Adaptive First-Order Optimization for Biobank-Scale Genetic Clustering

---

ADAMIXTURE is an unsupervised global ancestry inference method that scales the ADMIXTURE model to biobank-sized datasets. It combines the Expectation–Maximization (EM) framework with the Adam first-order optimizer, enabling parameter updates after a single EM step. This approach accelerates convergence while maintaining comparable or improved accuracy, substantially reducing runtime on large genotype datasets. For more information, we recommend reading our preprint.

The software can be invoked via CLI and has a similar interface to ADMIXTURE (e.g. the output format is completely interchangeable).

System requirements

Hardware requirements

The successful usage of this package requires a computer with enough RAM to be able to handle the large datasets the network has been designed to work with. Due to this, we recommend using compute clusters whenever available to avoid memory issues.

Software requirements

We recommend creating a fresh Python 3.10+ virtual environment. For a faster installation experience, we highly recommend using uv.

Important

If you plan to use GPU acceleration, ensure that the CUDA toolkit is correctly loaded (e.g., module load cuda) before starting the installation. This ensures that the dependencies and internal components are correctly configured for your hardware.

As an example, using uv (recommended):

$ uv venv --python 3.10
$ source .venv/bin/activate
$ uv pip install adamixture

Installation Guide

The package can be easily installed in at most a few minutes using pip (make sure to add the --upgrade flag if updating the version):

$ pip install adamixture

Running ADAMIXTURE

To train a model, simply invoke the following commands from the root directory of the project. For more info about all the arguments, please run adamixture --help. Note that BED, VCF and PGEN are supported:

As an example, the following ADMIXTURE call

$ ./admixture snps_data.bed 8 -s 42

would be equivalent in ADAMIXTURE by running

$ adamixture -k 8 --data_path snps_data.bed --save_dir SAVE_PATH --name snps_data -s 42

Two files will be output to the SAVE_PATH directory (the name parameter will be used to create the full filenames):

• A .P file, similar to ADMIXTURE.
• A .Q file, similar to ADMIXTURE.

Logs are printed to the stdout channel by default. If you want to save them to a file, you can use the command tee along with a pipe:

$ adamixture -k 8 ... | tee run.log

Running with multi-threading

To run ADAMIXTURE using multiple CPU threads, use the -t flag:

$ adamixture -k 8 --data_path data.bed --save_dir out/ --name test -t 8

Running with GPU acceleration

To leverage GPU acceleration (highly recommended for large datasets), use the --device flag:

• NVIDIA GPU (CUDA):

  $ adamixture -k 8 --data_path data.bed --save_dir out/ --name test --device gpu

• macOS Apple Silicon (MPS):

  $ adamixture -k 8 --data_path data.bed --save_dir out/ --name test --device mps

Tip

GPU Acceleration: Using GPUs greatly speeds up processing and is highly recommended for large datasets. You can specify the hardware to use with the --device parameter:

• For NVIDIA GPUs, use --device gpu (requires CUDA).
• For macOS users with Apple Silicon (M1/M2/M3/M4/M5), use --device mps to enable Metal Performance Shaders (MPS) acceleration.
• Note that biobank-scale datasets are best handled on dedicated CUDA-capable GPUs due to high RAM requirements.

Multi-K Sweep

Instead of running ADAMIXTURE for a single K, you can automatically sweep over a range of K values using --min_k and --max_k. The data is loaded once, and each K is trained sequentially:

$ adamixture --min_k 2 --max_k 10 --data_path snps_data.bed --save_dir SAVE_PATH --name snps_sweep

Cross-validation

Use --cv to estimate the optimal K by masking a fraction of genotype entries and measuring prediction error. → Full documentation

$ adamixture -k 8 --cv --data_path data.bed --save_dir out/ --name test

Plotting

Native high-quality visualizations with hierarchical population labels (--labels, --labels2, --labels3) and multi-run alignment. → Full documentation

$ adamixture -k 8 --data_path data.bed --save_dir out/ --name test --plot pdf 300

Projection Mode

Estimate ancestry proportions for new samples using a pre-trained, fixed P matrix (Q-only optimisation). K is detected automatically from P. → Full documentation

$ adamixture-project \
    --data_path new_samples.bed \
    --p_path trained_model/results.8.P \
    --save_dir projection_out/ \
    --name projected

Supervised Mode

Anchor the model with known population labels for a subset of samples while estimating Q freely for unlabeled ones. Labels use the same format as --labels (population name or -). → Full documentation

$ adamixture-supervised \
    --data_path all_samples.bed \
    --labels labels.txt \
    --save_dir supervised_out/ \
    --name supervised_run \
    -k 8

Other options

All hyperparameters and flags can be explored with:

$ adamixture --help

Key optimizer arguments:

Argument	Default	Description
--lr	0.005	Adam learning rate
--beta1	0.80	Adam β₁
--beta2	0.88	Adam β₂
--reg_adam	1e-8	Adam ε (numerical stability)
--lr_decay	0.5	Learning rate decay factor
--min_lr	1e-4	Minimum learning rate
--patience_adam	3	Checks without improvement before decaying lr
--tol_adam	0.1	Convergence tolerance
--max_iter	10000	Maximum Adam-EM iterations
--check	5	Log-likelihood evaluation frequency
-t	1	Number of CPU threads
-s	42	Random seed
--chunk_size	4096	Number of SNPs in chunk operations

Troubleshooting and Tips

→ Full documentation

License

This project is licensed under the BSD 3-Clause License - see the LICENSE file for details.

Cite

When using this software, please cite the following preprint:

@article{saurina2026adamixture,
  title={ADAMIXTURE: Adaptive First-Order Optimization for Biobank-Scale Genetic Clustering},
  author={Saurina-i-Ricos, Joan and Mas Monserrat, Daniel and Ioannidis, Alexander G.},
  journal={bioRxiv},
  year={2026},
  doi={10.64898/2026.02.13.700171},
  url={https://doi.org/10.64898/2026.02.13.700171}
}