Negative dataset selection impacts machine learning-based predictors for multiple bacterial species promoters

This repository contains the dataset and code for the study on how negative dataset selection impacts machine learning-based predictors for promoters in multiple bacterial species.

Published in: Bioinformatics, Volume 41, Issue 4, April 2025.

Prerequisites

• A Linux-based operating system (tested in Ubuntu 22.04)
• Docker installed on your system
• Docker Compose installed
• NVIDIA Container Toolkit for GPU support
• Weights & Biases account for experiment tracking
• AWS S3 bucket for storing training checkpoints and logs

Project structure

• data/: This directory contains all the data used in the project, divided into different subfolders.

  • processed/: Stores processed data, based on dataset type (CDS/SRS).
  • raw/: Holds unprocessed data files originally obtained from the National Center for Biotechnology Information (NCBI) for CDS strains and the Prokaryotic Promoter Database for promoters.

• docker/:

  • pytorch/: Includes configurations and scripts necessary to set up and run a PyTorch environment within a Docker container.
  • tensorflow/: Contains configurations and scripts to establish a TensorFlow environment in Docker. Additionally, specific entry point scripts are included in the entrypoints/ subfolder, such as build_dataset.sh for dataset generation and entrypoint.sh as the main startup script for the container for training/tuning RF and CNN models.

• notebooks/: Contains Jupyter notebooks used for exploratory data analysis, and DNABERT analysis (latent space and genoma evaluation)

• scripts/: Includes all scripts used in the project, including common scripts, model definitions, data preprocessing, experiment execution, and utility functions.

Setup

Environment variables

Create a .env file in the root directory with the following environment variables:

• WANDB_API_KEY: This is your Weights & Biases API key, which is necessary for logging experiments and results to the Weights & Biases platform.

• AWS_ACCESS_KEY_ID: This is your AWS Access Key ID, necessary for authenticating requests to AWS services (tested with S3).

• AWS_SECRET_ACCESS_KEY: This is your AWS Secret Access Key, which pairs with the AWS Access Key ID to securely authenticate your AWS requests.

Scripts

Dataset generation (optional)

If you want to generate the dataset from scratch, you can run the following command:

docker compose up dataset-build

Hyperparameter tuning and training

Random Forest and Convolutional Neural Network models

Edit /docker/tensorflow/entrypoints/entrypoint.sh based on the model and dataset and then run:

docker compose up training

BERT-based models

Edit /docker/pytorch/entrypoint.sh based on the dataset and then run:

docker compose up training-bert

Jupyter notebooks

If you want to run the Jupyter notebooks, execute the following command to start the Jupyter server:

docker compose up <jupyter-torch|jupyter-tf>

Model Availability

Trained BERT-based models are available on Zenodo. An example of their usage can be found in notebooks/bert_results.ipynb. The models are provided as models_cds.zip (trained on the CDS dataset) and models_random.zip (trained on the SRS dataset).

How to Cite This Work

If this work has contributed to your research, please consider citing the paper, the software, or both depending on your use:

Paper:

Marcelo González, Roberto E Durán, Michael Seeger, Mauricio Araya, Nicolás Jara,
Negative dataset selection impacts machine learning-based predictors for multiple
bacterial species promoters, Bioinformatics, Volume 41, Issue 4, April 2025, btaf135,
https://doi.org/10.1093/bioinformatics/btaf135

Software:

Marcelo González, «maigonzalezh/MultispeciesPromoterClassifier: v1.0.0». Zenodo, mar. 13, 2025.
doi: 10.5281/zenodo.15016403.