arDCA

auto-regressive Direct Coupling Analysis (arDCA) 2.0.

Overview

This package is the GPU-accelerated version of the original version that can be found at ArDCA.jl. The current implementation also aims at providing a user-friendly command line interface for training and sampling from an autoregressive DCA model.

Installation

During the installation, arDCA will also install adabmDCA and all its dependencies.

Option 1:

Install the package via PyPl:

python -m pip install arDCA

Option 2

Clone this repository:

git clone https://github.com/spqb/arDCA.git
cd arDCA
python -m pip install .

Using the package

We provide a Colab notebook where it is shown hot to train and sample an arDCA model using RNA sequences.

Alternatively, one can install the package locally and run from the command line one of the two implemented routines:

Train arDCA from the command line

Once installed, you can launch the package routing by using the command arDCA. All the training options can be listed via

arDCA train -h

To launch a training with default arguments, use

arDCA train -d <path_data> -o <output_folder> -l <label>

where path_data is the path to the input multi-sequence alignment in fasta format and label is an identifier for the output files. The parameters of the trained model are saved in the file output_folder/<label>_params.pth, and can be easily loaded afterwrds using the Pytorch methods.

By default, the program assumes that the input data are protein sequences. If you want to use RNA sequences, you should use the argument --alphabet rna.

Warning

Depending on the dataset, the default regularization parameters reg_h and reg_J may not work properly. If the training does not converge or the model's generation capabilities are poor, you may want to increase these values.

Sample arDCA from the command line

To generate new sequences using the command line, the minimal input command is

arDCA sample -p <path_params> -o <output_folder> -l <label> --ngen <num_sequences>

where num_sequences is the number of sequences to be generated. The output will be saved in fasta format at output_folder/<label>_samples.fasta.

If the argument -d <path_data> is provided, the script will also compute the Pearson correlation coefficient and the slope between the two-sites correlation matrix of the data and the generated samples.

License

This package is open-sourced under the Apache License 2.0.

Citation

If you use this package in your research, please cite

Trinquier, J., Uguzzoni, G., Pagnani, A. et al. Efficient generative modeling of protein sequences using simple autoregressive models. Nat Commun 12, 5800 (2021).