cohorts differing only in sequencing depth at AUC 0.855, and ps_slope carries a coefficient of +1.51 against −0.59 for the next feature. Practical consequence: never compare simulated to real chromatin on P(s) slope without depth-matching first — it will dominate the comparison and report a sim-to-real gap that is really a depth difference.
• Compartment strength does not converge with depth (slope −0.00017 at 500 contacts/cell, then +0.000068 → +0.000030 from 1.5k to 20k). It sign-flips at low depth and decays rather than stabilising — the same failure the AAC of 1.23 reports, seen from a second direction.

Guards

run_all.py refuses to proceed past Tier 1. A simulator that does not produce a P(s) slope in [−1.5, −0.8], a PC1 correlated with its own compartment prior, and insulation minima at CTCF sites is not a chromatin model, whatever else it reproduces.

Tier 0 checks determinism under a fixed seed — added after a real bug in which replication-timing features were derived from Python's hash(), which is salted per process, making the simulator silently non-reproducible across runs.