Datasets:

Larrivhan
/

NeCast-benchmark

---
license: cc-by-4.0
pretty_name: NeCast benchmark
tags:
- population-genetics
- demographic-inference
- effective-population-size
- benchmark
- msprime
- simulation
size_categories:
- 1K<n<10K
---
# NeCast benchmark

The simulated benchmark on which the NeCast paper compares NeCast with phlash, SMC++, FitCoal
and MSMC2: the true histories, the simulated windows, the exact inputs each method received,
every estimate, the scores, the timing runs, eight-window joint fits and twelve scenarios with
population structure.

## Contents

| path | what it holds | size |
|---|---|---|
| `histories/` | the 33 size histories: definition and true curve of each | small |
| `windows/{a1,a2,b}/<history>/` | the simulated windows of each setting | 2.0 GB |
| `necast_inputs/{a1,a2,b}/<history>/` | summary statistics and PSMC+ posteriors NeCast read | 0.15 GB |
| `method_inputs/{a1,a2}/<history>/` | the files given to phlash, SMC++, FitCoal and MSMC2 | 7.0 GB |
| `estimates/<method>/{a1,a2}/<history>/` | every fit, with its settings and receipt; `estimates/necast/` also covers `b` | 1.8 GB |
| `map_variants/<history>/<variant>/` | setting `b` run by NeCast with five versions of the maps | 5.0 GB |
| `scores/{a1,a2,b}/<history>/` | per-window scores and every method's curve on the 64 intervals | small |
| `results/` | the tables behind the paper's benchmark figures (`main/`, `supplement/`) | small |
| `timing/` | the timing runs on CPU and GPU nodes | 0.5 GB |
| `joint_fits/` | phlash fitted to eight windows at once, and NeCast averaged over the same eight | small |
| `structure/` | 12 scenarios with population structure, 11 replicates each | 7.4 GB |
| `figures/` | the figure on this card | small |
| `SHA256SUMS` | a checksum for every file except this card | small |

## The three settings

All windows are 30 Mb long, hold 64 diploid individuals, and were made with the same simulator
as the NeCast training data (msprime 1.4.2, the coalescent with recombination), with seeds that
no training window used. Every method read the same 32 individuals of each window; they are
listed in each input's `meta.json` (`columns`).

| setting | histories | rates and maps | errors | methods |
|---|---|---|---|---|
| `a1` | 33 | uniform, μ = 1.25 × 10⁻⁸ and r = 10⁻⁸ per bp per generation | none | NeCast, phlash, SMC++, FitCoal, MSMC2 |
| `a2` | 12 | as `a1` | genotype errors, missing calls, false sites and phase switches at the levels of the training data, added to the `a1` windows | the same |
| `b` | 33 | slices of the deCODE 2019 recombination map and the Roulette mutation map | as `a2` | NeCast |

Each history folder holds 12 windows; 8 are scored, and the scores list which ones. In setting
`b`, NeCast read each window five times (`map_variants/`): with the stored maps (`v1_true`),
with both maps flat at the genome-wide means (`v3_flat_genomewide`) or at the window's own means
(`v3b_flat_rowmean`), with the recombination map and a flat mutation map
(`v4_rec_observed_mut_genomewide`), and with the mutation map and a flat recombination map
(`v5_mut_observed_rec_genomewide`).

The 33 histories fall into five families: 15 human and other realistic models (empirical human
curves, out-of-Africa models and the stdpopsim single-population models), 7 bottlenecks, 6 steps
and slow changes, 3 constant sizes and 2 oscillations. `histories/<name>/definition.json` gives
each one; `trajectories.csv` and `histories/scenarios_v2_files/v2_all_targets.npz` give the truth.
`histories/scenario_ids.json` lists the histories in each setting.

![The 33 histories](https://huggingface.co/datasets/Larrivhan/NeCast-benchmark/resolve/main/figures/benchmark_histories.png)

**The 33 histories**, as the 64-interval curves every method is scored against, ordered by
family: human and other realistic models, bottlenecks, steps and slow changes, constant sizes,
oscillations. The 12 marked a2 are also in setting `a2`; all 33 are in `a1` and `b`.

