Infer demographic parameters from Ancestral Recombination Graphs (ARGs), by extracting pairwise coalescence counts for use in a maximum likelihood model. For detail on the method and results, see the preprint:

DeHaas, Drew, Zhibai Jia, Leo Speidel, and Xinzhu Wei. "Inference of complex demographic history using composite likelihood based on whole-genome genealogies." bioRxiv (2025): 2025-10.

See the documentation for commands, examples, and concepts.

See PAPER_EXPERIMENTS.md for the configurations that were used in the paper.

Python 3.8 or newer is supported.

Install from PyPi:

pip install mrpast

On Linux, this will use prebuilt binaries. On MacOS, this will trigger a source code build, which requires CMake and gcc or clang (C++17 support required).

You can also install the conda package via the bioconda channel: conda install mrpast.

Recommend using a virtual environment, the below creates and activates one:

python3 -m venv MyEnv
source MyEnv/bin/activate

Clone repo, then build and install:

git clone --recursive https://github.com/aprilweilab/mrpast.git
pip install mrpast/

There are three primary subcommands to mrpast, and they are usually run in this order:

1. mrpast simulate
2. mrpast process
3. mrpast solve

These steps describe the "Simulated ARG" workflow, where no ARG inference is performed. See the documentation for workflows making use of inferred ARGs.

In order to test out a demographic model, it is recommended that you start out by simulating that model and verifying that mrpast can recover the model parameters with the necessary accuracy. The simulation is done via msprime and produces ancestral recombination graphs (ARGs) in the form of a tskit tree-sequence file (.trees).

Example:

# Simulate the model 10 times, using a DNA sequence length of 100Kbp and the default recombination rate (1e-8)
mrpast simulate --replicates 10 --seq-len 100000 --debug-demo examples/5deme1epoch.yaml 5de1

This creates 10 tree-sequence files (ARGs) that are named like 5de1*.trees, using the given model.

Given an ARG in tree-sequence format, either from simulation (see above) or from ARG inference, we then extract coalescence information.

Example:

# Use 10 CPU threads to process the data and produce 10 replicates (expanded models) to be solved (later).
# `--bootstrap` creates 100 bootstrap samples by default, the average of which is used for input the maximum
# likelihood function
mrpast process --jobs 10 --replicates 10 --suffix trial1 --bootstrap coalcounts examples/5deme1epoch.yaml 5de1

See mrpast process --help for more options that control time discretization, distance between sampled trees, etc.

If we want, we could use --solve to run the solver as soon as processing completed. Otherwise, see the next section.

If you didn't pass --solve to mrpast process then you can run the solver via:

mrpast solve --jobs 10 5deme1epoch.*.solve_in.*.json

The resulting output files will be listed, and the best output (maximum likelihood) will be listed as well. The JSON files for the output contains the parameter values, their bounds, their initialized values, and (if present) their ground truth values. Assuming the best result was 5deme1epoch.trial1.solve_in.0.out.json, we can quickly view the results via

mrpast show -n 5deme1epoch.trial1.solve_in.0.out.json

The simulated data, inferred ARG workflow is:

1. mrpast simulate: Simulate your model with some ground-truth parameter values.
2. mrpast sim2vcf -p: Convert all .trees files with the given prefix to VCF files, and emit the corresponding .popmap.json files (which maps each sample to a population).
3. mrpast arginfer: Infer ARG from the VCF files, and then attach the population IDs to the ARG (.trees files) using the .popmap.json
4. mrpast process: Process and solve the inferred ARGs

The real data workflow is:

1. Manually create a .popmap.json file for your VCF dataset. See the documentation for more details.
2. mrpast arginfer: Infer ARG from the VCF files, and then attach the population IDs to the ARG (.trees files) using the .popmap.json
3. mrpast process: Process and solve the inferred ARGs

The demographic model is specified via YAML. See the examples directory for example models. See the documentation for details on model syntax and behavior.

1. Compile for the native CPU; this can speed up the numerical solver, but makes the resulting package less portable.

MRPAST_ENABLE_NATIVE=1 pip install mrpast/

2. Build the solver in debug mode, so GDB can be attached.

MRPAST_DEBUG=1 pip install mrpast/