Skip to content

AI-sandbox/aegen

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ægen DOI - 10.1101/2023.09.27.558320

Entropy analysis

ægen is a meta-autoencoder which allows to customize the shape of the autoencoder (regular, window-based, or hybrid) and specify the desired latent space distribution (Gaussian, Multi-Bernoulli, or VQ-VAE). Additionally, it allows to use conditioning and/or denoising modes. If you find this paper or implementation useful, please consider citing our bioRxiv preprint!

@article{geleta2023aegen,
	author = {Margarita Geleta and Daniel Mas Montserrat and Xavier Giro-i-Nieto and Alexander G Ioannidis},
	title = {Deep Variational Autoencoders for Population Genetics},
	year = {2023},
	doi = {10.1101/2023.09.27.558320},
	URL = {https://www.biorxiv.org/content/early/2023/09/28/2023.09.27.558320},
	eprint = {https://www.biorxiv.org/content/early/2023/09/28/2023.09.27.558320.full.pdf},
	journal = {bioRxiv}
}

Dependencies

Environment setup

Assuming a Python virtual environment is set up, the dependencies can be installed with:

$ pip3 install -r requirements.txt

The whole project has been developed with Python 3.6.1 and PyTorch 1.4.0.

Data

Founders' dataset

Human dataset – the human dataset is composed of public-use human whole genome sequences collected from real world-wide populations. The three sources are the listed below:

The dataset was pruned to contain only single-ancestry origin individuals, i.e., individuals whose four grandparents self-reported belonging to the same ancestral group. After pruning, the dataset resulted in 2965 single-ancestry phased human genomes, each containing a maternal and paternal copy. For that reason, each sequence could be expanded into two, doubling the number of sequences to 5930, to which we refer as founders.

Dataset filesystem

The whole code is mounted on a specific data filesystem tree. First, it is needed to define the enviroment variables with the data paths. Those paths can be defined in the scripts/ini.sh script: $USER_PATH is the environment variable pointing to the root of this repository, $IN_PATH is the path for incoming data (training, validation and test sets), and $OUT_PATH is the path for outgoing data (logs).

The input data filesystem is defined as follows:

$IN_PATH
└─ data
    ├─ human
    │   ├─ sample maps -> EUR, EAS, AMR, SAS, AFR, OCE, WAS
    │   ├─ human HapMap genetic map (.gmap)
    │   ├─ reference panel metadata (.tsv)
    │   ├─ chr22
    │   │   ├─ VCF file (.vcf)
    │   │   └─ prepared
    │   │       ├─ train -> HDF5 files with single-ancestry simulated data
    │   │       ├─ valid -> HDF5 files with single-ancestry simulated data
    │   │       └─ test  -> HDF5 files with single-ancestry simulated data
    │   └─ other chr can be added
    └─ other species can be added

Data augmentation

Two types of data augmentation have been used: (1) offline and (2) online. Offline data augmentation precomputes the training set data before starting training the model, whereas online simulation simulates new data samples on-the-fly. In this section, steps to perform offline simulation are explained. In order to specify which type of simulation to use, define that in the params.yaml file in the root folder of this repository.

Founders have been split in three non-overlapping groups with proportions 80%, 10% and 10%, to generate the training, validation and test sets. For each set, several datasets have been simulated with the corresponding subset of founders using Wright-Fisher simulation within each population separately and basing the recombination on the human HapMap genetic map.

In order to run the offline simulation, execute the following commands:

$ cd scripts
$ source ini.sh
## Create single-ancestry maps and simulate single-ancestry individuals within each split.
$ ./simulate.sh species=human generations=[desired num of generations] individuals=[desired num of ind/gen]
## Create HDF5 datasets for each split with the specified number of SNPs.
$ ./create.sh snps=[desired num snps]

Training

The proposed method consists of a highly-adaptable and modular autoencoder that accepts flags to switch to conditioning mode, use different encoder/decoder architectures and specifiy the distribution at the bottleneck of the model. Furthermore, the model accepts two sets: (1) a set of fixed parameters, which defines the shape of the network, conditioning, number and size of layers in the encoder/decoder, dropouts, batch normalization and activation functions; (2) a set of hyperparameters, which defines optimizer flags and values, such as, the learning rate, weight decay, data augmentation simulation mode, among others. All of those parameters and hyperparameters are defined in the params.yaml file in the root folder of this repository. Once both sets have been specifies, a training session can be started by using:

$ cd scripts
$ source ini.sh

## Store the params.yaml file used in this experiment in $OUT_PATH
$ rm -rf $OUT_PATH/experiments/exp[number]
$ mkdir -p $OUT_PATH/experiments/exp[number]
$ cp $USER_PATH/params.yaml $OUT_PATH/experiments/exp[number]/
$ touch $OUT_PATH/experiments/exp[number]/exp[number].log
$ chmod +rwx $OUT_PATH/experiments/exp[number]/exp[number].log

$ python3 $USER_PATH/src/trainer.py \
--species human \
--chr 22 \
--params $OUT_PATH/experiments/exp[number]/params.yaml \
--num [number] \
--verbose False \
--evolution False

Or, if using a Slurm queue, running ./submit.sh experiment=[number] in the scripts folder.

Evaluation

Compression benchmarks

To run the compression benchmarks, please refer to the script under src/eval/compressors.py.

Pre-trained models

Not public yet.

License

NOTICE: This software is available for use free of charge for academic research use only. Academic users may fork this repository and modify and improve to suit their research needs, but also inherit these terms and must include a licensing notice to that effect. Commercial users, for profit companies or consultants, and non-profit institutions not qualifying as "academic research" should contact geleta@berkeley.edu. This applies to this repository directly and any other repository that includes source, executables, or git commands that pull/clone this repository as part of its function. Such repositories, whether ours or others, must include this notice.