## Methods and settings

| folder | method | setting |
|---|---|---|
| `estimates/necast/` | NeCast, final weights | as released in `Larrivhan/NeCast` |
| `estimates/phlash/` | phlash | its defaults |
| `estimates/smcpp/` | SMC++ | the setting `tune_d4_rp3`, chosen on two tuning histories, at the default stopping tolerance |
| `estimates/smcpp_converged/` | SMC++ | the same, with stopping tolerance 10⁻⁶ (supplement) |
| `estimates/fitcoal/` | FitCoal | its defaults |
| `estimates/msmc2/` | MSMC2 | time patterning `1*2+20*1+1*10` (supplement) |

A fit folder (`<history>/K1/g<window>/n32/`) holds the estimated curve (`curve.npz`: time
points `t_left`, sizes `ne`, the method's version and runtime), the fit settings
(`fit_spec.json`), a receipt, and the method's own output where it has one (for phlash, its
posterior samples). NeCast's `predictions.npz` holds log10 Ne on the 64 intervals for the 8
scored windows, with the truth and the interval edges.

## Inputs given to the methods

`method_inputs/<setting>/<history>/r<window>/n32/` holds one window as each method reads it:
`view.bcf` (32 diploid individuals), `multihetsep.txt` (MSMC2), `phlash_contig.npz` (phlash) and
`sfs_unfolded.npy`, with `meta.json` recording the individuals and checksums. Where two sites
share a position, the first is exported and the others dropped, since the methods require
distinct positions.

## Scores

`scores/<setting>/<history>/v2score_20260917/` holds the main scores and `v2supp_20260918/` the
supplement (converged SMC++ and MSMC2): `model_metrics.csv` and `tool_metrics.csv` (the RMSE of
log10 Ne over the 64 intervals, `m1_rmse64`, and further measures), `tool_curves64.npz` (every
method's curve on the 64 intervals) and `runtime.tsv`. `truth/` has the true curves. The 64
intervals run from 50 to 500,000 generations ago, evenly spaced in log time.

`results/main/` and `results/supplement/` hold the tables built from these scores
(`scores_rows.tsv`, `summary_by_scenario.tsv`, `summary_by_family.tsv`, `paired_contrasts.tsv`,
`B_map_effect.tsv`, `timing_C1_C2.tsv`, `curves_bundle.json`), each with a checksum manifest.

## Joint fits and structure

`joint_fits/` has phlash fitted to eight windows of a history at once, with its scripts and
summary (`results/E3_summary.json`). `structure/` has the 12 structure scenarios (a panmictic
control, island models, changing migration, admixture and ancient structure): the windows, the
phlash fits, NeCast's runs, and the results (`phlash_and_results/results/`).

## Licence and sources

The files here are released under CC BY 4.0. They were made with msprime (Baumdicker et al.
2022, *Genetics* 220:iyab229), the deCODE 2019 map (Halldorsson et al. 2019, *Science*
363:eaau1043), the Roulette mutation rates (Seplyarskiy et al. 2023, *Nature Genetics*) and
stdpopsim (Adrion et al. 2020, *eLife* 9:e54967), and compare phlash (Terhorst 2025, *Nature
Genetics* 57:2570), SMC++ (Terhorst, Kamm and Song 2017, *Nature Genetics* 49:303), FitCoal
(Hu et al. 2023, *Science* 381:979) and MSMC2 (Schiffels and Wang 2020, *Methods Mol Biol*
2090:147).

## Citation

The NeCast paper is in preparation; its reference will be added here.

