generator/code/README.md

DLCoalSim

Coalescent simulation of demographic histories, for training and evaluating methods that infer effective population size through time, N_e(t), from sequence data.

A row is one simulated sample: bit-packed haplotypes, variant positions, the recombination and mutation maps as a method would receive them rather than as they truly were, and a 64-bin log-time target. Nothing is precomputed into summary statistics — the point is that a method computes its own.

pip install -e .
dlcoalsim --preset mini --out-dir runs/first

That writes 512 rows in a few minutes. --preset core_v01 is 20,000 rows at 10 Mb; --preset full_v01 is 100,000 rows at 20 Mb.

Reading a dataset

runs/first/
  samples/shard_00000.npz        numerical arrays, 128 rows a shard
  samples/shard_00000.jsonl.gz   one metadata object per row
  metadata/samples.csv           the same metadata, whole dataset
  config.json  manifest.json     what produced it
  scripts/loader.py              a minimal reader, written beside the data

import numpy as np

shard = np.load("runs/first/samples/shard_00000.npz")
row = 0
lo, hi = shard["variant_offsets"][row], shard["variant_offsets"][row + 1]

target    = shard["target_log10_ne"][row]              # (64,) log10 Ne per time bin
positions = shard["variant_positions_bp"][lo:hi]       # uint32
packed    = shard["genotype_packed"][lo:hi]            # (variants, 8) uint8
n_haplotypes = int(shard["n_haplotypes"][0])
haplotypes = np.unpackbits(packed, axis=1, bitorder="little")[:, :n_haplotypes]

Time bins are log-spaced from 50 to 500,000 generations; time_edges in manifest.json gives the boundaries. Genotypes are packed one bit per haplotype, n_haplotypes to a row. Missing calls are a sparse flat index (missing_flat_idx, missing_offsets) rather than a mask.

A map stores one offsets array, not two. Its *_pos array holds the breakpoints and *_rate the rates between them, so a row holds one fewer rate than positions — none of either where a row has no map:

offsets = shard["obs_rec_offsets"]
lengths = np.diff(offsets.astype(np.int64))
rate_at = np.concatenate(([0], np.cumsum(np.maximum(lengths - 1, 0))))

breakpoints = shard["obs_rec_pos"][offsets[row]:offsets[row + 1]]
rates       = shard["obs_rec_rate"][rate_at[row]:rate_at[row + 1]]

The maps come in two forms. obs_rec_* and obs_mut_* are what a method is allowed to see: a published map's resolution, its error, and its overall scale offset. true_rec_* and true_mut_* are what the sequence was actually simulated under, written for --truth-fraction of rows — 2% by default — so the gap between the two can be measured on a subsample. Training on the true maps is cheating.

What a row can be

Two mixtures are drawn independently, so any history can appear under any map.

Demographic history

recipe	weight	what it is
spectral_walk	0.200	Band-limited Fourier series on continuous log time. Spectral slope sets roughness, an explicit log10 range sets amplitude, an envelope tilt sets the recent–ancient asymmetry.
empirical_human_ne_template	0.190	191 curves inferred from real human data by SMC++, phlash, MSMC2 and MSMC-IM, with time warp, level shift, contrast and smoothed noise.
size_event	0.100	One size change: time, duration, signed depth, and the level it settles at. A duration running past the window is a permanent step.
bottleneck	0.100	A drop of 0.4 to 3 decades that ends and recovers.
event_on_background	0.086	A quieter spectral walk with one or two sharp events on it.
multi_event	0.080	Two to four events with independent signs.
ooa_chain	0.080	Ancestral size, ancient expansion, a bottleneck held open below the ancestral level, then recent exponential growth — ordered.
constant	0.040	A single effective size. The control.
exponential	0.040	A signed rate through msprime's own growth_rate, so the trajectory is exponential in linear time.
sawtooth	0.030	Periodic growth and decline, with strength, placement, extent, regularity and amplitude drift all drawn.
literature_sawtooth	0.030	The Zigzag_1S14 character: a drawn number of half cycles, spaced about fourfold, each about a decade deep.
stdpopsim_single_pop	0.024	All 18 single-population stdpopsim models across 8 species, verbatim. Rows carry held_out_evaluation.

Sizes span 10² to 10⁶, and the level prior is near-uniform across that range — the bottom decade is real (a vaquita's effective size is under three thousand) and a method that has never seen it cannot be trusted there.

Maps

source	weight	what it is
custom_heterogeneous	0.55	Synthetic: hotspots at 60 per Mb, median width 498 bp, carrying 2.0% of sequence above ten times the mean, on an Ornstein–Uhlenbeck background. Calibrated against deCODE 2019 measured on chr20.
empirical_map_slice	0.20	A random interval of the real deCODE and Roulette maps, taken at the same chromosomal coordinates so the two stay in register.
random_global_mu_r	0.15	Flat maps, recombination drawn 0.3–3× baseline and mutation 0.5–2×.
clean_constant_map	0.10	Flat at baseline, no observation noise. The control.

Observation error is a fraction of each map's own spread, split between a regional component correlated over 80 kb–1.5 Mb and a per-cell one. The regional part is what makes the error survive averaging, which is what a real map's disagreement does: two published maps of one region correlate 0.687 at a kilobase and 0.890 at a hundred.

Configuration

Every field of simulator.config.Config is a command-line flag with the underscores turned into dashes. --preset sets a coherent group of them:

preset	rows	length	haplotypes
mini	512	1 Mb	24
core_v01	20,000	10 Mb	64
full_v01	100,000	20 Mb	64

Useful ones beyond the preset:

• --truth-fraction — what share of rows carry the true maps
• --arg-detail — none, summary (default: pair coalescences and branch length per time bin, 512 bytes a row) or full (the tree sequence itself, which runs 0.8–28 MB a row)
• --workers, --shard-size, --compute-chunk-size — parallelism and layout

Generation is resumable and content-addressed: an interrupted run picks up from its receipts rather than starting over, and a completed dataset is verified against them before it is trusted.

Data that ships with the package

Everything the simulator needs is inside simulator/assets, resolved against the package rather than the working directory, so an installed copy works from anywhere.

	source
maps/decode2019_avg_hg38.bed	deCODE 2019, Halldórsson et al., Science 363:eaau1043. 4,531,535 localised crossovers, 682 bp average resolution, native GRCh38. A pedigree map, so it carries no demographic assumption — the LD-based alternatives are inferred under one, and this project infers demography from linkage.
maps/roulette_2023_...bed	Roulette, Seplyarskiy et al., Nature Genetics 2023. 1 kb bins.
templates/human_ne/	191 inferred human curves: SMC++ on 1000 Genomes, phlash, MSMC2 and MSMC-IM.
templates/stdpopsim_single_pop/	The 18 single-population models in the stdpopsim catalogue, across 8 species.

simulator/assets/SOURCES.md carries the full provenance. The build scripts that produced each frozen file sit beside it; rebuilding needs pip install -e ".[build-assets]".

Tests

pip install -e ".[dev]"
pytest

The suite is calibration as much as correctness. Where a number in the generator was measured from real data, a test pins it to that measurement and says which one — the synthetic map's hotspot width against deCODE chr20, the demography mixture's turning points against the 191 inferred curves, the observed map's correlation with the truth against two published maps of one region.

License

MIT. The maps and templates under simulator/assets are third-party data redistributed under their own terms; see SOURCES.md.