Larrivhan/dlcoalsim-110k

DLCoalSim — 109,248 simulated genomes with known effective population size

Coalescent simulations of 30 Mb, 128 haplotypes each, labelled with the effective population size trajectory that generated them: 64 log-spaced time bins from 53.7 to 465,286 generations. The task is to recover the trajectory from the sequence.

What a row is

array	shape	meaning
genotype_packed	bit-packed	biallelic genotypes, 128 haplotypes
variant_positions_bp	per row	segregating site positions
target_log10_ne	64	the label: log10 Ne per time bin
obs_rec_rate / obs_rec_pos	per row	the recombination map as observed
obs_mut_rate / obs_mut_pos	per row	the mutation map as observed
source_id	1	which of four map sources drew this row
demography_type_id	1	coarse label of the history's shape
pair_coalescences_per_bin	2016	ground truth pairwise coalescence counts
branch_length_per_bin	64	ground truth branch length per bin

Rows are stored four to a fragment; the *_offsets arrays slice them.

Read this before you use it

The large end of the distribution is very slightly truncated. 110,000 rows were planned and 109,248 delivered (99.31%). The 752 missing rows are the ones whose simulations exceeded the compute budget, and they are not a random sample: the fragments written last have a median size 50% above the median of the whole cohort, because what makes a row expensive is the same thing that makes it large -- many segregating sites. Anyone studying the extreme end of the parameter space should know that end is thinned by well under a percent.

Nothing was substituted for them. Filling the gap with cheaper draws would have replaced the hardest rows with easier ones and quietly changed the source mixture declared below.

Ground truth is included. pair_coalescences_per_bin and branch_length_per_bin are quantities from inside the simulation, not observables. They are here so the dataset can be used to study what is recoverable in principle. Training on them is not the task.

The observed maps are not the true maps, except for 27% of rows. Four map sources are mixed, and they do not share an observation model:

source_id	source	share	observed map
0	constant at the baseline rate	10%	exact
1	constant at a per-row random rate	15%	exact
2	synthetic heterogeneous	55%	noised, σ=0.55 (rec) / 0.33 (mut)
3	deCODE slice	20%	noised, σ=0.8

Sources 0 and 1 are constant maps, and a constant map has no spatial structure to mis-measure, so the observed copy is the truth — the rate is taken as known, which is what methods in this field assume of mu. That is a modelling choice rather than an oversight, but it makes the cohort heterogeneous: those rows hand a model the exact global rate and the rest hand it a noised one. Since Ne is inferred through theta = 4·Ne·mu, source 1 in particular is an easier inference problem than sources 2 and 3.

Stratify by source_id when you report. A number averaged over all four sources mixes two different problems.

Provenance of the third-party maps

The simulated genotypes are original. The observed recombination and mutation maps are derived from published data, redistributed under its own terms:

• Recombination — Halldórsson et al., Science 363:eaau1043 (2019). 4,531,535 crossovers in Icelandic pedigrees, native GRCh38, obtained as recombAvg.bw from UCSC. A pedigree map is used rather than an LD-based one on purpose: HapMap II and pyrho are inferred under a demographic model, and inferring demography from linkage with them would be circular.
• Mutation — Seplyarskiy et al., Nature Genetics (2023), Roulette, 1 kb bins, pseudo-R² 0.86 against observed variation.
• Demographic templates — 191 published human Ne curves (smc++, phlash, MSMC2, MSMC-IM), used as the one part of the mixture nobody designed and as the calibration target for every synthetic recipe.

Please cite the two map papers alongside this dataset.

Constants

Sequence length 30 Mb; 128 haplotypes (64 diploids); 64 time bins spaced by a factor of 1.1548; baseline recombination 8.749e-9 and mutation 7.493e-9 per bp per generation.