	---
	license: cc-by-4.0
	pretty_name: NeCast benchmark
	tags:
	- population-genetics
	- demographic-inference
	- effective-population-size
	- benchmark
	- msprime
	- simulation
	size_categories:
	- 1K<n<10K
	---

	# NeCast benchmark

	The simulated benchmark on which the NeCast paper compares NeCast with phlash, SMC++, FitCoal
	and MSMC2: the true histories, the simulated windows, the exact inputs each method received,
	every estimate, the scores, the timing runs, eight-window joint fits and twelve scenarios with
	population structure.

	## Contents

	\| path \| what it holds \| size \|
	\|---\|---\|---\|
	\| `histories/` \| the 33 size histories: definition and true curve of each \| small \|
	\| `windows/{a1,a2,b}/<history>/` \| the simulated windows of each setting \| 2.0 GB \|
	\| `necast_inputs/{a1,a2,b}/<history>/` \| summary statistics and PSMC+ posteriors NeCast read \| 0.15 GB \|
	\| `method_inputs/{a1,a2}/<history>/` \| the files given to phlash, SMC++, FitCoal and MSMC2 \| 7.0 GB \|
	\| `estimates/<method>/{a1,a2}/<history>/` \| every fit, with its settings and receipt; `estimates/necast/` also covers `b` \| 1.8 GB \|
	\| `map_variants/<history>/<variant>/` \| setting `b` run by NeCast with five versions of the maps \| 5.0 GB \|
	\| `scores/{a1,a2,b}/<history>/` \| per-window scores and every method's curve on the 64 intervals \| small \|
	\| `results/` \| the tables behind the paper's benchmark figures (`main/`, `supplement/`) \| small \|
	\| `timing/` \| the timing runs on CPU and GPU nodes \| 0.5 GB \|
	\| `joint_fits/` \| phlash fitted to eight windows at once, and NeCast averaged over the same eight \| small \|
	\| `structure/` \| 12 scenarios with population structure, 11 replicates each \| 7.4 GB \|
	\| `figures/` \| the figure on this card \| small \|
	\| `SHA256SUMS` \| a checksum for every file except this card \| small \|

	## The three settings

	All windows are 30 Mb long, hold 64 diploid individuals, and were made with the same simulator
	as the NeCast training data (msprime 1.4.2, the coalescent with recombination), with seeds that
	no training window used. Every method read the same 32 individuals of each window; they are
	listed in each input's `meta.json` (`columns`).

	\| setting \| histories \| rates and maps \| errors \| methods \|
	\|---\|---\|---\|---\|---\|
	\| `a1` \| 33 \| uniform, μ = 1.25 × 10⁻⁸ and r = 10⁻⁸ per bp per generation \| none \| NeCast, phlash, SMC++, FitCoal, MSMC2 \|
	\| `a2` \| 12 \| as `a1` \| genotype errors, missing calls, false sites and phase switches at the levels of the training data, added to the `a1` windows \| the same \|
	\| `b` \| 33 \| slices of the deCODE 2019 recombination map and the Roulette mutation map \| as `a2` \| NeCast \|

	Each history folder holds 12 windows; 8 are scored, and the scores list which ones. In setting
	`b`, NeCast read each window five times (`map_variants/`): with the stored maps (`v1_true`),
	with both maps flat at the genome-wide means (`v3_flat_genomewide`) or at the window's own means
	(`v3b_flat_rowmean`), with the recombination map and a flat mutation map
	(`v4_rec_observed_mut_genomewide`), and with the mutation map and a flat recombination map
	(`v5_mut_observed_rec_genomewide`).

	The 33 histories fall into five families: 15 human and other realistic models (empirical human
	curves, out-of-Africa models and the stdpopsim single-population models), 7 bottlenecks, 6 steps
	and slow changes, 3 constant sizes and 2 oscillations. `histories/<name>/definition.json` gives
	each one; `trajectories.csv` and `histories/scenarios_v2_files/v2_all_targets.npz` give the truth.
	`histories/scenario_ids.json` lists the histories in each setting.

	![The 33 histories](https://huggingface.co/datasets/Larrivhan/NeCast-benchmark/resolve/main/figures/benchmark_histories.png)

	The 33 histories, as the 64-interval curves every method is scored against, ordered by
	family: human and other realistic models, bottlenecks, steps and slow changes, constant sizes,
	oscillations. The 12 marked a2 are also in setting `a2`; all 33 are in `a1` and `b`.

	## Methods and settings

	\| folder \| method \| setting \|
	\|---\|---\|---\|
	\| `estimates/necast/` \| NeCast, final weights \| as released in `Larrivhan/NeCast` \|
	\| `estimates/phlash/` \| phlash \| its defaults \|
	\| `estimates/smcpp/` \| SMC++ \| the setting `tune_d4_rp3`, chosen on two tuning histories, at the default stopping tolerance \|
	\| `estimates/smcpp_converged/` \| SMC++ \| the same, with stopping tolerance 10⁻⁶ (supplement) \|
	\| `estimates/fitcoal/` \| FitCoal \| its defaults \|
	\| `estimates/msmc2/` \| MSMC2 \| time patterning `12+201+1*10` (supplement) \|

	A fit folder (`<history>/K1/g<window>/n32/`) holds the estimated curve (`curve.npz`: time
	points `t_left`, sizes `ne`, the method's version and runtime), the fit settings
	(`fit_spec.json`), a receipt, and the method's own output where it has one (for phlash, its
	posterior samples). NeCast's `predictions.npz` holds log10 Ne on the 64 intervals for the 8
	scored windows, with the truth and the interval edges.

	## Inputs given to the methods

	`method_inputs/<setting>/<history>/r<window>/n32/` holds one window as each method reads it:
	`view.bcf` (32 diploid individuals), `multihetsep.txt` (MSMC2), `phlash_contig.npz` (phlash) and
	`sfs_unfolded.npy`, with `meta.json` recording the individuals and checksums. Where two sites
	share a position, the first is exported and the others dropped, since the methods require
	distinct positions.

	## Scores

	`scores/<setting>/<history>/v2score_20260917/` holds the main scores and `v2supp_20260918/` the
	supplement (converged SMC++ and MSMC2): `model_metrics.csv` and `tool_metrics.csv` (the RMSE of
	log10 Ne over the 64 intervals, `m1_rmse64`, and further measures), `tool_curves64.npz` (every
	method's curve on the 64 intervals) and `runtime.tsv`. `truth/` has the true curves. The 64
	intervals run from 50 to 500,000 generations ago, evenly spaced in log time.

	`results/main/` and `results/supplement/` hold the tables built from these scores
	(`scores_rows.tsv`, `summary_by_scenario.tsv`, `summary_by_family.tsv`, `paired_contrasts.tsv`,
	`B_map_effect.tsv`, `timing_C1_C2.tsv`, `curves_bundle.json`), each with a checksum manifest.

	## Joint fits and structure

	`joint_fits/` has phlash fitted to eight windows of a history at once, with its scripts and
	summary (`results/E3_summary.json`). `structure/` has the 12 structure scenarios (a panmictic
	control, island models, changing migration, admixture and ancient structure): the windows, the
	phlash fits, NeCast's runs, and the results (`phlash_and_results/results/`).

	## Licence and sources

	The files here are released under CC BY 4.0. They were made with msprime (Baumdicker et al.
	2022, Genetics 220:iyab229), the deCODE 2019 map (Halldorsson et al. 2019, Science
	363:eaau1043), the Roulette mutation rates (Seplyarskiy et al. 2023, Nature Genetics) and
	stdpopsim (Adrion et al. 2020, eLife 9:e54967), and compare phlash (Terhorst 2025, *Nature
	Genetics* 57:2570), SMC++ (Terhorst, Kamm and Song 2017, Nature Genetics 49:303), FitCoal
	(Hu et al. 2023, Science 381:979) and MSMC2 (Schiffels and Wang 2020, Methods Mol Biol
	2090:147).

	## Citation

	The NeCast paper is in preparation; its reference will be added here.